* [PATCH for-8.0 01/29] include/qemu/cpuid: Introduce xgetbv_low
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-21 12:15 ` Philippe Mathieu-Daudé
2022-11-18 9:47 ` [PATCH for-8.0 02/29] include/exec/memop: Add bits describing atomicity Richard Henderson
` (27 subsequent siblings)
28 siblings, 1 reply; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
Replace the two uses of asm to expand xgetbv with an inline function.
Since one of the two has been using the mnemonic, assume that the
comment about "older versions of the assember" is obsolete, as even
that is 4 years old.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
include/qemu/cpuid.h | 7 +++++++
util/bufferiszero.c | 3 +--
tcg/i386/tcg-target.c.inc | 11 ++++-------
3 files changed, 12 insertions(+), 9 deletions(-)
diff --git a/include/qemu/cpuid.h b/include/qemu/cpuid.h
index 7adb12d320..1451e8ef2f 100644
--- a/include/qemu/cpuid.h
+++ b/include/qemu/cpuid.h
@@ -71,4 +71,11 @@
#define bit_LZCNT (1 << 5)
#endif
+static inline unsigned xgetbv_low(unsigned c)
+{
+ unsigned a, d;
+ asm("xgetbv" : "=a"(a), "=d"(d) : "c"(c));
+ return a;
+}
+
#endif /* QEMU_CPUID_H */
diff --git a/util/bufferiszero.c b/util/bufferiszero.c
index ec3cd4ca15..b0660d484d 100644
--- a/util/bufferiszero.c
+++ b/util/bufferiszero.c
@@ -287,8 +287,7 @@ static void __attribute__((constructor)) init_cpuid_cache(void)
/* We must check that AVX is not just available, but usable. */
if ((c & bit_OSXSAVE) && (c & bit_AVX) && max >= 7) {
- int bv;
- __asm("xgetbv" : "=a"(bv), "=d"(d) : "c"(0));
+ unsigned bv = xgetbv_low(0);
__cpuid_count(7, 0, a, b, c, d);
if ((bv & 0x6) == 0x6 && (b & bit_AVX2)) {
cache |= CACHE_AVX2;
diff --git a/tcg/i386/tcg-target.c.inc b/tcg/i386/tcg-target.c.inc
index c96b5a6f43..1361960156 100644
--- a/tcg/i386/tcg-target.c.inc
+++ b/tcg/i386/tcg-target.c.inc
@@ -4148,12 +4148,9 @@ static void tcg_target_init(TCGContext *s)
/* There are a number of things we must check before we can be
sure of not hitting invalid opcode. */
if (c & bit_OSXSAVE) {
- unsigned xcrl, xcrh;
- /* The xgetbv instruction is not available to older versions of
- * the assembler, so we encode the instruction manually.
- */
- asm(".byte 0x0f, 0x01, 0xd0" : "=a" (xcrl), "=d" (xcrh) : "c" (0));
- if ((xcrl & 6) == 6) {
+ unsigned bv = xgetbv_low(0);
+
+ if ((bv & 6) == 6) {
have_avx1 = (c & bit_AVX) != 0;
have_avx2 = (b7 & bit_AVX2) != 0;
@@ -4164,7 +4161,7 @@ static void tcg_target_init(TCGContext *s)
* check that OPMASK and all extended ZMM state are enabled
* even if we're not using them -- the insns will fault.
*/
- if ((xcrl & 0xe0) == 0xe0
+ if ((bv & 0xe0) == 0xe0
&& (b7 & bit_AVX512F)
&& (b7 & bit_AVX512VL)) {
have_avx512vl = true;
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* Re: [PATCH for-8.0 01/29] include/qemu/cpuid: Introduce xgetbv_low
2022-11-18 9:47 ` [PATCH for-8.0 01/29] include/qemu/cpuid: Introduce xgetbv_low Richard Henderson
@ 2022-11-21 12:15 ` Philippe Mathieu-Daudé
0 siblings, 0 replies; 48+ messages in thread
From: Philippe Mathieu-Daudé @ 2022-11-21 12:15 UTC (permalink / raw)
To: Richard Henderson, qemu-devel
On 18/11/22 10:47, Richard Henderson wrote:
> Replace the two uses of asm to expand xgetbv with an inline function.
> Since one of the two has been using the mnemonic, assume that the
> comment about "older versions of the assember" is obsolete, as even
> that is 4 years old.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
> include/qemu/cpuid.h | 7 +++++++
> util/bufferiszero.c | 3 +--
> tcg/i386/tcg-target.c.inc | 11 ++++-------
> 3 files changed, 12 insertions(+), 9 deletions(-)
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
^ permalink raw reply [flat|nested] 48+ messages in thread
* [PATCH for-8.0 02/29] include/exec/memop: Add bits describing atomicity
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
2022-11-18 9:47 ` [PATCH for-8.0 01/29] include/qemu/cpuid: Introduce xgetbv_low Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-18 9:47 ` [PATCH for-8.0 03/29] accel/tcg: Add cpu_in_serial_context Richard Henderson
` (26 subsequent siblings)
28 siblings, 0 replies; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
These bits may be used to describe the precise atomicity
requirements of the guest, which may then be used to
constrain the methods by which it may be emulated by the host.
For instance, the AArch64 LDP (32-bit) instruction changes
semantics with ARMv8.4 LSE2, from
MO_64 | MO_ATMAX_4 | MO_ATOM_IFALIGN
(64-bits, single-copy atomic only on 4 byte units,
nonatomic if not aligned by 4),
to
MO_64 | MO_ATMAX_SIZE | MO_ATOM_WITHIN16
(64-bits, single-copy atomic within a 16 byte block)
The former may be implemented with two 4 byte loads, or
a single 8 byte load if that happens to be efficient on
the host. The latter may not, and may also require a
helper when misaligned.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
include/exec/memop.h | 36 ++++++++++++++++++++++++++++++++++++
1 file changed, 36 insertions(+)
diff --git a/include/exec/memop.h b/include/exec/memop.h
index 25d027434a..04e4048f0b 100644
--- a/include/exec/memop.h
+++ b/include/exec/memop.h
@@ -81,6 +81,42 @@ typedef enum MemOp {
MO_ALIGN_32 = 5 << MO_ASHIFT,
MO_ALIGN_64 = 6 << MO_ASHIFT,
+ /*
+ * MO_ATOM_* describes that atomicity requirements of the operation:
+ * MO_ATOM_IFALIGN: the operation must be single-copy atomic if and
+ * only if it is aligned; if unaligned there is no atomicity.
+ * MO_ATOM_NONE: the operation has no atomicity requirements.
+ * MO_ATOM_SUBALIGN: the operation is single-copy atomic by parts
+ * by the alignment. E.g. if the address is 0 mod 4, then each
+ * 4-byte subobject is single-copy atomic.
+ * This is the atomicity of IBM Power and S390X processors.
+ * MO_ATOM_WITHIN16: the operation is single-copy atomic, even if it
+ * is unaligned, so long as it does not cross a 16-byte boundary;
+ * if it crosses a 16-byte boundary there is no atomicity.
+ * This is the atomicity of Arm FEAT_LSE2.
+ *
+ * MO_ATMAX_* describes the maximum atomicity unit required:
+ * MO_ATMAX_SIZE: the entire operation, i.e. MO_SIZE.
+ * MO_ATMAX_[248]: units of N bytes.
+ *
+ * Note the default (i.e. 0) values are single-copy atomic to the
+ * size of the operation, if aligned. This retains the behaviour
+ * from before these were introduced.
+ */
+ MO_ATOM_SHIFT = 8,
+ MO_ATOM_MASK = 0x3 << MO_ATOM_SHIFT,
+ MO_ATOM_IFALIGN = 0 << MO_ATOM_SHIFT,
+ MO_ATOM_NONE = 1 << MO_ATOM_SHIFT,
+ MO_ATOM_SUBALIGN = 2 << MO_ATOM_SHIFT,
+ MO_ATOM_WITHIN16 = 3 << MO_ATOM_SHIFT,
+
+ MO_ATMAX_SHIFT = 10,
+ MO_ATMAX_MASK = 0x3 << MO_ATMAX_SHIFT,
+ MO_ATMAX_SIZE = 0 << MO_ATMAX_SHIFT,
+ MO_ATMAX_2 = 1 << MO_ATMAX_SHIFT,
+ MO_ATMAX_4 = 2 << MO_ATMAX_SHIFT,
+ MO_ATMAX_8 = 3 << MO_ATMAX_SHIFT,
+
/* Combinations of the above, for ease of use. */
MO_UB = MO_8,
MO_UW = MO_16,
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* [PATCH for-8.0 03/29] accel/tcg: Add cpu_in_serial_context
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
2022-11-18 9:47 ` [PATCH for-8.0 01/29] include/qemu/cpuid: Introduce xgetbv_low Richard Henderson
2022-11-18 9:47 ` [PATCH for-8.0 02/29] include/exec/memop: Add bits describing atomicity Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-18 9:47 ` [PATCH for-8.0 04/29] accel/tcg: Introduce tlb_read_idx Richard Henderson
` (25 subsequent siblings)
28 siblings, 0 replies; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
Like cpu_in_exclusive_context, but also true if
there is no other cpu against which we could race.
Use it in tb_flush as a direct replacement.
Use it in cpu_loop_exit_atomic to ensure that there
is no loop against cpu_exec_step_atomic.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
accel/tcg/internal.h | 5 +++++
accel/tcg/cpu-exec-common.c | 3 +++
accel/tcg/tb-maint.c | 2 +-
3 files changed, 9 insertions(+), 1 deletion(-)
diff --git a/accel/tcg/internal.h b/accel/tcg/internal.h
index cb13bade4f..f06bf58e7a 100644
--- a/accel/tcg/internal.h
+++ b/accel/tcg/internal.h
@@ -119,4 +119,9 @@ static inline target_ulong log_pc(CPUState *cpu, const TranslationBlock *tb)
#endif
}
+static inline bool cpu_in_serial_context(CPUState *cs)
+{
+ return !(cs->tcg_cflags & CF_PARALLEL) || cpu_in_exclusive_context(cs);
+}
+
#endif /* ACCEL_TCG_INTERNAL_H */
diff --git a/accel/tcg/cpu-exec-common.c b/accel/tcg/cpu-exec-common.c
index c7bc8c6efa..2fb4454c7a 100644
--- a/accel/tcg/cpu-exec-common.c
+++ b/accel/tcg/cpu-exec-common.c
@@ -21,6 +21,7 @@
#include "sysemu/cpus.h"
#include "sysemu/tcg.h"
#include "exec/exec-all.h"
+#include "internal.h"
bool tcg_allowed;
@@ -78,6 +79,8 @@ void cpu_loop_exit_restore(CPUState *cpu, uintptr_t pc)
void cpu_loop_exit_atomic(CPUState *cpu, uintptr_t pc)
{
+ /* Prevent looping if already executing in a serial context. */
+ g_assert(!cpu_in_serial_context(cpu));
cpu->exception_index = EXCP_ATOMIC;
cpu_loop_exit_restore(cpu, pc);
}
diff --git a/accel/tcg/tb-maint.c b/accel/tcg/tb-maint.c
index 0cdb35548c..a7c067628c 100644
--- a/accel/tcg/tb-maint.c
+++ b/accel/tcg/tb-maint.c
@@ -119,7 +119,7 @@ void tb_flush(CPUState *cpu)
if (tcg_enabled()) {
unsigned tb_flush_count = qatomic_mb_read(&tb_ctx.tb_flush_count);
- if (cpu_in_exclusive_context(cpu)) {
+ if (cpu_in_serial_context(cpu)) {
do_tb_flush(cpu, RUN_ON_CPU_HOST_INT(tb_flush_count));
} else {
async_safe_run_on_cpu(cpu, do_tb_flush,
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* [PATCH for-8.0 04/29] accel/tcg: Introduce tlb_read_idx
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (2 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 03/29] accel/tcg: Add cpu_in_serial_context Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-21 12:25 ` Philippe Mathieu-Daudé
2022-11-18 9:47 ` [PATCH for-8.0 05/29] accel/tcg: Reorg system mode load helpers Richard Henderson
` (24 subsequent siblings)
28 siblings, 1 reply; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
Instead of playing with offsetof in various places, use
MMUAccessType to index an array. This is easily defined
instead of the previous dummy padding array in the union.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
include/exec/cpu-defs.h | 7 ++-
include/exec/cpu_ldst.h | 26 ++++++++--
accel/tcg/cputlb.c | 104 +++++++++++++---------------------------
3 files changed, 59 insertions(+), 78 deletions(-)
diff --git a/include/exec/cpu-defs.h b/include/exec/cpu-defs.h
index 21309cf567..7ce3bcb06b 100644
--- a/include/exec/cpu-defs.h
+++ b/include/exec/cpu-defs.h
@@ -128,8 +128,11 @@ typedef struct CPUTLBEntry {
use the corresponding iotlb value. */
uintptr_t addend;
};
- /* padding to get a power of two size */
- uint8_t dummy[1 << CPU_TLB_ENTRY_BITS];
+ /*
+ * Padding to get a power of two size, as well as index
+ * access to addr_{read,write,code}.
+ */
+ target_ulong addr_idx[(1 << CPU_TLB_ENTRY_BITS) / TARGET_LONG_SIZE];
};
} CPUTLBEntry;
diff --git a/include/exec/cpu_ldst.h b/include/exec/cpu_ldst.h
index 09b55cc0ee..fad6efc0ad 100644
--- a/include/exec/cpu_ldst.h
+++ b/include/exec/cpu_ldst.h
@@ -360,13 +360,29 @@ static inline void clear_helper_retaddr(void)
/* Needed for TCG_OVERSIZED_GUEST */
#include "tcg/tcg.h"
+static inline target_ulong tlb_read_idx(const CPUTLBEntry *entry,
+ MMUAccessType access_type)
+{
+ /* Do not rearrange the CPUTLBEntry structure members. */
+ QEMU_BUILD_BUG_ON(offsetof(CPUTLBEntry, addr_read) !=
+ MMU_DATA_LOAD * TARGET_LONG_SIZE);
+ QEMU_BUILD_BUG_ON(offsetof(CPUTLBEntry, addr_write) !=
+ MMU_DATA_STORE * TARGET_LONG_SIZE);
+ QEMU_BUILD_BUG_ON(offsetof(CPUTLBEntry, addr_code) !=
+ MMU_INST_FETCH * TARGET_LONG_SIZE);
+
+ const target_ulong *ptr = &entry->addr_idx[access_type];
+#if TCG_OVERSIZED_GUEST
+ return *ptr;
+#else
+ /* ofs might correspond to .addr_write, so use qatomic_read */
+ return qatomic_read(ptr);
+#endif
+}
+
static inline target_ulong tlb_addr_write(const CPUTLBEntry *entry)
{
-#if TCG_OVERSIZED_GUEST
- return entry->addr_write;
-#else
- return qatomic_read(&entry->addr_write);
-#endif
+ return tlb_read_idx(entry, MMU_DATA_STORE);
}
/* Find the TLB index corresponding to the mmu_idx + address pair. */
diff --git a/accel/tcg/cputlb.c b/accel/tcg/cputlb.c
index d177afcad6..00a2b217e5 100644
--- a/accel/tcg/cputlb.c
+++ b/accel/tcg/cputlb.c
@@ -1437,34 +1437,17 @@ static void io_writex(CPUArchState *env, CPUTLBEntryFull *full,
}
}
-static inline target_ulong tlb_read_ofs(CPUTLBEntry *entry, size_t ofs)
-{
-#if TCG_OVERSIZED_GUEST
- return *(target_ulong *)((uintptr_t)entry + ofs);
-#else
- /* ofs might correspond to .addr_write, so use qatomic_read */
- return qatomic_read((target_ulong *)((uintptr_t)entry + ofs));
-#endif
-}
-
/* Return true if ADDR is present in the victim tlb, and has been copied
back to the main tlb. */
static bool victim_tlb_hit(CPUArchState *env, size_t mmu_idx, size_t index,
- size_t elt_ofs, target_ulong page)
+ MMUAccessType access_type, target_ulong page)
{
size_t vidx;
assert_cpu_is_self(env_cpu(env));
for (vidx = 0; vidx < CPU_VTLB_SIZE; ++vidx) {
CPUTLBEntry *vtlb = &env_tlb(env)->d[mmu_idx].vtable[vidx];
- target_ulong cmp;
-
- /* elt_ofs might correspond to .addr_write, so use qatomic_read */
-#if TCG_OVERSIZED_GUEST
- cmp = *(target_ulong *)((uintptr_t)vtlb + elt_ofs);
-#else
- cmp = qatomic_read((target_ulong *)((uintptr_t)vtlb + elt_ofs));
-#endif
+ target_ulong cmp = tlb_read_idx(vtlb, access_type);
if (cmp == page) {
/* Found entry in victim tlb, swap tlb and iotlb. */
@@ -1486,11 +1469,6 @@ static bool victim_tlb_hit(CPUArchState *env, size_t mmu_idx, size_t index,
return false;
}
-/* Macro to call the above, with local variables from the use context. */
-#define VICTIM_TLB_HIT(TY, ADDR) \
- victim_tlb_hit(env, mmu_idx, index, offsetof(CPUTLBEntry, TY), \
- (ADDR) & TARGET_PAGE_MASK)
-
static void notdirty_write(CPUState *cpu, vaddr mem_vaddr, unsigned size,
CPUTLBEntryFull *full, uintptr_t retaddr)
{
@@ -1526,29 +1504,12 @@ static int probe_access_internal(CPUArchState *env, target_ulong addr,
{
uintptr_t index = tlb_index(env, mmu_idx, addr);
CPUTLBEntry *entry = tlb_entry(env, mmu_idx, addr);
- target_ulong tlb_addr, page_addr;
- size_t elt_ofs;
- int flags;
+ target_ulong tlb_addr = tlb_read_idx(entry, access_type);
+ target_ulong page_addr = addr & TARGET_PAGE_MASK;
+ int flags = TLB_FLAGS_MASK;
- switch (access_type) {
- case MMU_DATA_LOAD:
- elt_ofs = offsetof(CPUTLBEntry, addr_read);
- break;
- case MMU_DATA_STORE:
- elt_ofs = offsetof(CPUTLBEntry, addr_write);
- break;
- case MMU_INST_FETCH:
- elt_ofs = offsetof(CPUTLBEntry, addr_code);
- break;
- default:
- g_assert_not_reached();
- }
- tlb_addr = tlb_read_ofs(entry, elt_ofs);
-
- flags = TLB_FLAGS_MASK;
- page_addr = addr & TARGET_PAGE_MASK;
if (!tlb_hit_page(tlb_addr, page_addr)) {
- if (!victim_tlb_hit(env, mmu_idx, index, elt_ofs, page_addr)) {
+ if (!victim_tlb_hit(env, mmu_idx, index, access_type, page_addr)) {
CPUState *cs = env_cpu(env);
if (!cs->cc->tcg_ops->tlb_fill(cs, addr, fault_size, access_type,
@@ -1570,7 +1531,7 @@ static int probe_access_internal(CPUArchState *env, target_ulong addr,
*/
flags &= ~TLB_INVALID_MASK;
}
- tlb_addr = tlb_read_ofs(entry, elt_ofs);
+ tlb_addr = tlb_read_idx(entry, access_type);
}
flags &= tlb_addr;
@@ -1784,7 +1745,8 @@ static void *atomic_mmu_lookup(CPUArchState *env, target_ulong addr,
if (prot & PAGE_WRITE) {
tlb_addr = tlb_addr_write(tlbe);
if (!tlb_hit(tlb_addr, addr)) {
- if (!VICTIM_TLB_HIT(addr_write, addr)) {
+ if (!victim_tlb_hit(env, mmu_idx, index, MMU_DATA_STORE,
+ addr & TARGET_PAGE_MASK)) {
tlb_fill(env_cpu(env), addr, size,
MMU_DATA_STORE, mmu_idx, retaddr);
index = tlb_index(env, mmu_idx, addr);
@@ -1808,7 +1770,8 @@ static void *atomic_mmu_lookup(CPUArchState *env, target_ulong addr,
} else /* if (prot & PAGE_READ) */ {
tlb_addr = tlbe->addr_read;
if (!tlb_hit(tlb_addr, addr)) {
- if (!VICTIM_TLB_HIT(addr_write, addr)) {
+ if (!victim_tlb_hit(env, mmu_idx, index, MMU_DATA_LOAD,
+ addr & TARGET_PAGE_MASK)) {
tlb_fill(env_cpu(env), addr, size,
MMU_DATA_LOAD, mmu_idx, retaddr);
index = tlb_index(env, mmu_idx, addr);
@@ -1894,13 +1857,9 @@ load_memop(const void *haddr, MemOp op)
static inline uint64_t QEMU_ALWAYS_INLINE
load_helper(CPUArchState *env, target_ulong addr, MemOpIdx oi,
- uintptr_t retaddr, MemOp op, bool code_read,
+ uintptr_t retaddr, MemOp op, MMUAccessType access_type,
FullLoadHelper *full_load)
{
- const size_t tlb_off = code_read ?
- offsetof(CPUTLBEntry, addr_code) : offsetof(CPUTLBEntry, addr_read);
- const MMUAccessType access_type =
- code_read ? MMU_INST_FETCH : MMU_DATA_LOAD;
const unsigned a_bits = get_alignment_bits(get_memop(oi));
const size_t size = memop_size(op);
uintptr_t mmu_idx = get_mmuidx(oi);
@@ -1920,18 +1879,18 @@ load_helper(CPUArchState *env, target_ulong addr, MemOpIdx oi,
index = tlb_index(env, mmu_idx, addr);
entry = tlb_entry(env, mmu_idx, addr);
- tlb_addr = code_read ? entry->addr_code : entry->addr_read;
+ tlb_addr = tlb_read_idx(entry, access_type);
/* If the TLB entry is for a different page, reload and try again. */
if (!tlb_hit(tlb_addr, addr)) {
- if (!victim_tlb_hit(env, mmu_idx, index, tlb_off,
+ if (!victim_tlb_hit(env, mmu_idx, index, access_type,
addr & TARGET_PAGE_MASK)) {
tlb_fill(env_cpu(env), addr, size,
access_type, mmu_idx, retaddr);
index = tlb_index(env, mmu_idx, addr);
entry = tlb_entry(env, mmu_idx, addr);
}
- tlb_addr = code_read ? entry->addr_code : entry->addr_read;
+ tlb_addr = tlb_read_idx(entry, access_type);
tlb_addr &= ~TLB_INVALID_MASK;
}
@@ -2017,7 +1976,8 @@ static uint64_t full_ldub_mmu(CPUArchState *env, target_ulong addr,
MemOpIdx oi, uintptr_t retaddr)
{
validate_memop(oi, MO_UB);
- return load_helper(env, addr, oi, retaddr, MO_UB, false, full_ldub_mmu);
+ return load_helper(env, addr, oi, retaddr, MO_UB, MMU_DATA_LOAD,
+ full_ldub_mmu);
}
tcg_target_ulong helper_ret_ldub_mmu(CPUArchState *env, target_ulong addr,
@@ -2030,7 +1990,7 @@ static uint64_t full_le_lduw_mmu(CPUArchState *env, target_ulong addr,
MemOpIdx oi, uintptr_t retaddr)
{
validate_memop(oi, MO_LEUW);
- return load_helper(env, addr, oi, retaddr, MO_LEUW, false,
+ return load_helper(env, addr, oi, retaddr, MO_LEUW, MMU_DATA_LOAD,
full_le_lduw_mmu);
}
@@ -2044,7 +2004,7 @@ static uint64_t full_be_lduw_mmu(CPUArchState *env, target_ulong addr,
MemOpIdx oi, uintptr_t retaddr)
{
validate_memop(oi, MO_BEUW);
- return load_helper(env, addr, oi, retaddr, MO_BEUW, false,
+ return load_helper(env, addr, oi, retaddr, MO_BEUW, MMU_DATA_LOAD,
full_be_lduw_mmu);
}
@@ -2058,7 +2018,7 @@ static uint64_t full_le_ldul_mmu(CPUArchState *env, target_ulong addr,
MemOpIdx oi, uintptr_t retaddr)
{
validate_memop(oi, MO_LEUL);
- return load_helper(env, addr, oi, retaddr, MO_LEUL, false,
+ return load_helper(env, addr, oi, retaddr, MO_LEUL, MMU_DATA_LOAD,
full_le_ldul_mmu);
}
@@ -2072,7 +2032,7 @@ static uint64_t full_be_ldul_mmu(CPUArchState *env, target_ulong addr,
MemOpIdx oi, uintptr_t retaddr)
{
validate_memop(oi, MO_BEUL);
- return load_helper(env, addr, oi, retaddr, MO_BEUL, false,
+ return load_helper(env, addr, oi, retaddr, MO_BEUL, MMU_DATA_LOAD,
full_be_ldul_mmu);
}
@@ -2086,7 +2046,7 @@ uint64_t helper_le_ldq_mmu(CPUArchState *env, target_ulong addr,
MemOpIdx oi, uintptr_t retaddr)
{
validate_memop(oi, MO_LEUQ);
- return load_helper(env, addr, oi, retaddr, MO_LEUQ, false,
+ return load_helper(env, addr, oi, retaddr, MO_LEUQ, MMU_DATA_LOAD,
helper_le_ldq_mmu);
}
@@ -2094,7 +2054,7 @@ uint64_t helper_be_ldq_mmu(CPUArchState *env, target_ulong addr,
MemOpIdx oi, uintptr_t retaddr)
{
validate_memop(oi, MO_BEUQ);
- return load_helper(env, addr, oi, retaddr, MO_BEUQ, false,
+ return load_helper(env, addr, oi, retaddr, MO_BEUQ, MMU_DATA_LOAD,
helper_be_ldq_mmu);
}
@@ -2290,7 +2250,6 @@ store_helper_unaligned(CPUArchState *env, target_ulong addr, uint64_t val,
uintptr_t retaddr, size_t size, uintptr_t mmu_idx,
bool big_endian)
{
- const size_t tlb_off = offsetof(CPUTLBEntry, addr_write);
uintptr_t index, index2;
CPUTLBEntry *entry, *entry2;
target_ulong page1, page2, tlb_addr, tlb_addr2;
@@ -2312,7 +2271,7 @@ store_helper_unaligned(CPUArchState *env, target_ulong addr, uint64_t val,
tlb_addr2 = tlb_addr_write(entry2);
if (page1 != page2 && !tlb_hit_page(tlb_addr2, page2)) {
- if (!victim_tlb_hit(env, mmu_idx, index2, tlb_off, page2)) {
+ if (!victim_tlb_hit(env, mmu_idx, index2, MMU_DATA_STORE, page2)) {
tlb_fill(env_cpu(env), page2, size2, MMU_DATA_STORE,
mmu_idx, retaddr);
index2 = tlb_index(env, mmu_idx, page2);
@@ -2365,7 +2324,6 @@ static inline void QEMU_ALWAYS_INLINE
store_helper(CPUArchState *env, target_ulong addr, uint64_t val,
MemOpIdx oi, uintptr_t retaddr, MemOp op)
{
- const size_t tlb_off = offsetof(CPUTLBEntry, addr_write);
const unsigned a_bits = get_alignment_bits(get_memop(oi));
const size_t size = memop_size(op);
uintptr_t mmu_idx = get_mmuidx(oi);
@@ -2388,7 +2346,7 @@ store_helper(CPUArchState *env, target_ulong addr, uint64_t val,
/* If the TLB entry is for a different page, reload and try again. */
if (!tlb_hit(tlb_addr, addr)) {
- if (!victim_tlb_hit(env, mmu_idx, index, tlb_off,
+ if (!victim_tlb_hit(env, mmu_idx, index, MMU_DATA_STORE,
addr & TARGET_PAGE_MASK)) {
tlb_fill(env_cpu(env), addr, size, MMU_DATA_STORE,
mmu_idx, retaddr);
@@ -2694,7 +2652,8 @@ void cpu_st16_le_mmu(CPUArchState *env, abi_ptr addr, Int128 val,
static uint64_t full_ldub_code(CPUArchState *env, target_ulong addr,
MemOpIdx oi, uintptr_t retaddr)
{
- return load_helper(env, addr, oi, retaddr, MO_8, true, full_ldub_code);
+ return load_helper(env, addr, oi, retaddr, MO_8,
+ MMU_INST_FETCH, full_ldub_code);
}
uint32_t cpu_ldub_code(CPUArchState *env, abi_ptr addr)
@@ -2706,7 +2665,8 @@ uint32_t cpu_ldub_code(CPUArchState *env, abi_ptr addr)
static uint64_t full_lduw_code(CPUArchState *env, target_ulong addr,
MemOpIdx oi, uintptr_t retaddr)
{
- return load_helper(env, addr, oi, retaddr, MO_TEUW, true, full_lduw_code);
+ return load_helper(env, addr, oi, retaddr, MO_TEUW,
+ MMU_INST_FETCH, full_lduw_code);
}
uint32_t cpu_lduw_code(CPUArchState *env, abi_ptr addr)
@@ -2718,7 +2678,8 @@ uint32_t cpu_lduw_code(CPUArchState *env, abi_ptr addr)
static uint64_t full_ldl_code(CPUArchState *env, target_ulong addr,
MemOpIdx oi, uintptr_t retaddr)
{
- return load_helper(env, addr, oi, retaddr, MO_TEUL, true, full_ldl_code);
+ return load_helper(env, addr, oi, retaddr, MO_TEUL,
+ MMU_INST_FETCH, full_ldl_code);
}
uint32_t cpu_ldl_code(CPUArchState *env, abi_ptr addr)
@@ -2730,7 +2691,8 @@ uint32_t cpu_ldl_code(CPUArchState *env, abi_ptr addr)
static uint64_t full_ldq_code(CPUArchState *env, target_ulong addr,
MemOpIdx oi, uintptr_t retaddr)
{
- return load_helper(env, addr, oi, retaddr, MO_TEUQ, true, full_ldq_code);
+ return load_helper(env, addr, oi, retaddr, MO_TEUQ,
+ MMU_INST_FETCH, full_ldq_code);
}
uint64_t cpu_ldq_code(CPUArchState *env, abi_ptr addr)
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* Re: [PATCH for-8.0 04/29] accel/tcg: Introduce tlb_read_idx
2022-11-18 9:47 ` [PATCH for-8.0 04/29] accel/tcg: Introduce tlb_read_idx Richard Henderson
@ 2022-11-21 12:25 ` Philippe Mathieu-Daudé
0 siblings, 0 replies; 48+ messages in thread
From: Philippe Mathieu-Daudé @ 2022-11-21 12:25 UTC (permalink / raw)
To: Richard Henderson, qemu-devel
On 18/11/22 10:47, Richard Henderson wrote:
> Instead of playing with offsetof in various places, use
> MMUAccessType to index an array. This is easily defined
> instead of the previous dummy padding array in the union.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
> include/exec/cpu-defs.h | 7 ++-
> include/exec/cpu_ldst.h | 26 ++++++++--
> accel/tcg/cputlb.c | 104 +++++++++++++---------------------------
> 3 files changed, 59 insertions(+), 78 deletions(-)
Nice.
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
^ permalink raw reply [flat|nested] 48+ messages in thread
* [PATCH for-8.0 05/29] accel/tcg: Reorg system mode load helpers
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (3 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 04/29] accel/tcg: Introduce tlb_read_idx Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-18 9:47 ` [PATCH for-8.0 06/29] accel/tcg: Reorg system mode store helpers Richard Henderson
` (23 subsequent siblings)
28 siblings, 0 replies; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
Instead of trying to unify all operations on uint64_t, pull out
mmu_lookup() to perform the basic tlb hit and resolution.
Create individual functions to handle access by size.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
accel/tcg/cputlb.c | 606 ++++++++++++++++++++++++++++++---------------
1 file changed, 413 insertions(+), 193 deletions(-)
diff --git a/accel/tcg/cputlb.c b/accel/tcg/cputlb.c
index 00a2b217e5..c05647f1ba 100644
--- a/accel/tcg/cputlb.c
+++ b/accel/tcg/cputlb.c
@@ -1699,6 +1699,182 @@ bool tlb_plugin_lookup(CPUState *cpu, target_ulong addr, int mmu_idx,
#endif
+/*
+ * Probe for a load/store operation.
+ * Return the host address and into @flags.
+ */
+
+typedef struct MMULookupPageData {
+ CPUTLBEntryFull *full;
+ void *haddr;
+ target_ulong addr;
+ int flags;
+ int size;
+} MMULookupPageData;
+
+typedef struct MMULookupLocals {
+ MMULookupPageData page[2];
+ MemOp memop;
+ int mmu_idx;
+} MMULookupLocals;
+
+/**
+ * mmu_lookup1: translate one page
+ * @env: cpu context
+ * @data: lookup parameters
+ * @mmu_idx: virtual address context
+ * @access_type: load/store/code
+ * @ra: return address into tcg generated code, or 0
+ *
+ * Resolve the translation for the one page at @data.addr, filling in
+ * the rest of @data with the results. If the translation fails,
+ * tlb_fill will longjmp out. Return true if the softmmu tlb for
+ * @mmu_idx may have resized.
+ */
+static bool mmu_lookup1(CPUArchState *env, MMULookupPageData *data,
+ int mmu_idx, MMUAccessType access_type, uintptr_t ra)
+{
+ target_ulong addr = data->addr;
+ uintptr_t index = tlb_index(env, mmu_idx, addr);
+ CPUTLBEntry *entry = tlb_entry(env, mmu_idx, addr);
+ target_ulong tlb_addr = tlb_read_idx(entry, access_type);
+ bool maybe_resized = false;
+
+ /* If the TLB entry is for a different page, reload and try again. */
+ if (!tlb_hit(tlb_addr, addr)) {
+ if (!victim_tlb_hit(env, mmu_idx, index, access_type,
+ addr & TARGET_PAGE_MASK)) {
+ tlb_fill(env_cpu(env), addr, data->size, access_type, mmu_idx, ra);
+ maybe_resized = true;
+ index = tlb_index(env, mmu_idx, addr);
+ entry = tlb_entry(env, mmu_idx, addr);
+ }
+ tlb_addr = tlb_read_idx(entry, access_type) & ~TLB_INVALID_MASK;
+ }
+
+ data->flags = tlb_addr & TLB_FLAGS_MASK;
+ data->full = &env_tlb(env)->d[mmu_idx].fulltlb[index];
+ /* Compute haddr speculatively; depending on flags it might be invalid. */
+ data->haddr = (void *)((uintptr_t)addr + entry->addend);
+
+ return maybe_resized;
+}
+
+/**
+ * mmu_watch_or_dirty
+ * @env: cpu context
+ * @data: lookup parameters
+ * @access_type: load/store/code
+ * @ra: return address into tcg generated code, or 0
+ *
+ * Trigger watchpoints for @data.addr:@data.size;
+ * record writes to protected clean pages.
+ */
+static void mmu_watch_or_dirty(CPUArchState *env, MMULookupPageData *data,
+ MMUAccessType access_type, uintptr_t ra)
+{
+ CPUTLBEntryFull *full = data->full;
+ target_ulong addr = data->addr;
+ int flags = data->flags;
+ int size = data->size;
+
+ /* On watchpoint hit, this will longjmp out. */
+ if (flags & TLB_WATCHPOINT) {
+ int wp = access_type == MMU_DATA_STORE ? BP_MEM_WRITE : BP_MEM_READ;
+ cpu_check_watchpoint(env_cpu(env), addr, size, full->attrs, wp, ra);
+ flags &= ~TLB_WATCHPOINT;
+ }
+
+ if (flags & TLB_NOTDIRTY) {
+ notdirty_write(env_cpu(env), addr, size, full, ra);
+ flags &= ~TLB_NOTDIRTY;
+ }
+ data->flags = flags;
+}
+
+/**
+ * mmu_lookup: translate page(s)
+ * @env: cpu context
+ * @addr: virtual address
+ * @oi: combined mmu_idx and MemOp
+ * @ra: return address into tcg generated code, or 0
+ * @access_type: load/store/code
+ * @l: output result
+ *
+ * Resolve the translation for the page(s) beginning at @addr, for MemOp.size
+ * bytes. Return true if the lookup crosses a page boundary.
+ */
+static inline bool QEMU_ALWAYS_INLINE
+mmu_lookup(CPUArchState *env, target_ulong addr, MemOpIdx oi, uintptr_t ra,
+ MMUAccessType access_type, MMULookupLocals *l)
+{
+ unsigned a_bits;
+ bool crosspage;
+ int flags;
+
+ l->memop = get_memop(oi);
+ l->mmu_idx = get_mmuidx(oi);
+
+ tcg_debug_assert(l->mmu_idx < NB_MMU_MODES);
+
+ /* Handle CPU specific unaligned behaviour */
+ a_bits = get_alignment_bits(l->memop);
+ if (addr & ((1 << a_bits) - 1)) {
+ cpu_unaligned_access(env_cpu(env), addr, access_type, l->mmu_idx, ra);
+ }
+
+ l->page[0].addr = addr;
+ l->page[0].size = memop_size(l->memop);
+ l->page[1].addr = (addr + l->page[0].size - 1) & TARGET_PAGE_MASK;
+ l->page[1].size = 0;
+ crosspage = (addr ^ l->page[1].addr) & TARGET_PAGE_MASK;
+
+ if (likely(!crosspage)) {
+ mmu_lookup1(env, &l->page[0], l->mmu_idx, access_type, ra);
+
+ flags = l->page[0].flags;
+ if (unlikely(flags & (TLB_WATCHPOINT | TLB_NOTDIRTY))) {
+ mmu_watch_or_dirty(env, &l->page[0], access_type, ra);
+ }
+ } else {
+ /* Finish compute of page crossing. */
+ int size1 = l->page[1].addr - addr;
+ l->page[1].size = l->page[0].size - size1;
+ l->page[0].size = size1;
+
+ /*
+ * Lookup both pages, recognizing exceptions from either. If the
+ * second lookup potentially resized, refresh first CPUTLBEntryFull.
+ */
+ mmu_lookup1(env, &l->page[0], l->mmu_idx, access_type, ra);
+ if (mmu_lookup1(env, &l->page[1], l->mmu_idx, access_type, ra)) {
+ uintptr_t index = tlb_index(env, l->mmu_idx, addr);
+ l->page[0].full = &env_tlb(env)->d[l->mmu_idx].fulltlb[index];
+ }
+
+ flags = l->page[0].flags | l->page[1].flags;
+ if (unlikely(flags & (TLB_WATCHPOINT | TLB_NOTDIRTY))) {
+ mmu_watch_or_dirty(env, &l->page[0], access_type, ra);
+ mmu_watch_or_dirty(env, &l->page[1], access_type, ra);
+ }
+ }
+
+ return crosspage;
+}
+
+/*
+ * Since target/sparc is the only user of TLB_BSWAP, and all
+ * Sparc accesses are aligned, any treatment across two pages
+ * would be arbitrary. Refuse it until there's a use.
+ */
+#define assert_no_tlb_bswap_(F) \
+ tcg_debug_assert((F & TLB_BSWAP) == 0)
+#define assert_no_tlb_bswap \
+ do { \
+ assert_no_tlb_bswap_(l.page[0].flags); \
+ assert_no_tlb_bswap_(l.page[1].flags); \
+ } while (0)
+
/*
* Probe for an atomic operation. Do not allow unaligned operations,
* or io operations to proceed. Return the host address.
@@ -1855,113 +2031,6 @@ load_memop(const void *haddr, MemOp op)
}
}
-static inline uint64_t QEMU_ALWAYS_INLINE
-load_helper(CPUArchState *env, target_ulong addr, MemOpIdx oi,
- uintptr_t retaddr, MemOp op, MMUAccessType access_type,
- FullLoadHelper *full_load)
-{
- const unsigned a_bits = get_alignment_bits(get_memop(oi));
- const size_t size = memop_size(op);
- uintptr_t mmu_idx = get_mmuidx(oi);
- uintptr_t index;
- CPUTLBEntry *entry;
- target_ulong tlb_addr;
- void *haddr;
- uint64_t res;
-
- tcg_debug_assert(mmu_idx < NB_MMU_MODES);
-
- /* Handle CPU specific unaligned behaviour */
- if (addr & ((1 << a_bits) - 1)) {
- cpu_unaligned_access(env_cpu(env), addr, access_type,
- mmu_idx, retaddr);
- }
-
- index = tlb_index(env, mmu_idx, addr);
- entry = tlb_entry(env, mmu_idx, addr);
- tlb_addr = tlb_read_idx(entry, access_type);
-
- /* If the TLB entry is for a different page, reload and try again. */
- if (!tlb_hit(tlb_addr, addr)) {
- if (!victim_tlb_hit(env, mmu_idx, index, access_type,
- addr & TARGET_PAGE_MASK)) {
- tlb_fill(env_cpu(env), addr, size,
- access_type, mmu_idx, retaddr);
- index = tlb_index(env, mmu_idx, addr);
- entry = tlb_entry(env, mmu_idx, addr);
- }
- tlb_addr = tlb_read_idx(entry, access_type);
- tlb_addr &= ~TLB_INVALID_MASK;
- }
-
- /* Handle anything that isn't just a straight memory access. */
- if (unlikely(tlb_addr & ~TARGET_PAGE_MASK)) {
- CPUTLBEntryFull *full;
- bool need_swap;
-
- /* For anything that is unaligned, recurse through full_load. */
- if ((addr & (size - 1)) != 0) {
- goto do_unaligned_access;
- }
-
- full = &env_tlb(env)->d[mmu_idx].fulltlb[index];
-
- /* Handle watchpoints. */
- if (unlikely(tlb_addr & TLB_WATCHPOINT)) {
- /* On watchpoint hit, this will longjmp out. */
- cpu_check_watchpoint(env_cpu(env), addr, size,
- full->attrs, BP_MEM_READ, retaddr);
- }
-
- need_swap = size > 1 && (tlb_addr & TLB_BSWAP);
-
- /* Handle I/O access. */
- if (likely(tlb_addr & TLB_MMIO)) {
- return io_readx(env, full, mmu_idx, addr, retaddr,
- access_type, op ^ (need_swap * MO_BSWAP));
- }
-
- haddr = (void *)((uintptr_t)addr + entry->addend);
-
- /*
- * Keep these two load_memop separate to ensure that the compiler
- * is able to fold the entire function to a single instruction.
- * There is a build-time assert inside to remind you of this. ;-)
- */
- if (unlikely(need_swap)) {
- return load_memop(haddr, op ^ MO_BSWAP);
- }
- return load_memop(haddr, op);
- }
-
- /* Handle slow unaligned access (it spans two pages or IO). */
- if (size > 1
- && unlikely((addr & ~TARGET_PAGE_MASK) + size - 1
- >= TARGET_PAGE_SIZE)) {
- target_ulong addr1, addr2;
- uint64_t r1, r2;
- unsigned shift;
- do_unaligned_access:
- addr1 = addr & ~((target_ulong)size - 1);
- addr2 = addr1 + size;
- r1 = full_load(env, addr1, oi, retaddr);
- r2 = full_load(env, addr2, oi, retaddr);
- shift = (addr & (size - 1)) * 8;
-
- if (memop_big_endian(op)) {
- /* Big-endian combine. */
- res = (r1 << shift) | (r2 >> ((size * 8) - shift));
- } else {
- /* Little-endian combine. */
- res = (r1 >> shift) | (r2 << ((size * 8) - shift));
- }
- return res & MAKE_64BIT_MASK(0, size * 8);
- }
-
- haddr = (void *)((uintptr_t)addr + entry->addend);
- return load_memop(haddr, op);
-}
-
/*
* For the benefit of TCG generated code, we want to avoid the
* complication of ABI-specific return type promotion and always
@@ -1972,90 +2041,240 @@ load_helper(CPUArchState *env, target_ulong addr, MemOpIdx oi,
* We don't bother with this widened value for SOFTMMU_CODE_ACCESS.
*/
-static uint64_t full_ldub_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr)
+/**
+ * do_ld_mmio_beN:
+ * @env: cpu context
+ * @p: translation parameters
+ * @ret_be: accumulated data
+ * @mmu_idx: virtual address context
+ * @ra: return address into tcg generated code, or 0
+ *
+ * Load @p->size bytes from @p->addr, which is memory-mapped i/o.
+ * The bytes are concatenated with in big-endian order with @ret_be.
+ */
+static uint64_t do_ld_mmio_beN(CPUArchState *env, MMULookupPageData *p,
+ uint64_t ret_be, int mmu_idx,
+ MMUAccessType type, uintptr_t ra)
{
- validate_memop(oi, MO_UB);
- return load_helper(env, addr, oi, retaddr, MO_UB, MMU_DATA_LOAD,
- full_ldub_mmu);
+ CPUTLBEntryFull *full = p->full;
+ target_ulong addr = p->addr;
+ int i, size = p->size;
+
+ QEMU_IOTHREAD_LOCK_GUARD();
+ for (i = 0; i < size; i++) {
+ uint8_t x = io_readx(env, full, mmu_idx, addr + i, ra, type, MO_UB);
+ ret_be = (ret_be << 8) | x;
+ }
+ return ret_be;
+}
+
+/**
+ * do_ld_bytes_beN
+ * @p: translation parameters
+ * @ret_be: accumulated data
+ *
+ * Load @p->size bytes from @p->haddr, which is RAM.
+ * The bytes to concatenated in big-endian order with @ret_be.
+ */
+static uint64_t do_ld_bytes_beN(MMULookupPageData *p, uint64_t ret_be)
+{
+ uint8_t *haddr = p->haddr;
+ int i, size = p->size;
+
+ for (i = 0; i < size; i++) {
+ ret_be = (ret_be << 8) | haddr[i];
+ }
+ return ret_be;
+}
+
+/*
+ * Wrapper for the above.
+ */
+static uint64_t do_ld_beN(CPUArchState *env, MMULookupPageData *p,
+ uint64_t ret_be, int mmu_idx,
+ MMUAccessType type, uintptr_t ra)
+{
+ if (unlikely(p->flags & TLB_MMIO)) {
+ return do_ld_mmio_beN(env, p, ret_be, mmu_idx, type, ra);
+ } else {
+ return do_ld_bytes_beN(p, ret_be);
+ }
+}
+
+static uint8_t do_ld_1(CPUArchState *env, MMULookupPageData *p, int mmu_idx,
+ MMUAccessType type, uintptr_t ra)
+{
+ if (unlikely(p->flags & TLB_MMIO)) {
+ return io_readx(env, p->full, mmu_idx, p->addr, ra, type, MO_UB);
+ } else {
+ return *(uint8_t *)p->haddr;
+ }
+}
+
+static uint8_t do_ld1_mmu(CPUArchState *env, target_ulong addr, MemOpIdx oi,
+ uintptr_t ra, MMUAccessType access_type)
+{
+ MMULookupLocals l;
+ bool crosspage;
+
+ crosspage = mmu_lookup(env, addr, oi, ra, access_type, &l);
+ tcg_debug_assert(!crosspage);
+
+ return do_ld_1(env, &l.page[0], l.mmu_idx, access_type, ra);
}
tcg_target_ulong helper_ret_ldub_mmu(CPUArchState *env, target_ulong addr,
MemOpIdx oi, uintptr_t retaddr)
{
- return full_ldub_mmu(env, addr, oi, retaddr);
+ validate_memop(oi, MO_UB);
+ return do_ld1_mmu(env, addr, oi, retaddr, MMU_DATA_LOAD);
}
-static uint64_t full_le_lduw_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr)
+static uint16_t do_ld2_mmu(CPUArchState *env, target_ulong addr, MemOpIdx oi,
+ uintptr_t ra, MMUAccessType access_type)
{
- validate_memop(oi, MO_LEUW);
- return load_helper(env, addr, oi, retaddr, MO_LEUW, MMU_DATA_LOAD,
- full_le_lduw_mmu);
+ MMULookupLocals l;
+ bool crosspage;
+ uint16_t ret;
+
+ crosspage = mmu_lookup(env, addr, oi, ra, access_type, &l);
+ if (likely(!crosspage)) {
+ if (l.page[0].flags & TLB_BSWAP) {
+ l.memop ^= MO_BSWAP;
+ }
+ if (unlikely(l.page[0].flags & TLB_MMIO)) {
+ ret = io_readx(env, l.page[0].full, l.mmu_idx, addr, ra,
+ access_type, l.memop);
+ } else {
+ /* Perform the load host endian, then swap if necessary. */
+ ret = load_memop(l.page[0].haddr, MO_UW);
+ if (l.memop & MO_BSWAP) {
+ ret = bswap16(ret);
+ }
+ }
+ } else {
+ uint8_t a, b;
+
+ assert_no_tlb_bswap;
+
+ a = do_ld_1(env, &l.page[0], l.mmu_idx, access_type, ra);
+ b = do_ld_1(env, &l.page[1], l.mmu_idx, access_type, ra);
+
+ if ((l.memop & MO_BSWAP) == MO_LE) {
+ ret = a | (b << 8);
+ } else {
+ ret = b | (a << 8);
+ }
+ }
+ return ret;
}
tcg_target_ulong helper_le_lduw_mmu(CPUArchState *env, target_ulong addr,
MemOpIdx oi, uintptr_t retaddr)
{
- return full_le_lduw_mmu(env, addr, oi, retaddr);
-}
-
-static uint64_t full_be_lduw_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr)
-{
- validate_memop(oi, MO_BEUW);
- return load_helper(env, addr, oi, retaddr, MO_BEUW, MMU_DATA_LOAD,
- full_be_lduw_mmu);
+ validate_memop(oi, MO_LEUW);
+ return do_ld2_mmu(env, addr, oi, retaddr, MMU_DATA_LOAD);
}
tcg_target_ulong helper_be_lduw_mmu(CPUArchState *env, target_ulong addr,
MemOpIdx oi, uintptr_t retaddr)
{
- return full_be_lduw_mmu(env, addr, oi, retaddr);
+ validate_memop(oi, MO_BEUW);
+ return do_ld2_mmu(env, addr, oi, retaddr, MMU_DATA_LOAD);
}
-static uint64_t full_le_ldul_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr)
+static uint32_t do_ld4_mmu(CPUArchState *env, target_ulong addr, MemOpIdx oi,
+ uintptr_t ra, MMUAccessType access_type)
{
- validate_memop(oi, MO_LEUL);
- return load_helper(env, addr, oi, retaddr, MO_LEUL, MMU_DATA_LOAD,
- full_le_ldul_mmu);
+ MMULookupLocals l;
+ bool crosspage;
+ uint32_t ret;
+
+ crosspage = mmu_lookup(env, addr, oi, ra, access_type, &l);
+ if (likely(!crosspage)) {
+ if (l.page[0].flags & TLB_BSWAP) {
+ l.memop ^= MO_BSWAP;
+ }
+ if (unlikely(l.page[0].flags & TLB_MMIO)) {
+ ret = io_readx(env, l.page[0].full, l.mmu_idx, addr, ra,
+ access_type, l.memop);
+ } else {
+ /* Perform the load host endian. */
+ ret = load_memop(l.page[0].haddr, MO_UL);
+ if (l.memop & MO_BSWAP) {
+ ret = bswap32(ret);
+ }
+ }
+ } else {
+ assert_no_tlb_bswap;
+ ret = do_ld_beN(env, &l.page[0], 0, l.mmu_idx, access_type, ra);
+ ret = do_ld_beN(env, &l.page[1], ret, l.mmu_idx, access_type, ra);
+ if ((l.memop & MO_BSWAP) == MO_LE) {
+ ret = bswap32(ret);
+ }
+ }
+ return ret;
}
tcg_target_ulong helper_le_ldul_mmu(CPUArchState *env, target_ulong addr,
MemOpIdx oi, uintptr_t retaddr)
{
- return full_le_ldul_mmu(env, addr, oi, retaddr);
-}
-
-static uint64_t full_be_ldul_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr)
-{
- validate_memop(oi, MO_BEUL);
- return load_helper(env, addr, oi, retaddr, MO_BEUL, MMU_DATA_LOAD,
- full_be_ldul_mmu);
+ validate_memop(oi, MO_LEUL);
+ return do_ld4_mmu(env, addr, oi, retaddr, MMU_DATA_LOAD);
}
tcg_target_ulong helper_be_ldul_mmu(CPUArchState *env, target_ulong addr,
MemOpIdx oi, uintptr_t retaddr)
{
- return full_be_ldul_mmu(env, addr, oi, retaddr);
+ validate_memop(oi, MO_BEUL);
+ return do_ld4_mmu(env, addr, oi, retaddr, MMU_DATA_LOAD);
+}
+
+static uint64_t do_ld8_mmu(CPUArchState *env, target_ulong addr, MemOpIdx oi,
+ uintptr_t ra, MMUAccessType access_type)
+{
+ MMULookupLocals l;
+ bool crosspage;
+ uint64_t ret;
+
+ crosspage = mmu_lookup(env, addr, oi, ra, access_type, &l);
+ if (likely(!crosspage)) {
+ if (l.page[0].flags & TLB_BSWAP) {
+ l.memop ^= MO_BSWAP;
+ }
+ if (unlikely(l.page[0].flags & TLB_MMIO)) {
+ ret = io_readx(env, l.page[0].full, l.mmu_idx, addr, ra,
+ access_type, l.memop);
+ } else {
+ /* Perform the load host endian. */
+ ret = load_memop(l.page[0].haddr, MO_UQ);
+ if (l.memop & MO_BSWAP) {
+ ret = bswap64(ret);
+ }
+ }
+ } else {
+ assert_no_tlb_bswap;
+ ret = do_ld_beN(env, &l.page[0], 0, l.mmu_idx, access_type, ra);
+ ret = do_ld_beN(env, &l.page[1], ret, l.mmu_idx, access_type, ra);
+ if ((l.memop & MO_BSWAP) == MO_LE) {
+ ret = bswap64(ret);
+ }
+ }
+ return ret;
}
uint64_t helper_le_ldq_mmu(CPUArchState *env, target_ulong addr,
MemOpIdx oi, uintptr_t retaddr)
{
validate_memop(oi, MO_LEUQ);
- return load_helper(env, addr, oi, retaddr, MO_LEUQ, MMU_DATA_LOAD,
- helper_le_ldq_mmu);
+ return do_ld8_mmu(env, addr, oi, retaddr, MMU_DATA_LOAD);
}
uint64_t helper_be_ldq_mmu(CPUArchState *env, target_ulong addr,
MemOpIdx oi, uintptr_t retaddr)
{
validate_memop(oi, MO_BEUQ);
- return load_helper(env, addr, oi, retaddr, MO_BEUQ, MMU_DATA_LOAD,
- helper_be_ldq_mmu);
+ return do_ld8_mmu(env, addr, oi, retaddr, MMU_DATA_LOAD);
}
/*
@@ -2098,56 +2317,85 @@ tcg_target_ulong helper_be_ldsl_mmu(CPUArchState *env, target_ulong addr,
* Load helpers for cpu_ldst.h.
*/
-static inline uint64_t cpu_load_helper(CPUArchState *env, abi_ptr addr,
- MemOpIdx oi, uintptr_t retaddr,
- FullLoadHelper *full_load)
+static void plugin_load_cb(CPUArchState *env, abi_ptr addr, MemOpIdx oi)
{
- uint64_t ret;
-
- ret = full_load(env, addr, oi, retaddr);
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
- return ret;
}
uint8_t cpu_ldb_mmu(CPUArchState *env, abi_ptr addr, MemOpIdx oi, uintptr_t ra)
{
- return cpu_load_helper(env, addr, oi, ra, full_ldub_mmu);
+ uint8_t ret;
+
+ validate_memop(oi, MO_UB);
+ ret = do_ld1_mmu(env, addr, oi, ra, MMU_DATA_LOAD);
+ plugin_load_cb(env, addr, oi);
+ return ret;
}
uint16_t cpu_ldw_be_mmu(CPUArchState *env, abi_ptr addr,
MemOpIdx oi, uintptr_t ra)
{
- return cpu_load_helper(env, addr, oi, ra, full_be_lduw_mmu);
+ uint16_t ret;
+
+ validate_memop(oi, MO_BEUW);
+ ret = do_ld2_mmu(env, addr, oi, ra, MMU_DATA_LOAD);
+ plugin_load_cb(env, addr, oi);
+ return ret;
}
uint32_t cpu_ldl_be_mmu(CPUArchState *env, abi_ptr addr,
MemOpIdx oi, uintptr_t ra)
{
- return cpu_load_helper(env, addr, oi, ra, full_be_ldul_mmu);
+ uint32_t ret;
+
+ validate_memop(oi, MO_BEUL);
+ ret = do_ld4_mmu(env, addr, oi, ra, MMU_DATA_LOAD);
+ plugin_load_cb(env, addr, oi);
+ return ret;
}
uint64_t cpu_ldq_be_mmu(CPUArchState *env, abi_ptr addr,
MemOpIdx oi, uintptr_t ra)
{
- return cpu_load_helper(env, addr, oi, ra, helper_be_ldq_mmu);
+ uint64_t ret;
+
+ validate_memop(oi, MO_BEUQ);
+ ret = do_ld8_mmu(env, addr, oi, ra, MMU_DATA_LOAD);
+ plugin_load_cb(env, addr, oi);
+ return ret;
}
uint16_t cpu_ldw_le_mmu(CPUArchState *env, abi_ptr addr,
MemOpIdx oi, uintptr_t ra)
{
- return cpu_load_helper(env, addr, oi, ra, full_le_lduw_mmu);
+ uint16_t ret;
+
+ validate_memop(oi, MO_LEUW);
+ ret = do_ld2_mmu(env, addr, oi, ra, MMU_DATA_LOAD);
+ plugin_load_cb(env, addr, oi);
+ return ret;
}
uint32_t cpu_ldl_le_mmu(CPUArchState *env, abi_ptr addr,
MemOpIdx oi, uintptr_t ra)
{
- return cpu_load_helper(env, addr, oi, ra, full_le_ldul_mmu);
+ uint32_t ret;
+
+ validate_memop(oi, MO_LEUL);
+ ret = do_ld4_mmu(env, addr, oi, ra, MMU_DATA_LOAD);
+ plugin_load_cb(env, addr, oi);
+ return ret;
}
uint64_t cpu_ldq_le_mmu(CPUArchState *env, abi_ptr addr,
MemOpIdx oi, uintptr_t ra)
{
- return cpu_load_helper(env, addr, oi, ra, helper_le_ldq_mmu);
+ uint64_t ret;
+
+ validate_memop(oi, MO_LEUQ);
+ ret = do_ld8_mmu(env, addr, oi, ra, MMU_DATA_LOAD);
+ plugin_load_cb(env, addr, oi);
+ return ret;
}
Int128 cpu_ld16_be_mmu(CPUArchState *env, abi_ptr addr,
@@ -2649,54 +2897,26 @@ void cpu_st16_le_mmu(CPUArchState *env, abi_ptr addr, Int128 val,
/* Code access functions. */
-static uint64_t full_ldub_code(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr)
-{
- return load_helper(env, addr, oi, retaddr, MO_8,
- MMU_INST_FETCH, full_ldub_code);
-}
-
uint32_t cpu_ldub_code(CPUArchState *env, abi_ptr addr)
{
MemOpIdx oi = make_memop_idx(MO_UB, cpu_mmu_index(env, true));
- return full_ldub_code(env, addr, oi, 0);
-}
-
-static uint64_t full_lduw_code(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr)
-{
- return load_helper(env, addr, oi, retaddr, MO_TEUW,
- MMU_INST_FETCH, full_lduw_code);
+ return do_ld1_mmu(env, addr, oi, 0, MMU_INST_FETCH);
}
uint32_t cpu_lduw_code(CPUArchState *env, abi_ptr addr)
{
MemOpIdx oi = make_memop_idx(MO_TEUW, cpu_mmu_index(env, true));
- return full_lduw_code(env, addr, oi, 0);
-}
-
-static uint64_t full_ldl_code(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr)
-{
- return load_helper(env, addr, oi, retaddr, MO_TEUL,
- MMU_INST_FETCH, full_ldl_code);
+ return do_ld2_mmu(env, addr, oi, 0, MMU_INST_FETCH);
}
uint32_t cpu_ldl_code(CPUArchState *env, abi_ptr addr)
{
MemOpIdx oi = make_memop_idx(MO_TEUL, cpu_mmu_index(env, true));
- return full_ldl_code(env, addr, oi, 0);
-}
-
-static uint64_t full_ldq_code(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr)
-{
- return load_helper(env, addr, oi, retaddr, MO_TEUQ,
- MMU_INST_FETCH, full_ldq_code);
+ return do_ld4_mmu(env, addr, oi, 0, MMU_INST_FETCH);
}
uint64_t cpu_ldq_code(CPUArchState *env, abi_ptr addr)
{
MemOpIdx oi = make_memop_idx(MO_TEUQ, cpu_mmu_index(env, true));
- return full_ldq_code(env, addr, oi, 0);
+ return do_ld8_mmu(env, addr, oi, 0, MMU_INST_FETCH);
}
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* [PATCH for-8.0 06/29] accel/tcg: Reorg system mode store helpers
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (4 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 05/29] accel/tcg: Reorg system mode load helpers Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-18 9:47 ` [PATCH for-8.0 07/29] accel/tcg: Honor atomicity of loads Richard Henderson
` (22 subsequent siblings)
28 siblings, 0 replies; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
Instead of trying to unify all operations on uint64_t, use
mmu_lookup() to perform the basic tlb hit and resolution.
Create individual functions to handle access by size.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
accel/tcg/cputlb.c | 399 ++++++++++++++++++++-------------------------
1 file changed, 181 insertions(+), 218 deletions(-)
diff --git a/accel/tcg/cputlb.c b/accel/tcg/cputlb.c
index c05647f1ba..5562fb82d6 100644
--- a/accel/tcg/cputlb.c
+++ b/accel/tcg/cputlb.c
@@ -2490,322 +2490,285 @@ store_memop(void *haddr, uint64_t val, MemOp op)
}
}
-static void full_stb_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
- MemOpIdx oi, uintptr_t retaddr);
-
-static void __attribute__((noinline))
-store_helper_unaligned(CPUArchState *env, target_ulong addr, uint64_t val,
- uintptr_t retaddr, size_t size, uintptr_t mmu_idx,
- bool big_endian)
+/**
+ * do_st_mmio_leN:
+ * @env: cpu context
+ * @p: translation parameters
+ * @val_le: data to store
+ * @mmu_idx: virtual address context
+ * @ra: return address into tcg generated code, or 0
+ *
+ * Store @p->size bytes at @p->addr, which is memory-mapped i/o.
+ * The bytes to store are extracted in little-endian order from @val_le;
+ * return the bytes of @val_le beyond @p->size that have not been stored.
+ */
+static uint64_t do_st_mmio_leN(CPUArchState *env, MMULookupPageData *p,
+ uint64_t val_le, int mmu_idx, uintptr_t ra)
{
- uintptr_t index, index2;
- CPUTLBEntry *entry, *entry2;
- target_ulong page1, page2, tlb_addr, tlb_addr2;
- MemOpIdx oi;
- size_t size2;
- int i;
+ CPUTLBEntryFull *full = p->full;
+ target_ulong addr = p->addr;
+ int i, size = p->size;
- /*
- * Ensure the second page is in the TLB. Note that the first page
- * is already guaranteed to be filled, and that the second page
- * cannot evict the first. An exception to this rule is PAGE_WRITE_INV
- * handling: the first page could have evicted itself.
- */
- page1 = addr & TARGET_PAGE_MASK;
- page2 = (addr + size) & TARGET_PAGE_MASK;
- size2 = (addr + size) & ~TARGET_PAGE_MASK;
- index2 = tlb_index(env, mmu_idx, page2);
- entry2 = tlb_entry(env, mmu_idx, page2);
-
- tlb_addr2 = tlb_addr_write(entry2);
- if (page1 != page2 && !tlb_hit_page(tlb_addr2, page2)) {
- if (!victim_tlb_hit(env, mmu_idx, index2, MMU_DATA_STORE, page2)) {
- tlb_fill(env_cpu(env), page2, size2, MMU_DATA_STORE,
- mmu_idx, retaddr);
- index2 = tlb_index(env, mmu_idx, page2);
- entry2 = tlb_entry(env, mmu_idx, page2);
- }
- tlb_addr2 = tlb_addr_write(entry2);
+ QEMU_IOTHREAD_LOCK_GUARD();
+ for (i = 0; i < size; i++, val_le >>= 8) {
+ io_writex(env, full, mmu_idx, val_le, addr + i, ra, MO_UB);
}
+ return val_le;
+}
- index = tlb_index(env, mmu_idx, addr);
- entry = tlb_entry(env, mmu_idx, addr);
- tlb_addr = tlb_addr_write(entry);
+/**
+ * do_st_bytes_leN:
+ * @p: translation parameters
+ * @val_le: data to store
+ *
+ * Store @p->size bytes at @p->haddr, which is RAM.
+ * The bytes to store are extracted in little-endian order from @val_le;
+ * return the bytes of @val_le beyond @p->size that have not been stored.
+ */
+static uint64_t do_st_bytes_leN(MMULookupPageData *p, uint64_t val_le)
+{
+ uint8_t *haddr = p->haddr;
+ int i, size = p->size;
- /*
- * Handle watchpoints. Since this may trap, all checks
- * must happen before any store.
- */
- if (unlikely(tlb_addr & TLB_WATCHPOINT)) {
- cpu_check_watchpoint(env_cpu(env), addr, size - size2,
- env_tlb(env)->d[mmu_idx].fulltlb[index].attrs,
- BP_MEM_WRITE, retaddr);
- }
- if (unlikely(tlb_addr2 & TLB_WATCHPOINT)) {
- cpu_check_watchpoint(env_cpu(env), page2, size2,
- env_tlb(env)->d[mmu_idx].fulltlb[index2].attrs,
- BP_MEM_WRITE, retaddr);
+ for (i = 0; i < size; i++, val_le >>= 8) {
+ haddr[i] = val_le;
}
+ return val_le;
+}
- /*
- * XXX: not efficient, but simple.
- * This loop must go in the forward direction to avoid issues
- * with self-modifying code in Windows 64-bit.
- */
- oi = make_memop_idx(MO_UB, mmu_idx);
- if (big_endian) {
- for (i = 0; i < size; ++i) {
- /* Big-endian extract. */
- uint8_t val8 = val >> (((size - 1) * 8) - (i * 8));
- full_stb_mmu(env, addr + i, val8, oi, retaddr);
- }
+/*
+ * Wrapper for the above.
+ */
+static uint64_t do_st_leN(CPUArchState *env, MMULookupPageData *p,
+ uint64_t val_le, int mmu_idx, uintptr_t ra)
+{
+ if (unlikely(p->flags & TLB_MMIO)) {
+ return do_st_mmio_leN(env, p, val_le, mmu_idx, ra);
} else {
- for (i = 0; i < size; ++i) {
- /* Little-endian extract. */
- uint8_t val8 = val >> (i * 8);
- full_stb_mmu(env, addr + i, val8, oi, retaddr);
- }
+ return do_st_bytes_leN(p, val_le);
}
}
-static inline void QEMU_ALWAYS_INLINE
-store_helper(CPUArchState *env, target_ulong addr, uint64_t val,
- MemOpIdx oi, uintptr_t retaddr, MemOp op)
+static void do_st_1(CPUArchState *env, MMULookupPageData *p, uint8_t val,
+ int mmu_idx, uintptr_t ra)
{
- const unsigned a_bits = get_alignment_bits(get_memop(oi));
- const size_t size = memop_size(op);
- uintptr_t mmu_idx = get_mmuidx(oi);
- uintptr_t index;
- CPUTLBEntry *entry;
- target_ulong tlb_addr;
- void *haddr;
-
- tcg_debug_assert(mmu_idx < NB_MMU_MODES);
-
- /* Handle CPU specific unaligned behaviour */
- if (addr & ((1 << a_bits) - 1)) {
- cpu_unaligned_access(env_cpu(env), addr, MMU_DATA_STORE,
- mmu_idx, retaddr);
+ if (unlikely(p->flags & TLB_MMIO)) {
+ io_writex(env, p->full, mmu_idx, val, p->addr, ra, MO_UB);
+ } else {
+ *(uint8_t *)p->haddr = val;
}
-
- index = tlb_index(env, mmu_idx, addr);
- entry = tlb_entry(env, mmu_idx, addr);
- tlb_addr = tlb_addr_write(entry);
-
- /* If the TLB entry is for a different page, reload and try again. */
- if (!tlb_hit(tlb_addr, addr)) {
- if (!victim_tlb_hit(env, mmu_idx, index, MMU_DATA_STORE,
- addr & TARGET_PAGE_MASK)) {
- tlb_fill(env_cpu(env), addr, size, MMU_DATA_STORE,
- mmu_idx, retaddr);
- index = tlb_index(env, mmu_idx, addr);
- entry = tlb_entry(env, mmu_idx, addr);
- }
- tlb_addr = tlb_addr_write(entry) & ~TLB_INVALID_MASK;
- }
-
- /* Handle anything that isn't just a straight memory access. */
- if (unlikely(tlb_addr & ~TARGET_PAGE_MASK)) {
- CPUTLBEntryFull *full;
- bool need_swap;
-
- /* For anything that is unaligned, recurse through byte stores. */
- if ((addr & (size - 1)) != 0) {
- goto do_unaligned_access;
- }
-
- full = &env_tlb(env)->d[mmu_idx].fulltlb[index];
-
- /* Handle watchpoints. */
- if (unlikely(tlb_addr & TLB_WATCHPOINT)) {
- /* On watchpoint hit, this will longjmp out. */
- cpu_check_watchpoint(env_cpu(env), addr, size,
- full->attrs, BP_MEM_WRITE, retaddr);
- }
-
- need_swap = size > 1 && (tlb_addr & TLB_BSWAP);
-
- /* Handle I/O access. */
- if (tlb_addr & TLB_MMIO) {
- io_writex(env, full, mmu_idx, val, addr, retaddr,
- op ^ (need_swap * MO_BSWAP));
- return;
- }
-
- /* Ignore writes to ROM. */
- if (unlikely(tlb_addr & TLB_DISCARD_WRITE)) {
- return;
- }
-
- /* Handle clean RAM pages. */
- if (tlb_addr & TLB_NOTDIRTY) {
- notdirty_write(env_cpu(env), addr, size, full, retaddr);
- }
-
- haddr = (void *)((uintptr_t)addr + entry->addend);
-
- /*
- * Keep these two store_memop separate to ensure that the compiler
- * is able to fold the entire function to a single instruction.
- * There is a build-time assert inside to remind you of this. ;-)
- */
- if (unlikely(need_swap)) {
- store_memop(haddr, val, op ^ MO_BSWAP);
- } else {
- store_memop(haddr, val, op);
- }
- return;
- }
-
- /* Handle slow unaligned access (it spans two pages or IO). */
- if (size > 1
- && unlikely((addr & ~TARGET_PAGE_MASK) + size - 1
- >= TARGET_PAGE_SIZE)) {
- do_unaligned_access:
- store_helper_unaligned(env, addr, val, retaddr, size,
- mmu_idx, memop_big_endian(op));
- return;
- }
-
- haddr = (void *)((uintptr_t)addr + entry->addend);
- store_memop(haddr, val, op);
-}
-
-static void __attribute__((noinline))
-full_stb_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
- MemOpIdx oi, uintptr_t retaddr)
-{
- validate_memop(oi, MO_UB);
- store_helper(env, addr, val, oi, retaddr, MO_UB);
}
void helper_ret_stb_mmu(CPUArchState *env, target_ulong addr, uint8_t val,
- MemOpIdx oi, uintptr_t retaddr)
+ MemOpIdx oi, uintptr_t ra)
{
- full_stb_mmu(env, addr, val, oi, retaddr);
+ MMULookupLocals l;
+ bool crosspage;
+
+ validate_memop(oi, MO_UB);
+ crosspage = mmu_lookup(env, addr, oi, ra, MMU_DATA_STORE, &l);
+ tcg_debug_assert(!crosspage);
+
+ do_st_1(env, &l.page[0], val, l.mmu_idx, ra);
}
-static void full_le_stw_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
- MemOpIdx oi, uintptr_t retaddr)
+static void do_st2_mmu(CPUArchState *env, target_ulong addr, uint16_t val,
+ MemOpIdx oi, uintptr_t ra)
{
- validate_memop(oi, MO_LEUW);
- store_helper(env, addr, val, oi, retaddr, MO_LEUW);
+ MMULookupLocals l;
+ bool crosspage;
+
+ crosspage = mmu_lookup(env, addr, oi, ra, MMU_DATA_STORE, &l);
+ if (likely(!crosspage)) {
+ if (l.page[0].flags & TLB_BSWAP) {
+ l.memop ^= MO_BSWAP;
+ }
+ if (unlikely(l.page[0].flags & TLB_MMIO)) {
+ io_writex(env, l.page[0].full, l.mmu_idx, val, addr, ra, l.memop);
+ } else {
+ /* Swap to host endian if necessary, then store. */
+ if (l.memop & MO_BSWAP) {
+ val = bswap16(val);
+ }
+ store_memop(l.page[0].haddr, val, MO_UW);
+ }
+ } else {
+ uint8_t a, b;
+
+ assert_no_tlb_bswap;
+
+ if ((l.memop & MO_BSWAP) == MO_LE) {
+ a = val, b = val >> 8;
+ } else {
+ b = val, a = val >> 8;
+ }
+ do_st_1(env, &l.page[0], a, l.mmu_idx, ra);
+ do_st_1(env, &l.page[1], b, l.mmu_idx, ra);
+ }
}
void helper_le_stw_mmu(CPUArchState *env, target_ulong addr, uint16_t val,
MemOpIdx oi, uintptr_t retaddr)
{
- full_le_stw_mmu(env, addr, val, oi, retaddr);
-}
-
-static void full_be_stw_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
- MemOpIdx oi, uintptr_t retaddr)
-{
- validate_memop(oi, MO_BEUW);
- store_helper(env, addr, val, oi, retaddr, MO_BEUW);
+ validate_memop(oi, MO_LEUW);
+ do_st2_mmu(env, addr, val, oi, retaddr);
}
void helper_be_stw_mmu(CPUArchState *env, target_ulong addr, uint16_t val,
MemOpIdx oi, uintptr_t retaddr)
{
- full_be_stw_mmu(env, addr, val, oi, retaddr);
+ validate_memop(oi, MO_BEUW);
+ do_st2_mmu(env, addr, val, oi, retaddr);
}
-static void full_le_stl_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
- MemOpIdx oi, uintptr_t retaddr)
+static void do_st4_mmu(CPUArchState *env, target_ulong addr, uint32_t val,
+ MemOpIdx oi, uintptr_t ra)
{
- validate_memop(oi, MO_LEUL);
- store_helper(env, addr, val, oi, retaddr, MO_LEUL);
+ MMULookupLocals l;
+ bool crosspage;
+
+ crosspage = mmu_lookup(env, addr, oi, ra, MMU_DATA_STORE, &l);
+ if (likely(!crosspage)) {
+ if (l.page[0].flags & TLB_BSWAP) {
+ l.memop ^= MO_BSWAP;
+ }
+ if (unlikely(l.page[0].flags & TLB_MMIO)) {
+ io_writex(env, l.page[0].full, l.mmu_idx, val, addr, ra, l.memop);
+ } else {
+ /* Swap to host endian if necessary, then store. */
+ if (l.memop & MO_BSWAP) {
+ val = bswap32(val);
+ }
+ store_memop(l.page[0].haddr, val, MO_UL);
+ }
+ } else {
+ assert_no_tlb_bswap;
+
+ /* Swap to little endian for simplicity, then store by bytes. */
+ if ((l.memop & MO_BSWAP) != MO_LE) {
+ val = bswap32(val);
+ }
+ val = do_st_leN(env, &l.page[0], val, l.mmu_idx, ra);
+ (void) do_st_leN(env, &l.page[1], val, l.mmu_idx, ra);
+ }
}
void helper_le_stl_mmu(CPUArchState *env, target_ulong addr, uint32_t val,
MemOpIdx oi, uintptr_t retaddr)
{
- full_le_stl_mmu(env, addr, val, oi, retaddr);
-}
-
-static void full_be_stl_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
- MemOpIdx oi, uintptr_t retaddr)
-{
- validate_memop(oi, MO_BEUL);
- store_helper(env, addr, val, oi, retaddr, MO_BEUL);
+ validate_memop(oi, MO_LEUL);
+ do_st4_mmu(env, addr, val, oi, retaddr);
}
void helper_be_stl_mmu(CPUArchState *env, target_ulong addr, uint32_t val,
MemOpIdx oi, uintptr_t retaddr)
{
- full_be_stl_mmu(env, addr, val, oi, retaddr);
+ validate_memop(oi, MO_BEUL);
+ do_st4_mmu(env, addr, val, oi, retaddr);
+}
+
+static void do_st8_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
+ MemOpIdx oi, uintptr_t ra)
+{
+ MMULookupLocals l;
+ bool crosspage;
+
+ crosspage = mmu_lookup(env, addr, oi, ra, MMU_DATA_STORE, &l);
+ if (likely(!crosspage)) {
+ if (l.page[0].flags & TLB_BSWAP) {
+ l.memop ^= MO_BSWAP;
+ }
+ if (unlikely(l.page[0].flags & TLB_MMIO)) {
+ io_writex(env, l.page[0].full, l.mmu_idx, val, addr, ra, l.memop);
+ } else {
+ /* Swap to host endian if necessary, then store. */
+ if (l.memop & MO_BSWAP) {
+ val = bswap64(val);
+ }
+ store_memop(l.page[0].haddr, val, MO_UQ);
+ }
+ } else {
+ assert_no_tlb_bswap;
+
+ /* Swap to little endian for simplicity, then store by bytes. */
+ if ((l.memop & MO_BSWAP) != MO_LE) {
+ val = bswap64(val);
+ }
+ val = do_st_leN(env, &l.page[0], val, l.mmu_idx, ra);
+ (void) do_st_leN(env, &l.page[1], val, l.mmu_idx, ra);
+ }
}
void helper_le_stq_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
MemOpIdx oi, uintptr_t retaddr)
{
validate_memop(oi, MO_LEUQ);
- store_helper(env, addr, val, oi, retaddr, MO_LEUQ);
+ do_st8_mmu(env, addr, val, oi, retaddr);
}
void helper_be_stq_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
MemOpIdx oi, uintptr_t retaddr)
{
validate_memop(oi, MO_BEUQ);
- store_helper(env, addr, val, oi, retaddr, MO_BEUQ);
+ do_st8_mmu(env, addr, val, oi, retaddr);
}
/*
* Store Helpers for cpu_ldst.h
*/
-typedef void FullStoreHelper(CPUArchState *env, target_ulong addr,
- uint64_t val, MemOpIdx oi, uintptr_t retaddr);
-
-static inline void cpu_store_helper(CPUArchState *env, target_ulong addr,
- uint64_t val, MemOpIdx oi, uintptr_t ra,
- FullStoreHelper *full_store)
+static void plugin_store_cb(CPUArchState *env, abi_ptr addr, MemOpIdx oi)
{
- full_store(env, addr, val, oi, ra);
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
}
void cpu_stb_mmu(CPUArchState *env, target_ulong addr, uint8_t val,
MemOpIdx oi, uintptr_t retaddr)
{
- cpu_store_helper(env, addr, val, oi, retaddr, full_stb_mmu);
+ helper_ret_stb_mmu(env, addr, val, oi, retaddr);
+ plugin_store_cb(env, addr, oi);
}
void cpu_stw_be_mmu(CPUArchState *env, target_ulong addr, uint16_t val,
MemOpIdx oi, uintptr_t retaddr)
{
- cpu_store_helper(env, addr, val, oi, retaddr, full_be_stw_mmu);
+ helper_be_stw_mmu(env, addr, val, oi, retaddr);
+ plugin_store_cb(env, addr, oi);
}
void cpu_stl_be_mmu(CPUArchState *env, target_ulong addr, uint32_t val,
MemOpIdx oi, uintptr_t retaddr)
{
- cpu_store_helper(env, addr, val, oi, retaddr, full_be_stl_mmu);
+ helper_be_stl_mmu(env, addr, val, oi, retaddr);
+ plugin_store_cb(env, addr, oi);
}
void cpu_stq_be_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
MemOpIdx oi, uintptr_t retaddr)
{
- cpu_store_helper(env, addr, val, oi, retaddr, helper_be_stq_mmu);
+ helper_be_stq_mmu(env, addr, val, oi, retaddr);
+ plugin_store_cb(env, addr, oi);
}
void cpu_stw_le_mmu(CPUArchState *env, target_ulong addr, uint16_t val,
MemOpIdx oi, uintptr_t retaddr)
{
- cpu_store_helper(env, addr, val, oi, retaddr, full_le_stw_mmu);
+ helper_le_stw_mmu(env, addr, val, oi, retaddr);
+ plugin_store_cb(env, addr, oi);
}
void cpu_stl_le_mmu(CPUArchState *env, target_ulong addr, uint32_t val,
MemOpIdx oi, uintptr_t retaddr)
{
- cpu_store_helper(env, addr, val, oi, retaddr, full_le_stl_mmu);
+ helper_le_stl_mmu(env, addr, val, oi, retaddr);
+ plugin_store_cb(env, addr, oi);
}
void cpu_stq_le_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
MemOpIdx oi, uintptr_t retaddr)
{
- cpu_store_helper(env, addr, val, oi, retaddr, helper_le_stq_mmu);
+ helper_le_stq_mmu(env, addr, val, oi, retaddr);
+ plugin_store_cb(env, addr, oi);
}
void cpu_st16_be_mmu(CPUArchState *env, abi_ptr addr, Int128 val,
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* [PATCH for-8.0 07/29] accel/tcg: Honor atomicity of loads
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (5 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 06/29] accel/tcg: Reorg system mode store helpers Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-22 14:35 ` Peter Maydell
2022-11-18 9:47 ` [PATCH for-8.0 08/29] accel/tcg: Honor atomicity of stores Richard Henderson
` (21 subsequent siblings)
28 siblings, 1 reply; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
Create ldst_atomicity.c.inc.
Not required for user-only code loads, because we've ensured that
the page is read-only before beginning to translate code.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
accel/tcg/cputlb.c | 174 ++++++++---
accel/tcg/user-exec.c | 26 +-
accel/tcg/ldst_atomicity.c.inc | 546 +++++++++++++++++++++++++++++++++
3 files changed, 695 insertions(+), 51 deletions(-)
create mode 100644 accel/tcg/ldst_atomicity.c.inc
diff --git a/accel/tcg/cputlb.c b/accel/tcg/cputlb.c
index 5562fb82d6..cdc109b473 100644
--- a/accel/tcg/cputlb.c
+++ b/accel/tcg/cputlb.c
@@ -1651,6 +1651,9 @@ tb_page_addr_t get_page_addr_code_hostp(CPUArchState *env, target_ulong addr,
return qemu_ram_addr_from_host_nofail(p);
}
+/* Load/store with atomicity primitives. */
+#include "ldst_atomicity.c.inc"
+
#ifdef CONFIG_PLUGIN
/*
* Perform a TLB lookup and populate the qemu_plugin_hwaddr structure.
@@ -2003,35 +2006,7 @@ static void validate_memop(MemOpIdx oi, MemOp expected)
* specifically for reading instructions from system memory. It is
* called by the translation loop and in some helpers where the code
* is disassembled. It shouldn't be called directly by guest code.
- */
-
-typedef uint64_t FullLoadHelper(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr);
-
-static inline uint64_t QEMU_ALWAYS_INLINE
-load_memop(const void *haddr, MemOp op)
-{
- switch (op) {
- case MO_UB:
- return ldub_p(haddr);
- case MO_BEUW:
- return lduw_be_p(haddr);
- case MO_LEUW:
- return lduw_le_p(haddr);
- case MO_BEUL:
- return (uint32_t)ldl_be_p(haddr);
- case MO_LEUL:
- return (uint32_t)ldl_le_p(haddr);
- case MO_BEUQ:
- return ldq_be_p(haddr);
- case MO_LEUQ:
- return ldq_le_p(haddr);
- default:
- qemu_build_not_reached();
- }
-}
-
-/*
+ *
* For the benefit of TCG generated code, we want to avoid the
* complication of ABI-specific return type promotion and always
* return a value extended to the register size of the host. This is
@@ -2087,17 +2062,134 @@ static uint64_t do_ld_bytes_beN(MMULookupPageData *p, uint64_t ret_be)
return ret_be;
}
+/**
+ * do_ld_parts_beN
+ * @p: translation parameters
+ * @ret_be: accumulated data
+ *
+ * As do_ld_bytes_beN, but atomically on each aligned part.
+ */
+static uint64_t do_ld_parts_beN(MMULookupPageData *p, uint64_t ret_be)
+{
+ void *haddr = p->haddr;
+ int size = p->size;
+
+ do {
+ uint64_t x;
+ int n;
+
+ /*
+ * Find minimum of alignment and size.
+ * This is slightly stronger than required by MO_ATOM_SUBALIGN, which
+ * would have only checked the low bits of addr|size once at the start,
+ * but is just as easy.
+ */
+ switch (((uintptr_t)haddr | size) & 7) {
+ case 4:
+ x = cpu_to_be32(load_atomic4(haddr));
+ ret_be = (ret_be << 32) | x;
+ n = 4;
+ break;
+ case 2:
+ case 6:
+ x = cpu_to_be16(load_atomic2(haddr));
+ ret_be = (ret_be << 16) | x;
+ n = 2;
+ break;
+ default:
+ x = *(uint8_t *)haddr;
+ ret_be = (ret_be << 8) | x;
+ n = 1;
+ break;
+ case 0:
+ g_assert_not_reached();
+ }
+ haddr += n;
+ size -= n;
+ } while (size != 0);
+ return ret_be;
+}
+
+/**
+ * do_ld_parts_be4
+ * @p: translation parameters
+ * @ret_be: accumulated data
+ *
+ * As do_ld_bytes_beN, but with one atomic load.
+ * Four aligned bytes are guaranteed to cover the load.
+ */
+static uint64_t do_ld_whole_be4(MMULookupPageData *p, uint64_t ret_be)
+{
+ int o = p->addr & 3;
+ uint32_t x = load_atomic4(p->haddr - o);
+
+ x = cpu_to_be32(x);
+ x <<= o * 8;
+ x >>= (4 - p->size) * 8;
+ return (ret_be << (p->size * 8)) | x;
+}
+
+/**
+ * do_ld_parts_be8
+ * @p: translation parameters
+ * @ret_be: accumulated data
+ *
+ * As do_ld_bytes_beN, but with one atomic load.
+ * Eight aligned bytes are guaranteed to cover the load.
+ */
+static uint64_t do_ld_whole_be8(CPUArchState *env, uintptr_t ra,
+ MMULookupPageData *p, uint64_t ret_be)
+{
+ int o = p->addr & 7;
+ uint64_t x = load_atomic8_or_exit(env, ra, p->haddr - o);
+
+ x = cpu_to_be64(x);
+ x <<= o * 8;
+ x >>= (8 - p->size) * 8;
+ return (ret_be << (p->size * 8)) | x;
+}
+
/*
* Wrapper for the above.
*/
static uint64_t do_ld_beN(CPUArchState *env, MMULookupPageData *p,
- uint64_t ret_be, int mmu_idx,
- MMUAccessType type, uintptr_t ra)
+ uint64_t ret_be, int mmu_idx, MMUAccessType type,
+ MemOp mop, uintptr_t ra)
{
+ MemOp atmax;
+
if (unlikely(p->flags & TLB_MMIO)) {
return do_ld_mmio_beN(env, p, ret_be, mmu_idx, type, ra);
- } else {
+ }
+
+ switch (mop & MO_ATOM_MASK) {
+ case MO_ATOM_WITHIN16:
+ /*
+ * It is a given that we cross a page and therefore there is no
+ * atomicity for the load as a whole, but there may be a subobject
+ * as defined by ATMAX which does not cross a 16-byte boundary.
+ */
+ atmax = mop & MO_ATMAX_MASK;
+ if (atmax == MO_ATMAX_SIZE) {
+ atmax = mop & MO_SIZE;
+ } else {
+ atmax >>= MO_ATMAX_SHIFT;
+ }
+ if (unlikely(p->size >= (1 << atmax))) {
+ if (!HAVE_al8_fast && p->size < 4) {
+ return do_ld_whole_be4(p, ret_be);
+ } else {
+ return do_ld_whole_be8(env, ra, p, ret_be);
+ }
+ }
+ /* fall through */
+ case MO_ATOM_IFALIGN:
+ case MO_ATOM_NONE:
return do_ld_bytes_beN(p, ret_be);
+ case MO_ATOM_SUBALIGN:
+ return do_ld_parts_beN(p, ret_be);
+ default:
+ g_assert_not_reached();
}
}
@@ -2147,7 +2239,7 @@ static uint16_t do_ld2_mmu(CPUArchState *env, target_ulong addr, MemOpIdx oi,
access_type, l.memop);
} else {
/* Perform the load host endian, then swap if necessary. */
- ret = load_memop(l.page[0].haddr, MO_UW);
+ ret = load_atom_2(env, ra, l.page[0].haddr, l.memop);
if (l.memop & MO_BSWAP) {
ret = bswap16(ret);
}
@@ -2200,15 +2292,17 @@ static uint32_t do_ld4_mmu(CPUArchState *env, target_ulong addr, MemOpIdx oi,
access_type, l.memop);
} else {
/* Perform the load host endian. */
- ret = load_memop(l.page[0].haddr, MO_UL);
+ ret = load_atom_4(env, ra, l.page[0].haddr, l.memop);
if (l.memop & MO_BSWAP) {
ret = bswap32(ret);
}
}
} else {
assert_no_tlb_bswap;
- ret = do_ld_beN(env, &l.page[0], 0, l.mmu_idx, access_type, ra);
- ret = do_ld_beN(env, &l.page[1], ret, l.mmu_idx, access_type, ra);
+ ret = do_ld_beN(env, &l.page[0], 0, l.mmu_idx,
+ access_type, l.memop, ra);
+ ret = do_ld_beN(env, &l.page[1], ret, l.mmu_idx,
+ access_type, l.memop, ra);
if ((l.memop & MO_BSWAP) == MO_LE) {
ret = bswap32(ret);
}
@@ -2247,15 +2341,17 @@ static uint64_t do_ld8_mmu(CPUArchState *env, target_ulong addr, MemOpIdx oi,
access_type, l.memop);
} else {
/* Perform the load host endian. */
- ret = load_memop(l.page[0].haddr, MO_UQ);
+ ret = load_atom_8(env, ra, l.page[0].haddr, l.memop);
if (l.memop & MO_BSWAP) {
ret = bswap64(ret);
}
}
} else {
assert_no_tlb_bswap;
- ret = do_ld_beN(env, &l.page[0], 0, l.mmu_idx, access_type, ra);
- ret = do_ld_beN(env, &l.page[1], ret, l.mmu_idx, access_type, ra);
+ ret = do_ld_beN(env, &l.page[0], 0, l.mmu_idx,
+ access_type, l.memop, ra);
+ ret = do_ld_beN(env, &l.page[1], ret, l.mmu_idx,
+ access_type, l.memop, ra);
if ((l.memop & MO_BSWAP) == MO_LE) {
ret = bswap64(ret);
}
diff --git a/accel/tcg/user-exec.c b/accel/tcg/user-exec.c
index a52c7ef826..ec721e5097 100644
--- a/accel/tcg/user-exec.c
+++ b/accel/tcg/user-exec.c
@@ -296,6 +296,8 @@ static void *cpu_mmu_lookup(CPUArchState *env, target_ulong addr,
return ret;
}
+#include "ldst_atomicity.c.inc"
+
uint8_t cpu_ldb_mmu(CPUArchState *env, abi_ptr addr,
MemOpIdx oi, uintptr_t ra)
{
@@ -318,10 +320,10 @@ uint16_t cpu_ldw_be_mmu(CPUArchState *env, abi_ptr addr,
validate_memop(oi, MO_BEUW);
haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_LOAD);
- ret = lduw_be_p(haddr);
+ ret = load_atom_2(env, ra, haddr, get_memop(oi));
clear_helper_retaddr();
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
- return ret;
+ return cpu_to_be16(ret);
}
uint32_t cpu_ldl_be_mmu(CPUArchState *env, abi_ptr addr,
@@ -332,10 +334,10 @@ uint32_t cpu_ldl_be_mmu(CPUArchState *env, abi_ptr addr,
validate_memop(oi, MO_BEUL);
haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_LOAD);
- ret = ldl_be_p(haddr);
+ ret = load_atom_4(env, ra, haddr, get_memop(oi));
clear_helper_retaddr();
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
- return ret;
+ return cpu_to_be32(ret);
}
uint64_t cpu_ldq_be_mmu(CPUArchState *env, abi_ptr addr,
@@ -346,10 +348,10 @@ uint64_t cpu_ldq_be_mmu(CPUArchState *env, abi_ptr addr,
validate_memop(oi, MO_BEUQ);
haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_LOAD);
- ret = ldq_be_p(haddr);
+ ret = load_atom_8(env, ra, haddr, get_memop(oi));
clear_helper_retaddr();
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
- return ret;
+ return cpu_to_be64(ret);
}
uint16_t cpu_ldw_le_mmu(CPUArchState *env, abi_ptr addr,
@@ -360,10 +362,10 @@ uint16_t cpu_ldw_le_mmu(CPUArchState *env, abi_ptr addr,
validate_memop(oi, MO_LEUW);
haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_LOAD);
- ret = lduw_le_p(haddr);
+ ret = load_atom_2(env, ra, haddr, get_memop(oi));
clear_helper_retaddr();
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
- return ret;
+ return cpu_to_le16(ret);
}
uint32_t cpu_ldl_le_mmu(CPUArchState *env, abi_ptr addr,
@@ -374,10 +376,10 @@ uint32_t cpu_ldl_le_mmu(CPUArchState *env, abi_ptr addr,
validate_memop(oi, MO_LEUL);
haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_LOAD);
- ret = ldl_le_p(haddr);
+ ret = load_atom_4(env, ra, haddr, get_memop(oi));
clear_helper_retaddr();
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
- return ret;
+ return cpu_to_le32(ret);
}
uint64_t cpu_ldq_le_mmu(CPUArchState *env, abi_ptr addr,
@@ -388,10 +390,10 @@ uint64_t cpu_ldq_le_mmu(CPUArchState *env, abi_ptr addr,
validate_memop(oi, MO_LEUQ);
haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_LOAD);
- ret = ldq_le_p(haddr);
+ ret = load_atom_8(env, ra, haddr, get_memop(oi));
clear_helper_retaddr();
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
- return ret;
+ return cpu_to_le64(ret);
}
Int128 cpu_ld16_be_mmu(CPUArchState *env, abi_ptr addr,
diff --git a/accel/tcg/ldst_atomicity.c.inc b/accel/tcg/ldst_atomicity.c.inc
new file mode 100644
index 0000000000..decc9a2a16
--- /dev/null
+++ b/accel/tcg/ldst_atomicity.c.inc
@@ -0,0 +1,546 @@
+/*
+ * Routines common to user and system emulation of load/store.
+ *
+ * Copyright (c) 2022 Linaro, Ltd.
+ *
+ * SPDX-License-Identifier: GPL-2.0-or-later
+ *
+ * This work is licensed under the terms of the GNU GPL, version 2 or later.
+ * See the COPYING file in the top-level directory.
+ */
+
+#ifdef CONFIG_ATOMIC64
+# define HAVE_al8 true
+#else
+# define HAVE_al8 false
+#endif
+#define HAVE_al8_fast (ATOMIC_REG_SIZE >= 8)
+
+#if defined(CONFIG_ATOMIC128)
+# define HAVE_al16_fast true
+#else
+# define HAVE_al16_fast false
+#endif
+
+/**
+ * required_atomicity:
+ *
+ * Return the lg2 bytes of atomicity required by @memop for @p.
+ * If the operation must be split into two operations to be
+ * examined separately for atomicity, return -lg2.
+ */
+static int required_atomicity(CPUArchState *env, uintptr_t p, MemOp memop)
+{
+ int atmax = memop & MO_ATMAX_MASK;
+ int size = memop & MO_SIZE;
+ unsigned tmp;
+
+ if (atmax == MO_ATMAX_SIZE) {
+ atmax = size;
+ } else {
+ atmax >>= MO_ATMAX_SHIFT;
+ }
+
+ switch (memop & MO_ATOM_MASK) {
+ case MO_ATOM_IFALIGN:
+ tmp = (1 << atmax) - 1;
+ if (p & tmp) {
+ return MO_8;
+ }
+ break;
+ case MO_ATOM_NONE:
+ return MO_8;
+ case MO_ATOM_SUBALIGN:
+ tmp = p & -p;
+ if (tmp != 0 && tmp < atmax) {
+ atmax = tmp;
+ }
+ break;
+ case MO_ATOM_WITHIN16:
+ tmp = p & 15;
+ if (tmp + (1 << size) <= 16) {
+ atmax = size;
+ } else if (atmax < size && tmp + (1 << atmax) != 16) {
+ /*
+ * Paired load/store, where the pairs aren't aligned.
+ * One of the two must still be handled atomically.
+ */
+ atmax = -atmax;
+ }
+ break;
+ default:
+ g_assert_not_reached();
+ }
+
+ /*
+ * Here we have the architectural atomicity of the operation.
+ * However, when executing in a serial context, we need no extra
+ * host atomicity in order to avoid racing. This reduction
+ * avoids looping with cpu_loop_exit_atomic.
+ */
+ if (cpu_in_serial_context(env_cpu(env))) {
+ return MO_8;
+ }
+ return atmax;
+}
+
+/**
+ * load_atomic2:
+ * @pv: host address
+ *
+ * Atomically load 2 aligned bytes from @pv.
+ */
+static inline uint16_t load_atomic2(void *pv)
+{
+ uint16_t *p = __builtin_assume_aligned(pv, 2);
+ return qatomic_read(p);
+}
+
+/**
+ * load_atomic4:
+ * @pv: host address
+ *
+ * Atomically load 4 aligned bytes from @pv.
+ */
+static inline uint32_t load_atomic4(void *pv)
+{
+ uint32_t *p = __builtin_assume_aligned(pv, 4);
+ return qatomic_read(p);
+}
+
+/**
+ * load_atomic8:
+ * @pv: host address
+ *
+ * Atomically load 8 aligned bytes from @pv.
+ */
+static inline uint64_t load_atomic8(void *pv)
+{
+ uint64_t *p = __builtin_assume_aligned(pv, 8);
+
+ qemu_build_assert(HAVE_al8);
+ return qatomic_read__nocheck(p);
+}
+
+/**
+ * load_atomic16:
+ * @pv: host address
+ *
+ * Atomically load 16 aligned bytes from @pv.
+ */
+static inline Int128 load_atomic16(void *pv)
+{
+#ifdef CONFIG_ATOMIC128
+ __uint128_t *p = __builtin_assume_aligned(pv, 16);
+ Int128Alias r;
+
+ r.u = qatomic_read__nocheck(p);
+ return r.s;
+#else
+ qemu_build_not_reached();
+#endif
+}
+
+/**
+ * load_atomic8_or_exit:
+ * @env: cpu context
+ * @ra: host unwind address
+ * @pv: host address
+ *
+ * Atomically load 8 aligned bytes from @pv.
+ * If this is not possible, longjmp out to restart serially.
+ */
+static uint64_t load_atomic8_or_exit(CPUArchState *env, uintptr_t ra, void *pv)
+{
+ uint64_t *p = __builtin_assume_aligned(pv, 8);
+
+ if (HAVE_al8) {
+ return load_atomic8(p);
+ }
+
+#ifdef CONFIG_USER_ONLY
+ /*
+ * If the page is not writable, then assume the value is immutable
+ * and requires no locking. This ignores the case of MAP_SHARED with
+ * another process, because the fallback start_exclusive solution
+ * provides no protection across processes.
+ */
+ if (!page_check_range(h2g(p), 8, PAGE_WRITE)) {
+ return *p;
+ }
+#endif
+
+ /* Ultimate fallback: re-execute in serial context. */
+ cpu_loop_exit_atomic(env_cpu(env), ra);
+}
+
+/**
+ * load_atomic16_or_exit:
+ * @env: cpu context
+ * @ra: host unwind address
+ * @pv: host address
+ *
+ * Atomically load 16 aligned bytes from @pv.
+ * If this is not possible, longjmp out to restart serially.
+ */
+static Int128 load_atomic16_or_exit(CPUArchState *env, uintptr_t ra, void *pv)
+{
+ Int128 *p = __builtin_assume_aligned(pv, 16);
+
+ if (HAVE_al16_fast) {
+ return load_atomic16(p);
+ }
+
+#ifdef CONFIG_USER_ONLY
+ /*
+ * We can only use cmpxchg to emulate a load if the page is writable.
+ * If the page is not writable, then assume the value is immutable
+ * and requires no locking. This ignores the case of MAP_SHARED with
+ * another process, because the fallback start_exclusive solution
+ * provides no protection across processes.
+ */
+ if (!page_check_range(h2g(p), 16, PAGE_WRITE)) {
+ return *p;
+ }
+#endif
+
+ /*
+ * In system mode all guest pages are writable, and for user-only
+ * we have just checked writability. Try cmpxchg.
+ */
+#if defined(CONFIG_CMPXCHG128)
+ /* Swap 0 with 0, with the side-effect of returning the old value. */
+ {
+ Int128Alias r;
+ r.u = __sync_val_compare_and_swap_16((__uint128_t *)p, 0, 0);
+ return r.s;
+ }
+#endif
+
+ /* Ultimate fallback: re-execute in serial context. */
+ cpu_loop_exit_atomic(env_cpu(env), ra);
+}
+
+/**
+ * load_atom_extract_al4x2:
+ * @pv: host address
+ *
+ * Load 4 bytes from @p, from two sequential atomic 4-byte loads.
+ */
+static uint32_t load_atom_extract_al4x2(void *pv)
+{
+ uintptr_t pi = (uintptr_t)pv;
+ int sh = (pi & 3) * 8;
+ uint32_t a, b;
+
+ pv = (void *)(pi & ~3);
+ a = load_atomic4(pv);
+ b = load_atomic4(pv + 4);
+
+ if (HOST_BIG_ENDIAN) {
+ return (a << sh) | (b >> (-sh & 31));
+ } else {
+ return (a >> sh) | (b << (-sh & 31));
+ }
+}
+
+/**
+ * load_atom_extract_al8x2:
+ * @pv: host address
+ *
+ * Load 8 bytes from @p, from two sequential atomic 8-byte loads.
+ */
+static uint64_t load_atom_extract_al8x2(void *pv)
+{
+ uintptr_t pi = (uintptr_t)pv;
+ int sh = (pi & 7) * 8;
+ uint64_t a, b;
+
+ pv = (void *)(pi & ~7);
+ a = load_atomic8(pv);
+ b = load_atomic8(pv + 8);
+
+ if (HOST_BIG_ENDIAN) {
+ return (a << sh) | (b >> (-sh & 63));
+ } else {
+ return (a >> sh) | (b << (-sh & 63));
+ }
+}
+
+/**
+ * load_atom_extract_al8:
+ * @pv: host address
+ * @s: object size in bytes, @s <= 4.
+ *
+ * Atomically load @s bytes from @p, when p % s != 0, and [p, p+s-1] does
+ * not cross an 8-byte boundary. This means that we can perform an atomic
+ * 8-byte load and extract.
+ * The value is returned in the low bits of a uint32_t.
+ */
+static uint32_t load_atom_extract_al8(void *pv, int s)
+{
+ uintptr_t pi = (uintptr_t)pv;
+ int o = pi & 7;
+ int shr = (HOST_BIG_ENDIAN ? 8 - s - o : o) * 8;
+
+ pv = (void *)(pi & ~7);
+ return load_atomic8(pv) >> shr;
+}
+
+/**
+ * load_atom_extract_al16_or_exit:
+ * @env: cpu context
+ * @ra: host unwind address
+ * @p: host address
+ * @s: object size in bytes, @s <= 8.
+ *
+ * Atomically load @s bytes from @p, when p % 16 < 8
+ * and p % 16 + s > 8. I.e. does not cross a 16-byte
+ * boundary, but *does* cross an 8-byte boundary.
+ * This is the slow version, so we must have eliminated
+ * any faster load_atom_extract_al8 case.
+ *
+ * If this is not possible, longjmp out to restart serially.
+ */
+static uint64_t load_atom_extract_al16_or_exit(CPUArchState *env, uintptr_t ra,
+ void *pv, int s)
+{
+ uintptr_t pi = (uintptr_t)pv;
+ int o = pi & 7;
+ int shr = (HOST_BIG_ENDIAN ? 16 - s - o : o) * 8;
+ Int128 r;
+
+ /*
+ * Note constraints above: p & 8 must be clear.
+ * Provoke SIGBUS if possible otherwise.
+ */
+ pv = (void *)(pi & ~7);
+ r = load_atomic16_or_exit(env, ra, pv);
+
+ r = int128_urshift(r, shr);
+ return int128_getlo(r);
+}
+
+/**
+ * load_atom_extract_al16_or_al8:
+ * @p: host address
+ * @s: object size in bytes, @s <= 8.
+ *
+ * Load @s bytes from @p, when p % s != 0. If [p, p+s-1] does not
+ * cross an 16-byte boundary then the access must be 16-byte atomic,
+ * otherwise the access must be 8-byte atomic.
+ */
+static inline uint64_t load_atom_extract_al16_or_al8(void *pv, int s)
+{
+#if defined(CONFIG_ATOMIC128)
+ uintptr_t pi = (uintptr_t)pv;
+ int o = pi & 7;
+ int shr = (HOST_BIG_ENDIAN ? 16 - s - o : o) * 8;
+ __uint128_t r;
+
+ pv = (void *)(pi & ~7);
+ if (pi & 8) {
+ uint64_t *p8 = __builtin_assume_aligned(pv, 16, 8);
+ uint64_t a = qatomic_read__nocheck(p8);
+ uint64_t b = qatomic_read__nocheck(p8 + 1);
+
+ if (HOST_BIG_ENDIAN) {
+ r = ((__uint128_t)a << 64) | b;
+ } else {
+ r = ((__uint128_t)b << 64) | a;
+ }
+ } else {
+ __uint128_t *p16 = __builtin_assume_aligned(pv, 16, 0);
+ r = qatomic_read__nocheck(p16);
+ }
+ return r >> shr;
+#else
+ qemu_build_not_reached();
+#endif
+}
+
+/**
+ * load_atom_4_by_2:
+ * @pv: host address
+ *
+ * Load 4 bytes from @pv, with two 2-byte atomic loads.
+ */
+static inline uint32_t load_atom_4_by_2(void *pv)
+{
+ uint32_t a = load_atomic2(pv);
+ uint32_t b = load_atomic2(pv + 2);
+
+ if (HOST_BIG_ENDIAN) {
+ return (a << 16) | b;
+ } else {
+ return (b << 16) | a;
+ }
+}
+
+/**
+ * load_atom_8_by_2:
+ * @pv: host address
+ *
+ * Load 8 bytes from @pv, with four 2-byte atomic loads.
+ */
+static inline uint64_t load_atom_8_by_2(void *pv)
+{
+ uint32_t a = load_atom_4_by_2(pv);
+ uint32_t b = load_atom_4_by_2(pv + 4);
+
+ if (HOST_BIG_ENDIAN) {
+ return ((uint64_t)a << 32) | b;
+ } else {
+ return ((uint64_t)b << 32) | a;
+ }
+}
+
+/**
+ * load_atom_8_by_4:
+ * @pv: host address
+ *
+ * Load 8 bytes from @pv, with two 4-byte atomic loads.
+ */
+static inline uint64_t load_atom_8_by_4(void *pv)
+{
+ uint32_t a = load_atomic4(pv);
+ uint32_t b = load_atomic4(pv + 4);
+
+ if (HOST_BIG_ENDIAN) {
+ return ((uint64_t)a << 32) | b;
+ } else {
+ return ((uint64_t)b << 32) | a;
+ }
+}
+
+/**
+ * load_atom_2:
+ * @p: host address
+ * @memop: the full memory op
+ *
+ * Load 2 bytes from @p, honoring the atomicity of @memop.
+ */
+static uint16_t load_atom_2(CPUArchState *env, uintptr_t ra,
+ void *pv, MemOp memop)
+{
+ uintptr_t pi = (uintptr_t)pv;
+ int atmax;
+
+ if (likely((pi & 1) == 0)) {
+ return load_atomic2(pv);
+ }
+ if (HAVE_al16_fast) {
+ return load_atom_extract_al16_or_al8(pv, 2);
+ }
+
+ atmax = required_atomicity(env, pi, memop);
+ switch (atmax) {
+ case MO_8:
+ return lduw_he_p(pv);
+ case MO_16:
+ /* The only case remaining is MO_ATOM_WITHIN16. */
+ if (!HAVE_al8_fast && (pi & 3) == 1) {
+ /* Big or little endian, we want the middle two bytes. */
+ return load_atomic4(pv - 1) >> 8;
+ }
+ if (unlikely((pi & 15) != 7)) {
+ return load_atom_extract_al8(pv, 2);
+ }
+ return load_atom_extract_al16_or_exit(env, ra, pv, 2);
+ default:
+ g_assert_not_reached();
+ }
+}
+
+/**
+ * load_atom_4:
+ * @p: host address
+ * @memop: the full memory op
+ *
+ * Load 4 bytes from @p, honoring the atomicity of @memop.
+ */
+static uint32_t load_atom_4(CPUArchState *env, uintptr_t ra,
+ void *pv, MemOp memop)
+{
+ uintptr_t pi = (uintptr_t)pv;
+ int atmax;
+
+ if (likely((pi & 3) == 0)) {
+ return load_atomic4(pv);
+ }
+ if (HAVE_al16_fast) {
+ return load_atom_extract_al16_or_al8(pv, 4);
+ }
+
+ atmax = required_atomicity(env, pi, memop);
+ switch (atmax) {
+ case MO_8:
+ case MO_16:
+ case -MO_16:
+ /*
+ * For MO_ATOM_IFALIGN, this is more atomicity than required,
+ * but it's trivially supported on all hosts, better than 4
+ * individual byte loads (when the host requires alignment),
+ * and overlaps with the MO_ATOM_SUBALIGN case of p % 2 == 0.
+ */
+ return load_atom_extract_al4x2(pv);
+ case MO_32:
+ if (!(pi & 4)) {
+ return load_atom_extract_al8(pv, 4);
+ }
+ return load_atom_extract_al16_or_exit(env, ra, pv, 4);
+ default:
+ g_assert_not_reached();
+ }
+}
+
+/**
+ * load_atom_8:
+ * @p: host address
+ * @memop: the full memory op
+ *
+ * Load 8 bytes from @p, honoring the atomicity of @memop.
+ */
+static uint64_t load_atom_8(CPUArchState *env, uintptr_t ra,
+ void *pv, MemOp memop)
+{
+ uintptr_t pi = (uintptr_t)pv;
+ int atmax;
+
+ /*
+ * If the host does not support 8-byte atomics, wait until we have
+ * examined the atomicity parameters below.
+ */
+ if (HAVE_al8 && likely((pi & 7) == 0)) {
+ return load_atomic8(pv);
+ }
+ if (HAVE_al16_fast) {
+ return load_atom_extract_al16_or_al8(pv, 8);
+ }
+
+ atmax = required_atomicity(env, pi, memop);
+ if (atmax == MO_64) {
+ if (!HAVE_al8 && (pi & 7) == 0) {
+ load_atomic8_or_exit(env, ra, pv);
+ }
+ return load_atom_extract_al16_or_exit(env, ra, pv, 8);
+ }
+ if (HAVE_al8_fast) {
+ return load_atom_extract_al8x2(pv);
+ }
+ switch (atmax) {
+ case MO_8:
+ return ldq_he_p(pv);
+ case MO_16:
+ return load_atom_8_by_2(pv);
+ case MO_32:
+ return load_atom_8_by_4(pv);
+ case -MO_32:
+ if (HAVE_al8) {
+ return load_atom_extract_al8x2(pv);
+ }
+ cpu_loop_exit_atomic(env_cpu(env), ra);
+ default:
+ g_assert_not_reached();
+ }
+}
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* Re: [PATCH for-8.0 07/29] accel/tcg: Honor atomicity of loads
2022-11-18 9:47 ` [PATCH for-8.0 07/29] accel/tcg: Honor atomicity of loads Richard Henderson
@ 2022-11-22 14:35 ` Peter Maydell
2022-11-22 18:04 ` Richard Henderson
0 siblings, 1 reply; 48+ messages in thread
From: Peter Maydell @ 2022-11-22 14:35 UTC (permalink / raw)
To: Richard Henderson; +Cc: qemu-devel
On Fri, 18 Nov 2022 at 09:50, Richard Henderson
<richard.henderson@linaro.org> wrote:
>
> Create ldst_atomicity.c.inc.
>
> Not required for user-only code loads, because we've ensured that
> the page is read-only before beginning to translate code.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> +/**
> + * required_atomicity:
> + *
> + * Return the lg2 bytes of atomicity required by @memop for @p.
> + * If the operation must be split into two operations to be
> + * examined separately for atomicity, return -lg2.
> + */
> +static int required_atomicity(CPUArchState *env, uintptr_t p, MemOp memop)
> +{
> + int atmax = memop & MO_ATMAX_MASK;
> + int size = memop & MO_SIZE;
> + unsigned tmp;
> +
> + if (atmax == MO_ATMAX_SIZE) {
> + atmax = size;
> + } else {
> + atmax >>= MO_ATMAX_SHIFT;
> + }
> +
> + switch (memop & MO_ATOM_MASK) {
> + case MO_ATOM_IFALIGN:
> + tmp = (1 << atmax) - 1;
> + if (p & tmp) {
> + return MO_8;
> + }
> + break;
> + case MO_ATOM_NONE:
> + return MO_8;
> + case MO_ATOM_SUBALIGN:
> + tmp = p & -p;
> + if (tmp != 0 && tmp < atmax) {
> + atmax = tmp;
> + }
> + break;
> + case MO_ATOM_WITHIN16:
> + tmp = p & 15;
> + if (tmp + (1 << size) <= 16) {
> + atmax = size;
> + } else if (atmax < size && tmp + (1 << atmax) != 16) {
> + /*
> + * Paired load/store, where the pairs aren't aligned.
> + * One of the two must still be handled atomically.
> + */
> + atmax = -atmax;
> + }
> + break;
> + default:
> + g_assert_not_reached();
> + }
> +
> + /*
> + * Here we have the architectural atomicity of the operation.
> + * However, when executing in a serial context, we need no extra
> + * host atomicity in order to avoid racing. This reduction
> + * avoids looping with cpu_loop_exit_atomic.
> + */
> + if (cpu_in_serial_context(env_cpu(env))) {
Is it OK to use cpu_in_serial_context() here ? Even if
there's no other vCPU executing in parallel, there might
be device model code doing a memory write in the iothread,
I think.
> + return MO_8;
> + }
> + return atmax;
> +}
thanks
-- PMM
^ permalink raw reply [flat|nested] 48+ messages in thread* Re: [PATCH for-8.0 07/29] accel/tcg: Honor atomicity of loads
2022-11-22 14:35 ` Peter Maydell
@ 2022-11-22 18:04 ` Richard Henderson
0 siblings, 0 replies; 48+ messages in thread
From: Richard Henderson @ 2022-11-22 18:04 UTC (permalink / raw)
To: Peter Maydell; +Cc: qemu-devel
On 11/22/22 06:35, Peter Maydell wrote:
>> + /*
>> + * Here we have the architectural atomicity of the operation.
>> + * However, when executing in a serial context, we need no extra
>> + * host atomicity in order to avoid racing. This reduction
>> + * avoids looping with cpu_loop_exit_atomic.
>> + */
>> + if (cpu_in_serial_context(env_cpu(env))) {
>
> Is it OK to use cpu_in_serial_context() here ? Even if
> there's no other vCPU executing in parallel, there might
> be device model code doing a memory write in the iothread,
> I think.
Well, it's no different from how we currently treat compare-and-swap expansion. But you
have a point -- we should probably be doing something with the iothread lock for both
EXCP_ATOMIC and round-robin mode.
r~
^ permalink raw reply [flat|nested] 48+ messages in thread
* [PATCH for-8.0 08/29] accel/tcg: Honor atomicity of stores
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (6 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 07/29] accel/tcg: Honor atomicity of loads Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-18 9:47 ` [PATCH for-8.0 09/29] tcg/tci: Use cpu_{ld,st}_mmu Richard Henderson
` (20 subsequent siblings)
28 siblings, 0 replies; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
accel/tcg/cputlb.c | 177 +++++++++++++----
accel/tcg/user-exec.c | 12 +-
accel/tcg/ldst_atomicity.c.inc | 336 +++++++++++++++++++++++++++++++++
3 files changed, 480 insertions(+), 45 deletions(-)
diff --git a/accel/tcg/cputlb.c b/accel/tcg/cputlb.c
index cdc109b473..69f8a25a7f 100644
--- a/accel/tcg/cputlb.c
+++ b/accel/tcg/cputlb.c
@@ -2556,36 +2556,6 @@ Int128 cpu_ld16_le_mmu(CPUArchState *env, abi_ptr addr,
* Store Helpers
*/
-static inline void QEMU_ALWAYS_INLINE
-store_memop(void *haddr, uint64_t val, MemOp op)
-{
- switch (op) {
- case MO_UB:
- stb_p(haddr, val);
- break;
- case MO_BEUW:
- stw_be_p(haddr, val);
- break;
- case MO_LEUW:
- stw_le_p(haddr, val);
- break;
- case MO_BEUL:
- stl_be_p(haddr, val);
- break;
- case MO_LEUL:
- stl_le_p(haddr, val);
- break;
- case MO_BEUQ:
- stq_be_p(haddr, val);
- break;
- case MO_LEUQ:
- stq_le_p(haddr, val);
- break;
- default:
- qemu_build_not_reached();
- }
-}
-
/**
* do_st_mmio_leN:
* @env: cpu context
@@ -2632,16 +2602,145 @@ static uint64_t do_st_bytes_leN(MMULookupPageData *p, uint64_t val_le)
return val_le;
}
+/**
+ * do_st_parts_leN
+ * @p: translation parameters
+ * @val_le: data to store
+ *
+ * As do_st_bytes_leN, but atomically on each aligned part.
+ */
+static uint64_t do_st_parts_leN(MMULookupPageData *p, uint64_t val_le)
+{
+ void *haddr = p->haddr;
+ int size = p->size;
+
+ do {
+ int n;
+
+ /* Find minimum of alignment and size */
+ switch (((uintptr_t)haddr | size) & 7) {
+ case 4:
+ store_atomic4(haddr, le32_to_cpu(val_le));
+ val_le >>= 32;
+ n = 4;
+ break;
+ case 2:
+ case 6:
+ store_atomic2(haddr, le16_to_cpu(val_le));
+ val_le >>= 16;
+ n = 2;
+ break;
+ default:
+ stb_p(haddr, val_le);
+ val_le >>= 8;
+ n = 1;
+ break;
+ case 0:
+ g_assert_not_reached();
+ }
+ haddr += n;
+ size -= n;
+ } while (size != 0);
+ return val_le;
+}
+
+/**
+ * do_st_whole_le4
+ * @p: translation parameters
+ * @val_le: data to store
+ *
+ * As do_st_bytes_leN, but atomically on each aligned part.
+ * Four aligned bytes are guaranteed to cover the store.
+ */
+static uint64_t do_st_whole_le4(MMULookupPageData *p, uint64_t val_le)
+{
+ int sz = p->size * 8;
+ int o = p->addr & 3;
+ int sh = o * 8;
+ uint32_t m = MAKE_64BIT_MASK(0, sz);
+ uint32_t v;
+
+ if (HOST_BIG_ENDIAN) {
+ v = bswap32(val_le) >> sh;
+ m = bswap32(m) >> sh;
+ } else {
+ v = val_le << sh;
+ m <<= sh;
+ }
+ store_atom_insert_al4(p->haddr - o, v, m);
+ return val_le >> sz;
+}
+
+/**
+ * do_st_whole_le8
+ * @p: translation parameters
+ * @val_le: data to store
+ *
+ * As do_st_bytes_leN, but atomically on each aligned part.
+ * Eight aligned bytes are guaranteed to cover the store.
+ */
+static uint64_t do_st_whole_le8(MMULookupPageData *p, uint64_t val_le)
+{
+ int sz = p->size * 8;
+ int o = p->addr & 7;
+ int sh = o * 8;
+ uint64_t m = MAKE_64BIT_MASK(0, sz);
+ uint64_t v;
+
+ if (HOST_BIG_ENDIAN) {
+ v = bswap64(val_le) >> sh;
+ m = bswap64(m) >> sh;
+ } else {
+ v = val_le << sh;
+ m <<= sh;
+ }
+ store_atom_insert_al8(p->haddr - o, v, m);
+ return val_le >> sz;
+}
+
/*
* Wrapper for the above.
*/
static uint64_t do_st_leN(CPUArchState *env, MMULookupPageData *p,
- uint64_t val_le, int mmu_idx, uintptr_t ra)
+ uint64_t val_le, int mmu_idx,
+ MemOp mop, uintptr_t ra)
{
+ MemOp atmax;
+
if (unlikely(p->flags & TLB_MMIO)) {
return do_st_mmio_leN(env, p, val_le, mmu_idx, ra);
- } else {
+ }
+
+ switch (mop & MO_ATOM_MASK) {
+ case MO_ATOM_WITHIN16:
+ /*
+ * It is a given that we cross a page and therefore there is no
+ * atomicity for the load as a whole, but there may be a subobject
+ * as defined by ATMAX which does not cross a 16-byte boundary.
+ */
+ atmax = mop & MO_ATMAX_MASK;
+ if (atmax == MO_ATMAX_SIZE) {
+ atmax = mop & MO_SIZE;
+ } else {
+ atmax >>= MO_ATMAX_SHIFT;
+ }
+ if (unlikely(p->size >= (1 << atmax))) {
+ if (!HAVE_al8_fast && p->size <= 4) {
+ return do_st_whole_le4(p, val_le);
+ } else if (HAVE_al8) {
+ return do_st_whole_le8(p, val_le);
+ } else {
+ cpu_loop_exit_atomic(env_cpu(env), ra);
+ }
+ }
+ /* fall through */
+ case MO_ATOM_IFALIGN:
+ case MO_ATOM_NONE:
return do_st_bytes_leN(p, val_le);
+ case MO_ATOM_SUBALIGN:
+ return do_st_parts_leN(p, val_le);
+ default:
+ g_assert_not_reached();
}
}
@@ -2686,7 +2785,7 @@ static void do_st2_mmu(CPUArchState *env, target_ulong addr, uint16_t val,
if (l.memop & MO_BSWAP) {
val = bswap16(val);
}
- store_memop(l.page[0].haddr, val, MO_UW);
+ store_atom_2(env, ra, l.page[0].haddr, l.memop, val);
}
} else {
uint8_t a, b;
@@ -2735,7 +2834,7 @@ static void do_st4_mmu(CPUArchState *env, target_ulong addr, uint32_t val,
if (l.memop & MO_BSWAP) {
val = bswap32(val);
}
- store_memop(l.page[0].haddr, val, MO_UL);
+ store_atom_4(env, ra, l.page[0].haddr, l.memop, val);
}
} else {
assert_no_tlb_bswap;
@@ -2744,8 +2843,8 @@ static void do_st4_mmu(CPUArchState *env, target_ulong addr, uint32_t val,
if ((l.memop & MO_BSWAP) != MO_LE) {
val = bswap32(val);
}
- val = do_st_leN(env, &l.page[0], val, l.mmu_idx, ra);
- (void) do_st_leN(env, &l.page[1], val, l.mmu_idx, ra);
+ val = do_st_leN(env, &l.page[0], val, l.mmu_idx, l.memop, ra);
+ (void) do_st_leN(env, &l.page[1], val, l.mmu_idx, l.memop, ra);
}
}
@@ -2781,7 +2880,7 @@ static void do_st8_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
if (l.memop & MO_BSWAP) {
val = bswap64(val);
}
- store_memop(l.page[0].haddr, val, MO_UQ);
+ store_atom_8(env, ra, l.page[0].haddr, l.memop, val);
}
} else {
assert_no_tlb_bswap;
@@ -2790,8 +2889,8 @@ static void do_st8_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
if ((l.memop & MO_BSWAP) != MO_LE) {
val = bswap64(val);
}
- val = do_st_leN(env, &l.page[0], val, l.mmu_idx, ra);
- (void) do_st_leN(env, &l.page[1], val, l.mmu_idx, ra);
+ val = do_st_leN(env, &l.page[0], val, l.mmu_idx, l.memop, ra);
+ (void) do_st_leN(env, &l.page[1], val, l.mmu_idx, l.memop, ra);
}
}
diff --git a/accel/tcg/user-exec.c b/accel/tcg/user-exec.c
index ec721e5097..ddba8c9dd7 100644
--- a/accel/tcg/user-exec.c
+++ b/accel/tcg/user-exec.c
@@ -451,7 +451,7 @@ void cpu_stw_be_mmu(CPUArchState *env, abi_ptr addr, uint16_t val,
validate_memop(oi, MO_BEUW);
haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_STORE);
- stw_be_p(haddr, val);
+ store_atom_2(env, ra, haddr, get_memop(oi), be16_to_cpu(val));
clear_helper_retaddr();
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
}
@@ -463,7 +463,7 @@ void cpu_stl_be_mmu(CPUArchState *env, abi_ptr addr, uint32_t val,
validate_memop(oi, MO_BEUL);
haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_STORE);
- stl_be_p(haddr, val);
+ store_atom_4(env, ra, haddr, get_memop(oi), be32_to_cpu(val));
clear_helper_retaddr();
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
}
@@ -475,7 +475,7 @@ void cpu_stq_be_mmu(CPUArchState *env, abi_ptr addr, uint64_t val,
validate_memop(oi, MO_BEUQ);
haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_STORE);
- stq_be_p(haddr, val);
+ store_atom_8(env, ra, haddr, get_memop(oi), be64_to_cpu(val));
clear_helper_retaddr();
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
}
@@ -487,7 +487,7 @@ void cpu_stw_le_mmu(CPUArchState *env, abi_ptr addr, uint16_t val,
validate_memop(oi, MO_LEUW);
haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_STORE);
- stw_le_p(haddr, val);
+ store_atom_2(env, ra, haddr, get_memop(oi), le16_to_cpu(val));
clear_helper_retaddr();
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
}
@@ -499,7 +499,7 @@ void cpu_stl_le_mmu(CPUArchState *env, abi_ptr addr, uint32_t val,
validate_memop(oi, MO_LEUL);
haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_STORE);
- stl_le_p(haddr, val);
+ store_atom_4(env, ra, haddr, get_memop(oi), le32_to_cpu(val));
clear_helper_retaddr();
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
}
@@ -511,7 +511,7 @@ void cpu_stq_le_mmu(CPUArchState *env, abi_ptr addr, uint64_t val,
validate_memop(oi, MO_LEUQ);
haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_STORE);
- stq_le_p(haddr, val);
+ store_atom_8(env, ra, haddr, get_memop(oi), le64_to_cpu(val));
clear_helper_retaddr();
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
}
diff --git a/accel/tcg/ldst_atomicity.c.inc b/accel/tcg/ldst_atomicity.c.inc
index decc9a2a16..8876c16371 100644
--- a/accel/tcg/ldst_atomicity.c.inc
+++ b/accel/tcg/ldst_atomicity.c.inc
@@ -21,6 +21,12 @@
#else
# define HAVE_al16_fast false
#endif
+#if defined(CONFIG_ATOMIC128) || defined(CONFIG_CMPXCHG128)
+# define HAVE_al16 true
+#else
+# define HAVE_al16 false
+#endif
+
/**
* required_atomicity:
@@ -544,3 +550,333 @@ static uint64_t load_atom_8(CPUArchState *env, uintptr_t ra,
g_assert_not_reached();
}
}
+
+/**
+ * store_atomic2:
+ * @pv: host address
+ * @val: value to store
+ *
+ * Atomically store 2 aligned bytes to @pv.
+ */
+static inline void store_atomic2(void *pv, uint16_t val)
+{
+ uint16_t *p = __builtin_assume_aligned(pv, 2);
+ qatomic_set(p, val);
+}
+
+/**
+ * store_atomic4:
+ * @pv: host address
+ * @val: value to store
+ *
+ * Atomically store 4 aligned bytes to @pv.
+ */
+static inline void store_atomic4(void *pv, uint32_t val)
+{
+ uint32_t *p = __builtin_assume_aligned(pv, 4);
+ qatomic_set(p, val);
+}
+
+/**
+ * store_atomic8:
+ * @pv: host address
+ * @val: value to store
+ *
+ * Atomically store 8 aligned bytes to @pv.
+ */
+static inline void store_atomic8(void *pv, uint64_t val)
+{
+ uint64_t *p = __builtin_assume_aligned(pv, 8);
+
+ qemu_build_assert(HAVE_al8);
+ qatomic_set__nocheck(p, val);
+}
+
+/**
+ * store_atom_4x2
+ */
+static inline void store_atom_4_by_2(void *pv, uint32_t val)
+{
+ uint16_t *p = __builtin_assume_aligned(pv, 2);
+ qatomic_set(p, val >> (HOST_BIG_ENDIAN ? 16 : 0));
+ qatomic_set(p + 2, val >> (HOST_BIG_ENDIAN ? 0 : 16));
+}
+
+/**
+ * store_atom_8_by_2
+ */
+static inline void store_atom_8_by_2(void *pv, uint64_t val)
+{
+ uint32_t *p = __builtin_assume_aligned(pv, 4);
+ qatomic_set(p, val >> (HOST_BIG_ENDIAN ? 32 : 0));
+ qatomic_set(p + 4, val >> (HOST_BIG_ENDIAN ? 0 : 32));
+}
+
+/**
+ * store_atom_8_by_4
+ */
+static inline void store_atom_8_by_4(void *pv, uint64_t val)
+{
+ uint16_t *p = __builtin_assume_aligned(pv, 2);
+ qatomic_set(p, val >> (HOST_BIG_ENDIAN ? 48 : 0));
+ qatomic_set(p + 2, val >> (HOST_BIG_ENDIAN ? 32 : 16));
+ qatomic_set(p + 4, val >> (HOST_BIG_ENDIAN ? 16 : 32));
+ qatomic_set(p + 6, val >> (HOST_BIG_ENDIAN ? 0 : 48));
+}
+
+/**
+ * store_atom_insert_al4:
+ * @p: host address
+ * @val: shifted value to store
+ * @msk: mask for value to store
+ *
+ * Atomically store @val to @p, masked by @msk.
+ */
+static void store_atom_insert_al4(uint32_t *p, uint32_t val, uint32_t msk)
+{
+ uint32_t old, new;
+
+ p = __builtin_assume_aligned(p, 4);
+ old = qatomic_read(p);
+ do {
+ new = (old & ~msk) | val;
+ } while (!__atomic_compare_exchange_n(p, &old, new, true,
+ __ATOMIC_RELAXED, __ATOMIC_RELAXED));
+}
+
+/**
+ * store_atom_insert_al8:
+ * @p: host address
+ * @val: shifted value to store
+ * @msk: mask for value to store
+ *
+ * Atomically store @val to @p masked by @msk.
+ */
+static void store_atom_insert_al8(uint64_t *p, uint64_t val, uint64_t msk)
+{
+ uint64_t old, new;
+
+ qemu_build_assert(HAVE_al8);
+ p = __builtin_assume_aligned(p, 8);
+ old = qatomic_read__nocheck(p);
+ do {
+ new = (old & ~msk) | val;
+ } while (!__atomic_compare_exchange_n(p, &old, new, true,
+ __ATOMIC_RELAXED, __ATOMIC_RELAXED));
+}
+
+/**
+ * store_atom_insert_al16:
+ * @p: host address
+ * @val: shifted value to store
+ * @msk: mask for value to store
+ *
+ * Atomically store @val to @p masked by @msk.
+ */
+static void store_atom_insert_al16(Int128 *ps, Int128Alias val, Int128Alias msk)
+{
+#if defined(CONFIG_ATOMIC128)
+ __uint128_t *pu, old, new;
+
+ /* With CONFIG_ATOMIC128, we can avoid the memory barriers. */
+ pu = __builtin_assume_aligned(ps, 16);
+ old = *pu;
+ do {
+ new = (old & ~msk.u) | val.u;
+ } while (!__atomic_compare_exchange_n(pu, &old, new, true,
+ __ATOMIC_RELAXED, __ATOMIC_RELAXED));
+#elif defined(CONFIG_CMPXCHG128)
+ __uint128_t *pu, old, new;
+
+ /*
+ * Without CONFIG_ATOMIC128, __atomic_compare_exchange_n will always
+ * defer to libatomic, so we must use __sync_val_compare_and_swap_16
+ * and accept the sequential consistency that comes with it.
+ */
+ pu = __builtin_assume_aligned(ps, 16);
+ do {
+ old = *pu;
+ new = (old & ~msk.u) | val.u;
+ } while (!__sync_bool_compare_and_swap_16(pu, old, new));
+#else
+ qemu_build_not_reached();
+#endif
+}
+
+/**
+ * store_atom_2:
+ * @p: host address
+ * @val: the value to store
+ * @memop: the full memory op
+ *
+ * Store 2 bytes to @p, honoring the atomicity of @memop.
+ */
+static void store_atom_2(CPUArchState *env, uintptr_t ra,
+ void *pv, MemOp memop, uint16_t val)
+{
+ uintptr_t pi = (uintptr_t)pv;
+ MemOp atmax;
+
+ if (likely((pi & 1) == 0)) {
+ store_atomic2(pv, val);
+ return;
+ }
+
+ atmax = required_atomicity(env, pi, memop);
+ if (atmax == MO_8) {
+ stw_he_p(pv, val);
+ return;
+ }
+
+ /* The only case remaining is MO_ATOM_WITHIN16. */
+ if (!HAVE_al8_fast && (pi & 3) == 1) {
+ /* Big or little endian, we want the middle two bytes. */
+ store_atom_insert_al4(pv - 1, val << 8, 0x00ffff00);
+ return;
+ }
+
+ if ((pi & 15) != 7) {
+ if (HAVE_al8) {
+ int sh = (pi & 7) * 8;
+ uint64_t v, m;
+
+ pv = (void *)(pi & ~7);
+ if (HOST_BIG_ENDIAN) {
+ v = (uint64_t)val << (48 - sh);
+ m = 0xffffull << (48 - sh);
+ } else {
+ v = (uint64_t)val << sh;
+ m = 0xffffull << sh;
+ }
+ store_atom_insert_al8(pv, v, m);
+ return;
+ }
+ } else {
+ if (HAVE_al16) {
+ Int128 v, m;
+
+ /* Big or little endian, we want the middle two bytes. */
+ v = int128_lshift(int128_make64(val), 56);
+ m = int128_lshift(int128_make64(0xffff), 56);
+ store_atom_insert_al16(pv - 7, v, m);
+ return;
+ }
+ }
+
+ cpu_loop_exit_atomic(env_cpu(env), ra);
+}
+
+/**
+ * store_atom_4:
+ * @p: host address
+ * @val: the value to store
+ * @memop: the full memory op
+ *
+ * Store 4 bytes to @p, honoring the atomicity of @memop.
+ */
+static void store_atom_4(CPUArchState *env, uintptr_t ra,
+ void *pv, MemOp memop, uint32_t val)
+{
+ uintptr_t pi = (uintptr_t)pv;
+ MemOp atmax;
+
+ if (likely((pi & 3) == 0)) {
+ store_atomic4(pv, val);
+ return;
+ }
+
+ atmax = required_atomicity(env, pi, memop);
+ switch (atmax) {
+ case MO_8:
+ stl_he_p(pv, val);
+ return;
+ case MO_16:
+ store_atom_4_by_2(pv, val);
+ return;
+ case MO_32:
+ if ((pi & 7) < 4) {
+ if (HAVE_al8) {
+ int sh = (pi & 7) * 8;
+ uint64_t v, m;
+
+ pv = (void *)(pi & ~7);
+ if (HOST_BIG_ENDIAN) {
+ v = (uint64_t)val << (32 - sh);
+ m = 0xffffffffull << (32 - sh);
+ } else {
+ v = (uint64_t)val << sh;
+ m = 0xffffffffull << sh;
+ }
+ store_atom_insert_al8(pv, v, m);
+ return;
+ }
+ } else {
+ if (HAVE_al16) {
+ int sh = (pi & 7) * 8;
+ Int128 v, m;
+
+ v = int128_make64(val);
+ m = int128_make64(0xffffffffull);
+ v = int128_lshift(v, HOST_BIG_ENDIAN ? 96 - sh : sh);
+ m = int128_lshift(m, HOST_BIG_ENDIAN ? 96 - sh : sh);
+
+ pv = (void *)(pi & ~15);
+ store_atom_insert_al16(pv, v, m);
+ return;
+ }
+ }
+ cpu_loop_exit_atomic(env_cpu(env), ra);
+ default:
+ g_assert_not_reached();
+ }
+}
+
+/**
+ * store_atom_8:
+ * @p: host address
+ * @val: the value to store
+ * @memop: the full memory op
+ *
+ * Store 8 bytes to @p, honoring the atomicity of @memop.
+ */
+static void store_atom_8(CPUArchState *env, uintptr_t ra,
+ void *pv, MemOp memop, uint64_t val)
+{
+ uintptr_t pi = (uintptr_t)pv;
+ MemOp atmax;
+
+ if (HAVE_al8 && likely((pi & 7) == 0)) {
+ store_atomic8(pv, val);
+ return;
+ }
+
+ atmax = required_atomicity(env, pi, memop);
+ switch (atmax) {
+ case MO_8:
+ stq_he_p(pv, val);
+ return;
+ case MO_16:
+ store_atom_8_by_2(pv, val);
+ return;
+ case MO_32:
+ store_atom_8_by_4(pv, val);
+ return;
+ case MO_64:
+ if (HAVE_al16) {
+ int sh = (pi & 7) * 8;
+ Int128 v, m;
+
+ v = int128_make64(val);
+ m = int128_make64(-1ull);
+ v = int128_lshift(v, HOST_BIG_ENDIAN ? 64 - sh : sh);
+ m = int128_lshift(m, HOST_BIG_ENDIAN ? 64 - sh : sh);
+
+ pv = (void *)(pi & ~15);
+ store_atom_insert_al16(pv, v, m);
+ return;
+ }
+ cpu_loop_exit_atomic(env_cpu(env), ra);
+ default:
+ g_assert_not_reached();
+ }
+}
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* [PATCH for-8.0 09/29] tcg/tci: Use cpu_{ld,st}_mmu
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (7 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 08/29] accel/tcg: Honor atomicity of stores Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-21 12:40 ` Philippe Mathieu-Daudé
2022-11-18 9:47 ` [PATCH for-8.0 10/29] tcg: Unify helper_{be,le}_{ld,st}* Richard Henderson
` (19 subsequent siblings)
28 siblings, 1 reply; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
Unify the softmmu and the user-only paths by using the
official memory interface. Avoid double logging of memory
operations to plugins by relying on the ones within the
cpu_*_mmu functions.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
tcg/tcg-op.c | 9 +++-
tcg/tci.c | 127 ++++++++-------------------------------------------
2 files changed, 26 insertions(+), 110 deletions(-)
diff --git a/tcg/tcg-op.c b/tcg/tcg-op.c
index e7e4951a3c..1f81c3dbb3 100644
--- a/tcg/tcg-op.c
+++ b/tcg/tcg-op.c
@@ -2914,7 +2914,12 @@ static void tcg_gen_req_mo(TCGBar type)
static inline TCGv plugin_prep_mem_callbacks(TCGv vaddr)
{
-#ifdef CONFIG_PLUGIN
+ /*
+ * With TCI, we get memory tracing via cpu_{ld,st}_mmu.
+ * No need to instrument memory operations inline, and
+ * we don't want to log the same memory operation twice.
+ */
+#if defined(CONFIG_PLUGIN) && !defined(CONFIG_TCG_INTERPRETER)
if (tcg_ctx->plugin_insn != NULL) {
/* Save a copy of the vaddr for use after a load. */
TCGv temp = tcg_temp_new();
@@ -2928,7 +2933,7 @@ static inline TCGv plugin_prep_mem_callbacks(TCGv vaddr)
static void plugin_gen_mem_callbacks(TCGv vaddr, MemOpIdx oi,
enum qemu_plugin_mem_rw rw)
{
-#ifdef CONFIG_PLUGIN
+#if defined(CONFIG_PLUGIN) && !defined(CONFIG_TCG_INTERPRETER)
if (tcg_ctx->plugin_insn != NULL) {
qemu_plugin_meminfo_t info = make_plugin_meminfo(oi, rw);
plugin_gen_empty_mem_callback(vaddr, info);
diff --git a/tcg/tci.c b/tcg/tci.c
index 022fe9d0f8..52fdd3f5ec 100644
--- a/tcg/tci.c
+++ b/tcg/tci.c
@@ -293,87 +293,34 @@ static uint64_t tci_qemu_ld(CPUArchState *env, target_ulong taddr,
MemOp mop = get_memop(oi);
uintptr_t ra = (uintptr_t)tb_ptr;
-#ifdef CONFIG_SOFTMMU
switch (mop & (MO_BSWAP | MO_SSIZE)) {
case MO_UB:
- return helper_ret_ldub_mmu(env, taddr, oi, ra);
+ return cpu_ldb_mmu(env, taddr, oi, ra);
case MO_SB:
- return helper_ret_ldsb_mmu(env, taddr, oi, ra);
+ return (int8_t)cpu_ldb_mmu(env, taddr, oi, ra);
case MO_LEUW:
- return helper_le_lduw_mmu(env, taddr, oi, ra);
+ return cpu_ldw_le_mmu(env, taddr, oi, ra);
case MO_LESW:
- return helper_le_ldsw_mmu(env, taddr, oi, ra);
+ return (int16_t)cpu_ldw_le_mmu(env, taddr, oi, ra);
case MO_LEUL:
- return helper_le_ldul_mmu(env, taddr, oi, ra);
+ return cpu_ldl_le_mmu(env, taddr, oi, ra);
case MO_LESL:
- return helper_le_ldsl_mmu(env, taddr, oi, ra);
+ return (int32_t)cpu_ldl_le_mmu(env, taddr, oi, ra);
case MO_LEUQ:
- return helper_le_ldq_mmu(env, taddr, oi, ra);
+ return cpu_ldq_le_mmu(env, taddr, oi, ra);
case MO_BEUW:
- return helper_be_lduw_mmu(env, taddr, oi, ra);
+ return cpu_ldw_be_mmu(env, taddr, oi, ra);
case MO_BESW:
- return helper_be_ldsw_mmu(env, taddr, oi, ra);
+ return (int16_t)cpu_ldw_be_mmu(env, taddr, oi, ra);
case MO_BEUL:
- return helper_be_ldul_mmu(env, taddr, oi, ra);
+ return cpu_ldl_be_mmu(env, taddr, oi, ra);
case MO_BESL:
- return helper_be_ldsl_mmu(env, taddr, oi, ra);
+ return (int32_t)cpu_ldl_be_mmu(env, taddr, oi, ra);
case MO_BEUQ:
- return helper_be_ldq_mmu(env, taddr, oi, ra);
+ return cpu_ldq_be_mmu(env, taddr, oi, ra);
default:
g_assert_not_reached();
}
-#else
- void *haddr = g2h(env_cpu(env), taddr);
- unsigned a_mask = (1u << get_alignment_bits(mop)) - 1;
- uint64_t ret;
-
- set_helper_retaddr(ra);
- if (taddr & a_mask) {
- helper_unaligned_ld(env, taddr);
- }
- switch (mop & (MO_BSWAP | MO_SSIZE)) {
- case MO_UB:
- ret = ldub_p(haddr);
- break;
- case MO_SB:
- ret = ldsb_p(haddr);
- break;
- case MO_LEUW:
- ret = lduw_le_p(haddr);
- break;
- case MO_LESW:
- ret = ldsw_le_p(haddr);
- break;
- case MO_LEUL:
- ret = (uint32_t)ldl_le_p(haddr);
- break;
- case MO_LESL:
- ret = (int32_t)ldl_le_p(haddr);
- break;
- case MO_LEUQ:
- ret = ldq_le_p(haddr);
- break;
- case MO_BEUW:
- ret = lduw_be_p(haddr);
- break;
- case MO_BESW:
- ret = ldsw_be_p(haddr);
- break;
- case MO_BEUL:
- ret = (uint32_t)ldl_be_p(haddr);
- break;
- case MO_BESL:
- ret = (int32_t)ldl_be_p(haddr);
- break;
- case MO_BEUQ:
- ret = ldq_be_p(haddr);
- break;
- default:
- g_assert_not_reached();
- }
- clear_helper_retaddr();
- return ret;
-#endif
}
static void tci_qemu_st(CPUArchState *env, target_ulong taddr, uint64_t val,
@@ -382,67 +329,31 @@ static void tci_qemu_st(CPUArchState *env, target_ulong taddr, uint64_t val,
MemOp mop = get_memop(oi);
uintptr_t ra = (uintptr_t)tb_ptr;
-#ifdef CONFIG_SOFTMMU
switch (mop & (MO_BSWAP | MO_SIZE)) {
case MO_UB:
- helper_ret_stb_mmu(env, taddr, val, oi, ra);
+ cpu_stb_mmu(env, taddr, val, oi, ra);
break;
case MO_LEUW:
- helper_le_stw_mmu(env, taddr, val, oi, ra);
+ cpu_stw_le_mmu(env, taddr, val, oi, ra);
break;
case MO_LEUL:
- helper_le_stl_mmu(env, taddr, val, oi, ra);
+ cpu_stl_le_mmu(env, taddr, val, oi, ra);
break;
case MO_LEUQ:
- helper_le_stq_mmu(env, taddr, val, oi, ra);
+ cpu_stq_le_mmu(env, taddr, val, oi, ra);
break;
case MO_BEUW:
- helper_be_stw_mmu(env, taddr, val, oi, ra);
+ cpu_stw_be_mmu(env, taddr, val, oi, ra);
break;
case MO_BEUL:
- helper_be_stl_mmu(env, taddr, val, oi, ra);
+ cpu_stl_be_mmu(env, taddr, val, oi, ra);
break;
case MO_BEUQ:
- helper_be_stq_mmu(env, taddr, val, oi, ra);
+ cpu_stq_be_mmu(env, taddr, val, oi, ra);
break;
default:
g_assert_not_reached();
}
-#else
- void *haddr = g2h(env_cpu(env), taddr);
- unsigned a_mask = (1u << get_alignment_bits(mop)) - 1;
-
- set_helper_retaddr(ra);
- if (taddr & a_mask) {
- helper_unaligned_st(env, taddr);
- }
- switch (mop & (MO_BSWAP | MO_SIZE)) {
- case MO_UB:
- stb_p(haddr, val);
- break;
- case MO_LEUW:
- stw_le_p(haddr, val);
- break;
- case MO_LEUL:
- stl_le_p(haddr, val);
- break;
- case MO_LEUQ:
- stq_le_p(haddr, val);
- break;
- case MO_BEUW:
- stw_be_p(haddr, val);
- break;
- case MO_BEUL:
- stl_be_p(haddr, val);
- break;
- case MO_BEUQ:
- stq_be_p(haddr, val);
- break;
- default:
- g_assert_not_reached();
- }
- clear_helper_retaddr();
-#endif
}
#if TCG_TARGET_REG_BITS == 64
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* Re: [PATCH for-8.0 09/29] tcg/tci: Use cpu_{ld,st}_mmu
2022-11-18 9:47 ` [PATCH for-8.0 09/29] tcg/tci: Use cpu_{ld,st}_mmu Richard Henderson
@ 2022-11-21 12:40 ` Philippe Mathieu-Daudé
0 siblings, 0 replies; 48+ messages in thread
From: Philippe Mathieu-Daudé @ 2022-11-21 12:40 UTC (permalink / raw)
To: Richard Henderson, qemu-devel
On 18/11/22 10:47, Richard Henderson wrote:
> Unify the softmmu and the user-only paths by using the
> official memory interface. Avoid double logging of memory
> operations to plugins by relying on the ones within the
> cpu_*_mmu functions.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
> tcg/tcg-op.c | 9 +++-
> tcg/tci.c | 127 ++++++++-------------------------------------------
> 2 files changed, 26 insertions(+), 110 deletions(-)
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
^ permalink raw reply [flat|nested] 48+ messages in thread
* [PATCH for-8.0 10/29] tcg: Unify helper_{be,le}_{ld,st}*
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (8 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 09/29] tcg/tci: Use cpu_{ld,st}_mmu Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-21 12:48 ` Philippe Mathieu-Daudé
2022-11-18 9:47 ` [PATCH for-8.0 11/29] accel/tcg: Implement helper_{ld, st}*_mmu for user-only Richard Henderson
` (18 subsequent siblings)
28 siblings, 1 reply; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
With the current structure of cputlb.c, there is no difference
between the little-endian and big-endian entry points, aside
from the assert. Unify the pairs of functions.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
include/tcg/tcg-ldst.h | 60 ++++------
accel/tcg/cputlb.c | 190 ++++++++++---------------------
docs/devel/loads-stores.rst | 36 ++----
tcg/aarch64/tcg-target.c.inc | 39 +++----
tcg/arm/tcg-target.c.inc | 45 +++-----
tcg/i386/tcg-target.c.inc | 40 +++----
tcg/loongarch64/tcg-target.c.inc | 25 ++--
tcg/mips/tcg-target.c.inc | 40 +++----
tcg/ppc/tcg-target.c.inc | 30 ++---
tcg/riscv/tcg-target.c.inc | 51 +++------
tcg/s390x/tcg-target.c.inc | 38 +++----
tcg/sparc64/tcg-target.c.inc | 37 +++---
12 files changed, 226 insertions(+), 405 deletions(-)
diff --git a/include/tcg/tcg-ldst.h b/include/tcg/tcg-ldst.h
index 2ba22bd5fe..56fa7afe5e 100644
--- a/include/tcg/tcg-ldst.h
+++ b/include/tcg/tcg-ldst.h
@@ -28,47 +28,31 @@
#ifdef CONFIG_SOFTMMU
/* Value zero-extended to tcg register size. */
-tcg_target_ulong helper_ret_ldub_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr);
-tcg_target_ulong helper_le_lduw_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr);
-tcg_target_ulong helper_le_ldul_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr);
-uint64_t helper_le_ldq_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr);
-tcg_target_ulong helper_be_lduw_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr);
-tcg_target_ulong helper_be_ldul_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr);
-uint64_t helper_be_ldq_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr);
+tcg_target_ulong helper_ldub_mmu(CPUArchState *env, target_ulong addr,
+ MemOpIdx oi, uintptr_t retaddr);
+tcg_target_ulong helper_lduw_mmu(CPUArchState *env, target_ulong addr,
+ MemOpIdx oi, uintptr_t retaddr);
+tcg_target_ulong helper_ldul_mmu(CPUArchState *env, target_ulong addr,
+ MemOpIdx oi, uintptr_t retaddr);
+uint64_t helper_ldq_mmu(CPUArchState *env, target_ulong addr,
+ MemOpIdx oi, uintptr_t retaddr);
/* Value sign-extended to tcg register size. */
-tcg_target_ulong helper_ret_ldsb_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr);
-tcg_target_ulong helper_le_ldsw_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr);
-tcg_target_ulong helper_le_ldsl_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr);
-tcg_target_ulong helper_be_ldsw_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr);
-tcg_target_ulong helper_be_ldsl_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr);
+tcg_target_ulong helper_ldsb_mmu(CPUArchState *env, target_ulong addr,
+ MemOpIdx oi, uintptr_t retaddr);
+tcg_target_ulong helper_ldsw_mmu(CPUArchState *env, target_ulong addr,
+ MemOpIdx oi, uintptr_t retaddr);
+tcg_target_ulong helper_ldsl_mmu(CPUArchState *env, target_ulong addr,
+ MemOpIdx oi, uintptr_t retaddr);
-void helper_ret_stb_mmu(CPUArchState *env, target_ulong addr, uint8_t val,
- MemOpIdx oi, uintptr_t retaddr);
-void helper_le_stw_mmu(CPUArchState *env, target_ulong addr, uint16_t val,
- MemOpIdx oi, uintptr_t retaddr);
-void helper_le_stl_mmu(CPUArchState *env, target_ulong addr, uint32_t val,
- MemOpIdx oi, uintptr_t retaddr);
-void helper_le_stq_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
- MemOpIdx oi, uintptr_t retaddr);
-void helper_be_stw_mmu(CPUArchState *env, target_ulong addr, uint16_t val,
- MemOpIdx oi, uintptr_t retaddr);
-void helper_be_stl_mmu(CPUArchState *env, target_ulong addr, uint32_t val,
- MemOpIdx oi, uintptr_t retaddr);
-void helper_be_stq_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
- MemOpIdx oi, uintptr_t retaddr);
+void helper_stb_mmu(CPUArchState *env, target_ulong addr, uint8_t val,
+ MemOpIdx oi, uintptr_t retaddr);
+void helper_stw_mmu(CPUArchState *env, target_ulong addr, uint16_t val,
+ MemOpIdx oi, uintptr_t retaddr);
+void helper_stl_mmu(CPUArchState *env, target_ulong addr, uint32_t val,
+ MemOpIdx oi, uintptr_t retaddr);
+void helper_stq_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
+ MemOpIdx oi, uintptr_t retaddr);
#else
diff --git a/accel/tcg/cputlb.c b/accel/tcg/cputlb.c
index 69f8a25a7f..3d32adc0e7 100644
--- a/accel/tcg/cputlb.c
+++ b/accel/tcg/cputlb.c
@@ -1980,25 +1980,6 @@ static void *atomic_mmu_lookup(CPUArchState *env, target_ulong addr,
cpu_loop_exit_atomic(env_cpu(env), retaddr);
}
-/*
- * Verify that we have passed the correct MemOp to the correct function.
- *
- * In the case of the helper_*_mmu functions, we will have done this by
- * using the MemOp to look up the helper during code generation.
- *
- * In the case of the cpu_*_mmu functions, this is up to the caller.
- * We could present one function to target code, and dispatch based on
- * the MemOp, but so far we have worked hard to avoid an indirect function
- * call along the memory path.
- */
-static void validate_memop(MemOpIdx oi, MemOp expected)
-{
-#ifdef CONFIG_DEBUG_TCG
- MemOp have = get_memop(oi) & (MO_SIZE | MO_BSWAP);
- assert(have == expected);
-#endif
-}
-
/*
* Load Helpers
*
@@ -2215,10 +2196,10 @@ static uint8_t do_ld1_mmu(CPUArchState *env, target_ulong addr, MemOpIdx oi,
return do_ld_1(env, &l.page[0], l.mmu_idx, access_type, ra);
}
-tcg_target_ulong helper_ret_ldub_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr)
+tcg_target_ulong helper_ldub_mmu(CPUArchState *env, target_ulong addr,
+ MemOpIdx oi, uintptr_t retaddr)
{
- validate_memop(oi, MO_UB);
+ tcg_debug_assert((get_memop(oi) & MO_SIZE) == MO_8);
return do_ld1_mmu(env, addr, oi, retaddr, MMU_DATA_LOAD);
}
@@ -2261,17 +2242,10 @@ static uint16_t do_ld2_mmu(CPUArchState *env, target_ulong addr, MemOpIdx oi,
return ret;
}
-tcg_target_ulong helper_le_lduw_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr)
+tcg_target_ulong helper_lduw_mmu(CPUArchState *env, target_ulong addr,
+ MemOpIdx oi, uintptr_t retaddr)
{
- validate_memop(oi, MO_LEUW);
- return do_ld2_mmu(env, addr, oi, retaddr, MMU_DATA_LOAD);
-}
-
-tcg_target_ulong helper_be_lduw_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr)
-{
- validate_memop(oi, MO_BEUW);
+ tcg_debug_assert((get_memop(oi) & MO_SIZE) == MO_16);
return do_ld2_mmu(env, addr, oi, retaddr, MMU_DATA_LOAD);
}
@@ -2310,17 +2284,10 @@ static uint32_t do_ld4_mmu(CPUArchState *env, target_ulong addr, MemOpIdx oi,
return ret;
}
-tcg_target_ulong helper_le_ldul_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr)
+tcg_target_ulong helper_ldul_mmu(CPUArchState *env, target_ulong addr,
+ MemOpIdx oi, uintptr_t retaddr)
{
- validate_memop(oi, MO_LEUL);
- return do_ld4_mmu(env, addr, oi, retaddr, MMU_DATA_LOAD);
-}
-
-tcg_target_ulong helper_be_ldul_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr)
-{
- validate_memop(oi, MO_BEUL);
+ tcg_debug_assert((get_memop(oi) & MO_SIZE) == MO_32);
return do_ld4_mmu(env, addr, oi, retaddr, MMU_DATA_LOAD);
}
@@ -2359,17 +2326,10 @@ static uint64_t do_ld8_mmu(CPUArchState *env, target_ulong addr, MemOpIdx oi,
return ret;
}
-uint64_t helper_le_ldq_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr)
+uint64_t helper_ldq_mmu(CPUArchState *env, target_ulong addr,
+ MemOpIdx oi, uintptr_t retaddr)
{
- validate_memop(oi, MO_LEUQ);
- return do_ld8_mmu(env, addr, oi, retaddr, MMU_DATA_LOAD);
-}
-
-uint64_t helper_be_ldq_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr)
-{
- validate_memop(oi, MO_BEUQ);
+ tcg_debug_assert((get_memop(oi) & MO_SIZE) == MO_64);
return do_ld8_mmu(env, addr, oi, retaddr, MMU_DATA_LOAD);
}
@@ -2378,35 +2338,22 @@ uint64_t helper_be_ldq_mmu(CPUArchState *env, target_ulong addr,
* avoid this for 64-bit data, or for 32-bit data on 32-bit host.
*/
-
-tcg_target_ulong helper_ret_ldsb_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr)
+tcg_target_ulong helper_ldsb_mmu(CPUArchState *env, target_ulong addr,
+ MemOpIdx oi, uintptr_t retaddr)
{
- return (int8_t)helper_ret_ldub_mmu(env, addr, oi, retaddr);
+ return (int8_t)helper_ldub_mmu(env, addr, oi, retaddr);
}
-tcg_target_ulong helper_le_ldsw_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr)
+tcg_target_ulong helper_ldsw_mmu(CPUArchState *env, target_ulong addr,
+ MemOpIdx oi, uintptr_t retaddr)
{
- return (int16_t)helper_le_lduw_mmu(env, addr, oi, retaddr);
+ return (int16_t)helper_lduw_mmu(env, addr, oi, retaddr);
}
-tcg_target_ulong helper_be_ldsw_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr)
+tcg_target_ulong helper_ldsl_mmu(CPUArchState *env, target_ulong addr,
+ MemOpIdx oi, uintptr_t retaddr)
{
- return (int16_t)helper_be_lduw_mmu(env, addr, oi, retaddr);
-}
-
-tcg_target_ulong helper_le_ldsl_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr)
-{
- return (int32_t)helper_le_ldul_mmu(env, addr, oi, retaddr);
-}
-
-tcg_target_ulong helper_be_ldsl_mmu(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t retaddr)
-{
- return (int32_t)helper_be_ldul_mmu(env, addr, oi, retaddr);
+ return (int32_t)helper_ldul_mmu(env, addr, oi, retaddr);
}
/*
@@ -2422,7 +2369,7 @@ uint8_t cpu_ldb_mmu(CPUArchState *env, abi_ptr addr, MemOpIdx oi, uintptr_t ra)
{
uint8_t ret;
- validate_memop(oi, MO_UB);
+ tcg_debug_assert((get_memop(oi) & MO_SIZE) == MO_UB);
ret = do_ld1_mmu(env, addr, oi, ra, MMU_DATA_LOAD);
plugin_load_cb(env, addr, oi);
return ret;
@@ -2433,7 +2380,7 @@ uint16_t cpu_ldw_be_mmu(CPUArchState *env, abi_ptr addr,
{
uint16_t ret;
- validate_memop(oi, MO_BEUW);
+ tcg_debug_assert((get_memop(oi) & (MO_BSWAP | MO_SIZE)) == MO_BEUW);
ret = do_ld2_mmu(env, addr, oi, ra, MMU_DATA_LOAD);
plugin_load_cb(env, addr, oi);
return ret;
@@ -2444,7 +2391,7 @@ uint32_t cpu_ldl_be_mmu(CPUArchState *env, abi_ptr addr,
{
uint32_t ret;
- validate_memop(oi, MO_BEUL);
+ tcg_debug_assert((get_memop(oi) & (MO_BSWAP | MO_SIZE)) == MO_BEUL);
ret = do_ld4_mmu(env, addr, oi, ra, MMU_DATA_LOAD);
plugin_load_cb(env, addr, oi);
return ret;
@@ -2455,7 +2402,7 @@ uint64_t cpu_ldq_be_mmu(CPUArchState *env, abi_ptr addr,
{
uint64_t ret;
- validate_memop(oi, MO_BEUQ);
+ tcg_debug_assert((get_memop(oi) & (MO_BSWAP | MO_SIZE)) == MO_BEUQ);
ret = do_ld8_mmu(env, addr, oi, ra, MMU_DATA_LOAD);
plugin_load_cb(env, addr, oi);
return ret;
@@ -2466,7 +2413,7 @@ uint16_t cpu_ldw_le_mmu(CPUArchState *env, abi_ptr addr,
{
uint16_t ret;
- validate_memop(oi, MO_LEUW);
+ tcg_debug_assert((get_memop(oi) & (MO_BSWAP | MO_SIZE)) == MO_LEUW);
ret = do_ld2_mmu(env, addr, oi, ra, MMU_DATA_LOAD);
plugin_load_cb(env, addr, oi);
return ret;
@@ -2477,7 +2424,7 @@ uint32_t cpu_ldl_le_mmu(CPUArchState *env, abi_ptr addr,
{
uint32_t ret;
- validate_memop(oi, MO_LEUL);
+ tcg_debug_assert((get_memop(oi) & (MO_BSWAP | MO_SIZE)) == MO_LEUL);
ret = do_ld4_mmu(env, addr, oi, ra, MMU_DATA_LOAD);
plugin_load_cb(env, addr, oi);
return ret;
@@ -2488,7 +2435,7 @@ uint64_t cpu_ldq_le_mmu(CPUArchState *env, abi_ptr addr,
{
uint64_t ret;
- validate_memop(oi, MO_LEUQ);
+ tcg_debug_assert((get_memop(oi) & (MO_BSWAP | MO_SIZE)) == MO_LEUQ);
ret = do_ld8_mmu(env, addr, oi, ra, MMU_DATA_LOAD);
plugin_load_cb(env, addr, oi);
return ret;
@@ -2516,8 +2463,8 @@ Int128 cpu_ld16_be_mmu(CPUArchState *env, abi_ptr addr,
mop = (mop & ~(MO_SIZE | MO_AMASK)) | MO_64 | MO_UNALN;
new_oi = make_memop_idx(mop, mmu_idx);
- h = helper_be_ldq_mmu(env, addr, new_oi, ra);
- l = helper_be_ldq_mmu(env, addr + 8, new_oi, ra);
+ h = helper_ldq_mmu(env, addr, new_oi, ra);
+ l = helper_ldq_mmu(env, addr + 8, new_oi, ra);
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
return int128_make128(l, h);
@@ -2545,8 +2492,8 @@ Int128 cpu_ld16_le_mmu(CPUArchState *env, abi_ptr addr,
mop = (mop & ~(MO_SIZE | MO_AMASK)) | MO_64 | MO_UNALN;
new_oi = make_memop_idx(mop, mmu_idx);
- l = helper_le_ldq_mmu(env, addr, new_oi, ra);
- h = helper_le_ldq_mmu(env, addr + 8, new_oi, ra);
+ l = helper_ldq_mmu(env, addr, new_oi, ra);
+ h = helper_ldq_mmu(env, addr + 8, new_oi, ra);
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
return int128_make128(l, h);
@@ -2754,13 +2701,13 @@ static void do_st_1(CPUArchState *env, MMULookupPageData *p, uint8_t val,
}
}
-void helper_ret_stb_mmu(CPUArchState *env, target_ulong addr, uint8_t val,
- MemOpIdx oi, uintptr_t ra)
+void helper_stb_mmu(CPUArchState *env, target_ulong addr, uint8_t val,
+ MemOpIdx oi, uintptr_t ra)
{
MMULookupLocals l;
bool crosspage;
- validate_memop(oi, MO_UB);
+ tcg_debug_assert((get_memop(oi) & MO_SIZE) == MO_8);
crosspage = mmu_lookup(env, addr, oi, ra, MMU_DATA_STORE, &l);
tcg_debug_assert(!crosspage);
@@ -2802,17 +2749,10 @@ static void do_st2_mmu(CPUArchState *env, target_ulong addr, uint16_t val,
}
}
-void helper_le_stw_mmu(CPUArchState *env, target_ulong addr, uint16_t val,
- MemOpIdx oi, uintptr_t retaddr)
+void helper_stw_mmu(CPUArchState *env, target_ulong addr, uint16_t val,
+ MemOpIdx oi, uintptr_t retaddr)
{
- validate_memop(oi, MO_LEUW);
- do_st2_mmu(env, addr, val, oi, retaddr);
-}
-
-void helper_be_stw_mmu(CPUArchState *env, target_ulong addr, uint16_t val,
- MemOpIdx oi, uintptr_t retaddr)
-{
- validate_memop(oi, MO_BEUW);
+ tcg_debug_assert((get_memop(oi) & MO_SIZE) == MO_16);
do_st2_mmu(env, addr, val, oi, retaddr);
}
@@ -2848,17 +2788,10 @@ static void do_st4_mmu(CPUArchState *env, target_ulong addr, uint32_t val,
}
}
-void helper_le_stl_mmu(CPUArchState *env, target_ulong addr, uint32_t val,
- MemOpIdx oi, uintptr_t retaddr)
+void helper_stl_mmu(CPUArchState *env, target_ulong addr, uint32_t val,
+ MemOpIdx oi, uintptr_t retaddr)
{
- validate_memop(oi, MO_LEUL);
- do_st4_mmu(env, addr, val, oi, retaddr);
-}
-
-void helper_be_stl_mmu(CPUArchState *env, target_ulong addr, uint32_t val,
- MemOpIdx oi, uintptr_t retaddr)
-{
- validate_memop(oi, MO_BEUL);
+ tcg_debug_assert((get_memop(oi) & MO_SIZE) == MO_32);
do_st4_mmu(env, addr, val, oi, retaddr);
}
@@ -2894,17 +2827,10 @@ static void do_st8_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
}
}
-void helper_le_stq_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
- MemOpIdx oi, uintptr_t retaddr)
+void helper_stq_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
+ MemOpIdx oi, uintptr_t retaddr)
{
- validate_memop(oi, MO_LEUQ);
- do_st8_mmu(env, addr, val, oi, retaddr);
-}
-
-void helper_be_stq_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
- MemOpIdx oi, uintptr_t retaddr)
-{
- validate_memop(oi, MO_BEUQ);
+ tcg_debug_assert((get_memop(oi) & MO_SIZE) == MO_64);
do_st8_mmu(env, addr, val, oi, retaddr);
}
@@ -2920,49 +2846,55 @@ static void plugin_store_cb(CPUArchState *env, abi_ptr addr, MemOpIdx oi)
void cpu_stb_mmu(CPUArchState *env, target_ulong addr, uint8_t val,
MemOpIdx oi, uintptr_t retaddr)
{
- helper_ret_stb_mmu(env, addr, val, oi, retaddr);
+ helper_stb_mmu(env, addr, val, oi, retaddr);
plugin_store_cb(env, addr, oi);
}
void cpu_stw_be_mmu(CPUArchState *env, target_ulong addr, uint16_t val,
MemOpIdx oi, uintptr_t retaddr)
{
- helper_be_stw_mmu(env, addr, val, oi, retaddr);
+ tcg_debug_assert((get_memop(oi) & (MO_BSWAP | MO_SIZE)) == MO_BEUW);
+ do_st2_mmu(env, addr, val, oi, retaddr);
plugin_store_cb(env, addr, oi);
}
void cpu_stl_be_mmu(CPUArchState *env, target_ulong addr, uint32_t val,
MemOpIdx oi, uintptr_t retaddr)
{
- helper_be_stl_mmu(env, addr, val, oi, retaddr);
+ tcg_debug_assert((get_memop(oi) & (MO_BSWAP | MO_SIZE)) == MO_BEUL);
+ do_st4_mmu(env, addr, val, oi, retaddr);
plugin_store_cb(env, addr, oi);
}
void cpu_stq_be_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
MemOpIdx oi, uintptr_t retaddr)
{
- helper_be_stq_mmu(env, addr, val, oi, retaddr);
+ tcg_debug_assert((get_memop(oi) & (MO_BSWAP | MO_SIZE)) == MO_BEUQ);
+ do_st8_mmu(env, addr, val, oi, retaddr);
plugin_store_cb(env, addr, oi);
}
void cpu_stw_le_mmu(CPUArchState *env, target_ulong addr, uint16_t val,
MemOpIdx oi, uintptr_t retaddr)
{
- helper_le_stw_mmu(env, addr, val, oi, retaddr);
+ tcg_debug_assert((get_memop(oi) & (MO_BSWAP | MO_SIZE)) == MO_LEUW);
+ do_st2_mmu(env, addr, val, oi, retaddr);
plugin_store_cb(env, addr, oi);
}
void cpu_stl_le_mmu(CPUArchState *env, target_ulong addr, uint32_t val,
MemOpIdx oi, uintptr_t retaddr)
{
- helper_le_stl_mmu(env, addr, val, oi, retaddr);
+ tcg_debug_assert((get_memop(oi) & (MO_BSWAP | MO_SIZE)) == MO_LEUL);
+ do_st4_mmu(env, addr, val, oi, retaddr);
plugin_store_cb(env, addr, oi);
}
void cpu_stq_le_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
MemOpIdx oi, uintptr_t retaddr)
{
- helper_le_stq_mmu(env, addr, val, oi, retaddr);
+ tcg_debug_assert((get_memop(oi) & (MO_BSWAP | MO_SIZE)) == MO_LEUQ);
+ do_st8_mmu(env, addr, val, oi, retaddr);
plugin_store_cb(env, addr, oi);
}
@@ -2987,8 +2919,8 @@ void cpu_st16_be_mmu(CPUArchState *env, abi_ptr addr, Int128 val,
mop = (mop & ~(MO_SIZE | MO_AMASK)) | MO_64 | MO_UNALN;
new_oi = make_memop_idx(mop, mmu_idx);
- helper_be_stq_mmu(env, addr, int128_gethi(val), new_oi, ra);
- helper_be_stq_mmu(env, addr + 8, int128_getlo(val), new_oi, ra);
+ helper_stq_mmu(env, addr, int128_gethi(val), new_oi, ra);
+ helper_stq_mmu(env, addr + 8, int128_getlo(val), new_oi, ra);
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
}
@@ -3014,8 +2946,8 @@ void cpu_st16_le_mmu(CPUArchState *env, abi_ptr addr, Int128 val,
mop = (mop & ~(MO_SIZE | MO_AMASK)) | MO_64 | MO_UNALN;
new_oi = make_memop_idx(mop, mmu_idx);
- helper_le_stq_mmu(env, addr, int128_getlo(val), new_oi, ra);
- helper_le_stq_mmu(env, addr + 8, int128_gethi(val), new_oi, ra);
+ helper_stq_mmu(env, addr, int128_getlo(val), new_oi, ra);
+ helper_stq_mmu(env, addr + 8, int128_gethi(val), new_oi, ra);
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
}
diff --git a/docs/devel/loads-stores.rst b/docs/devel/loads-stores.rst
index ad5dfe133e..d2cefc77a2 100644
--- a/docs/devel/loads-stores.rst
+++ b/docs/devel/loads-stores.rst
@@ -297,31 +297,20 @@ swap: ``translator_ld{sign}{size}_swap(env, ptr, swap)``
Regexes for git grep
- ``\<translator_ld[us]\?[bwlq]\(_swap\)\?\>``
-``helper_*_{ld,st}*_mmu``
+``helper_{ld,st}*_mmu``
~~~~~~~~~~~~~~~~~~~~~~~~~
These functions are intended primarily to be called by the code
-generated by the TCG backend. They may also be called by target
-CPU helper function code. Like the ``cpu_{ld,st}_mmuidx_ra`` functions
-they perform accesses by guest virtual address, with a given ``mmuidx``.
+generated by the TCG backend. Like the ``cpu_{ld,st}_mmu`` functions
+they perform accesses by guest virtual address, with a given ``MemOpIdx``.
-These functions specify an ``opindex`` parameter which encodes
-(among other things) the mmu index to use for the access. This parameter
-should be created by calling ``make_memop_idx()``.
+They differ from ``cpu_{ld,st}_mmu`` in that they take the endianness
+of the operation only from the MemOpIdx, and loads extend the return
+value to the size of a host general register (``tcg_target_ulong``).
-The ``retaddr`` parameter should be the result of GETPC() called directly
-from the top level HELPER(foo) function (or 0 if no guest CPU state
-unwinding is required).
+load: ``helper_ld{sign}{size}_mmu(env, addr, opindex, retaddr)``
-**TODO** The names of these functions are a bit odd for historical
-reasons because they were originally expected to be called only from
-within generated code. We should rename them to bring them more in
-line with the other memory access functions. The explicit endianness
-is the only feature they have beyond ``*_mmuidx_ra``.
-
-load: ``helper_{endian}_ld{sign}{size}_mmu(env, addr, opindex, retaddr)``
-
-store: ``helper_{endian}_st{size}_mmu(env, addr, val, opindex, retaddr)``
+store: ``helper_{size}_mmu(env, addr, val, opindex, retaddr)``
``sign``
- (empty) : for 32 or 64 bit sizes
@@ -334,14 +323,9 @@ store: ``helper_{endian}_st{size}_mmu(env, addr, val, opindex, retaddr)``
- ``l`` : 32 bits
- ``q`` : 64 bits
-``endian``
- - ``le`` : little endian
- - ``be`` : big endian
- - ``ret`` : target endianness
-
Regexes for git grep
- - ``\<helper_\(le\|be\|ret\)_ld[us]\?[bwlq]_mmu\>``
- - ``\<helper_\(le\|be\|ret\)_st[bwlq]_mmu\>``
+ - ``\<helper_ld[us]\?[bwlq]_mmu\>``
+ - ``\<helper_st[bwlq]_mmu\>``
``address_space_*``
~~~~~~~~~~~~~~~~~~~
diff --git a/tcg/aarch64/tcg-target.c.inc b/tcg/aarch64/tcg-target.c.inc
index dfe569dd8c..001a71bbc0 100644
--- a/tcg/aarch64/tcg-target.c.inc
+++ b/tcg/aarch64/tcg-target.c.inc
@@ -1564,37 +1564,26 @@ static void tcg_out_adr(TCGContext *s, TCGReg rd, const void *target)
}
#ifdef CONFIG_SOFTMMU
-/* helper signature: helper_ret_ld_mmu(CPUState *env, target_ulong addr,
- * MemOpIdx oi, uintptr_t ra)
+/*
+ * helper signature: helper_ld*_mmu(CPUState *env, target_ulong addr,
+ * MemOpIdx oi, uintptr_t ra)
*/
static void * const qemu_ld_helpers[MO_SIZE + 1] = {
- [MO_8] = helper_ret_ldub_mmu,
-#if HOST_BIG_ENDIAN
- [MO_16] = helper_be_lduw_mmu,
- [MO_32] = helper_be_ldul_mmu,
- [MO_64] = helper_be_ldq_mmu,
-#else
- [MO_16] = helper_le_lduw_mmu,
- [MO_32] = helper_le_ldul_mmu,
- [MO_64] = helper_le_ldq_mmu,
-#endif
+ [MO_8] = helper_ldub_mmu,
+ [MO_16] = helper_lduw_mmu,
+ [MO_32] = helper_ldul_mmu,
+ [MO_64] = helper_ldq_mmu,
};
-/* helper signature: helper_ret_st_mmu(CPUState *env, target_ulong addr,
- * uintxx_t val, MemOpIdx oi,
- * uintptr_t ra)
+/*
+ * helper signature: helper_st*_mmu(CPUState *env, target_ulong addr,
+ * uintxx_t val, MemOpIdx oi, uintptr_t ra)
*/
static void * const qemu_st_helpers[MO_SIZE + 1] = {
- [MO_8] = helper_ret_stb_mmu,
-#if HOST_BIG_ENDIAN
- [MO_16] = helper_be_stw_mmu,
- [MO_32] = helper_be_stl_mmu,
- [MO_64] = helper_be_stq_mmu,
-#else
- [MO_16] = helper_le_stw_mmu,
- [MO_32] = helper_le_stl_mmu,
- [MO_64] = helper_le_stq_mmu,
-#endif
+ [MO_8] = helper_stb_mmu,
+ [MO_16] = helper_stw_mmu,
+ [MO_32] = helper_stl_mmu,
+ [MO_64] = helper_stq_mmu,
};
static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
diff --git a/tcg/arm/tcg-target.c.inc b/tcg/arm/tcg-target.c.inc
index add90ddeb4..1f89745c86 100644
--- a/tcg/arm/tcg-target.c.inc
+++ b/tcg/arm/tcg-target.c.inc
@@ -1300,41 +1300,26 @@ static void tcg_out_vldst(TCGContext *s, ARMInsn insn,
}
#ifdef CONFIG_SOFTMMU
-/* helper signature: helper_ret_ld_mmu(CPUState *env, target_ulong addr,
- * int mmu_idx, uintptr_t ra)
+/*
+ * helper signature: helper_ld*_mmu(CPUState *env, target_ulong addr,
+ * int mmu_idx, uintptr_t ra)
*/
-static void * const qemu_ld_helpers[MO_SSIZE + 1] = {
- [MO_UB] = helper_ret_ldub_mmu,
- [MO_SB] = helper_ret_ldsb_mmu,
-#if HOST_BIG_ENDIAN
- [MO_UW] = helper_be_lduw_mmu,
- [MO_UL] = helper_be_ldul_mmu,
- [MO_UQ] = helper_be_ldq_mmu,
- [MO_SW] = helper_be_ldsw_mmu,
- [MO_SL] = helper_be_ldul_mmu,
-#else
- [MO_UW] = helper_le_lduw_mmu,
- [MO_UL] = helper_le_ldul_mmu,
- [MO_UQ] = helper_le_ldq_mmu,
- [MO_SW] = helper_le_ldsw_mmu,
- [MO_SL] = helper_le_ldul_mmu,
-#endif
+static void * const qemu_ld_helpers[MO_SIZE + 1] = {
+ [MO_UB] = helper_ldub_mmu,
+ [MO_UW] = helper_lduw_mmu,
+ [MO_UL] = helper_ldul_mmu,
+ [MO_UQ] = helper_ldq_mmu,
};
-/* helper signature: helper_ret_st_mmu(CPUState *env, target_ulong addr,
- * uintxx_t val, int mmu_idx, uintptr_t ra)
+/*
+ * helper signature: helper_st*_mmu(CPUState *env, target_ulong addr,
+ * uintxx_t val, int mmu_idx, uintptr_t ra)
*/
static void * const qemu_st_helpers[MO_SIZE + 1] = {
- [MO_8] = helper_ret_stb_mmu,
-#if HOST_BIG_ENDIAN
- [MO_16] = helper_be_stw_mmu,
- [MO_32] = helper_be_stl_mmu,
- [MO_64] = helper_be_stq_mmu,
-#else
- [MO_16] = helper_le_stw_mmu,
- [MO_32] = helper_le_stl_mmu,
- [MO_64] = helper_le_stq_mmu,
-#endif
+ [MO_8] = helper_stb_mmu,
+ [MO_16] = helper_stw_mmu,
+ [MO_32] = helper_stl_mmu,
+ [MO_64] = helper_stq_mmu,
};
/* Helper routines for marshalling helper function arguments into
diff --git a/tcg/i386/tcg-target.c.inc b/tcg/i386/tcg-target.c.inc
index 1361960156..24e9efe631 100644
--- a/tcg/i386/tcg-target.c.inc
+++ b/tcg/i386/tcg-target.c.inc
@@ -1728,30 +1728,26 @@ static void tcg_out_nopn(TCGContext *s, int n)
}
#if defined(CONFIG_SOFTMMU)
-/* helper signature: helper_ret_ld_mmu(CPUState *env, target_ulong addr,
- * int mmu_idx, uintptr_t ra)
+/*
+ * helper signature: helper_ld*_mmu(CPUState *env, target_ulong addr,
+ * int mmu_idx, uintptr_t ra)
*/
-static void * const qemu_ld_helpers[(MO_SIZE | MO_BSWAP) + 1] = {
- [MO_UB] = helper_ret_ldub_mmu,
- [MO_LEUW] = helper_le_lduw_mmu,
- [MO_LEUL] = helper_le_ldul_mmu,
- [MO_LEUQ] = helper_le_ldq_mmu,
- [MO_BEUW] = helper_be_lduw_mmu,
- [MO_BEUL] = helper_be_ldul_mmu,
- [MO_BEUQ] = helper_be_ldq_mmu,
+static void * const qemu_ld_helpers[MO_SIZE + 1] = {
+ [MO_UB] = helper_ldub_mmu,
+ [MO_UW] = helper_lduw_mmu,
+ [MO_UL] = helper_ldul_mmu,
+ [MO_UQ] = helper_ldq_mmu,
};
-/* helper signature: helper_ret_st_mmu(CPUState *env, target_ulong addr,
- * uintxx_t val, int mmu_idx, uintptr_t ra)
+/*
+ * helper signature: helper_st*_mmu(CPUState *env, target_ulong addr,
+ * uintxx_t val, int mmu_idx, uintptr_t ra)
*/
-static void * const qemu_st_helpers[(MO_SIZE | MO_BSWAP) + 1] = {
- [MO_UB] = helper_ret_stb_mmu,
- [MO_LEUW] = helper_le_stw_mmu,
- [MO_LEUL] = helper_le_stl_mmu,
- [MO_LEUQ] = helper_le_stq_mmu,
- [MO_BEUW] = helper_be_stw_mmu,
- [MO_BEUL] = helper_be_stl_mmu,
- [MO_BEUQ] = helper_be_stq_mmu,
+static void * const qemu_st_helpers[MO_SIZE + 1] = {
+ [MO_UB] = helper_stb_mmu,
+ [MO_UW] = helper_stw_mmu,
+ [MO_UL] = helper_stl_mmu,
+ [MO_UQ] = helper_stq_mmu,
};
/* Perform the TLB load and compare.
@@ -1926,7 +1922,7 @@ static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
(uintptr_t)l->raddr);
}
- tcg_out_branch(s, 1, qemu_ld_helpers[opc & (MO_BSWAP | MO_SIZE)]);
+ tcg_out_branch(s, 1, qemu_ld_helpers[opc & MO_SIZE]);
data_reg = l->datalo_reg;
switch (opc & MO_SSIZE) {
@@ -2033,7 +2029,7 @@ static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
/* "Tail call" to the helper, with the return address back inline. */
tcg_out_push(s, retaddr);
- tcg_out_jmp(s, qemu_st_helpers[opc & (MO_BSWAP | MO_SIZE)]);
+ tcg_out_jmp(s, qemu_st_helpers[opc & MO_SIZE]);
return true;
}
#else
diff --git a/tcg/loongarch64/tcg-target.c.inc b/tcg/loongarch64/tcg-target.c.inc
index f6b0ed00bb..e73b48bd0f 100644
--- a/tcg/loongarch64/tcg-target.c.inc
+++ b/tcg/loongarch64/tcg-target.c.inc
@@ -655,26 +655,25 @@ static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
#if defined(CONFIG_SOFTMMU)
/*
- * helper signature: helper_ret_ld_mmu(CPUState *env, target_ulong addr,
- * MemOpIdx oi, uintptr_t ra)
+ * helper signature: helper_ld*_mmu(CPUState *env, target_ulong addr,
+ * MemOpIdx oi, uintptr_t ra)
*/
static void * const qemu_ld_helpers[4] = {
- [MO_8] = helper_ret_ldub_mmu,
- [MO_16] = helper_le_lduw_mmu,
- [MO_32] = helper_le_ldul_mmu,
- [MO_64] = helper_le_ldq_mmu,
+ [MO_8] = helper_ldub_mmu,
+ [MO_16] = helper_lduw_mmu,
+ [MO_32] = helper_ldul_mmu,
+ [MO_64] = helper_ldq_mmu,
};
/*
- * helper signature: helper_ret_st_mmu(CPUState *env, target_ulong addr,
- * uintxx_t val, MemOpIdx oi,
- * uintptr_t ra)
+ * helper signature: helper_st*_mmu(CPUState *env, target_ulong addr,
+ * uintxx_t val, MemOpIdx oi, uintptr_t ra)
*/
static void * const qemu_st_helpers[4] = {
- [MO_8] = helper_ret_stb_mmu,
- [MO_16] = helper_le_stw_mmu,
- [MO_32] = helper_le_stl_mmu,
- [MO_64] = helper_le_stq_mmu,
+ [MO_8] = helper_stb_mmu,
+ [MO_16] = helper_stw_mmu,
+ [MO_32] = helper_stl_mmu,
+ [MO_64] = helper_stq_mmu,
};
/* We expect to use a 12-bit negative offset from ENV. */
diff --git a/tcg/mips/tcg-target.c.inc b/tcg/mips/tcg-target.c.inc
index 92883176c6..a23e2c409f 100644
--- a/tcg/mips/tcg-target.c.inc
+++ b/tcg/mips/tcg-target.c.inc
@@ -1037,31 +1037,21 @@ static void tcg_out_call(TCGContext *s, const tcg_insn_unit *arg,
}
#if defined(CONFIG_SOFTMMU)
-static void * const qemu_ld_helpers[(MO_SSIZE | MO_BSWAP) + 1] = {
- [MO_UB] = helper_ret_ldub_mmu,
- [MO_SB] = helper_ret_ldsb_mmu,
- [MO_LEUW] = helper_le_lduw_mmu,
- [MO_LESW] = helper_le_ldsw_mmu,
- [MO_LEUL] = helper_le_ldul_mmu,
- [MO_LEUQ] = helper_le_ldq_mmu,
- [MO_BEUW] = helper_be_lduw_mmu,
- [MO_BESW] = helper_be_ldsw_mmu,
- [MO_BEUL] = helper_be_ldul_mmu,
- [MO_BEUQ] = helper_be_ldq_mmu,
-#if TCG_TARGET_REG_BITS == 64
- [MO_LESL] = helper_le_ldsl_mmu,
- [MO_BESL] = helper_be_ldsl_mmu,
-#endif
+static void * const qemu_ld_helpers[MO_SSIZE + 1] = {
+ [MO_UB] = helper_ldub_mmu,
+ [MO_SB] = helper_ldsb_mmu,
+ [MO_UW] = helper_lduw_mmu,
+ [MO_SW] = helper_ldsw_mmu,
+ [MO_UL] = helper_ldul_mmu,
+ [MO_SL] = helper_ldsl_mmu,
+ [MO_UQ] = helper_ldq_mmu,
};
-static void * const qemu_st_helpers[(MO_SIZE | MO_BSWAP) + 1] = {
- [MO_UB] = helper_ret_stb_mmu,
- [MO_LEUW] = helper_le_stw_mmu,
- [MO_LEUL] = helper_le_stl_mmu,
- [MO_LEUQ] = helper_le_stq_mmu,
- [MO_BEUW] = helper_be_stw_mmu,
- [MO_BEUL] = helper_be_stl_mmu,
- [MO_BEUQ] = helper_be_stq_mmu,
+static void * const qemu_st_helpers[MO_SIZE + 1] = {
+ [MO_UB] = helper_stb_mmu,
+ [MO_UW] = helper_stw_mmu,
+ [MO_UL] = helper_stl_mmu,
+ [MO_UQ] = helper_stq_mmu,
};
/* Helper routines for marshalling helper function arguments into
@@ -1267,7 +1257,7 @@ static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
}
i = tcg_out_call_iarg_imm(s, i, oi);
i = tcg_out_call_iarg_imm(s, i, (intptr_t)l->raddr);
- tcg_out_call_int(s, qemu_ld_helpers[opc & (MO_BSWAP | MO_SSIZE)], false);
+ tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SSIZE], false);
/* delay slot */
tcg_out_mov(s, TCG_TYPE_PTR, tcg_target_call_iarg_regs[0], TCG_AREG0);
@@ -1345,7 +1335,7 @@ static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
computation to take place in the return address register. */
tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_RA, (intptr_t)l->raddr);
i = tcg_out_call_iarg_reg(s, i, TCG_REG_RA);
- tcg_out_call_int(s, qemu_st_helpers[opc & (MO_BSWAP | MO_SIZE)], true);
+ tcg_out_call_int(s, qemu_st_helpers[opc & MO_SIZE], true);
/* delay slot */
tcg_out_mov(s, TCG_TYPE_PTR, tcg_target_call_iarg_regs[0], TCG_AREG0);
return true;
diff --git a/tcg/ppc/tcg-target.c.inc b/tcg/ppc/tcg-target.c.inc
index e86d4a5e78..3e9fc9bd25 100644
--- a/tcg/ppc/tcg-target.c.inc
+++ b/tcg/ppc/tcg-target.c.inc
@@ -2052,27 +2052,21 @@ static const uint32_t qemu_exts_opc[4] = {
/* helper signature: helper_ld_mmu(CPUState *env, target_ulong addr,
* int mmu_idx, uintptr_t ra)
*/
-static void * const qemu_ld_helpers[(MO_SIZE | MO_BSWAP) + 1] = {
- [MO_UB] = helper_ret_ldub_mmu,
- [MO_LEUW] = helper_le_lduw_mmu,
- [MO_LEUL] = helper_le_ldul_mmu,
- [MO_LEUQ] = helper_le_ldq_mmu,
- [MO_BEUW] = helper_be_lduw_mmu,
- [MO_BEUL] = helper_be_ldul_mmu,
- [MO_BEUQ] = helper_be_ldq_mmu,
+static void * const qemu_ld_helpers[MO_SIZE + 1] = {
+ [MO_UB] = helper_ldub_mmu,
+ [MO_UW] = helper_lduw_mmu,
+ [MO_UL] = helper_ldul_mmu,
+ [MO_UQ] = helper_ldq_mmu,
};
/* helper signature: helper_st_mmu(CPUState *env, target_ulong addr,
* uintxx_t val, int mmu_idx, uintptr_t ra)
*/
-static void * const qemu_st_helpers[(MO_SIZE | MO_BSWAP) + 1] = {
- [MO_UB] = helper_ret_stb_mmu,
- [MO_LEUW] = helper_le_stw_mmu,
- [MO_LEUL] = helper_le_stl_mmu,
- [MO_LEUQ] = helper_le_stq_mmu,
- [MO_BEUW] = helper_be_stw_mmu,
- [MO_BEUL] = helper_be_stl_mmu,
- [MO_BEUQ] = helper_be_stq_mmu,
+static void * const qemu_st_helpers[MO_SIZE + 1] = {
+ [MO_UB] = helper_stb_mmu,
+ [MO_UW] = helper_stw_mmu,
+ [MO_UL] = helper_stl_mmu,
+ [MO_UQ] = helper_stq_mmu,
};
/* We expect to use a 16-bit negative offset from ENV. */
@@ -2234,7 +2228,7 @@ static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
tcg_out_movi(s, TCG_TYPE_I32, arg++, oi);
tcg_out32(s, MFSPR | RT(arg) | LR);
- tcg_out_call_int(s, LK, qemu_ld_helpers[opc & (MO_BSWAP | MO_SIZE)]);
+ tcg_out_call_int(s, LK, qemu_ld_helpers[opc & MO_SIZE]);
lo = lb->datalo_reg;
hi = lb->datahi_reg;
@@ -2303,7 +2297,7 @@ static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
tcg_out_movi(s, TCG_TYPE_I32, arg++, oi);
tcg_out32(s, MFSPR | RT(arg) | LR);
- tcg_out_call_int(s, LK, qemu_st_helpers[opc & (MO_BSWAP | MO_SIZE)]);
+ tcg_out_call_int(s, LK, qemu_st_helpers[opc & MO_SIZE]);
tcg_out_b(s, 0, lb->raddr);
return true;
diff --git a/tcg/riscv/tcg-target.c.inc b/tcg/riscv/tcg-target.c.inc
index 417736cae7..7261f15197 100644
--- a/tcg/riscv/tcg-target.c.inc
+++ b/tcg/riscv/tcg-target.c.inc
@@ -858,46 +858,29 @@ static void tcg_out_mb(TCGContext *s, TCGArg a0)
*/
#if defined(CONFIG_SOFTMMU)
-/* helper signature: helper_ret_ld_mmu(CPUState *env, target_ulong addr,
- * MemOpIdx oi, uintptr_t ra)
+/*
+ * helper signature: helper_ld*_mmu(CPUState *env, target_ulong addr,
+ * MemOpIdx oi, uintptr_t ra)
*/
static void * const qemu_ld_helpers[MO_SSIZE + 1] = {
- [MO_UB] = helper_ret_ldub_mmu,
- [MO_SB] = helper_ret_ldsb_mmu,
-#if HOST_BIG_ENDIAN
- [MO_UW] = helper_be_lduw_mmu,
- [MO_SW] = helper_be_ldsw_mmu,
- [MO_UL] = helper_be_ldul_mmu,
-#if TCG_TARGET_REG_BITS == 64
- [MO_SL] = helper_be_ldsl_mmu,
-#endif
- [MO_UQ] = helper_be_ldq_mmu,
-#else
- [MO_UW] = helper_le_lduw_mmu,
- [MO_SW] = helper_le_ldsw_mmu,
- [MO_UL] = helper_le_ldul_mmu,
-#if TCG_TARGET_REG_BITS == 64
- [MO_SL] = helper_le_ldsl_mmu,
-#endif
- [MO_UQ] = helper_le_ldq_mmu,
-#endif
+ [MO_UB] = helper_ldub_mmu,
+ [MO_SB] = helper_ldsb_mmu,
+ [MO_UW] = helper_lduw_mmu,
+ [MO_SW] = helper_ldsw_mmu,
+ [MO_UL] = helper_ldul_mmu,
+ [MO_SL] = helper_ldsl_mmu,
+ [MO_UQ] = helper_ldq_mmu,
};
-/* helper signature: helper_ret_st_mmu(CPUState *env, target_ulong addr,
- * uintxx_t val, MemOpIdx oi,
- * uintptr_t ra)
+/*
+ * helper signature: helper_st*_mmu(CPUState *env, target_ulong addr,
+ * uintxx_t val, MemOpIdx oi, uintptr_t ra)
*/
static void * const qemu_st_helpers[MO_SIZE + 1] = {
- [MO_8] = helper_ret_stb_mmu,
-#if HOST_BIG_ENDIAN
- [MO_16] = helper_be_stw_mmu,
- [MO_32] = helper_be_stl_mmu,
- [MO_64] = helper_be_stq_mmu,
-#else
- [MO_16] = helper_le_stw_mmu,
- [MO_32] = helper_le_stl_mmu,
- [MO_64] = helper_le_stq_mmu,
-#endif
+ [MO_8] = helper_stb_mmu,
+ [MO_16] = helper_stw_mmu,
+ [MO_32] = helper_stl_mmu,
+ [MO_64] = helper_stq_mmu,
};
/* We don't support oversize guests */
diff --git a/tcg/s390x/tcg-target.c.inc b/tcg/s390x/tcg-target.c.inc
index 50655e9d1d..30556c430f 100644
--- a/tcg/s390x/tcg-target.c.inc
+++ b/tcg/s390x/tcg-target.c.inc
@@ -438,29 +438,21 @@ static const uint8_t tcg_cond_to_ltr_cond[] = {
};
#ifdef CONFIG_SOFTMMU
-static void * const qemu_ld_helpers[(MO_SSIZE | MO_BSWAP) + 1] = {
- [MO_UB] = helper_ret_ldub_mmu,
- [MO_SB] = helper_ret_ldsb_mmu,
- [MO_LEUW] = helper_le_lduw_mmu,
- [MO_LESW] = helper_le_ldsw_mmu,
- [MO_LEUL] = helper_le_ldul_mmu,
- [MO_LESL] = helper_le_ldsl_mmu,
- [MO_LEUQ] = helper_le_ldq_mmu,
- [MO_BEUW] = helper_be_lduw_mmu,
- [MO_BESW] = helper_be_ldsw_mmu,
- [MO_BEUL] = helper_be_ldul_mmu,
- [MO_BESL] = helper_be_ldsl_mmu,
- [MO_BEUQ] = helper_be_ldq_mmu,
+static void * const qemu_ld_helpers[MO_SSIZE + 1] = {
+ [MO_UB] = helper_ldub_mmu,
+ [MO_SB] = helper_ldsb_mmu,
+ [MO_UW] = helper_lduw_mmu,
+ [MO_SW] = helper_ldsw_mmu,
+ [MO_UL] = helper_ldul_mmu,
+ [MO_SL] = helper_ldsl_mmu,
+ [MO_UQ] = helper_ldq_mmu,
};
-static void * const qemu_st_helpers[(MO_SIZE | MO_BSWAP) + 1] = {
- [MO_UB] = helper_ret_stb_mmu,
- [MO_LEUW] = helper_le_stw_mmu,
- [MO_LEUL] = helper_le_stl_mmu,
- [MO_LEUQ] = helper_le_stq_mmu,
- [MO_BEUW] = helper_be_stw_mmu,
- [MO_BEUL] = helper_be_stl_mmu,
- [MO_BEUQ] = helper_be_stq_mmu,
+static void * const qemu_st_helpers[MO_SIZE + 1] = {
+ [MO_UB] = helper_stb_mmu,
+ [MO_UW] = helper_stw_mmu,
+ [MO_UL] = helper_stl_mmu,
+ [MO_UQ] = helper_stq_mmu,
};
#endif
@@ -1913,7 +1905,7 @@ static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
}
tcg_out_movi(s, TCG_TYPE_I32, TCG_REG_R4, oi);
tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R5, (uintptr_t)lb->raddr);
- tcg_out_call_int(s, qemu_ld_helpers[opc & (MO_BSWAP | MO_SSIZE)]);
+ tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SSIZE]);
tcg_out_mov(s, TCG_TYPE_I64, data_reg, TCG_REG_R2);
tgen_gotoi(s, S390_CC_ALWAYS, lb->raddr);
@@ -1954,7 +1946,7 @@ static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
}
tcg_out_movi(s, TCG_TYPE_I32, TCG_REG_R5, oi);
tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R6, (uintptr_t)lb->raddr);
- tcg_out_call_int(s, qemu_st_helpers[opc & (MO_BSWAP | MO_SIZE)]);
+ tcg_out_call_int(s, qemu_st_helpers[opc & MO_SIZE]);
tgen_gotoi(s, S390_CC_ALWAYS, lb->raddr);
return true;
diff --git a/tcg/sparc64/tcg-target.c.inc b/tcg/sparc64/tcg-target.c.inc
index 9b5afb8248..f9334b7c56 100644
--- a/tcg/sparc64/tcg-target.c.inc
+++ b/tcg/sparc64/tcg-target.c.inc
@@ -880,8 +880,8 @@ static void tcg_out_mb(TCGContext *s, TCGArg a0)
}
#ifdef CONFIG_SOFTMMU
-static const tcg_insn_unit *qemu_ld_trampoline[(MO_SSIZE | MO_BSWAP) + 1];
-static const tcg_insn_unit *qemu_st_trampoline[(MO_SIZE | MO_BSWAP) + 1];
+static const tcg_insn_unit *qemu_ld_trampoline[MO_SSIZE + 1];
+static const tcg_insn_unit *qemu_st_trampoline[MO_SIZE + 1];
static void emit_extend(TCGContext *s, TCGReg r, int op)
{
@@ -907,25 +907,18 @@ static void emit_extend(TCGContext *s, TCGReg r, int op)
static void build_trampolines(TCGContext *s)
{
static void * const qemu_ld_helpers[] = {
- [MO_UB] = helper_ret_ldub_mmu,
- [MO_SB] = helper_ret_ldsb_mmu,
- [MO_LEUW] = helper_le_lduw_mmu,
- [MO_LESW] = helper_le_ldsw_mmu,
- [MO_LEUL] = helper_le_ldul_mmu,
- [MO_LEUQ] = helper_le_ldq_mmu,
- [MO_BEUW] = helper_be_lduw_mmu,
- [MO_BESW] = helper_be_ldsw_mmu,
- [MO_BEUL] = helper_be_ldul_mmu,
- [MO_BEUQ] = helper_be_ldq_mmu,
+ [MO_UB] = helper_ldub_mmu,
+ [MO_SB] = helper_ldsb_mmu,
+ [MO_UW] = helper_lduw_mmu,
+ [MO_SW] = helper_ldsw_mmu,
+ [MO_UL] = helper_ldul_mmu,
+ [MO_UQ] = helper_ldq_mmu,
};
static void * const qemu_st_helpers[] = {
- [MO_UB] = helper_ret_stb_mmu,
- [MO_LEUW] = helper_le_stw_mmu,
- [MO_LEUL] = helper_le_stl_mmu,
- [MO_LEUQ] = helper_le_stq_mmu,
- [MO_BEUW] = helper_be_stw_mmu,
- [MO_BEUL] = helper_be_stl_mmu,
- [MO_BEUQ] = helper_be_stq_mmu,
+ [MO_UB] = helper_stb_mmu,
+ [MO_UW] = helper_stw_mmu,
+ [MO_UL] = helper_stl_mmu,
+ [MO_UQ] = helper_stq_mmu,
};
int i;
@@ -1196,9 +1189,9 @@ static void tcg_out_qemu_ld(TCGContext *s, TCGReg data, TCGReg addr,
/* We use the helpers to extend SB and SW data, leaving the case
of SL needing explicit extending below. */
if ((memop & MO_SSIZE) == MO_SL) {
- func = qemu_ld_trampoline[memop & (MO_BSWAP | MO_SIZE)];
+ func = qemu_ld_trampoline[MO_UL];
} else {
- func = qemu_ld_trampoline[memop & (MO_BSWAP | MO_SSIZE)];
+ func = qemu_ld_trampoline[memop & MO_SSIZE];
}
tcg_debug_assert(func != NULL);
tcg_out_call_nodelay(s, func, false);
@@ -1338,7 +1331,7 @@ static void tcg_out_qemu_st(TCGContext *s, TCGReg data, TCGReg addr,
tcg_out_mov(s, TCG_TYPE_REG, TCG_REG_O1, addrz);
tcg_out_mov(s, TCG_TYPE_REG, TCG_REG_O2, data);
- func = qemu_st_trampoline[memop & (MO_BSWAP | MO_SIZE)];
+ func = qemu_st_trampoline[memop & MO_SIZE];
tcg_debug_assert(func != NULL);
tcg_out_call_nodelay(s, func, false);
/* delay slot */
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* Re: [PATCH for-8.0 10/29] tcg: Unify helper_{be,le}_{ld,st}*
2022-11-18 9:47 ` [PATCH for-8.0 10/29] tcg: Unify helper_{be,le}_{ld,st}* Richard Henderson
@ 2022-11-21 12:48 ` Philippe Mathieu-Daudé
0 siblings, 0 replies; 48+ messages in thread
From: Philippe Mathieu-Daudé @ 2022-11-21 12:48 UTC (permalink / raw)
To: Richard Henderson, qemu-devel
On 18/11/22 10:47, Richard Henderson wrote:
> With the current structure of cputlb.c, there is no difference
> between the little-endian and big-endian entry points, aside
> from the assert. Unify the pairs of functions.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
> include/tcg/tcg-ldst.h | 60 ++++------
> accel/tcg/cputlb.c | 190 ++++++++++---------------------
> docs/devel/loads-stores.rst | 36 ++----
> tcg/aarch64/tcg-target.c.inc | 39 +++----
> tcg/arm/tcg-target.c.inc | 45 +++-----
> tcg/i386/tcg-target.c.inc | 40 +++----
> tcg/loongarch64/tcg-target.c.inc | 25 ++--
> tcg/mips/tcg-target.c.inc | 40 +++----
> tcg/ppc/tcg-target.c.inc | 30 ++---
> tcg/riscv/tcg-target.c.inc | 51 +++------
> tcg/s390x/tcg-target.c.inc | 38 +++----
> tcg/sparc64/tcg-target.c.inc | 37 +++---
> 12 files changed, 226 insertions(+), 405 deletions(-)
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
^ permalink raw reply [flat|nested] 48+ messages in thread
* [PATCH for-8.0 11/29] accel/tcg: Implement helper_{ld, st}*_mmu for user-only
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (9 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 10/29] tcg: Unify helper_{be,le}_{ld,st}* Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-18 9:47 ` [PATCH for-8.0 12/29] tcg: Add 128-bit guest memory primitives Richard Henderson
` (17 subsequent siblings)
28 siblings, 0 replies; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
TCG backends may need to defer to a helper to implement
the atomicity required by a given operation. Mirror the
interface used in system mode.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
include/tcg/tcg-ldst.h | 6 +-
accel/tcg/user-exec.c | 392 ++++++++++++++++++++++++++++-------------
2 files changed, 276 insertions(+), 122 deletions(-)
diff --git a/include/tcg/tcg-ldst.h b/include/tcg/tcg-ldst.h
index 56fa7afe5e..c1d945fd66 100644
--- a/include/tcg/tcg-ldst.h
+++ b/include/tcg/tcg-ldst.h
@@ -25,8 +25,6 @@
#ifndef TCG_LDST_H
#define TCG_LDST_H
-#ifdef CONFIG_SOFTMMU
-
/* Value zero-extended to tcg register size. */
tcg_target_ulong helper_ldub_mmu(CPUArchState *env, target_ulong addr,
MemOpIdx oi, uintptr_t retaddr);
@@ -54,10 +52,10 @@ void helper_stl_mmu(CPUArchState *env, target_ulong addr, uint32_t val,
void helper_stq_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
MemOpIdx oi, uintptr_t retaddr);
-#else
+#ifdef CONFIG_USER_ONLY
G_NORETURN void helper_unaligned_ld(CPUArchState *env, target_ulong addr);
G_NORETURN void helper_unaligned_st(CPUArchState *env, target_ulong addr);
-#endif /* CONFIG_SOFTMMU */
+#endif /* CONFIG_USER_ONLY*/
#endif /* TCG_LDST_H */
diff --git a/accel/tcg/user-exec.c b/accel/tcg/user-exec.c
index ddba8c9dd7..3455ff45a4 100644
--- a/accel/tcg/user-exec.c
+++ b/accel/tcg/user-exec.c
@@ -254,21 +254,6 @@ void *page_get_target_data(target_ulong address)
/* The softmmu versions of these helpers are in cputlb.c. */
-/*
- * Verify that we have passed the correct MemOp to the correct function.
- *
- * We could present one function to target code, and dispatch based on
- * the MemOp, but so far we have worked hard to avoid an indirect function
- * call along the memory path.
- */
-static void validate_memop(MemOpIdx oi, MemOp expected)
-{
-#ifdef CONFIG_DEBUG_TCG
- MemOp have = get_memop(oi) & (MO_SIZE | MO_BSWAP);
- assert(have == expected);
-#endif
-}
-
void helper_unaligned_ld(CPUArchState *env, target_ulong addr)
{
cpu_loop_exit_sigbus(env_cpu(env), addr, MMU_DATA_LOAD, GETPC());
@@ -279,10 +264,9 @@ void helper_unaligned_st(CPUArchState *env, target_ulong addr)
cpu_loop_exit_sigbus(env_cpu(env), addr, MMU_DATA_STORE, GETPC());
}
-static void *cpu_mmu_lookup(CPUArchState *env, target_ulong addr,
- MemOpIdx oi, uintptr_t ra, MMUAccessType type)
+static void *cpu_mmu_lookup(CPUArchState *env, abi_ptr addr,
+ MemOp mop, uintptr_t ra, MMUAccessType type)
{
- MemOp mop = get_memop(oi);
int a_bits = get_alignment_bits(mop);
void *ret;
@@ -298,100 +282,206 @@ static void *cpu_mmu_lookup(CPUArchState *env, target_ulong addr,
#include "ldst_atomicity.c.inc"
-uint8_t cpu_ldb_mmu(CPUArchState *env, abi_ptr addr,
- MemOpIdx oi, uintptr_t ra)
+static uint8_t do_ld1_mmu(CPUArchState *env, abi_ptr addr,
+ MemOp mop, uintptr_t ra)
{
void *haddr;
uint8_t ret;
- validate_memop(oi, MO_UB);
- haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_LOAD);
+ tcg_debug_assert((mop & MO_SIZE) == MO_8);
+ haddr = cpu_mmu_lookup(env, addr, mop, ra, MMU_DATA_LOAD);
ret = ldub_p(haddr);
clear_helper_retaddr();
+ return ret;
+}
+
+tcg_target_ulong helper_ldub_mmu(CPUArchState *env, target_ulong addr,
+ MemOpIdx oi, uintptr_t ra)
+{
+ return do_ld1_mmu(env, addr, get_memop(oi), ra);
+}
+
+tcg_target_ulong helper_ldsb_mmu(CPUArchState *env, target_ulong addr,
+ MemOpIdx oi, uintptr_t ra)
+{
+ return (int8_t)do_ld1_mmu(env, addr, get_memop(oi), ra);
+}
+
+uint8_t cpu_ldb_mmu(CPUArchState *env, abi_ptr addr,
+ MemOpIdx oi, uintptr_t ra)
+{
+ uint8_t ret = do_ld1_mmu(env, addr, get_memop(oi), ra);
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
return ret;
}
+static uint16_t do_ld2_he_mmu(CPUArchState *env, abi_ptr addr,
+ MemOp mop, uintptr_t ra)
+{
+ void *haddr;
+ uint16_t ret;
+
+ tcg_debug_assert((mop & MO_SIZE) == MO_16);
+ haddr = cpu_mmu_lookup(env, addr, mop, ra, MMU_DATA_LOAD);
+ ret = load_atom_2(env, ra, haddr, mop);
+ clear_helper_retaddr();
+ return ret;
+}
+
+tcg_target_ulong helper_lduw_mmu(CPUArchState *env, target_ulong addr,
+ MemOpIdx oi, uintptr_t ra)
+{
+ MemOp mop = get_memop(oi);
+ uint16_t ret = do_ld2_he_mmu(env, addr, mop, ra);
+
+ if (mop & MO_BSWAP) {
+ ret = bswap16(ret);
+ }
+ return ret;
+}
+
+tcg_target_ulong helper_ldsw_mmu(CPUArchState *env, target_ulong addr,
+ MemOpIdx oi, uintptr_t ra)
+{
+ MemOp mop = get_memop(oi);
+ int16_t ret = do_ld2_he_mmu(env, addr, mop, ra);
+
+ if (mop & MO_BSWAP) {
+ ret = bswap16(ret);
+ }
+ return ret;
+}
+
uint16_t cpu_ldw_be_mmu(CPUArchState *env, abi_ptr addr,
MemOpIdx oi, uintptr_t ra)
{
- void *haddr;
+ MemOp mop = get_memop(oi);
uint16_t ret;
- validate_memop(oi, MO_BEUW);
- haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_LOAD);
- ret = load_atom_2(env, ra, haddr, get_memop(oi));
- clear_helper_retaddr();
+ tcg_debug_assert((mop & MO_BSWAP) == MO_BE);
+ ret = do_ld2_he_mmu(env, addr, mop, ra);
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
return cpu_to_be16(ret);
}
-uint32_t cpu_ldl_be_mmu(CPUArchState *env, abi_ptr addr,
- MemOpIdx oi, uintptr_t ra)
-{
- void *haddr;
- uint32_t ret;
-
- validate_memop(oi, MO_BEUL);
- haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_LOAD);
- ret = load_atom_4(env, ra, haddr, get_memop(oi));
- clear_helper_retaddr();
- qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
- return cpu_to_be32(ret);
-}
-
-uint64_t cpu_ldq_be_mmu(CPUArchState *env, abi_ptr addr,
- MemOpIdx oi, uintptr_t ra)
-{
- void *haddr;
- uint64_t ret;
-
- validate_memop(oi, MO_BEUQ);
- haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_LOAD);
- ret = load_atom_8(env, ra, haddr, get_memop(oi));
- clear_helper_retaddr();
- qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
- return cpu_to_be64(ret);
-}
-
uint16_t cpu_ldw_le_mmu(CPUArchState *env, abi_ptr addr,
MemOpIdx oi, uintptr_t ra)
{
- void *haddr;
+ MemOp mop = get_memop(oi);
uint16_t ret;
- validate_memop(oi, MO_LEUW);
- haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_LOAD);
- ret = load_atom_2(env, ra, haddr, get_memop(oi));
- clear_helper_retaddr();
+ tcg_debug_assert((mop & MO_BSWAP) == MO_LE);
+ ret = do_ld2_he_mmu(env, addr, mop, ra);
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
return cpu_to_le16(ret);
}
+static uint32_t do_ld4_he_mmu(CPUArchState *env, abi_ptr addr,
+ MemOp mop, uintptr_t ra)
+{
+ void *haddr;
+ uint32_t ret;
+
+ tcg_debug_assert((mop & MO_SIZE) == MO_32);
+ haddr = cpu_mmu_lookup(env, addr, mop, ra, MMU_DATA_LOAD);
+ ret = load_atom_4(env, ra, haddr, mop);
+ clear_helper_retaddr();
+ return ret;
+}
+
+tcg_target_ulong helper_ldul_mmu(CPUArchState *env, target_ulong addr,
+ MemOpIdx oi, uintptr_t ra)
+{
+ MemOp mop = get_memop(oi);
+ uint32_t ret = do_ld4_he_mmu(env, addr, mop, ra);
+
+ if (mop & MO_BSWAP) {
+ ret = bswap32(ret);
+ }
+ return ret;
+}
+
+tcg_target_ulong helper_ldsl_mmu(CPUArchState *env, target_ulong addr,
+ MemOpIdx oi, uintptr_t ra)
+{
+ MemOp mop = get_memop(oi);
+ int32_t ret = do_ld4_he_mmu(env, addr, mop, ra);
+
+ if (mop & MO_BSWAP) {
+ ret = bswap32(ret);
+ }
+ return ret;
+}
+
+uint32_t cpu_ldl_be_mmu(CPUArchState *env, abi_ptr addr,
+ MemOpIdx oi, uintptr_t ra)
+{
+ MemOp mop = get_memop(oi);
+ uint32_t ret;
+
+ tcg_debug_assert((mop & MO_BSWAP) == MO_BE);
+ ret = do_ld4_he_mmu(env, addr, mop, ra);
+ qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
+ return cpu_to_be32(ret);
+}
+
uint32_t cpu_ldl_le_mmu(CPUArchState *env, abi_ptr addr,
MemOpIdx oi, uintptr_t ra)
{
- void *haddr;
+ MemOp mop = get_memop(oi);
uint32_t ret;
- validate_memop(oi, MO_LEUL);
- haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_LOAD);
- ret = load_atom_4(env, ra, haddr, get_memop(oi));
- clear_helper_retaddr();
+ tcg_debug_assert((mop & MO_BSWAP) == MO_LE);
+ ret = do_ld4_he_mmu(env, addr, mop, ra);
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
return cpu_to_le32(ret);
}
+static uint64_t do_ld8_he_mmu(CPUArchState *env, abi_ptr addr,
+ MemOp mop, uintptr_t ra)
+{
+ void *haddr;
+ uint64_t ret;
+
+ tcg_debug_assert((mop & MO_SIZE) == MO_64);
+ haddr = cpu_mmu_lookup(env, addr, mop, ra, MMU_DATA_LOAD);
+ ret = load_atom_8(env, ra, haddr, mop);
+ clear_helper_retaddr();
+ return ret;
+}
+
+uint64_t helper_ldq_mmu(CPUArchState *env, target_ulong addr,
+ MemOpIdx oi, uintptr_t ra)
+{
+ MemOp mop = get_memop(oi);
+ uint64_t ret = do_ld8_he_mmu(env, addr, mop, ra);
+
+ if (mop & MO_BSWAP) {
+ ret = bswap64(ret);
+ }
+ return ret;
+}
+
+uint64_t cpu_ldq_be_mmu(CPUArchState *env, abi_ptr addr,
+ MemOpIdx oi, uintptr_t ra)
+{
+ MemOp mop = get_memop(oi);
+ uint64_t ret;
+
+ tcg_debug_assert((mop & MO_BSWAP) == MO_BE);
+ ret = do_ld8_he_mmu(env, addr, mop, ra);
+ qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
+ return cpu_to_be64(ret);
+}
+
uint64_t cpu_ldq_le_mmu(CPUArchState *env, abi_ptr addr,
MemOpIdx oi, uintptr_t ra)
{
- void *haddr;
+ MemOp mop = get_memop(oi);
uint64_t ret;
- validate_memop(oi, MO_LEUQ);
- haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_LOAD);
- ret = load_atom_8(env, ra, haddr, get_memop(oi));
- clear_helper_retaddr();
+ tcg_debug_assert((mop & MO_BSWAP) == MO_LE);
+ ret = do_ld8_he_mmu(env, addr, mop, ra);
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
return cpu_to_le64(ret);
}
@@ -402,7 +492,7 @@ Int128 cpu_ld16_be_mmu(CPUArchState *env, abi_ptr addr,
void *haddr;
Int128 ret;
- validate_memop(oi, MO_128 | MO_BE);
+ tcg_debug_assert((get_memop(oi) & (MO_BSWAP | MO_SIZE)) == (MO_128 | MO_BE));
haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_LOAD);
memcpy(&ret, haddr, 16);
clear_helper_retaddr();
@@ -420,7 +510,7 @@ Int128 cpu_ld16_le_mmu(CPUArchState *env, abi_ptr addr,
void *haddr;
Int128 ret;
- validate_memop(oi, MO_128 | MO_LE);
+ tcg_debug_assert((get_memop(oi) & (MO_BSWAP | MO_SIZE)) == (MO_128 | MO_LE));
haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_LOAD);
memcpy(&ret, haddr, 16);
clear_helper_retaddr();
@@ -432,87 +522,153 @@ Int128 cpu_ld16_le_mmu(CPUArchState *env, abi_ptr addr,
return ret;
}
-void cpu_stb_mmu(CPUArchState *env, abi_ptr addr, uint8_t val,
- MemOpIdx oi, uintptr_t ra)
+static void do_st1_mmu(CPUArchState *env, abi_ptr addr, uint8_t val,
+ MemOp mop, uintptr_t ra)
{
void *haddr;
- validate_memop(oi, MO_UB);
- haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_STORE);
+ tcg_debug_assert((mop & MO_SIZE) == MO_8);
+ haddr = cpu_mmu_lookup(env, addr, mop, ra, MMU_DATA_STORE);
stb_p(haddr, val);
clear_helper_retaddr();
+}
+
+void helper_stb_mmu(CPUArchState *env, target_ulong addr, uint8_t val,
+ MemOpIdx oi, uintptr_t ra)
+{
+ do_st1_mmu(env, addr, val, get_memop(oi), ra);
+}
+
+void cpu_stb_mmu(CPUArchState *env, abi_ptr addr, uint8_t val,
+ MemOpIdx oi, uintptr_t ra)
+{
+ do_st1_mmu(env, addr, val, get_memop(oi), ra);
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
}
+static void do_st2_he_mmu(CPUArchState *env, abi_ptr addr, uint16_t val,
+ MemOp mop, uintptr_t ra)
+{
+ void *haddr;
+
+ tcg_debug_assert((mop & MO_SIZE) == MO_16);
+ haddr = cpu_mmu_lookup(env, addr, mop, ra, MMU_DATA_STORE);
+ store_atom_2(env, ra, haddr, mop, val);
+ clear_helper_retaddr();
+}
+
+void helper_stw_mmu(CPUArchState *env, target_ulong addr, uint16_t val,
+ MemOpIdx oi, uintptr_t ra)
+{
+ MemOp mop = get_memop(oi);
+
+ if (mop & MO_BSWAP) {
+ val = bswap16(val);
+ }
+ do_st2_he_mmu(env, addr, val, mop, ra);
+}
+
void cpu_stw_be_mmu(CPUArchState *env, abi_ptr addr, uint16_t val,
MemOpIdx oi, uintptr_t ra)
{
- void *haddr;
+ MemOp mop = get_memop(oi);
- validate_memop(oi, MO_BEUW);
- haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_STORE);
- store_atom_2(env, ra, haddr, get_memop(oi), be16_to_cpu(val));
- clear_helper_retaddr();
- qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
-}
-
-void cpu_stl_be_mmu(CPUArchState *env, abi_ptr addr, uint32_t val,
- MemOpIdx oi, uintptr_t ra)
-{
- void *haddr;
-
- validate_memop(oi, MO_BEUL);
- haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_STORE);
- store_atom_4(env, ra, haddr, get_memop(oi), be32_to_cpu(val));
- clear_helper_retaddr();
- qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
-}
-
-void cpu_stq_be_mmu(CPUArchState *env, abi_ptr addr, uint64_t val,
- MemOpIdx oi, uintptr_t ra)
-{
- void *haddr;
-
- validate_memop(oi, MO_BEUQ);
- haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_STORE);
- store_atom_8(env, ra, haddr, get_memop(oi), be64_to_cpu(val));
- clear_helper_retaddr();
+ tcg_debug_assert((mop & MO_BSWAP) == MO_BE);
+ do_st2_he_mmu(env, addr, be16_to_cpu(val), mop, ra);
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
}
void cpu_stw_le_mmu(CPUArchState *env, abi_ptr addr, uint16_t val,
MemOpIdx oi, uintptr_t ra)
+{
+ MemOp mop = get_memop(oi);
+
+ tcg_debug_assert((mop & MO_BSWAP) == MO_LE);
+ do_st2_he_mmu(env, addr, le16_to_cpu(val), mop, ra);
+ qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
+}
+
+static void do_st4_he_mmu(CPUArchState *env, abi_ptr addr, uint32_t val,
+ MemOp mop, uintptr_t ra)
{
void *haddr;
- validate_memop(oi, MO_LEUW);
- haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_STORE);
- store_atom_2(env, ra, haddr, get_memop(oi), le16_to_cpu(val));
+ tcg_debug_assert((mop & MO_SIZE) == MO_32);
+ haddr = cpu_mmu_lookup(env, addr, mop, ra, MMU_DATA_STORE);
+ store_atom_4(env, ra, haddr, mop, val);
clear_helper_retaddr();
+}
+
+void helper_stl_mmu(CPUArchState *env, target_ulong addr, uint32_t val,
+ MemOpIdx oi, uintptr_t ra)
+{
+ MemOp mop = get_memop(oi);
+
+ if (mop & MO_BSWAP) {
+ val = bswap32(val);
+ }
+ do_st4_he_mmu(env, addr, val, mop, ra);
+}
+
+void cpu_stl_be_mmu(CPUArchState *env, abi_ptr addr, uint32_t val,
+ MemOpIdx oi, uintptr_t ra)
+{
+ MemOp mop = get_memop(oi);
+
+ tcg_debug_assert((mop & MO_BSWAP) == MO_BE);
+ do_st4_he_mmu(env, addr, be32_to_cpu(val), mop, ra);
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
}
void cpu_stl_le_mmu(CPUArchState *env, abi_ptr addr, uint32_t val,
MemOpIdx oi, uintptr_t ra)
+{
+ MemOp mop = get_memop(oi);
+
+ tcg_debug_assert((mop & MO_BSWAP) == MO_LE);
+ do_st4_he_mmu(env, addr, le32_to_cpu(val), mop, ra);
+ qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
+}
+
+static void do_st8_he_mmu(CPUArchState *env, abi_ptr addr, uint64_t val,
+ MemOp mop, uintptr_t ra)
{
void *haddr;
- validate_memop(oi, MO_LEUL);
- haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_STORE);
- store_atom_4(env, ra, haddr, get_memop(oi), le32_to_cpu(val));
+ tcg_debug_assert((mop & MO_SIZE) == MO_64);
+ haddr = cpu_mmu_lookup(env, addr, mop, ra, MMU_DATA_STORE);
+ store_atom_8(env, ra, haddr, mop, val);
clear_helper_retaddr();
+}
+
+void helper_stq_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
+ MemOpIdx oi, uintptr_t ra)
+{
+ MemOp mop = get_memop(oi);
+
+ if (mop & MO_BSWAP) {
+ val = bswap64(val);
+ }
+ do_st8_he_mmu(env, addr, val, mop, ra);
+}
+
+void cpu_stq_be_mmu(CPUArchState *env, abi_ptr addr, uint64_t val,
+ MemOpIdx oi, uintptr_t ra)
+{
+ MemOp mop = get_memop(oi);
+
+ tcg_debug_assert((mop & MO_BSWAP) == MO_BE);
+ do_st8_he_mmu(env, addr, cpu_to_be64(val), mop, ra);
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
}
void cpu_stq_le_mmu(CPUArchState *env, abi_ptr addr, uint64_t val,
MemOpIdx oi, uintptr_t ra)
{
- void *haddr;
+ MemOp mop = get_memop(oi);
- validate_memop(oi, MO_LEUQ);
- haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_STORE);
- store_atom_8(env, ra, haddr, get_memop(oi), le64_to_cpu(val));
- clear_helper_retaddr();
+ tcg_debug_assert((mop & MO_BSWAP) == MO_LE);
+ do_st8_he_mmu(env, addr, cpu_to_le64(val), mop, ra);
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
}
@@ -521,7 +677,7 @@ void cpu_st16_be_mmu(CPUArchState *env, abi_ptr addr,
{
void *haddr;
- validate_memop(oi, MO_128 | MO_BE);
+ tcg_debug_assert((get_memop(oi) & (MO_BSWAP | MO_SIZE)) == (MO_128 | MO_BE));
haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_STORE);
if (!HOST_BIG_ENDIAN) {
val = bswap128(val);
@@ -536,7 +692,7 @@ void cpu_st16_le_mmu(CPUArchState *env, abi_ptr addr,
{
void *haddr;
- validate_memop(oi, MO_128 | MO_LE);
+ tcg_debug_assert((get_memop(oi) & (MO_BSWAP | MO_SIZE)) == (MO_128 | MO_LE));
haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_STORE);
if (HOST_BIG_ENDIAN) {
val = bswap128(val);
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* [PATCH for-8.0 12/29] tcg: Add 128-bit guest memory primitives
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (10 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 11/29] accel/tcg: Implement helper_{ld, st}*_mmu for user-only Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-22 3:30 ` Richard Henderson
2022-11-18 9:47 ` [PATCH for-8.0 13/29] meson: Detect atomic128 support with optimization Richard Henderson
` (16 subsequent siblings)
28 siblings, 1 reply; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
accel/tcg/tcg-runtime.h | 3 +
include/tcg/tcg-ldst.h | 4 +
accel/tcg/cputlb.c | 480 +++++++++++++++++++++++++--------
accel/tcg/user-exec.c | 94 +++++--
tcg/tcg-op.c | 178 +++++++-----
accel/tcg/ldst_atomicity.c.inc | 175 +++++++++++-
6 files changed, 729 insertions(+), 205 deletions(-)
diff --git a/accel/tcg/tcg-runtime.h b/accel/tcg/tcg-runtime.h
index e141a6ab24..a7a2038901 100644
--- a/accel/tcg/tcg-runtime.h
+++ b/accel/tcg/tcg-runtime.h
@@ -39,6 +39,9 @@ DEF_HELPER_FLAGS_1(exit_atomic, TCG_CALL_NO_WG, noreturn, env)
DEF_HELPER_FLAGS_3(memset, TCG_CALL_NO_RWG, ptr, ptr, int, ptr)
#endif /* IN_HELPER_PROTO */
+DEF_HELPER_FLAGS_3(ld_i128, TCG_CALL_NO_WG, i128, env, tl, i32)
+DEF_HELPER_FLAGS_4(st_i128, TCG_CALL_NO_WG, void, env, tl, i128, i32)
+
DEF_HELPER_FLAGS_5(atomic_cmpxchgb, TCG_CALL_NO_WG,
i32, env, tl, i32, i32, i32)
DEF_HELPER_FLAGS_5(atomic_cmpxchgw_be, TCG_CALL_NO_WG,
diff --git a/include/tcg/tcg-ldst.h b/include/tcg/tcg-ldst.h
index c1d945fd66..3004e5292d 100644
--- a/include/tcg/tcg-ldst.h
+++ b/include/tcg/tcg-ldst.h
@@ -34,6 +34,8 @@ tcg_target_ulong helper_ldul_mmu(CPUArchState *env, target_ulong addr,
MemOpIdx oi, uintptr_t retaddr);
uint64_t helper_ldq_mmu(CPUArchState *env, target_ulong addr,
MemOpIdx oi, uintptr_t retaddr);
+Int128 helper_ld16_mmu(CPUArchState *env, target_ulong addr,
+ MemOpIdx oi, uintptr_t retaddr);
/* Value sign-extended to tcg register size. */
tcg_target_ulong helper_ldsb_mmu(CPUArchState *env, target_ulong addr,
@@ -51,6 +53,8 @@ void helper_stl_mmu(CPUArchState *env, target_ulong addr, uint32_t val,
MemOpIdx oi, uintptr_t retaddr);
void helper_stq_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
MemOpIdx oi, uintptr_t retaddr);
+void helper_st16_mmu(CPUArchState *env, target_ulong addr, Int128 val,
+ MemOpIdx oi, uintptr_t retaddr);
#ifdef CONFIG_USER_ONLY
diff --git a/accel/tcg/cputlb.c b/accel/tcg/cputlb.c
index 3d32adc0e7..314dbfa83d 100644
--- a/accel/tcg/cputlb.c
+++ b/accel/tcg/cputlb.c
@@ -40,6 +40,7 @@
#include "qemu/plugin-memory.h"
#endif
#include "tcg/tcg-ldst.h"
+#include "exec/helper-proto.h"
/* DEBUG defines, enable DEBUG_TLB_LOG to log to the CPU_LOG_MMU target */
/* #define DEBUG_TLB */
@@ -2130,6 +2131,31 @@ static uint64_t do_ld_whole_be8(CPUArchState *env, uintptr_t ra,
return (ret_be << (p->size * 8)) | x;
}
+/**
+ * do_ld_parts_be16
+ * @p: translation parameters
+ * @ret_be: accumulated data
+ *
+ * As do_ld_bytes_beN, but with one atomic load.
+ * 16 aligned bytes are guaranteed to cover the load.
+ */
+static Int128 do_ld_whole_be16(CPUArchState *env, uintptr_t ra,
+ MMULookupPageData *p, uint64_t ret_be)
+{
+ int o = p->addr & 15;
+ Int128 x, y = load_atomic16_or_exit(env, ra, p->haddr - o);
+ int size = p->size;
+
+ if (!HOST_BIG_ENDIAN) {
+ y = bswap128(y);
+ }
+ y = int128_lshift(y, o * 8);
+ y = int128_urshift(y, (16 - size) * 8);
+ x = int128_make64(ret_be);
+ x = int128_lshift(x, size * 8);
+ return int128_or(x, y);
+}
+
/*
* Wrapper for the above.
*/
@@ -2174,6 +2200,59 @@ static uint64_t do_ld_beN(CPUArchState *env, MMULookupPageData *p,
}
}
+/*
+ * Wrapper for the above, for 8 < size < 16.
+ */
+static Int128 do_ld16_beN(CPUArchState *env, MMULookupPageData *p,
+ uint64_t a, int mmu_idx, MemOp mop, uintptr_t ra)
+{
+ int size = p->size;
+ uint64_t b;
+ MemOp atmax;
+
+ if (unlikely(p->flags & TLB_MMIO)) {
+ p->size = size - 8;
+ a = do_ld_mmio_beN(env, p, a, mmu_idx, MMU_DATA_LOAD, ra);
+ p->addr += p->size;
+ p->size = 8;
+ b = do_ld_mmio_beN(env, p, 0, mmu_idx, MMU_DATA_LOAD, ra);
+ } else {
+ switch (mop & MO_ATOM_MASK) {
+ case MO_ATOM_WITHIN16:
+ /*
+ * It is a given that we cross a page and therefore there is no
+ * atomicity for the load as a whole, but there may be a subobject
+ * as defined by ATMAX which does not cross a 16-byte boundary.
+ */
+ atmax = mop & MO_ATMAX_MASK;
+ if (atmax != MO_ATMAX_SIZE) {
+ atmax >>= MO_ATMAX_SHIFT;
+ if (unlikely(size >= (1 << atmax))) {
+ return do_ld_whole_be16(env, ra, p, a);
+ }
+ }
+ /* fall through */
+ case MO_ATOM_IFALIGN:
+ case MO_ATOM_NONE:
+ p->size = size - 8;
+ a = do_ld_bytes_beN(p, a);
+ b = ldq_be_p(p->haddr + size - 8);
+ break;
+ case MO_ATOM_SUBALIGN:
+ p->size = size - 8;
+ a = do_ld_parts_beN(p, a);
+ p->haddr += size - 8;
+ p->size = 8;
+ b = do_ld_parts_beN(p, 0);
+ break;
+ default:
+ g_assert_not_reached();
+ }
+ }
+
+ return int128_make128(b, a);
+}
+
static uint8_t do_ld_1(CPUArchState *env, MMULookupPageData *p, int mmu_idx,
MMUAccessType type, uintptr_t ra)
{
@@ -2184,6 +2263,21 @@ static uint8_t do_ld_1(CPUArchState *env, MMULookupPageData *p, int mmu_idx,
}
}
+static uint64_t do_ld_8(CPUArchState *env, MMULookupPageData *p, int mmu_idx,
+ MMUAccessType type, MemOp memop, uintptr_t ra)
+{
+ if (unlikely(p->flags & TLB_MMIO)) {
+ return io_readx(env, p->full, mmu_idx, p->addr, ra, type, memop);
+ } else {
+ /* Perform the load host endian. */
+ uint64_t ret = load_atom_8(env, ra, p->haddr, memop);
+ if (memop & MO_BSWAP) {
+ ret = bswap64(ret);
+ }
+ return ret;
+ }
+}
+
static uint8_t do_ld1_mmu(CPUArchState *env, target_ulong addr, MemOpIdx oi,
uintptr_t ra, MMUAccessType access_type)
{
@@ -2303,16 +2397,7 @@ static uint64_t do_ld8_mmu(CPUArchState *env, target_ulong addr, MemOpIdx oi,
if (l.page[0].flags & TLB_BSWAP) {
l.memop ^= MO_BSWAP;
}
- if (unlikely(l.page[0].flags & TLB_MMIO)) {
- ret = io_readx(env, l.page[0].full, l.mmu_idx, addr, ra,
- access_type, l.memop);
- } else {
- /* Perform the load host endian. */
- ret = load_atom_8(env, ra, l.page[0].haddr, l.memop);
- if (l.memop & MO_BSWAP) {
- ret = bswap64(ret);
- }
- }
+ return do_ld_8(env, &l.page[0], l.mmu_idx, access_type, l.memop, ra);
} else {
assert_no_tlb_bswap;
ret = do_ld_beN(env, &l.page[0], 0, l.mmu_idx,
@@ -2356,6 +2441,83 @@ tcg_target_ulong helper_ldsl_mmu(CPUArchState *env, target_ulong addr,
return (int32_t)helper_ldul_mmu(env, addr, oi, retaddr);
}
+static Int128 do_ld16_mmu(CPUArchState *env, target_ulong addr,
+ MemOpIdx oi, uintptr_t ra)
+{
+ MMULookupLocals l;
+ bool crosspage;
+ uint64_t a, b;
+ Int128 ret;
+
+ crosspage = mmu_lookup(env, addr, oi, ra, MMU_DATA_LOAD, &l);
+ if (likely(!crosspage)) {
+ if (l.page[0].flags & TLB_BSWAP) {
+ l.memop ^= MO_BSWAP;
+ }
+ /* Perform the load host endian. */
+ if (unlikely(l.page[0].flags & TLB_MMIO)) {
+ QEMU_IOTHREAD_LOCK_GUARD();
+ a = io_readx(env, l.page[0].full, l.mmu_idx, addr,
+ ra, MMU_DATA_LOAD, MO_64);
+ b = io_readx(env, l.page[0].full, l.mmu_idx, addr + 8,
+ ra, MMU_DATA_LOAD, MO_64);
+ ret = int128_make128(HOST_BIG_ENDIAN ? b : a,
+ HOST_BIG_ENDIAN ? a : b);
+ } else {
+ ret = load_atom_16(env, ra, l.page[0].haddr, l.memop);
+ }
+ if (l.memop & MO_BSWAP) {
+ ret = bswap128(ret);
+ }
+ } else {
+ int first = l.page[0].size;
+
+ assert_no_tlb_bswap;
+
+ if (first == 8) {
+ MemOp mop8 = (l.memop & ~MO_SIZE) | MO_64;
+
+ a = do_ld_8(env, &l.page[0], l.mmu_idx, MMU_DATA_LOAD, mop8, ra);
+ b = do_ld_8(env, &l.page[1], l.mmu_idx, MMU_DATA_LOAD, mop8, ra);
+ if ((mop8 & MO_BSWAP) == MO_LE) {
+ ret = int128_make128(a, b);
+ } else {
+ ret = int128_make128(b, a);
+ }
+ } else {
+ if (first < 8) {
+ a = do_ld_beN(env, &l.page[0], 0, l.mmu_idx,
+ MMU_DATA_LOAD, l.memop, ra);
+ ret = do_ld16_beN(env, &l.page[1], a, l.mmu_idx, l.memop, ra);
+ } else {
+ ret = do_ld16_beN(env, &l.page[0], 0, l.mmu_idx, l.memop, ra);
+ b = int128_getlo(ret);
+ ret = int128_lshift(ret, l.page[1].size * 8);
+ a = int128_gethi(ret);
+ b = do_ld_beN(env, &l.page[0], b, l.mmu_idx,
+ MMU_DATA_LOAD, l.memop, ra);
+ ret = int128_make128(b, a);
+ }
+ if ((l.memop & MO_BSWAP) == MO_LE) {
+ ret = bswap128(ret);
+ }
+ }
+ }
+ return ret;
+}
+
+Int128 helper_ld16_mmu(CPUArchState *env, target_ulong addr,
+ uint32_t oi, uintptr_t retaddr)
+{
+ tcg_debug_assert((get_memop(oi) & MO_SIZE) == MO_128);
+ return do_ld16_mmu(env, addr, oi, retaddr);
+}
+
+Int128 helper_ld_i128(CPUArchState *env, target_ulong addr, uint32_t oi)
+{
+ return helper_ld16_mmu(env, addr, oi, GETPC());
+}
+
/*
* Load helpers for cpu_ldst.h.
*/
@@ -2444,59 +2606,23 @@ uint64_t cpu_ldq_le_mmu(CPUArchState *env, abi_ptr addr,
Int128 cpu_ld16_be_mmu(CPUArchState *env, abi_ptr addr,
MemOpIdx oi, uintptr_t ra)
{
- MemOp mop = get_memop(oi);
- int mmu_idx = get_mmuidx(oi);
- MemOpIdx new_oi;
- unsigned a_bits;
- uint64_t h, l;
+ Int128 ret;
- tcg_debug_assert((mop & (MO_BSWAP|MO_SSIZE)) == (MO_BE|MO_128));
- a_bits = get_alignment_bits(mop);
-
- /* Handle CPU specific unaligned behaviour */
- if (addr & ((1 << a_bits) - 1)) {
- cpu_unaligned_access(env_cpu(env), addr, MMU_DATA_LOAD,
- mmu_idx, ra);
- }
-
- /* Construct an unaligned 64-bit replacement MemOpIdx. */
- mop = (mop & ~(MO_SIZE | MO_AMASK)) | MO_64 | MO_UNALN;
- new_oi = make_memop_idx(mop, mmu_idx);
-
- h = helper_ldq_mmu(env, addr, new_oi, ra);
- l = helper_ldq_mmu(env, addr + 8, new_oi, ra);
-
- qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
- return int128_make128(l, h);
+ tcg_debug_assert((get_memop(oi) & (MO_BSWAP|MO_SIZE)) == (MO_BE|MO_128));
+ ret = do_ld16_mmu(env, addr, oi, ra);
+ plugin_load_cb(env, addr, oi);
+ return ret;
}
Int128 cpu_ld16_le_mmu(CPUArchState *env, abi_ptr addr,
MemOpIdx oi, uintptr_t ra)
{
- MemOp mop = get_memop(oi);
- int mmu_idx = get_mmuidx(oi);
- MemOpIdx new_oi;
- unsigned a_bits;
- uint64_t h, l;
+ Int128 ret;
- tcg_debug_assert((mop & (MO_BSWAP|MO_SSIZE)) == (MO_LE|MO_128));
- a_bits = get_alignment_bits(mop);
-
- /* Handle CPU specific unaligned behaviour */
- if (addr & ((1 << a_bits) - 1)) {
- cpu_unaligned_access(env_cpu(env), addr, MMU_DATA_LOAD,
- mmu_idx, ra);
- }
-
- /* Construct an unaligned 64-bit replacement MemOpIdx. */
- mop = (mop & ~(MO_SIZE | MO_AMASK)) | MO_64 | MO_UNALN;
- new_oi = make_memop_idx(mop, mmu_idx);
-
- l = helper_ldq_mmu(env, addr, new_oi, ra);
- h = helper_ldq_mmu(env, addr + 8, new_oi, ra);
-
- qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
- return int128_make128(l, h);
+ tcg_debug_assert((get_memop(oi) & (MO_BSWAP|MO_SIZE)) == (MO_LE|MO_128));
+ ret = do_ld16_mmu(env, addr, oi, ra);
+ plugin_load_cb(env, addr, oi);
+ return ret;
}
/*
@@ -2645,6 +2771,36 @@ static uint64_t do_st_whole_le8(MMULookupPageData *p, uint64_t val_le)
return val_le >> sz;
}
+/**
+ * do_st_whole_le16
+ * @p: translation parameters
+ * @val_le: data to store
+ *
+ * As do_st_bytes_leN, but atomically on each aligned part.
+ * 16 aligned bytes are guaranteed to cover the store.
+ */
+static uint64_t do_st_whole_le16(MMULookupPageData *p, Int128 val_le)
+{
+ int szm64 = (p->size * 8) - 64;
+ int o = p->addr & 15;
+ int sh = o * 8;
+ Int128 m, v;
+
+ /* Like MAKE_64BIT_MASK(0, sz), but larger. */
+ m = int128_make128(-1, MAKE_64BIT_MASK(0, szm64));
+
+ if (HOST_BIG_ENDIAN) {
+ v = int128_urshift(bswap128(val_le), sh);
+ m = int128_urshift(bswap128(m), sh);
+ } else {
+ v = int128_lshift(val_le, sh);
+ m = int128_lshift(m, sh);
+ }
+ store_atom_insert_al16(p->haddr - o, v, m);
+
+ return int128_gethi(val_le) >> szm64;
+}
+
/*
* Wrapper for the above.
*/
@@ -2691,6 +2847,60 @@ static uint64_t do_st_leN(CPUArchState *env, MMULookupPageData *p,
}
}
+/*
+ * Wrapper for the above, for 8 < size < 16.
+ */
+static uint64_t do_st16_leN(CPUArchState *env, MMULookupPageData *p,
+ Int128 val_le, int mmu_idx,
+ MemOp mop, uintptr_t ra)
+{
+ int size = p->size;
+ MemOp atmax;
+
+ if (unlikely(p->flags & TLB_MMIO)) {
+ p->size = 8;
+ do_st_mmio_leN(env, p, int128_getlo(val_le), mmu_idx, ra);
+ p->size = size - 8;
+ p->addr += 8;
+ return do_st_mmio_leN(env, p, int128_gethi(val_le), mmu_idx, ra);
+ }
+
+ switch (mop & MO_ATOM_MASK) {
+ case MO_ATOM_WITHIN16:
+ /*
+ * It is a given that we cross a page and therefore there is no
+ * atomicity for the store as a whole, but there may be a subobject
+ * as defined by ATMAX which does not cross a 16-byte boundary.
+ */
+ atmax = mop & MO_ATMAX_MASK;
+ if (atmax != MO_ATMAX_SIZE) {
+ atmax >>= MO_ATMAX_SHIFT;
+ if (unlikely(size >= (1 << atmax))) {
+ if (HAVE_al16) {
+ return do_st_whole_le16(p, val_le);
+ } else {
+ cpu_loop_exit_atomic(env_cpu(env), ra);
+ }
+ }
+ }
+ /* fall through */
+ case MO_ATOM_IFALIGN:
+ case MO_ATOM_NONE:
+ stq_le_p(p->haddr, int128_getlo(val_le));
+ p->size = size - 8;
+ p->haddr += 8;
+ return do_st_bytes_leN(p, int128_gethi(val_le));
+ case MO_ATOM_SUBALIGN:
+ p->size = 8;
+ do_st_parts_leN(p, int128_getlo(val_le));
+ p->size = size - 8;
+ p->haddr += 8;
+ return do_st_parts_leN(p, int128_gethi(val_le));
+ default:
+ g_assert_not_reached();
+ }
+}
+
static void do_st_1(CPUArchState *env, MMULookupPageData *p, uint8_t val,
int mmu_idx, uintptr_t ra)
{
@@ -2701,6 +2911,20 @@ static void do_st_1(CPUArchState *env, MMULookupPageData *p, uint8_t val,
}
}
+static void do_st_8(CPUArchState *env, MMULookupPageData *p, uint64_t val,
+ int mmu_idx, MemOp memop, uintptr_t ra)
+{
+ if (unlikely(p->flags & TLB_MMIO)) {
+ io_writex(env, p->full, mmu_idx, val, p->addr, ra, memop);
+ } else {
+ /* Swap to host endian if necessary, then store. */
+ if (memop & MO_BSWAP) {
+ val = bswap64(val);
+ }
+ store_atom_8(env, ra, p->haddr, memop, val);
+ }
+}
+
void helper_stb_mmu(CPUArchState *env, target_ulong addr, uint8_t val,
MemOpIdx oi, uintptr_t ra)
{
@@ -2806,15 +3030,7 @@ static void do_st8_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
if (l.page[0].flags & TLB_BSWAP) {
l.memop ^= MO_BSWAP;
}
- if (unlikely(l.page[0].flags & TLB_MMIO)) {
- io_writex(env, l.page[0].full, l.mmu_idx, val, addr, ra, l.memop);
- } else {
- /* Swap to host endian if necessary, then store. */
- if (l.memop & MO_BSWAP) {
- val = bswap64(val);
- }
- store_atom_8(env, ra, l.page[0].haddr, l.memop, val);
- }
+ do_st_8(env, &l.page[0], val, l.mmu_idx, l.memop, ra);
} else {
assert_no_tlb_bswap;
@@ -2834,6 +3050,82 @@ void helper_stq_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
do_st8_mmu(env, addr, val, oi, retaddr);
}
+static void do_st16_mmu(CPUArchState *env, target_ulong addr, Int128 val,
+ MemOpIdx oi, uintptr_t ra)
+{
+ MMULookupLocals l;
+ bool crosspage;
+ uint64_t a, b;
+
+ crosspage = mmu_lookup(env, addr, oi, ra, MMU_DATA_LOAD, &l);
+ if (likely(!crosspage)) {
+ if (l.page[0].flags & TLB_BSWAP) {
+ l.memop ^= MO_BSWAP;
+ }
+ /* Swap to host endian if necessary, then store. */
+ if (l.memop & MO_BSWAP) {
+ val = bswap128(val);
+ }
+ if (unlikely(l.page[0].flags & TLB_MMIO)) {
+ QEMU_IOTHREAD_LOCK_GUARD();
+ if (HOST_BIG_ENDIAN) {
+ b = int128_getlo(val), a = int128_gethi(val);
+ } else {
+ a = int128_getlo(val), b = int128_gethi(val);
+ }
+ io_writex(env, l.page[0].full, l.mmu_idx, a, addr, ra, MO_64);
+ io_writex(env, l.page[0].full, l.mmu_idx, b, addr + 8, ra, MO_64);
+ } else {
+ store_atom_16(env, ra, l.page[0].haddr, l.memop, val);
+ }
+ } else {
+ int first = l.page[0].size;
+
+ assert_no_tlb_bswap;
+
+ if (first == 8) {
+ MemOp mop8 = (l.memop & ~(MO_SIZE | MO_BSWAP)) | MO_64;
+
+ if (l.memop & MO_BSWAP) {
+ val = bswap128(val);
+ }
+ if (HOST_BIG_ENDIAN) {
+ b = int128_getlo(val), a = int128_gethi(val);
+ } else {
+ a = int128_getlo(val), b = int128_gethi(val);
+ }
+ do_st_8(env, &l.page[0], a, l.mmu_idx, mop8, ra);
+ do_st_8(env, &l.page[1], b, l.mmu_idx, mop8, ra);
+ } else {
+ if ((l.memop & MO_BSWAP) != MO_LE) {
+ val = bswap128(val);
+ }
+ if (first < 8) {
+ do_st_leN(env, &l.page[0], int128_getlo(val),
+ l.mmu_idx, l.memop, ra);
+ val = int128_urshift(val, first * 8);
+ do_st16_leN(env, &l.page[1], val, l.mmu_idx, l.memop, ra);
+ } else {
+ b = do_st16_leN(env, &l.page[0], val, l.mmu_idx, l.memop, ra);
+ do_st_leN(env, &l.page[1], b, l.mmu_idx, l.memop, ra);
+ }
+ }
+ }
+}
+
+void helper_st16_mmu(CPUArchState *env, target_ulong addr, Int128 val,
+ MemOpIdx oi, uintptr_t retaddr)
+{
+ tcg_debug_assert((get_memop(oi) & MO_SIZE) == MO_128);
+ do_st16_mmu(env, addr, val, oi, retaddr);
+}
+
+void helper_st_i128(CPUArchState *env, target_ulong addr, Int128 val,
+ MemOpIdx oi)
+{
+ helper_st16_mmu(env, addr, val, oi, GETPC());
+}
+
/*
* Store Helpers for cpu_ldst.h
*/
@@ -2898,58 +3190,20 @@ void cpu_stq_le_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
plugin_store_cb(env, addr, oi);
}
-void cpu_st16_be_mmu(CPUArchState *env, abi_ptr addr, Int128 val,
- MemOpIdx oi, uintptr_t ra)
+void cpu_st16_be_mmu(CPUArchState *env, target_ulong addr, Int128 val,
+ MemOpIdx oi, uintptr_t retaddr)
{
- MemOp mop = get_memop(oi);
- int mmu_idx = get_mmuidx(oi);
- MemOpIdx new_oi;
- unsigned a_bits;
-
- tcg_debug_assert((mop & (MO_BSWAP|MO_SSIZE)) == (MO_BE|MO_128));
- a_bits = get_alignment_bits(mop);
-
- /* Handle CPU specific unaligned behaviour */
- if (addr & ((1 << a_bits) - 1)) {
- cpu_unaligned_access(env_cpu(env), addr, MMU_DATA_STORE,
- mmu_idx, ra);
- }
-
- /* Construct an unaligned 64-bit replacement MemOpIdx. */
- mop = (mop & ~(MO_SIZE | MO_AMASK)) | MO_64 | MO_UNALN;
- new_oi = make_memop_idx(mop, mmu_idx);
-
- helper_stq_mmu(env, addr, int128_gethi(val), new_oi, ra);
- helper_stq_mmu(env, addr + 8, int128_getlo(val), new_oi, ra);
-
- qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
+ tcg_debug_assert((get_memop(oi) & (MO_BSWAP|MO_SIZE)) == (MO_BE|MO_128));
+ do_st16_mmu(env, addr, val, oi, retaddr);
+ plugin_store_cb(env, addr, oi);
}
-void cpu_st16_le_mmu(CPUArchState *env, abi_ptr addr, Int128 val,
- MemOpIdx oi, uintptr_t ra)
+void cpu_st16_le_mmu(CPUArchState *env, target_ulong addr, Int128 val,
+ MemOpIdx oi, uintptr_t retaddr)
{
- MemOp mop = get_memop(oi);
- int mmu_idx = get_mmuidx(oi);
- MemOpIdx new_oi;
- unsigned a_bits;
-
- tcg_debug_assert((mop & (MO_BSWAP|MO_SSIZE)) == (MO_LE|MO_128));
- a_bits = get_alignment_bits(mop);
-
- /* Handle CPU specific unaligned behaviour */
- if (addr & ((1 << a_bits) - 1)) {
- cpu_unaligned_access(env_cpu(env), addr, MMU_DATA_STORE,
- mmu_idx, ra);
- }
-
- /* Construct an unaligned 64-bit replacement MemOpIdx. */
- mop = (mop & ~(MO_SIZE | MO_AMASK)) | MO_64 | MO_UNALN;
- new_oi = make_memop_idx(mop, mmu_idx);
-
- helper_stq_mmu(env, addr, int128_getlo(val), new_oi, ra);
- helper_stq_mmu(env, addr + 8, int128_gethi(val), new_oi, ra);
-
- qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
+ tcg_debug_assert((get_memop(oi) & (MO_BSWAP|MO_SIZE)) == (MO_LE|MO_128));
+ do_st16_mmu(env, addr, val, oi, retaddr);
+ plugin_store_cb(env, addr, oi);
}
#include "ldst_common.c.inc"
diff --git a/accel/tcg/user-exec.c b/accel/tcg/user-exec.c
index 3455ff45a4..7ae88ccff1 100644
--- a/accel/tcg/user-exec.c
+++ b/accel/tcg/user-exec.c
@@ -486,18 +486,45 @@ uint64_t cpu_ldq_le_mmu(CPUArchState *env, abi_ptr addr,
return cpu_to_le64(ret);
}
-Int128 cpu_ld16_be_mmu(CPUArchState *env, abi_ptr addr,
- MemOpIdx oi, uintptr_t ra)
+static Int128 do_ld16_he_mmu(CPUArchState *env, abi_ptr addr,
+ MemOp mop, uintptr_t ra)
{
void *haddr;
Int128 ret;
- tcg_debug_assert((get_memop(oi) & (MO_BSWAP | MO_SIZE)) == (MO_128 | MO_BE));
- haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_LOAD);
- memcpy(&ret, haddr, 16);
+ tcg_debug_assert((mop & MO_SIZE) == MO_128);
+ haddr = cpu_mmu_lookup(env, addr, mop, ra, MMU_DATA_LOAD);
+ ret = load_atom_16(env, ra, haddr, mop);
clear_helper_retaddr();
- qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
+ return ret;
+}
+Int128 helper_ld16_mmu(CPUArchState *env, target_ulong addr,
+ MemOpIdx oi, uintptr_t ra)
+{
+ MemOp mop = get_memop(oi);
+ Int128 ret = do_ld16_he_mmu(env, addr, mop, ra);
+
+ if (mop & MO_BSWAP) {
+ ret = bswap128(ret);
+ }
+ return ret;
+}
+
+Int128 helper_ld_i128(CPUArchState *env, target_ulong addr, MemOpIdx oi)
+{
+ return helper_ld16_mmu(env, addr, oi, GETPC());
+}
+
+Int128 cpu_ld16_be_mmu(CPUArchState *env, abi_ptr addr,
+ MemOpIdx oi, uintptr_t ra)
+{
+ MemOp mop = get_memop(oi);
+ Int128 ret;
+
+ tcg_debug_assert((mop & MO_BSWAP) == MO_BE);
+ ret = do_ld16_he_mmu(env, addr, mop, ra);
+ qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
if (!HOST_BIG_ENDIAN) {
ret = bswap128(ret);
}
@@ -507,15 +534,12 @@ Int128 cpu_ld16_be_mmu(CPUArchState *env, abi_ptr addr,
Int128 cpu_ld16_le_mmu(CPUArchState *env, abi_ptr addr,
MemOpIdx oi, uintptr_t ra)
{
- void *haddr;
+ MemOp mop = get_memop(oi);
Int128 ret;
- tcg_debug_assert((get_memop(oi) & (MO_BSWAP | MO_SIZE)) == (MO_128 | MO_LE));
- haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_LOAD);
- memcpy(&ret, haddr, 16);
- clear_helper_retaddr();
+ tcg_debug_assert((mop & MO_BSWAP) == MO_LE);
+ ret = do_ld16_he_mmu(env, addr, mop, ra);
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
-
if (HOST_BIG_ENDIAN) {
ret = bswap128(ret);
}
@@ -672,33 +696,57 @@ void cpu_stq_le_mmu(CPUArchState *env, abi_ptr addr, uint64_t val,
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
}
-void cpu_st16_be_mmu(CPUArchState *env, abi_ptr addr,
- Int128 val, MemOpIdx oi, uintptr_t ra)
+static void do_st16_he_mmu(CPUArchState *env, abi_ptr addr, Int128 val,
+ MemOp mop, uintptr_t ra)
{
void *haddr;
- tcg_debug_assert((get_memop(oi) & (MO_BSWAP | MO_SIZE)) == (MO_128 | MO_BE));
- haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_STORE);
+ tcg_debug_assert((mop & MO_SIZE) == MO_128);
+ haddr = cpu_mmu_lookup(env, addr, mop, ra, MMU_DATA_STORE);
+ store_atom_16(env, ra, haddr, mop, val);
+ clear_helper_retaddr();
+}
+
+void helper_st16_mmu(CPUArchState *env, target_ulong addr, Int128 val,
+ MemOpIdx oi, uintptr_t ra)
+{
+ MemOp mop = get_memop(oi);
+
+ if (mop & MO_BSWAP) {
+ val = bswap128(val);
+ }
+ do_st16_he_mmu(env, addr, val, mop, ra);
+}
+
+void helper_st_i128(CPUArchState *env, target_ulong addr,
+ Int128 val, MemOpIdx oi)
+{
+ helper_st16_mmu(env, addr, val, oi, GETPC());
+}
+
+void cpu_st16_be_mmu(CPUArchState *env, abi_ptr addr,
+ Int128 val, MemOpIdx oi, uintptr_t ra)
+{
+ MemOp mop = get_memop(oi);
+
+ tcg_debug_assert((mop & MO_BSWAP) == MO_BE);
if (!HOST_BIG_ENDIAN) {
val = bswap128(val);
}
- memcpy(haddr, &val, 16);
- clear_helper_retaddr();
+ do_st16_he_mmu(env, addr, val, mop, ra);
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
}
void cpu_st16_le_mmu(CPUArchState *env, abi_ptr addr,
Int128 val, MemOpIdx oi, uintptr_t ra)
{
- void *haddr;
+ MemOp mop = get_memop(oi);
- tcg_debug_assert((get_memop(oi) & (MO_BSWAP | MO_SIZE)) == (MO_128 | MO_LE));
- haddr = cpu_mmu_lookup(env, addr, oi, ra, MMU_DATA_STORE);
+ tcg_debug_assert((mop & MO_BSWAP) == MO_LE);
if (HOST_BIG_ENDIAN) {
val = bswap128(val);
}
- memcpy(haddr, &val, 16);
- clear_helper_retaddr();
+ do_st16_he_mmu(env, addr, val, mop, ra);
qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
}
diff --git a/tcg/tcg-op.c b/tcg/tcg-op.c
index 1f81c3dbb3..bbb29bed2b 100644
--- a/tcg/tcg-op.c
+++ b/tcg/tcg-op.c
@@ -3112,6 +3112,48 @@ void tcg_gen_qemu_st_i64(TCGv_i64 val, TCGv addr, TCGArg idx, MemOp memop)
}
}
+/*
+ * Return true if @mop, without knowledge of the pointer alignment,
+ * does not require 16-byte atomicity, and it would be adventagous
+ * to avoid a call to a helper function.
+ */
+static bool use_two_i64_for_i128(MemOp mop)
+{
+#ifdef CONFIG_SOFTMMU
+ /* Two softmmu tlb lookups is larger than one function call. */
+ return false;
+#else
+ /*
+ * For user-only, two 64-bit operations may well be smaller than a call.
+ * Determine if that would be legal for the requested atomicity.
+ */
+ MemOp atom = mop & MO_ATOM_MASK;
+ MemOp atmax = mop & MO_ATMAX_MASK;
+
+ /* In a serialized context, no atomicity is required. */
+ if (tcg_ctx->tb_cflags & CF_PARALLEL) {
+ return true;
+ }
+
+ if (atmax == MO_ATMAX_SIZE) {
+ atmax = mop & MO_SIZE;
+ } else {
+ atmax >>= MO_ATMAX_SHIFT;
+ }
+ switch (atom) {
+ case MO_ATOM_NONE:
+ return true;
+ case MO_ATOM_IFALIGN:
+ case MO_ATOM_SUBALIGN:
+ return atmax < MO_128;
+ case MO_ATOM_WITHIN16:
+ return atmax == MO_8;
+ default:
+ g_assert_not_reached();
+ }
+#endif
+}
+
static void canonicalize_memop_i128_as_i64(MemOp ret[2], MemOp orig)
{
MemOp mop_1 = orig, mop_2;
@@ -3159,91 +3201,105 @@ static void canonicalize_memop_i128_as_i64(MemOp ret[2], MemOp orig)
void tcg_gen_qemu_ld_i128(TCGv_i128 val, TCGv addr, TCGArg idx, MemOp memop)
{
- MemOp mop[2];
- TCGv addr_p8;
- TCGv_i64 x, y;
+ MemOpIdx oi = make_memop_idx(memop, idx);
- canonicalize_memop_i128_as_i64(mop, memop);
+ tcg_debug_assert((memop & MO_SIZE) == MO_128);
+ tcg_debug_assert((memop & MO_SIGN) == 0);
tcg_gen_req_mo(TCG_MO_LD_LD | TCG_MO_ST_LD);
addr = plugin_prep_mem_callbacks(addr);
- /* TODO: respect atomicity of the operation. */
/* TODO: allow the tcg backend to see the whole operation. */
- /*
- * Since there are no global TCGv_i128, there is no visible state
- * changed if the second load faults. Load directly into the two
- * subwords.
- */
- if ((memop & MO_BSWAP) == MO_LE) {
- x = TCGV128_LOW(val);
- y = TCGV128_HIGH(val);
+ if (use_two_i64_for_i128(memop)) {
+ MemOp mop[2];
+ TCGv addr_p8;
+ TCGv_i64 x, y;
+
+ canonicalize_memop_i128_as_i64(mop, memop);
+
+ /*
+ * Since there are no global TCGv_i128, there is no visible state
+ * changed if the second load faults. Load directly into the two
+ * subwords.
+ */
+ if ((memop & MO_BSWAP) == MO_LE) {
+ x = TCGV128_LOW(val);
+ y = TCGV128_HIGH(val);
+ } else {
+ x = TCGV128_HIGH(val);
+ y = TCGV128_LOW(val);
+ }
+
+ gen_ldst_i64(INDEX_op_qemu_ld_i64, x, addr, mop[0], idx);
+
+ if ((mop[0] ^ memop) & MO_BSWAP) {
+ tcg_gen_bswap64_i64(x, x);
+ }
+
+ addr_p8 = tcg_temp_new();
+ tcg_gen_addi_tl(addr_p8, addr, 8);
+ gen_ldst_i64(INDEX_op_qemu_ld_i64, y, addr_p8, mop[1], idx);
+ tcg_temp_free(addr_p8);
+
+ if ((mop[0] ^ memop) & MO_BSWAP) {
+ tcg_gen_bswap64_i64(y, y);
+ }
} else {
- x = TCGV128_HIGH(val);
- y = TCGV128_LOW(val);
+ gen_helper_ld_i128(val, cpu_env, addr, tcg_constant_i32(oi));
}
- gen_ldst_i64(INDEX_op_qemu_ld_i64, x, addr, mop[0], idx);
-
- if ((mop[0] ^ memop) & MO_BSWAP) {
- tcg_gen_bswap64_i64(x, x);
- }
-
- addr_p8 = tcg_temp_new();
- tcg_gen_addi_tl(addr_p8, addr, 8);
- gen_ldst_i64(INDEX_op_qemu_ld_i64, y, addr_p8, mop[1], idx);
- tcg_temp_free(addr_p8);
-
- if ((mop[0] ^ memop) & MO_BSWAP) {
- tcg_gen_bswap64_i64(y, y);
- }
-
- plugin_gen_mem_callbacks(addr, make_memop_idx(memop, idx),
- QEMU_PLUGIN_MEM_R);
+ plugin_gen_mem_callbacks(addr, oi, QEMU_PLUGIN_MEM_R);
}
void tcg_gen_qemu_st_i128(TCGv_i128 val, TCGv addr, TCGArg idx, MemOp memop)
{
- MemOp mop[2];
- TCGv addr_p8;
- TCGv_i64 x, y;
+ MemOpIdx oi = make_memop_idx(memop, idx);
- canonicalize_memop_i128_as_i64(mop, memop);
+ tcg_debug_assert((memop & MO_SIZE) == MO_128);
+ tcg_debug_assert((memop & MO_SIGN) == 0);
tcg_gen_req_mo(TCG_MO_ST_LD | TCG_MO_ST_ST);
addr = plugin_prep_mem_callbacks(addr);
- /* TODO: respect atomicity of the operation. */
/* TODO: allow the tcg backend to see the whole operation. */
- if ((memop & MO_BSWAP) == MO_LE) {
- x = TCGV128_LOW(val);
- y = TCGV128_HIGH(val);
+ if (use_two_i64_for_i128(memop)) {
+ MemOp mop[2];
+ TCGv addr_p8;
+ TCGv_i64 x, y;
+
+ canonicalize_memop_i128_as_i64(mop, memop);
+
+ if ((memop & MO_BSWAP) == MO_LE) {
+ x = TCGV128_LOW(val);
+ y = TCGV128_HIGH(val);
+ } else {
+ x = TCGV128_HIGH(val);
+ y = TCGV128_LOW(val);
+ }
+
+ addr_p8 = tcg_temp_new();
+ if ((mop[0] ^ memop) & MO_BSWAP) {
+ TCGv_i64 t = tcg_temp_new_i64();
+
+ tcg_gen_bswap64_i64(t, x);
+ gen_ldst_i64(INDEX_op_qemu_st_i64, t, addr, mop[0], idx);
+ tcg_gen_bswap64_i64(t, y);
+ tcg_gen_addi_tl(addr_p8, addr, 8);
+ gen_ldst_i64(INDEX_op_qemu_st_i64, t, addr_p8, mop[1], idx);
+ tcg_temp_free_i64(t);
+ } else {
+ gen_ldst_i64(INDEX_op_qemu_st_i64, x, addr, mop[0], idx);
+ tcg_gen_addi_tl(addr_p8, addr, 8);
+ gen_ldst_i64(INDEX_op_qemu_st_i64, y, addr_p8, mop[1], idx);
+ }
+ tcg_temp_free(addr_p8);
} else {
- x = TCGV128_HIGH(val);
- y = TCGV128_LOW(val);
+ gen_helper_st_i128(cpu_env, addr, val, tcg_constant_i32(oi));
}
- addr_p8 = tcg_temp_new();
- if ((mop[0] ^ memop) & MO_BSWAP) {
- TCGv_i64 t = tcg_temp_new_i64();
-
- tcg_gen_bswap64_i64(t, x);
- gen_ldst_i64(INDEX_op_qemu_st_i64, t, addr, mop[0], idx);
- tcg_gen_bswap64_i64(t, y);
- tcg_gen_addi_tl(addr_p8, addr, 8);
- gen_ldst_i64(INDEX_op_qemu_st_i64, t, addr_p8, mop[1], idx);
- tcg_temp_free_i64(t);
- } else {
- gen_ldst_i64(INDEX_op_qemu_st_i64, x, addr, mop[0], idx);
- tcg_gen_addi_tl(addr_p8, addr, 8);
- gen_ldst_i64(INDEX_op_qemu_st_i64, y, addr_p8, mop[1], idx);
- }
- tcg_temp_free(addr_p8);
-
- plugin_gen_mem_callbacks(addr, make_memop_idx(memop, idx),
- QEMU_PLUGIN_MEM_W);
+ plugin_gen_mem_callbacks(addr, oi, QEMU_PLUGIN_MEM_W);
}
static void tcg_gen_ext_i32(TCGv_i32 ret, TCGv_i32 val, MemOp opc)
diff --git a/accel/tcg/ldst_atomicity.c.inc b/accel/tcg/ldst_atomicity.c.inc
index 8876c16371..e6a7558399 100644
--- a/accel/tcg/ldst_atomicity.c.inc
+++ b/accel/tcg/ldst_atomicity.c.inc
@@ -419,6 +419,21 @@ static inline uint64_t load_atom_8_by_4(void *pv)
}
}
+/**
+ * load_atom_8_by_8_or_4:
+ * @pv: host address
+ *
+ * Load 8 bytes from aligned @pv, with at least 4-byte atomicity.
+ */
+static inline uint64_t load_atom_8_by_8_or_4(void *pv)
+{
+ if (HAVE_al8_fast) {
+ return load_atomic8(pv);
+ } else {
+ return load_atom_8_by_4(pv);
+ }
+}
+
/**
* load_atom_2:
* @p: host address
@@ -551,6 +566,64 @@ static uint64_t load_atom_8(CPUArchState *env, uintptr_t ra,
}
}
+/**
+ * load_atom_16:
+ * @p: host address
+ * @memop: the full memory op
+ *
+ * Load 16 bytes from @p, honoring the atomicity of @memop.
+ */
+static Int128 load_atom_16(CPUArchState *env, uintptr_t ra,
+ void *pv, MemOp memop)
+{
+ uintptr_t pi = (uintptr_t)pv;
+ int atmax;
+ Int128 r;
+ uint64_t a, b;
+
+ /*
+ * If the host does not support 8-byte atomics, wait until we have
+ * examined the atomicity parameters below.
+ */
+ if (HAVE_al16_fast && likely((pi & 15) == 0)) {
+ return load_atomic16(pv);
+ }
+
+ atmax = required_atomicity(env, pi, memop);
+ switch (atmax) {
+ case MO_8:
+ memcpy(&r, pv, 16);
+ return r;
+ case MO_16:
+ a = load_atom_8_by_2(pv);
+ b = load_atom_8_by_2(pv + 8);
+ break;
+ case MO_32:
+ a = load_atom_8_by_4(pv);
+ b = load_atom_8_by_4(pv + 8);
+ break;
+ case MO_64:
+ if (!HAVE_al8) {
+ cpu_loop_exit_atomic(env_cpu(env), ra);
+ }
+ a = load_atomic8(pv);
+ b = load_atomic8(pv + 8);
+ break;
+ case -MO_64:
+ if (!HAVE_al8) {
+ cpu_loop_exit_atomic(env_cpu(env), ra);
+ }
+ a = load_atom_extract_al8x2(pv);
+ b = load_atom_extract_al8x2(pv + 8);
+ break;
+ case MO_128:
+ return load_atomic16_or_exit(env, ra, pv);
+ default:
+ g_assert_not_reached();
+ }
+ return int128_make128(HOST_BIG_ENDIAN ? b : a, HOST_BIG_ENDIAN ? a : b);
+}
+
/**
* store_atomic2:
* @pv: host address
@@ -592,6 +665,40 @@ static inline void store_atomic8(void *pv, uint64_t val)
qatomic_set__nocheck(p, val);
}
+/**
+ * store_atomic16:
+ * @pv: host address
+ * @val: value to store
+ *
+ * Atomically store 16 aligned bytes to @pv.
+ */
+static inline void store_atomic16(void *pv, Int128 val)
+{
+#if defined(CONFIG_ATOMIC128)
+ __uint128_t *pu = __builtin_assume_aligned(pv, 16);
+ Int128Alias new;
+
+ new.s = val;
+ qatomic_set__nocheck(pu, new.u);
+#elif defined(CONFIG_CMPXCHG128)
+ __uint128_t *pu = __builtin_assume_aligned(pv, 16);
+ __uint128_t o;
+ Int128Alias n;
+
+ /*
+ * Without CONFIG_ATOMIC128, __atomic_compare_exchange_n will always
+ * defer to libatomic, so we must use __sync_val_compare_and_swap_16
+ * and accept the sequential consistency that comes with it.
+ */
+ n.s = val;
+ do {
+ o = *pu;
+ } while (!__sync_bool_compare_and_swap_16(pu, o, n.u));
+#else
+ qemu_build_not_reached();
+#endif
+}
+
/**
* store_atom_4x2
*/
@@ -607,9 +714,8 @@ static inline void store_atom_4_by_2(void *pv, uint32_t val)
*/
static inline void store_atom_8_by_2(void *pv, uint64_t val)
{
- uint32_t *p = __builtin_assume_aligned(pv, 4);
- qatomic_set(p, val >> (HOST_BIG_ENDIAN ? 32 : 0));
- qatomic_set(p + 4, val >> (HOST_BIG_ENDIAN ? 0 : 32));
+ store_atom_4_by_2(pv, val >> (HOST_BIG_ENDIAN ? 32 : 0));
+ store_atom_4_by_2(pv + 4, val >> (HOST_BIG_ENDIAN ? 0 : 32));
}
/**
@@ -617,11 +723,9 @@ static inline void store_atom_8_by_2(void *pv, uint64_t val)
*/
static inline void store_atom_8_by_4(void *pv, uint64_t val)
{
- uint16_t *p = __builtin_assume_aligned(pv, 2);
- qatomic_set(p, val >> (HOST_BIG_ENDIAN ? 48 : 0));
- qatomic_set(p + 2, val >> (HOST_BIG_ENDIAN ? 32 : 16));
- qatomic_set(p + 4, val >> (HOST_BIG_ENDIAN ? 16 : 32));
- qatomic_set(p + 6, val >> (HOST_BIG_ENDIAN ? 0 : 48));
+ uint32_t *p = __builtin_assume_aligned(pv, 4);
+ qatomic_set(p, val >> (HOST_BIG_ENDIAN ? 32 : 0));
+ qatomic_set(p + 4, val >> (HOST_BIG_ENDIAN ? 0 : 32));
}
/**
@@ -880,3 +984,58 @@ static void store_atom_8(CPUArchState *env, uintptr_t ra,
g_assert_not_reached();
}
}
+
+/**
+ * store_atom_16:
+ * @p: host address
+ * @val: the value to store
+ * @memop: the full memory op
+ *
+ * Store 16 bytes to @p, honoring the atomicity of @memop.
+ */
+static void store_atom_16(CPUArchState *env, uintptr_t ra,
+ void *pv, MemOp memop, Int128 val)
+{
+ uintptr_t pi = (uintptr_t)pv;
+ uint64_t a, b;
+ MemOp atmax;
+
+ if (HAVE_al16_fast && likely((pi & 15) == 0)) {
+ store_atomic16(pv, val);
+ return;
+ }
+
+ atmax = required_atomicity(env, pi, memop);
+
+ a = HOST_BIG_ENDIAN ? int128_gethi(val) : int128_getlo(val);
+ b = HOST_BIG_ENDIAN ? int128_getlo(val) : int128_gethi(val);
+ switch (atmax) {
+ case MO_8:
+ memcpy(pv, &val, 16);
+ return;
+ case MO_16:
+ store_atom_8_by_2(pv, a);
+ store_atom_8_by_2(pv + 8, b);
+ return;
+ case MO_32:
+ store_atom_8_by_4(pv, a);
+ store_atom_8_by_4(pv + 8, b);
+ return;
+ case MO_64:
+ if (HAVE_al8) {
+ store_atomic8(pv, a);
+ store_atomic8(pv + 8, b);
+ return;
+ }
+ break;
+ case MO_128:
+ if (HAVE_al16) {
+ store_atomic16(pv, val);
+ return;
+ }
+ break;
+ default:
+ g_assert_not_reached();
+ }
+ cpu_loop_exit_atomic(env_cpu(env), ra);
+}
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* Re: [PATCH for-8.0 12/29] tcg: Add 128-bit guest memory primitives
2022-11-18 9:47 ` [PATCH for-8.0 12/29] tcg: Add 128-bit guest memory primitives Richard Henderson
@ 2022-11-22 3:30 ` Richard Henderson
0 siblings, 0 replies; 48+ messages in thread
From: Richard Henderson @ 2022-11-22 3:30 UTC (permalink / raw)
To: qemu-devel
On 11/18/22 01:47, Richard Henderson wrote:
> + ret = do_ld16_beN(env, &l.page[0], 0, l.mmu_idx, l.memop, ra);
> + b = int128_getlo(ret);
> + ret = int128_lshift(ret, l.page[1].size * 8);
> + a = int128_gethi(ret);
> + b = do_ld_beN(env, &l.page[0], b, l.mmu_idx,
> + MMU_DATA_LOAD, l.memop, ra);
This last should have been l.page[1] -- fixed locally.
r~
^ permalink raw reply [flat|nested] 48+ messages in thread
* [PATCH for-8.0 13/29] meson: Detect atomic128 support with optimization
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (11 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 12/29] tcg: Add 128-bit guest memory primitives Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-18 9:47 ` [PATCH for-8.0 14/29] tcg/i386: Add have_atomic16 Richard Henderson
` (15 subsequent siblings)
28 siblings, 0 replies; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
There is an edge condition prior to gcc13 for which optimization
is required to generate 16-byte atomic sequences. Detect this.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
accel/tcg/ldst_atomicity.c.inc | 38 ++++++++++++++++++-------
meson.build | 52 ++++++++++++++++++++++------------
2 files changed, 61 insertions(+), 29 deletions(-)
diff --git a/accel/tcg/ldst_atomicity.c.inc b/accel/tcg/ldst_atomicity.c.inc
index e6a7558399..68edab4398 100644
--- a/accel/tcg/ldst_atomicity.c.inc
+++ b/accel/tcg/ldst_atomicity.c.inc
@@ -16,6 +16,23 @@
#endif
#define HAVE_al8_fast (ATOMIC_REG_SIZE >= 8)
+/*
+ * If __alignof(unsigned __int128) < 16, GCC may refuse to inline atomics
+ * that are supported by the host, e.g. s390x. We can force the pointer to
+ * have our known alignment with __builtin_assume_aligned, however prior to
+ * GCC 13 that was only reliable with optimization enabled. See
+ * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107389
+ */
+#if defined(CONFIG_ATOMIC128_OPT)
+# if !defined(__OPTIMIZE__)
+# define ATTRIBUTE_ATOMIC128_OPT __attribute__((optimize("O1")))
+# endif
+# define CONFIG_ATOMIC128
+#endif
+#ifndef ATTRIBUTE_ATOMIC128_OPT
+# define ATTRIBUTE_ATOMIC128_OPT
+#endif
+
#if defined(CONFIG_ATOMIC128)
# define HAVE_al16_fast true
#else
@@ -134,7 +151,8 @@ static inline uint64_t load_atomic8(void *pv)
*
* Atomically load 16 aligned bytes from @pv.
*/
-static inline Int128 load_atomic16(void *pv)
+static inline Int128 ATTRIBUTE_ATOMIC128_OPT
+load_atomic16(void *pv)
{
#ifdef CONFIG_ATOMIC128
__uint128_t *p = __builtin_assume_aligned(pv, 16);
@@ -336,7 +354,8 @@ static uint64_t load_atom_extract_al16_or_exit(CPUArchState *env, uintptr_t ra,
* cross an 16-byte boundary then the access must be 16-byte atomic,
* otherwise the access must be 8-byte atomic.
*/
-static inline uint64_t load_atom_extract_al16_or_al8(void *pv, int s)
+static inline uint64_t ATTRIBUTE_ATOMIC128_OPT
+load_atom_extract_al16_or_al8(void *pv, int s)
{
#if defined(CONFIG_ATOMIC128)
uintptr_t pi = (uintptr_t)pv;
@@ -672,28 +691,24 @@ static inline void store_atomic8(void *pv, uint64_t val)
*
* Atomically store 16 aligned bytes to @pv.
*/
-static inline void store_atomic16(void *pv, Int128 val)
+static inline void ATTRIBUTE_ATOMIC128_OPT
+store_atomic16(void *pv, Int128Alias val)
{
#if defined(CONFIG_ATOMIC128)
__uint128_t *pu = __builtin_assume_aligned(pv, 16);
- Int128Alias new;
-
- new.s = val;
- qatomic_set__nocheck(pu, new.u);
+ qatomic_set__nocheck(pu, val.u);
#elif defined(CONFIG_CMPXCHG128)
__uint128_t *pu = __builtin_assume_aligned(pv, 16);
__uint128_t o;
- Int128Alias n;
/*
* Without CONFIG_ATOMIC128, __atomic_compare_exchange_n will always
* defer to libatomic, so we must use __sync_val_compare_and_swap_16
* and accept the sequential consistency that comes with it.
*/
- n.s = val;
do {
o = *pu;
- } while (!__sync_bool_compare_and_swap_16(pu, o, n.u));
+ } while (!__sync_bool_compare_and_swap_16(pu, o, val.u));
#else
qemu_build_not_reached();
#endif
@@ -777,7 +792,8 @@ static void store_atom_insert_al8(uint64_t *p, uint64_t val, uint64_t msk)
*
* Atomically store @val to @p masked by @msk.
*/
-static void store_atom_insert_al16(Int128 *ps, Int128Alias val, Int128Alias msk)
+static void ATTRIBUTE_ATOMIC128_OPT
+store_atom_insert_al16(Int128 *ps, Int128Alias val, Int128Alias msk)
{
#if defined(CONFIG_ATOMIC128)
__uint128_t *pu, old, new;
diff --git a/meson.build b/meson.build
index 4984e80e71..503eeabd79 100644
--- a/meson.build
+++ b/meson.build
@@ -2215,23 +2215,21 @@ config_host_data.set('HAVE_BROKEN_SIZE_MAX', not cc.compiles('''
return printf("%zu", SIZE_MAX);
}''', args: ['-Werror']))
-atomic_test = '''
+# See if 64-bit atomic operations are supported.
+# Note that without __atomic builtins, we can only
+# assume atomic loads/stores max at pointer size.
+config_host_data.set('CONFIG_ATOMIC64', cc.links('''
#include <stdint.h>
int main(void)
{
- @0@ x = 0, y = 0;
+ uint64_t x = 0, y = 0;
y = __atomic_load_n(&x, __ATOMIC_RELAXED);
__atomic_store_n(&x, y, __ATOMIC_RELAXED);
__atomic_compare_exchange_n(&x, &y, x, 0, __ATOMIC_RELAXED, __ATOMIC_RELAXED);
__atomic_exchange_n(&x, y, __ATOMIC_RELAXED);
__atomic_fetch_add(&x, y, __ATOMIC_RELAXED);
return 0;
- }'''
-
-# See if 64-bit atomic operations are supported.
-# Note that without __atomic builtins, we can only
-# assume atomic loads/stores max at pointer size.
-config_host_data.set('CONFIG_ATOMIC64', cc.links(atomic_test.format('uint64_t')))
+ }'''))
has_int128 = cc.links('''
__int128_t a;
@@ -2249,21 +2247,39 @@ if has_int128
# "do we have 128-bit atomics which are handled inline and specifically not
# via libatomic". The reason we can't use libatomic is documented in the
# comment starting "GCC is a house divided" in include/qemu/atomic128.h.
- has_atomic128 = cc.links(atomic_test.format('unsigned __int128'))
+ # We only care about these operations on 16-byte aligned pointers, so
+ # force 16-byte alignment of the pointer, which may be greater than
+ # __alignof(unsigned __int128) for the host.
+ atomic_test_128 = '''
+ int main(int ac, char **av) {
+ unsigned __int128 *p = __builtin_assume_aligned(av[ac - 1], sizeof(16));
+ p[1] = __atomic_load_n(&p[0], __ATOMIC_RELAXED);
+ __atomic_store_n(&p[2], p[3], __ATOMIC_RELAXED);
+ __atomic_compare_exchange_n(&p[4], &p[5], p[6], 0, __ATOMIC_RELAXED, __ATOMIC_RELAXED);
+ return 0;
+ }'''
+ has_atomic128 = cc.links(atomic_test_128)
config_host_data.set('CONFIG_ATOMIC128', has_atomic128)
if not has_atomic128
- has_cmpxchg128 = cc.links('''
- int main(void)
- {
- unsigned __int128 x = 0, y = 0;
- __sync_val_compare_and_swap_16(&x, y, x);
- return 0;
- }
- ''')
+ # Even with __builtin_assume_aligned, the above test may have failed
+ # without optimization enabled. Try again with optimizations locally
+ # enabled for the function. See
+ # https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107389
+ has_atomic128_opt = cc.links('__attribute__((optimize("O1")))' + atomic_test_128)
+ config_host_data.set('CONFIG_ATOMIC128_OPT', has_atomic128_opt)
- config_host_data.set('CONFIG_CMPXCHG128', has_cmpxchg128)
+ if not has_atomic128_opt
+ config_host_data.set('CONFIG_CMPXCHG128', cc.links('''
+ int main(void)
+ {
+ unsigned __int128 x = 0, y = 0;
+ __sync_val_compare_and_swap_16(&x, y, x);
+ return 0;
+ }
+ '''))
+ endif
endif
endif
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* [PATCH for-8.0 14/29] tcg/i386: Add have_atomic16
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (12 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 13/29] meson: Detect atomic128 support with optimization Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-18 9:47 ` [PATCH for-8.0 15/29] include/qemu/int128: Add vector type to Int128Alias Richard Henderson
` (14 subsequent siblings)
28 siblings, 0 replies; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
Notice when Intel has guaranteed that vmovdqa is atomic.
The new variable will also be used in generated code.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
include/qemu/cpuid.h | 18 ++++++++++++++++++
tcg/i386/tcg-target.h | 1 +
tcg/i386/tcg-target.c.inc | 27 +++++++++++++++++++++++++++
3 files changed, 46 insertions(+)
diff --git a/include/qemu/cpuid.h b/include/qemu/cpuid.h
index 1451e8ef2f..35325f1995 100644
--- a/include/qemu/cpuid.h
+++ b/include/qemu/cpuid.h
@@ -71,6 +71,24 @@
#define bit_LZCNT (1 << 5)
#endif
+/*
+ * Signatures for different CPU implementations as returned from Leaf 0.
+ */
+
+#ifndef signature_INTEL_ecx
+/* "Genu" "ineI" "ntel" */
+#define signature_INTEL_ebx 0x756e6547
+#define signature_INTEL_edx 0x49656e69
+#define signature_INTEL_ecx 0x6c65746e
+#endif
+
+#ifndef signature_AMD_ecx
+/* "Auth" "enti" "cAMD" */
+#define signature_AMD_ebx 0x68747541
+#define signature_AMD_edx 0x69746e65
+#define signature_AMD_ecx 0x444d4163
+#endif
+
static inline unsigned xgetbv_low(unsigned c)
{
unsigned a, d;
diff --git a/tcg/i386/tcg-target.h b/tcg/i386/tcg-target.h
index 9e0e82d80a..5b037b1d2b 100644
--- a/tcg/i386/tcg-target.h
+++ b/tcg/i386/tcg-target.h
@@ -120,6 +120,7 @@ extern bool have_avx512dq;
extern bool have_avx512vbmi2;
extern bool have_avx512vl;
extern bool have_movbe;
+extern bool have_atomic16;
/* optional instructions */
#define TCG_TARGET_HAS_div2_i32 1
diff --git a/tcg/i386/tcg-target.c.inc b/tcg/i386/tcg-target.c.inc
index 24e9efe631..f4c0c7b8a2 100644
--- a/tcg/i386/tcg-target.c.inc
+++ b/tcg/i386/tcg-target.c.inc
@@ -185,6 +185,7 @@ bool have_avx512dq;
bool have_avx512vbmi2;
bool have_avx512vl;
bool have_movbe;
+bool have_atomic16;
#ifdef CONFIG_CPUID_H
static bool have_bmi2;
@@ -4165,6 +4166,32 @@ static void tcg_target_init(TCGContext *s)
have_avx512dq = (b7 & bit_AVX512DQ) != 0;
have_avx512vbmi2 = (c7 & bit_AVX512VBMI2) != 0;
}
+
+ /*
+ * The Intel SDM has added:
+ * Processors that enumerate support for Intel® AVX
+ * (by setting the feature flag CPUID.01H:ECX.AVX[bit 28])
+ * guarantee that the 16-byte memory operations performed
+ * by the following instructions will always be carried
+ * out atomically:
+ * - MOVAPD, MOVAPS, and MOVDQA.
+ * - VMOVAPD, VMOVAPS, and VMOVDQA when encoded with VEX.128.
+ * - VMOVAPD, VMOVAPS, VMOVDQA32, and VMOVDQA64 when encoded
+ * with EVEX.128 and k0 (masking disabled).
+ * Note that these instructions require the linear addresses
+ * of their memory operands to be 16-byte aligned.
+ *
+ * AMD has provided an even stronger guarantee that processors
+ * with AVX provide 16-byte atomicity for all cachable,
+ * naturally aligned single loads and stores, e.g. MOVDQU.
+ *
+ * See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104688
+ */
+ if (have_avx1) {
+ __cpuid(0, a, b, c, d);
+ have_atomic16 = (c == signature_INTEL_ecx ||
+ c == signature_AMD_ecx);
+ }
}
}
}
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* [PATCH for-8.0 15/29] include/qemu/int128: Add vector type to Int128Alias
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (13 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 14/29] tcg/i386: Add have_atomic16 Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-21 23:45 ` Philippe Mathieu-Daudé
2022-11-22 18:21 ` Philippe Mathieu-Daudé
2022-11-18 9:47 ` [PATCH for-8.0 16/29] accel/tcg: Use have_atomic16 in ldst_atomicity.c.inc Richard Henderson
` (13 subsequent siblings)
28 siblings, 2 replies; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
Adding a vector type will make it easier to handle i386
have_atomic16 via AVX.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
include/qemu/int128.h | 10 +++++-----
1 file changed, 5 insertions(+), 5 deletions(-)
diff --git a/include/qemu/int128.h b/include/qemu/int128.h
index f62a46b48c..f29f90e6f4 100644
--- a/include/qemu/int128.h
+++ b/include/qemu/int128.h
@@ -479,16 +479,16 @@ static inline void bswap128s(Int128 *s)
/*
* When compiler supports a 128-bit type, define a combination of
* a possible structure and the native types. Ease parameter passing
- * via use of the transparent union extension.
+ * via use of the transparent union extension. Provide a vector type
+ * for use in atomicity on some hosts.
*/
-#ifdef CONFIG_INT128
typedef union {
Int128 s;
+ uint64_t v __attribute__((vector_size(16)));
+#ifdef CONFIG_INT128
__int128_t i;
__uint128_t u;
-} Int128Alias __attribute__((transparent_union));
-#else
-typedef Int128 Int128Alias;
#endif /* CONFIG_INT128 */
+} Int128Alias __attribute__((transparent_union));
#endif /* INT128_H */
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* Re: [PATCH for-8.0 15/29] include/qemu/int128: Add vector type to Int128Alias
2022-11-18 9:47 ` [PATCH for-8.0 15/29] include/qemu/int128: Add vector type to Int128Alias Richard Henderson
@ 2022-11-21 23:45 ` Philippe Mathieu-Daudé
2022-11-22 18:21 ` Philippe Mathieu-Daudé
1 sibling, 0 replies; 48+ messages in thread
From: Philippe Mathieu-Daudé @ 2022-11-21 23:45 UTC (permalink / raw)
To: Richard Henderson, qemu-devel
On 18/11/22 10:47, Richard Henderson wrote:
> Adding a vector type will make it easier to handle i386
> have_atomic16 via AVX.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
> include/qemu/int128.h | 10 +++++-----
> 1 file changed, 5 insertions(+), 5 deletions(-)
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
^ permalink raw reply [flat|nested] 48+ messages in thread
* Re: [PATCH for-8.0 15/29] include/qemu/int128: Add vector type to Int128Alias
2022-11-18 9:47 ` [PATCH for-8.0 15/29] include/qemu/int128: Add vector type to Int128Alias Richard Henderson
2022-11-21 23:45 ` Philippe Mathieu-Daudé
@ 2022-11-22 18:21 ` Philippe Mathieu-Daudé
2022-11-22 18:31 ` Philippe Mathieu-Daudé
1 sibling, 1 reply; 48+ messages in thread
From: Philippe Mathieu-Daudé @ 2022-11-22 18:21 UTC (permalink / raw)
To: Richard Henderson, qemu-devel
On 18/11/22 10:47, Richard Henderson wrote:
> Adding a vector type will make it easier to handle i386
> have_atomic16 via AVX.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
> include/qemu/int128.h | 10 +++++-----
> 1 file changed, 5 insertions(+), 5 deletions(-)
>
> diff --git a/include/qemu/int128.h b/include/qemu/int128.h
> index f62a46b48c..f29f90e6f4 100644
> --- a/include/qemu/int128.h
> +++ b/include/qemu/int128.h
> @@ -479,16 +479,16 @@ static inline void bswap128s(Int128 *s)
> /*
> * When compiler supports a 128-bit type, define a combination of
> * a possible structure and the native types. Ease parameter passing
> - * via use of the transparent union extension.
> + * via use of the transparent union extension. Provide a vector type
> + * for use in atomicity on some hosts.
> */
> -#ifdef CONFIG_INT128
> typedef union {
> Int128 s;
> + uint64_t v __attribute__((vector_size(16)));
> +#ifdef CONFIG_INT128
> __int128_t i;
> __uint128_t u;
> -} Int128Alias __attribute__((transparent_union));
> -#else
> -typedef Int128 Int128Alias;
> #endif /* CONFIG_INT128 */
> +} Int128Alias __attribute__((transparent_union));
>
> #endif /* INT128_H */
This triggers a warning with GCC:
include/qemu/int128.h:487:14: warning: alignment of field 'v' (128 bits)
does not match the alignment of the first field in transparent union;
transparent_union attribute ignored [-Wignored-attributes]
uint64_t v __attribute__((vector_size(16)));
^
include/qemu/int128.h:486:12: note: alignment of first field is 64 bits
Int128 s;
^
Meson:
Project version: 7.1.91
C compiler for the host machine: gcc-12 (gcc 12.2.0 "gcc-12 (Homebrew
GCC 12.2.0) 12.2.0")
C linker for the host machine: gcc-12 ld64 819.6
Host machine cpu family: aarch64
Host machine cpu: arm64
^ permalink raw reply [flat|nested] 48+ messages in thread* Re: [PATCH for-8.0 15/29] include/qemu/int128: Add vector type to Int128Alias
2022-11-22 18:21 ` Philippe Mathieu-Daudé
@ 2022-11-22 18:31 ` Philippe Mathieu-Daudé
0 siblings, 0 replies; 48+ messages in thread
From: Philippe Mathieu-Daudé @ 2022-11-22 18:31 UTC (permalink / raw)
To: Richard Henderson, qemu-devel
On 22/11/22 19:21, Philippe Mathieu-Daudé wrote:
> On 18/11/22 10:47, Richard Henderson wrote:
>> Adding a vector type will make it easier to handle i386
>> have_atomic16 via AVX.
>>
>> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
>> ---
>> include/qemu/int128.h | 10 +++++-----
>> 1 file changed, 5 insertions(+), 5 deletions(-)
>>
>> diff --git a/include/qemu/int128.h b/include/qemu/int128.h
>> index f62a46b48c..f29f90e6f4 100644
>> --- a/include/qemu/int128.h
>> +++ b/include/qemu/int128.h
>> @@ -479,16 +479,16 @@ static inline void bswap128s(Int128 *s)
>> /*
>> * When compiler supports a 128-bit type, define a combination of
>> * a possible structure and the native types. Ease parameter passing
>> - * via use of the transparent union extension.
>> + * via use of the transparent union extension. Provide a vector type
>> + * for use in atomicity on some hosts.
>> */
>> -#ifdef CONFIG_INT128
>> typedef union {
>> Int128 s;
>> + uint64_t v __attribute__((vector_size(16)));
>> +#ifdef CONFIG_INT128
>> __int128_t i;
>> __uint128_t u;
>> -} Int128Alias __attribute__((transparent_union));
>> -#else
>> -typedef Int128 Int128Alias;
>> #endif /* CONFIG_INT128 */
>> +} Int128Alias __attribute__((transparent_union));
>> #endif /* INT128_H */
>
> This triggers a warning with GCC:
Ah no, looking closer, even configured as ''--cc=gcc-12 --host-cc=gcc-12
--cxx=/bin/false', Clang got selected for ObjC, and this warning comes
from it:
Compiling Objective-C object libcommon.fa.p/ui_cocoa.m.o
In file included from ../../ui/cocoa.m:36:
In file included from include/sysemu/sysemu.h:5:
In file included from include/qemu/timer.h:4:
In file included from include/qemu/bitops.h:16:
In file included from include/qemu/host-utils.h:35:
> include/qemu/int128.h:487:14: warning: alignment of field 'v' (128 bits)
> does not match the alignment of the first field in transparent union;
> transparent_union attribute ignored [-Wignored-attributes]
> uint64_t v __attribute__((vector_size(16)));
> ^
> include/qemu/int128.h:486:12: note: alignment of first field is 64 bits
> Int128 s;
> ^
> Meson:
>
> Project version: 7.1.91
> C compiler for the host machine: gcc-12 (gcc 12.2.0 "gcc-12 (Homebrew
> GCC 12.2.0) 12.2.0")
> C linker for the host machine: gcc-12 ld64 819.6
> Host machine cpu family: aarch64
> Host machine cpu: arm64
Objective-C compiler for the host machine: clang (clang 14.0.0)
Objective-C linker for the host machine: clang ld64 819.6
Regards,
Phil.
^ permalink raw reply [flat|nested] 48+ messages in thread
* [PATCH for-8.0 16/29] accel/tcg: Use have_atomic16 in ldst_atomicity.c.inc
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (14 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 15/29] include/qemu/int128: Add vector type to Int128Alias Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-18 9:47 ` [PATCH for-8.0 17/29] tcg/aarch64: Add have_lse, have_lse2 Richard Henderson
` (12 subsequent siblings)
28 siblings, 0 replies; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
Hosts using Intel and AMD AVX cpus are quite common.
Add fast paths through ldst_atomicity using this.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
accel/tcg/ldst_atomicity.c.inc | 76 +++++++++++++++++++++++++++-------
1 file changed, 60 insertions(+), 16 deletions(-)
diff --git a/accel/tcg/ldst_atomicity.c.inc b/accel/tcg/ldst_atomicity.c.inc
index 68edab4398..d2a3783193 100644
--- a/accel/tcg/ldst_atomicity.c.inc
+++ b/accel/tcg/ldst_atomicity.c.inc
@@ -35,6 +35,14 @@
#if defined(CONFIG_ATOMIC128)
# define HAVE_al16_fast true
+#elif defined(CONFIG_TCG_INTERPRETER)
+/*
+ * FIXME: host specific detection this is in tcg/$host/,
+ * but we're using tcg/tci/ instead.
+ */
+# define HAVE_al16_fast false
+#elif defined(__x86_64__)
+# define HAVE_al16_fast likely(have_atomic16)
#else
# define HAVE_al16_fast false
#endif
@@ -160,6 +168,12 @@ load_atomic16(void *pv)
r.u = qatomic_read__nocheck(p);
return r.s;
+#elif defined(__x86_64__)
+ Int128Alias r;
+
+ /* Via HAVE_al16_fast, have_atomic16 is true. */
+ asm("vmovdqa %1, %0" : "=x" (r.u) : "m" (*(Int128 *)pv));
+ return r.s;
#else
qemu_build_not_reached();
#endif
@@ -379,6 +393,24 @@ load_atom_extract_al16_or_al8(void *pv, int s)
r = qatomic_read__nocheck(p16);
}
return r >> shr;
+#elif defined(__x86_64__)
+ uintptr_t pi = (uintptr_t)pv;
+ int shr = (pi & 7) * 8;
+ uint64_t a, b;
+
+ /* Via HAVE_al16_fast, have_atomic16 is true. */
+ pv = (void *)(pi & ~7);
+ if (pi & 8) {
+ uint64_t *p8 = __builtin_assume_aligned(pv, 16, 8);
+ a = qatomic_read__nocheck(p8);
+ b = qatomic_read__nocheck(p8 + 1);
+ } else {
+ asm("vmovdqa %2, %0\n\tvpextrq $1, %0, %1"
+ : "=x"(a), "=r"(b) : "m" (*(__uint128_t *)pv));
+ }
+ asm("shrd %b2, %1, %0" : "+r"(a) : "r"(b), "c"(shr));
+
+ return a;
#else
qemu_build_not_reached();
#endif
@@ -695,23 +727,35 @@ static inline void ATTRIBUTE_ATOMIC128_OPT
store_atomic16(void *pv, Int128Alias val)
{
#if defined(CONFIG_ATOMIC128)
- __uint128_t *pu = __builtin_assume_aligned(pv, 16);
- qatomic_set__nocheck(pu, val.u);
-#elif defined(CONFIG_CMPXCHG128)
- __uint128_t *pu = __builtin_assume_aligned(pv, 16);
- __uint128_t o;
-
- /*
- * Without CONFIG_ATOMIC128, __atomic_compare_exchange_n will always
- * defer to libatomic, so we must use __sync_val_compare_and_swap_16
- * and accept the sequential consistency that comes with it.
- */
- do {
- o = *pu;
- } while (!__sync_bool_compare_and_swap_16(pu, o, val.u));
-#else
- qemu_build_not_reached();
+ {
+ __uint128_t *pu = __builtin_assume_aligned(pv, 16);
+ qatomic_set__nocheck(pu, val.u);
+ return;
+ }
#endif
+#if defined(__x86_64__)
+ if (HAVE_al16_fast) {
+ asm("vmovdqa %1, %0" : "=m"(*(__uint128_t *)pv) : "x" (val.u));
+ return;
+ }
+#endif
+#if defined(CONFIG_CMPXCHG128)
+ {
+ __uint128_t *pu = __builtin_assume_aligned(pv, 16);
+ __uint128_t o;
+
+ /*
+ * Without CONFIG_ATOMIC128, __atomic_compare_exchange_n will always
+ * defer to libatomic, so we must use __sync_val_compare_and_swap_16
+ * and accept the sequential consistency that comes with it.
+ */
+ do {
+ o = *pu;
+ } while (!__sync_bool_compare_and_swap_16(pu, o, val.u));
+ return;
+ }
+#endif
+ qemu_build_not_reached();
}
/**
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* [PATCH for-8.0 17/29] tcg/aarch64: Add have_lse, have_lse2
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (15 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 16/29] accel/tcg: Use have_atomic16 in ldst_atomicity.c.inc Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-21 23:10 ` Philippe Mathieu-Daudé
2022-11-18 9:47 ` [PATCH for-8.0 18/29] accel/tcg: Add aarch64 specific support in ldst_atomicity Richard Henderson
` (11 subsequent siblings)
28 siblings, 1 reply; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
Notice when the host has additional atomic instructions.
The new variables will also be used in generated code.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
tcg/aarch64/tcg-target.h | 3 +++
tcg/aarch64/tcg-target.c.inc | 10 ++++++++++
2 files changed, 13 insertions(+)
diff --git a/tcg/aarch64/tcg-target.h b/tcg/aarch64/tcg-target.h
index 0dff5807f6..b8f734f371 100644
--- a/tcg/aarch64/tcg-target.h
+++ b/tcg/aarch64/tcg-target.h
@@ -57,6 +57,9 @@ typedef enum {
#define TCG_TARGET_CALL_ARG_I128 TCG_CALL_ARG_NORMAL
#define TCG_TARGET_CALL_RET_I128 TCG_CALL_RET_NORMAL
+extern bool have_lse;
+extern bool have_lse2;
+
/* optional instructions */
#define TCG_TARGET_HAS_div_i32 1
#define TCG_TARGET_HAS_rem_i32 1
diff --git a/tcg/aarch64/tcg-target.c.inc b/tcg/aarch64/tcg-target.c.inc
index 001a71bbc0..cf5ee6f742 100644
--- a/tcg/aarch64/tcg-target.c.inc
+++ b/tcg/aarch64/tcg-target.c.inc
@@ -13,6 +13,8 @@
#include "../tcg-ldst.c.inc"
#include "../tcg-pool.c.inc"
#include "qemu/bitops.h"
+#include <asm/hwcap.h>
+
/* We're going to re-use TCGType in setting of the SF bit, which controls
the size of the operation performed. If we know the values match, it
@@ -71,6 +73,9 @@ static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
return TCG_REG_X0 + slot;
}
+bool have_lse;
+bool have_lse2;
+
#define TCG_REG_TMP TCG_REG_X30
#define TCG_VEC_TMP TCG_REG_V31
@@ -2918,6 +2923,11 @@ static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
static void tcg_target_init(TCGContext *s)
{
+ unsigned long hwcap = qemu_getauxval(AT_HWCAP);
+
+ have_lse = hwcap & HWCAP_ATOMICS;
+ have_lse2 = hwcap & HWCAP_USCAT;
+
tcg_target_available_regs[TCG_TYPE_I32] = 0xffffffffu;
tcg_target_available_regs[TCG_TYPE_I64] = 0xffffffffu;
tcg_target_available_regs[TCG_TYPE_V64] = 0xffffffff00000000ull;
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* Re: [PATCH for-8.0 17/29] tcg/aarch64: Add have_lse, have_lse2
2022-11-18 9:47 ` [PATCH for-8.0 17/29] tcg/aarch64: Add have_lse, have_lse2 Richard Henderson
@ 2022-11-21 23:10 ` Philippe Mathieu-Daudé
2022-11-21 23:14 ` Philippe Mathieu-Daudé
0 siblings, 1 reply; 48+ messages in thread
From: Philippe Mathieu-Daudé @ 2022-11-21 23:10 UTC (permalink / raw)
To: Richard Henderson, qemu-devel
On 18/11/22 10:47, Richard Henderson wrote:
> Notice when the host has additional atomic instructions.
> The new variables will also be used in generated code.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
> tcg/aarch64/tcg-target.h | 3 +++
> tcg/aarch64/tcg-target.c.inc | 10 ++++++++++
> 2 files changed, 13 insertions(+)
> diff --git a/tcg/aarch64/tcg-target.c.inc b/tcg/aarch64/tcg-target.c.inc
> index 001a71bbc0..cf5ee6f742 100644
> --- a/tcg/aarch64/tcg-target.c.inc
> +++ b/tcg/aarch64/tcg-target.c.inc
> @@ -13,6 +13,8 @@
> #include "../tcg-ldst.c.inc"
> #include "../tcg-pool.c.inc"
> #include "qemu/bitops.h"
> +#include <asm/hwcap.h>
This doesn't build on Darwin:
In file included from ../../tcg/tcg.c:426:
tcg/aarch64/tcg-target.c.inc:16:10: fatal error: 'asm/hwcap.h' file not
found
#include <asm/hwcap.h>
^~~~~~~~~~~~~
In file included from ../../accel/tcg/cputlb.c:1656:
../../accel/tcg/ldst_atomicity.c.inc:269:21: warning: value size does
not match register size specified by the constraint and modifier
[-Wasm-operand-widths]
: "=&r"(r.u), "=&r"(fail) : "Q"(*p));
^
../../accel/tcg/ldst_atomicity.c.inc:266:22: note: use constraint
modifier "w"
asm("0: ldxp %0, %R0, %2\n\t"
^~
%w0
../../accel/tcg/ldst_atomicity.c.inc:902:9: error: invalid operand in
inline asm: '0: ldxp $2, ${2:R}, $0 bic $2, $2, $4 bic ${2:R}, ${2:R},
${4:R} orr $2, $2, $3 orr ${2:R}, ${2:R}, ${3:R} stxp ${1:w}, $2,
${2:R}, $0 cbnz ${1:w}, 0b'
asm("0: ldxp %[t], %R[t], %[mem]\n\t"
^
../../accel/tcg/ldst_atomicity.c.inc:902:9: error: invalid operand in
inline asm: '0: ldxp $2, ${2:R}, $0 bic $2, $2, $4 bic ${2:R}, ${2:R},
${4:R} orr $2, $2, $3 orr ${2:R}, ${2:R}, ${3:R} stxp ${1:w}, $2,
${2:R}, $0 cbnz ${1:w}, 0b'
../../accel/tcg/ldst_atomicity.c.inc:902:9: error: invalid operand in
inline asm: '0: ldxp $2, ${2:R}, $0 bic $2, $2, $4 bic ${2:R}, ${2:R},
${4:R} orr $2, $2, $3 orr ${2:R}, ${2:R}, ${3:R} stxp ${1:w}, $2,
${2:R}, $0 cbnz ${1:w}, 0b'
../../accel/tcg/ldst_atomicity.c.inc:902:9: error: invalid operand in
inline asm: '0: ldxp $2, ${2:R}, $0 bic $2, $2, $4 bic ${2:R}, ${2:R},
${4:R} orr $2, $2, $3 orr ${2:R}, ${2:R}, ${3:R} stxp ${1:w}, $2,
${2:R}, $0 cbnz ${1:w}, 0b'
../../accel/tcg/ldst_atomicity.c.inc:902:9: error: invalid operand in
inline asm: '0: ldxp $2, ${2:R}, $0 bic $2, $2, $4 bic ${2:R}, ${2:R},
${4:R} orr $2, $2, $3 orr ${2:R}, ${2:R}, ${3:R} stxp ${1:w}, $2,
${2:R}, $0 cbnz ${1:w}, 0b'
../../accel/tcg/ldst_atomicity.c.inc:902:9: error: invalid operand in
inline asm: '0: ldxp $2, ${2:R}, $0 bic $2, $2, $4 bic ${2:R}, ${2:R},
${4:R} orr $2, $2, $3 orr ${2:R}, ${2:R}, ${3:R} stxp ${1:w}, $2,
${2:R}, $0 cbnz ${1:w}, 0b'
../../accel/tcg/ldst_atomicity.c.inc:902:9: error: invalid operand in
inline asm: '0: ldxp $2, ${2:R}, $0 bic $2, $2, $4 bic ${2:R}, ${2:R},
${4:R} orr $2, $2, $3 orr ${2:R}, ${2:R}, ${3:R} stxp ${1:w}, $2,
${2:R}, $0 cbnz ${1:w}, 0b'
../../accel/tcg/ldst_atomicity.c.inc:902:9: error: invalid operand in
inline asm: '0: ldxp $2, ${2:R}, $0 bic $2, $2, $4 bic ${2:R}, ${2:R},
${4:R} orr $2, $2, $3 orr ${2:R}, ${2:R}, ${3:R} stxp ${1:w}, $2,
${2:R}, $0 cbnz ${1:w}, 0b'
../../accel/tcg/ldst_atomicity.c.inc:902:9: error: unknown token in
expression
<inline asm>:1:15: note: instantiated into assembly here
0: ldxp x13, , [x9]
^
In file included from ../../accel/tcg/cputlb.c:1656:
../../accel/tcg/ldst_atomicity.c.inc:902:9: error: invalid operand
asm("0: ldxp %[t], %R[t], %[mem]\n\t"
^
<inline asm>:1:15: note: instantiated into assembly here
0: ldxp x13, , [x9]
^
In file included from ../../accel/tcg/cputlb.c:1656:
../../accel/tcg/ldst_atomicity.c.inc:903:32: error: unknown token in
expression
"bic %[t], %[t], %[m]\n\t"
^
<inline asm>:3:6: note: instantiated into assembly here
bic , ,
^
In file included from ../../accel/tcg/cputlb.c:1656:
../../accel/tcg/ldst_atomicity.c.inc:902:9: error: invalid operand in
inline asm: '0: ldxp $2, ${2:R}, $0 bic $2, $2, $4 bic ${2:R}, ${2:R},
${4:R} orr $2, $2, $3 orr ${2:R}, ${2:R}, ${3:R} stxp ${1:w}, $2,
${2:R}, $0 cbnz ${1:w}, 0b'
asm("0: ldxp %[t], %R[t], %[mem]\n\t"
^
../../accel/tcg/ldst_atomicity.c.inc:902:9: error: invalid operand in
inline asm: '0: ldxp $2, ${2:R}, $0 bic $2, $2, $4 bic ${2:R}, ${2:R},
${4:R} orr $2, $2, $3 orr ${2:R}, ${2:R}, ${3:R} stxp ${1:w}, $2,
${2:R}, $0 cbnz ${1:w}, 0b'
../../accel/tcg/ldst_atomicity.c.inc:902:9: error: invalid operand in
inline asm: '0: ldxp $2, ${2:R}, $0 bic $2, $2, $4 bic ${2:R}, ${2:R},
${4:R} orr $2, $2, $3 orr ${2:R}, ${2:R}, ${3:R} stxp ${1:w}, $2,
${2:R}, $0 cbnz ${1:w}, 0b'
fatal error: too many errors emitted, stopping now [-ferror-limit=]
^ permalink raw reply [flat|nested] 48+ messages in thread* Re: [PATCH for-8.0 17/29] tcg/aarch64: Add have_lse, have_lse2
2022-11-21 23:10 ` Philippe Mathieu-Daudé
@ 2022-11-21 23:14 ` Philippe Mathieu-Daudé
0 siblings, 0 replies; 48+ messages in thread
From: Philippe Mathieu-Daudé @ 2022-11-21 23:14 UTC (permalink / raw)
To: Richard Henderson, qemu-devel
On 22/11/22 00:10, Philippe Mathieu-Daudé wrote:
> On 18/11/22 10:47, Richard Henderson wrote:
>> Notice when the host has additional atomic instructions.
>> The new variables will also be used in generated code.
>>
>> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
>> ---
>> tcg/aarch64/tcg-target.h | 3 +++
>> tcg/aarch64/tcg-target.c.inc | 10 ++++++++++
>> 2 files changed, 13 insertions(+)
>
>
>> diff --git a/tcg/aarch64/tcg-target.c.inc b/tcg/aarch64/tcg-target.c.inc
>> index 001a71bbc0..cf5ee6f742 100644
>> --- a/tcg/aarch64/tcg-target.c.inc
>> +++ b/tcg/aarch64/tcg-target.c.inc
>> @@ -13,6 +13,8 @@
>> #include "../tcg-ldst.c.inc"
>> #include "../tcg-pool.c.inc"
>> #include "qemu/bitops.h"
>> +#include <asm/hwcap.h>
>
> This doesn't build on Darwin:
Project version: 7.1.91
C compiler for the host machine: clang (clang 14.0.0 "Apple clang
version 14.0.0 (clang-1400.0.29.102)")
C linker for the host machine: clang ld64 819.6
Host machine cpu family: aarch64
Host machine cpu: arm64
> In file included from ../../tcg/tcg.c:426:
> tcg/aarch64/tcg-target.c.inc:16:10: fatal error: 'asm/hwcap.h' file not
> found
> #include <asm/hwcap.h>
> ^~~~~~~~~~~~~
>
> In file included from ../../accel/tcg/cputlb.c:1656:
> ../../accel/tcg/ldst_atomicity.c.inc:269:21: warning: value size does
> not match register size specified by the constraint and modifier
> [-Wasm-operand-widths]
> : "=&r"(r.u), "=&r"(fail) : "Q"(*p));
> ^
^ permalink raw reply [flat|nested] 48+ messages in thread
* [PATCH for-8.0 18/29] accel/tcg: Add aarch64 specific support in ldst_atomicity
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (16 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 17/29] tcg/aarch64: Add have_lse, have_lse2 Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-18 9:47 ` [PATCH for-8.0 19/29] tcg: Introduce TCG_OPF_TYPE_MASK Richard Henderson
` (10 subsequent siblings)
28 siblings, 0 replies; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
We have code in atomic128.h noting that through GCC 8, there
was no support for atomic operations on __uint128. This has
been fixed in GCC 10. But we can still improve over any
basic compare-and-swap loop using the ldxp/stxp instructions.
Add fast paths for FEAT_LSE2, using the detection in tcg.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
accel/tcg/ldst_atomicity.c.inc | 75 ++++++++++++++++++++++++++++++++--
1 file changed, 72 insertions(+), 3 deletions(-)
diff --git a/accel/tcg/ldst_atomicity.c.inc b/accel/tcg/ldst_atomicity.c.inc
index d2a3783193..186862b5ec 100644
--- a/accel/tcg/ldst_atomicity.c.inc
+++ b/accel/tcg/ldst_atomicity.c.inc
@@ -41,6 +41,8 @@
* but we're using tcg/tci/ instead.
*/
# define HAVE_al16_fast false
+#elif defined(__aarch64__)
+# define HAVE_al16_fast likely(have_lse2)
#elif defined(__x86_64__)
# define HAVE_al16_fast likely(have_atomic16)
#else
@@ -48,6 +50,8 @@
#endif
#if defined(CONFIG_ATOMIC128) || defined(CONFIG_CMPXCHG128)
# define HAVE_al16 true
+#elif defined(__aarch64__)
+# define HAVE_al16 true
#else
# define HAVE_al16 false
#endif
@@ -168,6 +172,12 @@ load_atomic16(void *pv)
r.u = qatomic_read__nocheck(p);
return r.s;
+#elif defined(__aarch64__)
+ /* Via HAVE_al16_fast, FEAT_LSE2 is present: LDP becomes atomic. */
+ Int128Alias r;
+
+ asm("ldp %0, %R0, %1" : "=r"(r.u) : "m"(*(__uint128_t *)pv));
+ return r.s;
#elif defined(__x86_64__)
Int128Alias r;
@@ -246,7 +256,20 @@ static Int128 load_atomic16_or_exit(CPUArchState *env, uintptr_t ra, void *pv)
* In system mode all guest pages are writable, and for user-only
* we have just checked writability. Try cmpxchg.
*/
-#if defined(CONFIG_CMPXCHG128)
+#if defined(__aarch64__)
+ /* We can do better than cmpxchg for AArch64. */
+ {
+ Int128Alias r;
+ uint32_t fail;
+
+ /* The load must be paired with the store to guarantee not tearing. */
+ asm("0: ldxp %0, %R0, %2\n\t"
+ "stxp %w1, %0, %R0, %2\n\t"
+ "cbnz %w1, 0b"
+ : "=&r"(r.u), "=&r"(fail) : "Q"(*p));
+ return r.s;
+ }
+#elif defined(CONFIG_CMPXCHG128)
/* Swap 0 with 0, with the side-effect of returning the old value. */
{
Int128Alias r;
@@ -393,6 +416,18 @@ load_atom_extract_al16_or_al8(void *pv, int s)
r = qatomic_read__nocheck(p16);
}
return r >> shr;
+#elif defined(__aarch64__)
+ /*
+ * Via HAVE_al16_fast, FEAT_LSE2 is present.
+ * LDP becomes single-copy atomic if 16-byte aligned, and
+ * single-copy atomic on the parts if 8-byte aligned.
+ */
+ uintptr_t pi = (uintptr_t)pv;
+ int shr = (pi & 7) * 8;
+ uint64_t l, h;
+
+ asm("ldp %0, %1, %2" : "=r"(l), "=r"(h) : "m"(*(__uint128_t *)(pi & ~7)));
+ return (l >> shr) | (h << (-shr & 63));
#elif defined(__x86_64__)
uintptr_t pi = (uintptr_t)pv;
int shr = (pi & 7) * 8;
@@ -739,7 +774,23 @@ store_atomic16(void *pv, Int128Alias val)
return;
}
#endif
-#if defined(CONFIG_CMPXCHG128)
+#if defined(__aarch64__)
+ /* We can do better than cmpxchg for AArch64. */
+ __uint128_t *pu = __builtin_assume_aligned(pv, 16);
+ __uint128_t old;
+ uint32_t fail;
+
+ if (HAVE_al16_fast) {
+ /* Via HAVE_al16_fast, FEAT_LSE2 is present: STP becomes atomic. */
+ asm("stp %1, %R1, %0" : "=Q"(*pu) : "r"(val.u));
+ } else {
+ asm("0: ldxp %0, %R0, %1\n\t"
+ "stxp %w2, %3, %R3, %1\n\t"
+ "cbnz %w2, 0b"
+ : "=&r"(old), "=Q"(*pu), "=&r"(fail) : "r"(val.u));
+ }
+ return;
+#elif defined(CONFIG_CMPXCHG128)
{
__uint128_t *pu = __builtin_assume_aligned(pv, 16);
__uint128_t o;
@@ -839,7 +890,25 @@ static void store_atom_insert_al8(uint64_t *p, uint64_t val, uint64_t msk)
static void ATTRIBUTE_ATOMIC128_OPT
store_atom_insert_al16(Int128 *ps, Int128Alias val, Int128Alias msk)
{
-#if defined(CONFIG_ATOMIC128)
+#if defined(__aarch64__)
+ /*
+ * GCC only implements __sync* primitives for int128 on aarch64.
+ * We can do better without the barriers, and integrating the
+ * arithmetic into the load-exclusive/store-conditional pair.
+ */
+ __uint128_t tmp, *pu = __builtin_assume_aligned(ps, 16);
+ uint32_t fail;
+
+ asm("0: ldxp %[t], %R[t], %[mem]\n\t"
+ "bic %[t], %[t], %[m]\n\t"
+ "bic %R[t], %R[t], %R[m]\n\t"
+ "orr %[t], %[t], %[v]\n\t"
+ "orr %R[t], %R[t], %R[v]\n\t"
+ "stxp %w[f], %[t], %R[t], %[mem]\n\t"
+ "cbnz %w[f], 0b\n"
+ : [mem] "+Q"(*pu), [f] "=&r"(fail), [t] "=&r"(tmp)
+ : [v] "r"(val.u), [m] "r"(msk.u));
+#elif defined(CONFIG_ATOMIC128)
__uint128_t *pu, old, new;
/* With CONFIG_ATOMIC128, we can avoid the memory barriers. */
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* [PATCH for-8.0 19/29] tcg: Introduce TCG_OPF_TYPE_MASK
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (17 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 18/29] accel/tcg: Add aarch64 specific support in ldst_atomicity Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-21 16:12 ` Philippe Mathieu-Daudé
2022-11-18 9:47 ` [PATCH for-8.0 20/29] tcg: Add INDEX_op_qemu_{ld,st}_i128 Richard Henderson
` (9 subsequent siblings)
28 siblings, 1 reply; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
Reorg TCG_OPF_64BIT and TCG_OPF_VECTOR into a two-bit field so
that we can add TCG_OPF_128BIT without requiring another bit.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
include/tcg/tcg.h | 22 ++++++++++++----------
tcg/optimize.c | 15 ++++++++++++---
tcg/tcg.c | 4 ++--
tcg/aarch64/tcg-target.c.inc | 8 +++++---
tcg/tci/tcg-target.c.inc | 3 ++-
5 files changed, 33 insertions(+), 19 deletions(-)
diff --git a/include/tcg/tcg.h b/include/tcg/tcg.h
index a996da60b5..5874f1e30b 100644
--- a/include/tcg/tcg.h
+++ b/include/tcg/tcg.h
@@ -994,24 +994,26 @@ typedef struct TCGArgConstraint {
/* Bits for TCGOpDef->flags, 8 bits available, all used. */
enum {
+ /* Two bits describing the output type. */
+ TCG_OPF_TYPE_MASK = 0x03,
+ TCG_OPF_32BIT = 0x00,
+ TCG_OPF_64BIT = 0x01,
+ TCG_OPF_VECTOR = 0x02,
+ TCG_OPF_128BIT = 0x03,
/* Instruction exits the translation block. */
- TCG_OPF_BB_EXIT = 0x01,
+ TCG_OPF_BB_EXIT = 0x04,
/* Instruction defines the end of a basic block. */
- TCG_OPF_BB_END = 0x02,
+ TCG_OPF_BB_END = 0x08,
/* Instruction clobbers call registers and potentially update globals. */
- TCG_OPF_CALL_CLOBBER = 0x04,
+ TCG_OPF_CALL_CLOBBER = 0x10,
/* Instruction has side effects: it cannot be removed if its outputs
are not used, and might trigger exceptions. */
- TCG_OPF_SIDE_EFFECTS = 0x08,
- /* Instruction operands are 64-bits (otherwise 32-bits). */
- TCG_OPF_64BIT = 0x10,
+ TCG_OPF_SIDE_EFFECTS = 0x20,
/* Instruction is optional and not implemented by the host, or insn
is generic and should not be implemened by the host. */
- TCG_OPF_NOT_PRESENT = 0x20,
- /* Instruction operands are vectors. */
- TCG_OPF_VECTOR = 0x40,
+ TCG_OPF_NOT_PRESENT = 0x40,
/* Instruction is a conditional branch. */
- TCG_OPF_COND_BRANCH = 0x80
+ TCG_OPF_COND_BRANCH = 0x80,
};
typedef struct TCGOpDef {
diff --git a/tcg/optimize.c b/tcg/optimize.c
index 763bca9ea6..5c0bd6b6e6 100644
--- a/tcg/optimize.c
+++ b/tcg/optimize.c
@@ -2053,12 +2053,21 @@ void tcg_optimize(TCGContext *s)
copy_propagate(&ctx, op, def->nb_oargs, def->nb_iargs);
/* Pre-compute the type of the operation. */
- if (def->flags & TCG_OPF_VECTOR) {
+ switch (def->flags & TCG_OPF_TYPE_MASK) {
+ case TCG_OPF_VECTOR:
ctx.type = TCG_TYPE_V64 + TCGOP_VECL(op);
- } else if (def->flags & TCG_OPF_64BIT) {
+ break;
+ case TCG_OPF_128BIT:
+ ctx.type = TCG_TYPE_I128;
+ break;
+ case TCG_OPF_64BIT:
ctx.type = TCG_TYPE_I64;
- } else {
+ break;
+ case TCG_OPF_32BIT:
ctx.type = TCG_TYPE_I32;
+ break;
+ default:
+ qemu_build_not_reached();
}
/* Assume all bits affected, no bits known zero, no sign reps. */
diff --git a/tcg/tcg.c b/tcg/tcg.c
index 50db393594..d221f76366 100644
--- a/tcg/tcg.c
+++ b/tcg/tcg.c
@@ -2075,7 +2075,7 @@ static void tcg_dump_ops(TCGContext *s, FILE *f, bool have_prefs)
nb_iargs = def->nb_iargs;
nb_cargs = def->nb_cargs;
- if (def->flags & TCG_OPF_VECTOR) {
+ if ((def->flags & TCG_OPF_TYPE_MASK) == TCG_OPF_VECTOR) {
col += ne_fprintf(f, "v%d,e%d,", 64 << TCGOP_VECL(op),
8 << TCGOP_VECE(op));
}
@@ -4362,7 +4362,7 @@ static void tcg_reg_alloc_op(TCGContext *s, const TCGOp *op)
}
/* emit instruction */
- if (def->flags & TCG_OPF_VECTOR) {
+ if ((def->flags & TCG_OPF_TYPE_MASK) == TCG_OPF_VECTOR) {
tcg_out_vec_op(s, op->opc, TCGOP_VECL(op), TCGOP_VECE(op),
new_args, const_args);
} else {
diff --git a/tcg/aarch64/tcg-target.c.inc b/tcg/aarch64/tcg-target.c.inc
index cf5ee6f742..9ea1608015 100644
--- a/tcg/aarch64/tcg-target.c.inc
+++ b/tcg/aarch64/tcg-target.c.inc
@@ -1896,9 +1896,11 @@ static void tcg_out_op(TCGContext *s, TCGOpcode opc,
const TCGArg args[TCG_MAX_OP_ARGS],
const int const_args[TCG_MAX_OP_ARGS])
{
- /* 99% of the time, we can signal the use of extension registers
- by looking to see if the opcode handles 64-bit data. */
- TCGType ext = (tcg_op_defs[opc].flags & TCG_OPF_64BIT) != 0;
+ /*
+ * 99% of the time, we can signal the use of extension registers
+ * by looking to see if the opcode handles 32-bit data or not.
+ */
+ TCGType ext = (tcg_op_defs[opc].flags & TCG_OPF_TYPE_MASK) != TCG_OPF_32BIT;
/* Hoist the loads of the most common arguments. */
TCGArg a0 = args[0];
diff --git a/tcg/tci/tcg-target.c.inc b/tcg/tci/tcg-target.c.inc
index 357888a532..f8ec07839c 100644
--- a/tcg/tci/tcg-target.c.inc
+++ b/tcg/tci/tcg-target.c.inc
@@ -690,7 +690,8 @@ static void tcg_out_op(TCGContext *s, TCGOpcode opc,
CASE_32_64(sextract) /* Optional (TCG_TARGET_HAS_sextract_*). */
{
TCGArg pos = args[2], len = args[3];
- TCGArg max = tcg_op_defs[opc].flags & TCG_OPF_64BIT ? 64 : 32;
+ TCGArg max = ((tcg_op_defs[opc].flags & TCG_OPF_TYPE_MASK)
+ == TCG_OPF_32BIT ? 32 : 64);
tcg_debug_assert(pos < max);
tcg_debug_assert(pos + len <= max);
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* Re: [PATCH for-8.0 19/29] tcg: Introduce TCG_OPF_TYPE_MASK
2022-11-18 9:47 ` [PATCH for-8.0 19/29] tcg: Introduce TCG_OPF_TYPE_MASK Richard Henderson
@ 2022-11-21 16:12 ` Philippe Mathieu-Daudé
0 siblings, 0 replies; 48+ messages in thread
From: Philippe Mathieu-Daudé @ 2022-11-21 16:12 UTC (permalink / raw)
To: Richard Henderson, qemu-devel
On 18/11/22 10:47, Richard Henderson wrote:
> Reorg TCG_OPF_64BIT and TCG_OPF_VECTOR into a two-bit field so
> that we can add TCG_OPF_128BIT without requiring another bit.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
> include/tcg/tcg.h | 22 ++++++++++++----------
> tcg/optimize.c | 15 ++++++++++++---
> tcg/tcg.c | 4 ++--
> tcg/aarch64/tcg-target.c.inc | 8 +++++---
> tcg/tci/tcg-target.c.inc | 3 ++-
> 5 files changed, 33 insertions(+), 19 deletions(-)
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
^ permalink raw reply [flat|nested] 48+ messages in thread
* [PATCH for-8.0 20/29] tcg: Add INDEX_op_qemu_{ld,st}_i128
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (18 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 19/29] tcg: Introduce TCG_OPF_TYPE_MASK Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-21 22:59 ` Philippe Mathieu-Daudé
2022-11-18 9:47 ` [PATCH for-8.0 21/29] tcg/i386: Introduce tcg_out_mov2 Richard Henderson
` (8 subsequent siblings)
28 siblings, 1 reply; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
Add opcodes for backend support for 128-bit memory operations.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
include/tcg/tcg-opc.h | 8 +++++
tcg/aarch64/tcg-target.h | 2 ++
tcg/arm/tcg-target.h | 2 ++
tcg/i386/tcg-target.h | 2 ++
tcg/loongarch64/tcg-target.h | 2 ++
tcg/mips/tcg-target.h | 2 ++
tcg/ppc/tcg-target.h | 2 ++
tcg/riscv/tcg-target.h | 2 ++
tcg/s390x/tcg-target.h | 2 ++
tcg/sparc64/tcg-target.h | 2 ++
tcg/tci/tcg-target.h | 2 ++
tcg/tcg-op.c | 67 ++++++++++++++++++++++++++++++++----
tcg/tcg.c | 4 +++
tcg/README | 10 ++++--
14 files changed, 100 insertions(+), 9 deletions(-)
diff --git a/include/tcg/tcg-opc.h b/include/tcg/tcg-opc.h
index dd444734d9..94cf7c5d6a 100644
--- a/include/tcg/tcg-opc.h
+++ b/include/tcg/tcg-opc.h
@@ -213,6 +213,14 @@ DEF(qemu_st8_i32, 0, TLADDR_ARGS + 1, 1,
TCG_OPF_CALL_CLOBBER | TCG_OPF_SIDE_EFFECTS |
IMPL(TCG_TARGET_HAS_qemu_st8_i32))
+/* Only for 64-bit hosts at the moment. */
+DEF(qemu_ld_i128, 2, 1, 1,
+ TCG_OPF_CALL_CLOBBER | TCG_OPF_SIDE_EFFECTS | TCG_OPF_64BIT |
+ IMPL(TCG_TARGET_HAS_qemu_ldst_i128))
+DEF(qemu_st_i128, 0, 3, 1,
+ TCG_OPF_CALL_CLOBBER | TCG_OPF_SIDE_EFFECTS | TCG_OPF_64BIT |
+ IMPL(TCG_TARGET_HAS_qemu_ldst_i128))
+
/* Host vector support. */
#define IMPLVEC TCG_OPF_VECTOR | IMPL(TCG_TARGET_MAYBE_vec)
diff --git a/tcg/aarch64/tcg-target.h b/tcg/aarch64/tcg-target.h
index b8f734f371..b0fbf5b699 100644
--- a/tcg/aarch64/tcg-target.h
+++ b/tcg/aarch64/tcg-target.h
@@ -130,6 +130,8 @@ extern bool have_lse2;
#define TCG_TARGET_HAS_mulsh_i64 1
#define TCG_TARGET_HAS_direct_jump 1
+#define TCG_TARGET_HAS_qemu_ldst_i128 0
+
#define TCG_TARGET_HAS_v64 1
#define TCG_TARGET_HAS_v128 1
#define TCG_TARGET_HAS_v256 0
diff --git a/tcg/arm/tcg-target.h b/tcg/arm/tcg-target.h
index 6613d3d791..8bcab0ac9b 100644
--- a/tcg/arm/tcg-target.h
+++ b/tcg/arm/tcg-target.h
@@ -126,6 +126,8 @@ extern bool use_neon_instructions;
#define TCG_TARGET_HAS_direct_jump 0
#define TCG_TARGET_HAS_qemu_st8_i32 0
+#define TCG_TARGET_HAS_qemu_ldst_i128 0
+
#define TCG_TARGET_HAS_v64 use_neon_instructions
#define TCG_TARGET_HAS_v128 use_neon_instructions
#define TCG_TARGET_HAS_v256 0
diff --git a/tcg/i386/tcg-target.h b/tcg/i386/tcg-target.h
index 5b037b1d2b..53d2cb3412 100644
--- a/tcg/i386/tcg-target.h
+++ b/tcg/i386/tcg-target.h
@@ -195,6 +195,8 @@ extern bool have_atomic16;
#define TCG_TARGET_HAS_qemu_st8_i32 1
#endif
+#define TCG_TARGET_HAS_qemu_ldst_i128 0
+
/* We do not support older SSE systems, only beginning with AVX1. */
#define TCG_TARGET_HAS_v64 have_avx1
#define TCG_TARGET_HAS_v128 have_avx1
diff --git a/tcg/loongarch64/tcg-target.h b/tcg/loongarch64/tcg-target.h
index 9d0db8fdfe..6cb702a108 100644
--- a/tcg/loongarch64/tcg-target.h
+++ b/tcg/loongarch64/tcg-target.h
@@ -173,6 +173,8 @@ typedef enum {
#define TCG_TARGET_HAS_muluh_i64 1
#define TCG_TARGET_HAS_mulsh_i64 1
+#define TCG_TARGET_HAS_qemu_ldst_i128 0
+
void tb_target_set_jmp_target(uintptr_t, uintptr_t, uintptr_t, uintptr_t);
#define TCG_TARGET_DEFAULT_MO (0)
diff --git a/tcg/mips/tcg-target.h b/tcg/mips/tcg-target.h
index b235cba8ba..0897cfd8d5 100644
--- a/tcg/mips/tcg-target.h
+++ b/tcg/mips/tcg-target.h
@@ -204,6 +204,8 @@ extern bool use_mips32r2_instructions;
#define TCG_TARGET_HAS_ext16u_i64 0 /* andi rt, rs, 0xffff */
#endif
+#define TCG_TARGET_HAS_qemu_ldst_i128 0
+
#define TCG_TARGET_DEFAULT_MO (0)
#define TCG_TARGET_HAS_MEMORY_BSWAP 1
diff --git a/tcg/ppc/tcg-target.h b/tcg/ppc/tcg-target.h
index b5cd225cfa..920a746482 100644
--- a/tcg/ppc/tcg-target.h
+++ b/tcg/ppc/tcg-target.h
@@ -151,6 +151,8 @@ extern bool have_vsx;
#define TCG_TARGET_HAS_mulsh_i64 1
#endif
+#define TCG_TARGET_HAS_qemu_ldst_i128 0
+
/*
* While technically Altivec could support V64, it has no 64-bit store
* instruction and substituting two 32-bit stores makes the generated
diff --git a/tcg/riscv/tcg-target.h b/tcg/riscv/tcg-target.h
index d61ca902d3..205d513d08 100644
--- a/tcg/riscv/tcg-target.h
+++ b/tcg/riscv/tcg-target.h
@@ -168,6 +168,8 @@ typedef enum {
#define TCG_TARGET_HAS_mulsh_i64 1
#endif
+#define TCG_TARGET_HAS_qemu_ldst_i128 0
+
/* not defined -- call should be eliminated at compile time */
void tb_target_set_jmp_target(uintptr_t, uintptr_t, uintptr_t, uintptr_t);
diff --git a/tcg/s390x/tcg-target.h b/tcg/s390x/tcg-target.h
index 9a3856f0b3..f87905d1e4 100644
--- a/tcg/s390x/tcg-target.h
+++ b/tcg/s390x/tcg-target.h
@@ -139,6 +139,8 @@ extern uint64_t s390_facilities[3];
#define TCG_TARGET_HAS_muluh_i64 0
#define TCG_TARGET_HAS_mulsh_i64 0
+#define TCG_TARGET_HAS_qemu_ldst_i128 0
+
#define TCG_TARGET_HAS_v64 HAVE_FACILITY(VECTOR)
#define TCG_TARGET_HAS_v128 HAVE_FACILITY(VECTOR)
#define TCG_TARGET_HAS_v256 0
diff --git a/tcg/sparc64/tcg-target.h b/tcg/sparc64/tcg-target.h
index 53cfa843da..bfbfb51319 100644
--- a/tcg/sparc64/tcg-target.h
+++ b/tcg/sparc64/tcg-target.h
@@ -152,6 +152,8 @@ extern bool use_vis3_instructions;
#define TCG_TARGET_HAS_muluh_i64 use_vis3_instructions
#define TCG_TARGET_HAS_mulsh_i64 0
+#define TCG_TARGET_HAS_qemu_ldst_i128 0
+
#define TCG_AREG0 TCG_REG_I0
#define TCG_TARGET_DEFAULT_MO (0)
diff --git a/tcg/tci/tcg-target.h b/tcg/tci/tcg-target.h
index 9d569c9e04..e4899c7d02 100644
--- a/tcg/tci/tcg-target.h
+++ b/tcg/tci/tcg-target.h
@@ -128,6 +128,8 @@
#define TCG_TARGET_HAS_mulu2_i32 1
#endif /* TCG_TARGET_REG_BITS == 64 */
+#define TCG_TARGET_HAS_qemu_ldst_i128 0
+
/* Number of registers available. */
#define TCG_TARGET_NB_REGS 16
diff --git a/tcg/tcg-op.c b/tcg/tcg-op.c
index bbb29bed2b..6210577b85 100644
--- a/tcg/tcg-op.c
+++ b/tcg/tcg-op.c
@@ -3201,7 +3201,7 @@ static void canonicalize_memop_i128_as_i64(MemOp ret[2], MemOp orig)
void tcg_gen_qemu_ld_i128(TCGv_i128 val, TCGv addr, TCGArg idx, MemOp memop)
{
- MemOpIdx oi = make_memop_idx(memop, idx);
+ const MemOpIdx oi = make_memop_idx(memop, idx);
tcg_debug_assert((memop & MO_SIZE) == MO_128);
tcg_debug_assert((memop & MO_SIGN) == 0);
@@ -3209,9 +3209,35 @@ void tcg_gen_qemu_ld_i128(TCGv_i128 val, TCGv addr, TCGArg idx, MemOp memop)
tcg_gen_req_mo(TCG_MO_LD_LD | TCG_MO_ST_LD);
addr = plugin_prep_mem_callbacks(addr);
- /* TODO: allow the tcg backend to see the whole operation. */
+ /* TODO: For now, force 32-bit hosts to use the helper. */
+ if (TCG_TARGET_HAS_qemu_ldst_i128 && TCG_TARGET_REG_BITS == 64) {
+ TCGv_i64 lo, hi;
+ TCGArg addr_arg;
+ MemOpIdx adj_oi;
- if (use_two_i64_for_i128(memop)) {
+ /* TODO: Make TCG_TARGET_HAS_MEMORY_BSWAP fine grained. */
+ if (!TCG_TARGET_HAS_MEMORY_BSWAP && (memop & MO_BSWAP)) {
+ lo = TCGV128_HIGH(val);
+ hi = TCGV128_LOW(val);
+ adj_oi = make_memop_idx(memop & ~MO_BSWAP, idx);
+ } else {
+ lo = TCGV128_LOW(val);
+ hi = TCGV128_HIGH(val);
+ adj_oi = oi;
+ }
+
+#if TARGET_LONG_BITS == 32
+ addr_arg = tcgv_i32_arg(addr);
+#else
+ addr_arg = tcgv_i64_arg(addr);
+#endif
+ tcg_gen_op4ii_i64(INDEX_op_qemu_ld_i128, lo, hi, addr_arg, adj_oi);
+
+ if (!TCG_TARGET_HAS_MEMORY_BSWAP && (memop & MO_BSWAP)) {
+ tcg_gen_bswap64_i64(lo, lo);
+ tcg_gen_bswap64_i64(hi, hi);
+ }
+ } else if (use_two_i64_for_i128(memop)) {
MemOp mop[2];
TCGv addr_p8;
TCGv_i64 x, y;
@@ -3254,7 +3280,7 @@ void tcg_gen_qemu_ld_i128(TCGv_i128 val, TCGv addr, TCGArg idx, MemOp memop)
void tcg_gen_qemu_st_i128(TCGv_i128 val, TCGv addr, TCGArg idx, MemOp memop)
{
- MemOpIdx oi = make_memop_idx(memop, idx);
+ const MemOpIdx oi = make_memop_idx(memop, idx);
tcg_debug_assert((memop & MO_SIZE) == MO_128);
tcg_debug_assert((memop & MO_SIGN) == 0);
@@ -3262,9 +3288,38 @@ void tcg_gen_qemu_st_i128(TCGv_i128 val, TCGv addr, TCGArg idx, MemOp memop)
tcg_gen_req_mo(TCG_MO_ST_LD | TCG_MO_ST_ST);
addr = plugin_prep_mem_callbacks(addr);
- /* TODO: allow the tcg backend to see the whole operation. */
+ /* TODO: For now, force 32-bit hosts to use the helper. */
- if (use_two_i64_for_i128(memop)) {
+ if (TCG_TARGET_HAS_qemu_ldst_i128 && TCG_TARGET_REG_BITS == 64) {
+ TCGv_i64 lo, hi;
+ TCGArg addr_arg;
+ MemOpIdx adj_oi;
+
+ /* TODO: Make TCG_TARGET_HAS_MEMORY_BSWAP fine grained. */
+ if (!TCG_TARGET_HAS_MEMORY_BSWAP && (memop & MO_BSWAP)) {
+ lo = tcg_temp_new_i64();
+ hi = tcg_temp_new_i64();
+ tcg_gen_bswap64_i64(lo, TCGV128_HIGH(val));
+ tcg_gen_bswap64_i64(hi, TCGV128_LOW(val));
+ adj_oi = make_memop_idx(memop & ~MO_BSWAP, idx);
+ } else {
+ lo = TCGV128_LOW(val);
+ hi = TCGV128_HIGH(val);
+ adj_oi = oi;
+ }
+
+#if TARGET_LONG_BITS == 32
+ addr_arg = tcgv_i32_arg(addr);
+#else
+ addr_arg = tcgv_i64_arg(addr);
+#endif
+ tcg_gen_op4ii_i64(INDEX_op_qemu_st_i128, lo, hi, addr_arg, adj_oi);
+
+ if (!TCG_TARGET_HAS_MEMORY_BSWAP && (memop & MO_BSWAP)) {
+ tcg_temp_free_i64(lo);
+ tcg_temp_free_i64(hi);
+ }
+ } else if (use_two_i64_for_i128(memop)) {
MemOp mop[2];
TCGv addr_p8;
TCGv_i64 x, y;
diff --git a/tcg/tcg.c b/tcg/tcg.c
index d221f76366..9a000c55ed 100644
--- a/tcg/tcg.c
+++ b/tcg/tcg.c
@@ -1497,6 +1497,10 @@ bool tcg_op_supported(TCGOpcode op)
case INDEX_op_qemu_st8_i32:
return TCG_TARGET_HAS_qemu_st8_i32;
+ case INDEX_op_qemu_ld_i128:
+ case INDEX_op_qemu_st_i128:
+ return TCG_TARGET_HAS_qemu_ldst_i128;
+
case INDEX_op_mov_i32:
case INDEX_op_setcond_i32:
case INDEX_op_brcond_i32:
diff --git a/tcg/README b/tcg/README
index bc15cc3b32..b3f8578955 100644
--- a/tcg/README
+++ b/tcg/README
@@ -512,8 +512,8 @@ jump to the TCG epilogue to go back to the exec loop.
This operation is optional. If the TCG backend does not implement the
goto_ptr opcode, emitting this op is equivalent to emitting exit_tb(0).
-* qemu_ld_i32/i64 t0, t1, flags, memidx
-* qemu_st_i32/i64 t0, t1, flags, memidx
+* qemu_ld_i32/i64/i128 t0, t1, flags, memidx
+* qemu_st_i32/i64/i128 t0, t1, flags, memidx
* qemu_st8_i32 t0, t1, flags, memidx
Load data at the guest address t1 into t0, or store data in t0 at guest
@@ -522,7 +522,8 @@ register t0 only. The address t1 is always sized according to the guest,
and the width of the memory operation is controlled by flags.
Both t0 and t1 may be split into little-endian ordered pairs of registers
-if dealing with 64-bit quantities on a 32-bit host.
+if dealing with 64-bit quantities on a 32-bit host, or 128-bit quantities
+on a 64-bit host.
The memidx selects the qemu tlb index to use (e.g. user or kernel access).
The flags are the MemOp bits, selecting the sign, width, and endianness
@@ -531,6 +532,9 @@ of the memory access.
For a 32-bit host, qemu_ld/st_i64 is guaranteed to only be used with a
64-bit memory access specified in flags.
+For qemu_ld/st_i128, these are only supported for a 64-bit host, and are
+guaranteed to be used with the host memory ordering.
+
For i386, qemu_st8_i32 is exactly like qemu_st_i32, except the size of
the memory operation is known to be 8-bit. This allows the backend to
provide a different set of register constraints.
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* Re: [PATCH for-8.0 20/29] tcg: Add INDEX_op_qemu_{ld,st}_i128
2022-11-18 9:47 ` [PATCH for-8.0 20/29] tcg: Add INDEX_op_qemu_{ld,st}_i128 Richard Henderson
@ 2022-11-21 22:59 ` Philippe Mathieu-Daudé
0 siblings, 0 replies; 48+ messages in thread
From: Philippe Mathieu-Daudé @ 2022-11-21 22:59 UTC (permalink / raw)
To: Richard Henderson, qemu-devel; +Cc: Emanuele Giuseppe Esposito
On 18/11/22 10:47, Richard Henderson wrote:
> Add opcodes for backend support for 128-bit memory operations.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
> include/tcg/tcg-opc.h | 8 +++++
> tcg/aarch64/tcg-target.h | 2 ++
> tcg/arm/tcg-target.h | 2 ++
> tcg/i386/tcg-target.h | 2 ++
> tcg/loongarch64/tcg-target.h | 2 ++
> tcg/mips/tcg-target.h | 2 ++
> tcg/ppc/tcg-target.h | 2 ++
> tcg/riscv/tcg-target.h | 2 ++
> tcg/s390x/tcg-target.h | 2 ++
> tcg/sparc64/tcg-target.h | 2 ++
> tcg/tci/tcg-target.h | 2 ++
> tcg/tcg-op.c | 67 ++++++++++++++++++++++++++++++++----
> tcg/tcg.c | 4 +++
> tcg/README | 10 ++++--
> 14 files changed, 100 insertions(+), 9 deletions(-)
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
^ permalink raw reply [flat|nested] 48+ messages in thread
* [PATCH for-8.0 21/29] tcg/i386: Introduce tcg_out_mov2
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (19 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 20/29] tcg: Add INDEX_op_qemu_{ld,st}_i128 Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-21 16:21 ` Philippe Mathieu-Daudé
2022-11-18 9:47 ` [PATCH for-8.0 22/29] tcg/i386: Introduce tcg_out_testi Richard Henderson
` (7 subsequent siblings)
28 siblings, 1 reply; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
Create a helper for data movement minding register overlap.
Use the more general xchg instruction, which consumes one
extra byte, but simplifies the more general function.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
tcg/i386/tcg-target.c.inc | 27 +++++++++++++++++++++------
1 file changed, 21 insertions(+), 6 deletions(-)
diff --git a/tcg/i386/tcg-target.c.inc b/tcg/i386/tcg-target.c.inc
index f4c0c7b8a2..79568a3981 100644
--- a/tcg/i386/tcg-target.c.inc
+++ b/tcg/i386/tcg-target.c.inc
@@ -461,6 +461,7 @@ static bool tcg_target_const_match(int64_t val, TCGType type, int ct)
#define OPC_VPTERNLOGQ (0x25 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
#define OPC_VZEROUPPER (0x77 | P_EXT)
#define OPC_XCHG_ax_r32 (0x90)
+#define OPC_XCHG_EvGv (0x87)
#define OPC_GRP3_Eb (0xf6)
#define OPC_GRP3_Ev (0xf7)
@@ -1880,6 +1881,24 @@ static void add_qemu_ldst_label(TCGContext *s, bool is_ld, bool is_64,
}
}
+/* Move src1 to dst1 and src2 to dst2, minding possible overlap. */
+static void tcg_out_mov2(TCGContext *s,
+ TCGType type1, TCGReg dst1, TCGReg src1,
+ TCGType type2, TCGReg dst2, TCGReg src2)
+{
+ if (dst1 != src2) {
+ tcg_out_mov(s, type1, dst1, src1);
+ tcg_out_mov(s, type2, dst2, src2);
+ } else if (dst2 != src1) {
+ tcg_out_mov(s, type2, dst2, src2);
+ tcg_out_mov(s, type1, dst1, src1);
+ } else {
+ /* dst1 == src2 && dst2 == src1 -> xchg. */
+ int w = (type1 == TCG_TYPE_I32 && type2 == TCG_TYPE_I32 ? 0 : P_REXW);
+ tcg_out_modrm(s, OPC_XCHG_EvGv + w, dst1, dst2);
+ }
+}
+
/*
* Generate code for the slow path for a load at the end of block
*/
@@ -1947,13 +1966,9 @@ static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
case MO_UQ:
if (TCG_TARGET_REG_BITS == 64) {
tcg_out_mov(s, TCG_TYPE_I64, data_reg, TCG_REG_RAX);
- } else if (data_reg == TCG_REG_EDX) {
- /* xchg %edx, %eax */
- tcg_out_opc(s, OPC_XCHG_ax_r32 + TCG_REG_EDX, 0, 0, 0);
- tcg_out_mov(s, TCG_TYPE_I32, l->datahi_reg, TCG_REG_EAX);
} else {
- tcg_out_mov(s, TCG_TYPE_I32, data_reg, TCG_REG_EAX);
- tcg_out_mov(s, TCG_TYPE_I32, l->datahi_reg, TCG_REG_EDX);
+ tcg_out_mov2(s, TCG_TYPE_I32, data_reg, TCG_REG_EAX,
+ TCG_TYPE_I32, l->datahi_reg, TCG_REG_EDX);
}
break;
default:
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* Re: [PATCH for-8.0 21/29] tcg/i386: Introduce tcg_out_mov2
2022-11-18 9:47 ` [PATCH for-8.0 21/29] tcg/i386: Introduce tcg_out_mov2 Richard Henderson
@ 2022-11-21 16:21 ` Philippe Mathieu-Daudé
0 siblings, 0 replies; 48+ messages in thread
From: Philippe Mathieu-Daudé @ 2022-11-21 16:21 UTC (permalink / raw)
To: Richard Henderson, qemu-devel
On 18/11/22 10:47, Richard Henderson wrote:
> Create a helper for data movement minding register overlap.
> Use the more general xchg instruction, which consumes one
> extra byte, but simplifies the more general function.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
> tcg/i386/tcg-target.c.inc | 27 +++++++++++++++++++++------
> 1 file changed, 21 insertions(+), 6 deletions(-)
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
^ permalink raw reply [flat|nested] 48+ messages in thread
* [PATCH for-8.0 22/29] tcg/i386: Introduce tcg_out_testi
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (20 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 21/29] tcg/i386: Introduce tcg_out_mov2 Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-21 16:22 ` Philippe Mathieu-Daudé
2022-11-18 9:47 ` [PATCH for-8.0 23/29] tcg/i386: Use full load/store helpers in user-only mode Richard Henderson
` (6 subsequent siblings)
28 siblings, 1 reply; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
Split out a helper for choosing testb vs testl.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
tcg/i386/tcg-target.c.inc | 30 ++++++++++++++++++------------
1 file changed, 18 insertions(+), 12 deletions(-)
diff --git a/tcg/i386/tcg-target.c.inc b/tcg/i386/tcg-target.c.inc
index 79568a3981..5ddbbbaf18 100644
--- a/tcg/i386/tcg-target.c.inc
+++ b/tcg/i386/tcg-target.c.inc
@@ -1729,6 +1729,23 @@ static void tcg_out_nopn(TCGContext *s, int n)
tcg_out8(s, 0x90);
}
+/* Test register R vs immediate bits I, setting Z flag for EQ/NE. */
+static void __attribute__((unused))
+tcg_out_testi(TCGContext *s, TCGReg r, uint32_t i)
+{
+ /*
+ * This is used for testing alignment, so we can usually use testb.
+ * For i686, we have to use testl for %esi/%edi.
+ */
+ if (i <= 0xff && (TCG_TARGET_REG_BITS == 64 || r < 4)) {
+ tcg_out_modrm(s, OPC_GRP3_Eb | P_REXB_RM, EXT3_TESTi, r);
+ tcg_out8(s, i);
+ } else {
+ tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_TESTi, r);
+ tcg_out32(s, i);
+ }
+}
+
#if defined(CONFIG_SOFTMMU)
/*
* helper signature: helper_ld*_mmu(CPUState *env, target_ulong addr,
@@ -2056,18 +2073,7 @@ static void tcg_out_test_alignment(TCGContext *s, bool is_ld, TCGReg addrlo,
unsigned a_mask = (1 << a_bits) - 1;
TCGLabelQemuLdst *label;
- /*
- * We are expecting a_bits to max out at 7, so we can usually use testb.
- * For i686, we have to use testl for %esi/%edi.
- */
- if (a_mask <= 0xff && (TCG_TARGET_REG_BITS == 64 || addrlo < 4)) {
- tcg_out_modrm(s, OPC_GRP3_Eb | P_REXB_RM, EXT3_TESTi, addrlo);
- tcg_out8(s, a_mask);
- } else {
- tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_TESTi, addrlo);
- tcg_out32(s, a_mask);
- }
-
+ tcg_out_testi(s, addrlo, a_mask);
/* jne slow_path */
tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* [PATCH for-8.0 23/29] tcg/i386: Use full load/store helpers in user-only mode
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (21 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 22/29] tcg/i386: Introduce tcg_out_testi Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-18 9:47 ` [PATCH for-8.0 24/29] tcg/i386: Replace is64 with type in qemu_ld/st routines Richard Henderson
` (5 subsequent siblings)
28 siblings, 0 replies; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
Instead of using helper_unaligned_{ld,st}, use the full
load/store helpers. This will allow the fast path to
increase alignment to implement atomicity while not
immediately raising an alignment exception.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
tcg/i386/tcg-target.c.inc | 332 ++++++++++++++++----------------------
1 file changed, 142 insertions(+), 190 deletions(-)
diff --git a/tcg/i386/tcg-target.c.inc b/tcg/i386/tcg-target.c.inc
index 5ddbbbaf18..eb93807b5f 100644
--- a/tcg/i386/tcg-target.c.inc
+++ b/tcg/i386/tcg-target.c.inc
@@ -1746,7 +1746,6 @@ tcg_out_testi(TCGContext *s, TCGReg r, uint32_t i)
}
}
-#if defined(CONFIG_SOFTMMU)
/*
* helper signature: helper_ld*_mmu(CPUState *env, target_ulong addr,
* int mmu_idx, uintptr_t ra)
@@ -1769,108 +1768,6 @@ static void * const qemu_st_helpers[MO_SIZE + 1] = {
[MO_UQ] = helper_stq_mmu,
};
-/* Perform the TLB load and compare.
-
- Inputs:
- ADDRLO and ADDRHI contain the low and high part of the address.
-
- MEM_INDEX and S_BITS are the memory context and log2 size of the load.
-
- WHICH is the offset into the CPUTLBEntry structure of the slot to read.
- This should be offsetof addr_read or addr_write.
-
- Outputs:
- LABEL_PTRS is filled with 1 (32-bit addresses) or 2 (64-bit addresses)
- positions of the displacements of forward jumps to the TLB miss case.
-
- Second argument register is loaded with the low part of the address.
- In the TLB hit case, it has been adjusted as indicated by the TLB
- and so is a host address. In the TLB miss case, it continues to
- hold a guest address.
-
- First argument register is clobbered. */
-
-static inline void tcg_out_tlb_load(TCGContext *s, TCGReg addrlo, TCGReg addrhi,
- int mem_index, MemOp opc,
- tcg_insn_unit **label_ptr, int which)
-{
- const TCGReg r0 = TCG_REG_L0;
- const TCGReg r1 = TCG_REG_L1;
- TCGType ttype = TCG_TYPE_I32;
- TCGType tlbtype = TCG_TYPE_I32;
- int trexw = 0, hrexw = 0, tlbrexw = 0;
- unsigned a_bits = get_alignment_bits(opc);
- unsigned s_bits = opc & MO_SIZE;
- unsigned a_mask = (1 << a_bits) - 1;
- unsigned s_mask = (1 << s_bits) - 1;
- target_ulong tlb_mask;
-
- if (TCG_TARGET_REG_BITS == 64) {
- if (TARGET_LONG_BITS == 64) {
- ttype = TCG_TYPE_I64;
- trexw = P_REXW;
- }
- if (TCG_TYPE_PTR == TCG_TYPE_I64) {
- hrexw = P_REXW;
- if (TARGET_PAGE_BITS + CPU_TLB_DYN_MAX_BITS > 32) {
- tlbtype = TCG_TYPE_I64;
- tlbrexw = P_REXW;
- }
- }
- }
-
- tcg_out_mov(s, tlbtype, r0, addrlo);
- tcg_out_shifti(s, SHIFT_SHR + tlbrexw, r0,
- TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS);
-
- tcg_out_modrm_offset(s, OPC_AND_GvEv + trexw, r0, TCG_AREG0,
- TLB_MASK_TABLE_OFS(mem_index) +
- offsetof(CPUTLBDescFast, mask));
-
- tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, r0, TCG_AREG0,
- TLB_MASK_TABLE_OFS(mem_index) +
- offsetof(CPUTLBDescFast, table));
-
- /* If the required alignment is at least as large as the access, simply
- copy the address and mask. For lesser alignments, check that we don't
- cross pages for the complete access. */
- if (a_bits >= s_bits) {
- tcg_out_mov(s, ttype, r1, addrlo);
- } else {
- tcg_out_modrm_offset(s, OPC_LEA + trexw, r1, addrlo, s_mask - a_mask);
- }
- tlb_mask = (target_ulong)TARGET_PAGE_MASK | a_mask;
- tgen_arithi(s, ARITH_AND + trexw, r1, tlb_mask, 0);
-
- /* cmp 0(r0), r1 */
- tcg_out_modrm_offset(s, OPC_CMP_GvEv + trexw, r1, r0, which);
-
- /* Prepare for both the fast path add of the tlb addend, and the slow
- path function argument setup. */
- tcg_out_mov(s, ttype, r1, addrlo);
-
- /* jne slow_path */
- tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
- label_ptr[0] = s->code_ptr;
- s->code_ptr += 4;
-
- if (TARGET_LONG_BITS > TCG_TARGET_REG_BITS) {
- /* cmp 4(r0), addrhi */
- tcg_out_modrm_offset(s, OPC_CMP_GvEv, addrhi, r0, which + 4);
-
- /* jne slow_path */
- tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
- label_ptr[1] = s->code_ptr;
- s->code_ptr += 4;
- }
-
- /* TLB Hit. */
-
- /* add addend(r0), r1 */
- tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, r1, r0,
- offsetof(CPUTLBEntry, addend));
-}
-
/*
* Record the context of a call to the out of line helper code for the slow path
* for a load or store, so that we can later generate the correct helper code
@@ -1893,9 +1790,7 @@ static void add_qemu_ldst_label(TCGContext *s, bool is_ld, bool is_64,
label->addrhi_reg = addrhi;
label->raddr = tcg_splitwx_to_rx(raddr);
label->label_ptr[0] = label_ptr[0];
- if (TARGET_LONG_BITS > TCG_TARGET_REG_BITS) {
- label->label_ptr[1] = label_ptr[1];
- }
+ label->label_ptr[1] = label_ptr[1];
}
/* Move src1 to dst1 and src2 to dst2, minding possible overlap. */
@@ -1929,7 +1824,7 @@ static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
/* resolve label address */
tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
- if (TARGET_LONG_BITS > TCG_TARGET_REG_BITS) {
+ if (label_ptr[1]) {
tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
}
@@ -1952,8 +1847,9 @@ static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
tcg_out_sti(s, TCG_TYPE_PTR, (uintptr_t)l->raddr, TCG_REG_ESP, ofs);
} else {
+ tcg_out_mov(s, TCG_TYPE_TL, tcg_target_call_iarg_regs[1],
+ l->addrlo_reg);
tcg_out_mov(s, TCG_TYPE_PTR, tcg_target_call_iarg_regs[0], TCG_AREG0);
- /* The second argument is already loaded with addrlo. */
tcg_out_movi(s, TCG_TYPE_I32, tcg_target_call_iarg_regs[2], oi);
tcg_out_movi(s, TCG_TYPE_PTR, tcg_target_call_iarg_regs[3],
(uintptr_t)l->raddr);
@@ -2010,7 +1906,7 @@ static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
/* resolve label address */
tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
- if (TARGET_LONG_BITS > TCG_TARGET_REG_BITS) {
+ if (label_ptr[1]) {
tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
}
@@ -2043,10 +1939,11 @@ static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
tcg_out_movi(s, TCG_TYPE_PTR, retaddr, (uintptr_t)l->raddr);
tcg_out_st(s, TCG_TYPE_PTR, retaddr, TCG_REG_ESP, ofs);
} else {
+ tcg_out_mov2(s, TCG_TYPE_TL,
+ tcg_target_call_iarg_regs[1], l->addrlo_reg,
+ s_bits == MO_64 ? TCG_TYPE_I64 : TCG_TYPE_I32,
+ tcg_target_call_iarg_regs[2], l->datalo_reg);
tcg_out_mov(s, TCG_TYPE_PTR, tcg_target_call_iarg_regs[0], TCG_AREG0);
- /* The second argument is already loaded with addrlo. */
- tcg_out_mov(s, (s_bits == MO_64 ? TCG_TYPE_I64 : TCG_TYPE_I32),
- tcg_target_call_iarg_regs[2], l->datalo_reg);
tcg_out_movi(s, TCG_TYPE_I32, tcg_target_call_iarg_regs[3], oi);
if (ARRAY_SIZE(tcg_target_call_iarg_regs) > 4) {
@@ -2065,72 +1962,129 @@ static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
tcg_out_jmp(s, qemu_st_helpers[opc & MO_SIZE]);
return true;
}
+
+#if defined(CONFIG_SOFTMMU)
+/*
+ * Perform the TLB load and compare.
+ *
+ * Inputs:
+ * ADDRLO and ADDRHI contain the low and high part of the address.
+ *
+ * MEM_INDEX and S_BITS are the memory context and log2 size of the load.
+ *
+ * WHICH is the offset into the CPUTLBEntry structure of the slot to read.
+ * This should be offsetof addr_read or addr_write.
+ *
+ * Outputs:
+ * LABEL_PTRS is filled with 1 (32-bit addresses) or 2 (64-bit addresses)
+ * positions of the displacements of forward jumps to the TLB miss case.
+ *
+ * Second argument register is loaded with the low part of the address.
+ * In the TLB hit case, it has been adjusted as indicated by the TLB
+ * and so is a host address. In the TLB miss case, it continues to
+ * hold a guest address.
+ *
+ * First argument register is clobbered.
+ */
+static void tcg_out_tlb_load(TCGContext *s, TCGReg addrlo, TCGReg addrhi,
+ int mem_index, MemOp opc,
+ tcg_insn_unit **label_ptr, int which)
+{
+ const TCGReg r0 = TCG_REG_L0;
+ const TCGReg r1 = TCG_REG_L1;
+ TCGType ttype = TCG_TYPE_I32;
+ TCGType tlbtype = TCG_TYPE_I32;
+ int trexw = 0, hrexw = 0, tlbrexw = 0;
+ unsigned a_bits = get_alignment_bits(opc);
+ unsigned s_bits = opc & MO_SIZE;
+ unsigned a_mask = (1 << a_bits) - 1;
+ unsigned s_mask = (1 << s_bits) - 1;
+ target_ulong tlb_mask;
+
+ if (TCG_TARGET_REG_BITS == 64) {
+ if (TARGET_LONG_BITS == 64) {
+ ttype = TCG_TYPE_I64;
+ trexw = P_REXW;
+ }
+ if (TCG_TYPE_PTR == TCG_TYPE_I64) {
+ hrexw = P_REXW;
+ if (TARGET_PAGE_BITS + CPU_TLB_DYN_MAX_BITS > 32) {
+ tlbtype = TCG_TYPE_I64;
+ tlbrexw = P_REXW;
+ }
+ }
+ }
+
+ tcg_out_mov(s, tlbtype, r0, addrlo);
+ tcg_out_shifti(s, SHIFT_SHR + tlbrexw, r0,
+ TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS);
+
+ tcg_out_modrm_offset(s, OPC_AND_GvEv + trexw, r0, TCG_AREG0,
+ TLB_MASK_TABLE_OFS(mem_index) +
+ offsetof(CPUTLBDescFast, mask));
+
+ tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, r0, TCG_AREG0,
+ TLB_MASK_TABLE_OFS(mem_index) +
+ offsetof(CPUTLBDescFast, table));
+
+ /*
+ * If the required alignment is at least as large as the access, simply
+ * copy the address and mask. For lesser alignments, check that we don't
+ * cross pages for the complete access.
+ */
+ if (a_bits >= s_bits) {
+ tcg_out_mov(s, ttype, r1, addrlo);
+ } else {
+ tcg_out_modrm_offset(s, OPC_LEA + trexw, r1, addrlo, s_mask - a_mask);
+ }
+ tlb_mask = (target_ulong)TARGET_PAGE_MASK | a_mask;
+ tgen_arithi(s, ARITH_AND + trexw, r1, tlb_mask, 0);
+
+ /* cmp 0(r0), r1 */
+ tcg_out_modrm_offset(s, OPC_CMP_GvEv + trexw, r1, r0, which);
+
+ /*
+ * Prepare for both the fast path add of the tlb addend, and the slow
+ * path function argument setup.
+ */
+ tcg_out_mov(s, ttype, r1, addrlo);
+
+ /* jne slow_path */
+ tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
+ label_ptr[0] = s->code_ptr;
+ s->code_ptr += 4;
+
+ if (TARGET_LONG_BITS > TCG_TARGET_REG_BITS) {
+ /* cmp 4(r0), addrhi */
+ tcg_out_modrm_offset(s, OPC_CMP_GvEv, addrhi, r0, which + 4);
+
+ /* jne slow_path */
+ tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
+ label_ptr[1] = s->code_ptr;
+ s->code_ptr += 4;
+ }
+
+ /* TLB Hit. */
+
+ /* add addend(r0), r1 */
+ tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, r1, r0,
+ offsetof(CPUTLBEntry, addend));
+}
+
#else
-static void tcg_out_test_alignment(TCGContext *s, bool is_ld, TCGReg addrlo,
- TCGReg addrhi, unsigned a_bits)
+static void tcg_out_test_alignment(TCGContext *s, TCGReg addrlo,
+ unsigned a_bits, tcg_insn_unit **label_ptr)
{
unsigned a_mask = (1 << a_bits) - 1;
- TCGLabelQemuLdst *label;
tcg_out_testi(s, addrlo, a_mask);
/* jne slow_path */
tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
-
- label = new_ldst_label(s);
- label->is_ld = is_ld;
- label->addrlo_reg = addrlo;
- label->addrhi_reg = addrhi;
- label->raddr = tcg_splitwx_to_rx(s->code_ptr + 4);
- label->label_ptr[0] = s->code_ptr;
-
+ *label_ptr = s->code_ptr;
s->code_ptr += 4;
}
-static bool tcg_out_fail_alignment(TCGContext *s, TCGLabelQemuLdst *l)
-{
- /* resolve label address */
- tcg_patch32(l->label_ptr[0], s->code_ptr - l->label_ptr[0] - 4);
-
- if (TCG_TARGET_REG_BITS == 32) {
- int ofs = 0;
-
- tcg_out_st(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP, ofs);
- ofs += 4;
-
- tcg_out_st(s, TCG_TYPE_I32, l->addrlo_reg, TCG_REG_ESP, ofs);
- ofs += 4;
- if (TARGET_LONG_BITS == 64) {
- tcg_out_st(s, TCG_TYPE_I32, l->addrhi_reg, TCG_REG_ESP, ofs);
- ofs += 4;
- }
-
- tcg_out_pushi(s, (uintptr_t)l->raddr);
- } else {
- tcg_out_mov(s, TCG_TYPE_TL, tcg_target_call_iarg_regs[1],
- l->addrlo_reg);
- tcg_out_mov(s, TCG_TYPE_PTR, tcg_target_call_iarg_regs[0], TCG_AREG0);
-
- tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_RAX, (uintptr_t)l->raddr);
- tcg_out_push(s, TCG_REG_RAX);
- }
-
- /* "Tail call" to the helper, with the return address back inline. */
- tcg_out_jmp(s, (const void *)(l->is_ld ? helper_unaligned_ld
- : helper_unaligned_st));
- return true;
-}
-
-static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
-{
- return tcg_out_fail_alignment(s, l);
-}
-
-static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
-{
- return tcg_out_fail_alignment(s, l);
-}
-
#if TCG_TARGET_REG_BITS == 32
# define x86_guest_base_seg 0
# define x86_guest_base_index -1
@@ -2165,7 +2119,7 @@ static inline int setup_guest_base_seg(void)
return 0;
}
# endif
-#endif
+#endif /* TCG_TARGET_REG_BITS == 32 */
#endif /* SOFTMMU */
static void tcg_out_qemu_ld_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
@@ -2272,10 +2226,8 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args, bool is64)
TCGReg addrhi __attribute__((unused));
MemOpIdx oi;
MemOp opc;
-#if defined(CONFIG_SOFTMMU)
- int mem_index;
- tcg_insn_unit *label_ptr[2];
-#else
+ tcg_insn_unit *label_ptr[2] = { };
+#ifndef CONFIG_SOFTMMU
unsigned a_bits;
#endif
@@ -2287,26 +2239,27 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args, bool is64)
opc = get_memop(oi);
#if defined(CONFIG_SOFTMMU)
- mem_index = get_mmuidx(oi);
-
- tcg_out_tlb_load(s, addrlo, addrhi, mem_index, opc,
+ tcg_out_tlb_load(s, addrlo, addrhi, get_mmuidx(oi), opc,
label_ptr, offsetof(CPUTLBEntry, addr_read));
/* TLB Hit. */
tcg_out_qemu_ld_direct(s, datalo, datahi, TCG_REG_L1, -1, 0, 0, is64, opc);
/* Record the current context of a load into ldst label */
- add_qemu_ldst_label(s, true, is64, oi, datalo, datahi, addrlo, addrhi,
- s->code_ptr, label_ptr);
+ add_qemu_ldst_label(s, true, is64, oi, datalo, datahi,
+ TCG_REG_L1, addrhi, s->code_ptr, label_ptr);
#else
a_bits = get_alignment_bits(opc);
if (a_bits) {
- tcg_out_test_alignment(s, true, addrlo, addrhi, a_bits);
+ tcg_out_test_alignment(s, addrlo, a_bits, label_ptr);
}
-
tcg_out_qemu_ld_direct(s, datalo, datahi, addrlo, x86_guest_base_index,
x86_guest_base_offset, x86_guest_base_seg,
is64, opc);
+ if (a_bits) {
+ add_qemu_ldst_label(s, true, is64, oi, datalo, datahi,
+ addrlo, addrhi, s->code_ptr, label_ptr);
+ }
#endif
}
@@ -2368,10 +2321,8 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, bool is64)
TCGReg addrhi __attribute__((unused));
MemOpIdx oi;
MemOp opc;
-#if defined(CONFIG_SOFTMMU)
- int mem_index;
- tcg_insn_unit *label_ptr[2];
-#else
+ tcg_insn_unit *label_ptr[2] = { };
+#ifndef CONFIG_SOFTMMU
unsigned a_bits;
#endif
@@ -2383,25 +2334,26 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, bool is64)
opc = get_memop(oi);
#if defined(CONFIG_SOFTMMU)
- mem_index = get_mmuidx(oi);
-
- tcg_out_tlb_load(s, addrlo, addrhi, mem_index, opc,
+ tcg_out_tlb_load(s, addrlo, addrhi, get_mmuidx(oi), opc,
label_ptr, offsetof(CPUTLBEntry, addr_write));
/* TLB Hit. */
tcg_out_qemu_st_direct(s, datalo, datahi, TCG_REG_L1, -1, 0, 0, opc);
/* Record the current context of a store into ldst label */
- add_qemu_ldst_label(s, false, is64, oi, datalo, datahi, addrlo, addrhi,
- s->code_ptr, label_ptr);
+ add_qemu_ldst_label(s, false, is64, oi, datalo, datahi,
+ TCG_REG_L1, addrhi, s->code_ptr, label_ptr);
#else
a_bits = get_alignment_bits(opc);
if (a_bits) {
- tcg_out_test_alignment(s, false, addrlo, addrhi, a_bits);
+ tcg_out_test_alignment(s, addrlo, a_bits, label_ptr);
}
-
tcg_out_qemu_st_direct(s, datalo, datahi, addrlo, x86_guest_base_index,
x86_guest_base_offset, x86_guest_base_seg, opc);
+ if (a_bits) {
+ add_qemu_ldst_label(s, false, is64, oi, datalo, datahi,
+ addrlo, addrhi, s->code_ptr, label_ptr);
+ }
#endif
}
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* [PATCH for-8.0 24/29] tcg/i386: Replace is64 with type in qemu_ld/st routines
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (22 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 23/29] tcg/i386: Use full load/store helpers in user-only mode Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-21 16:27 ` Philippe Mathieu-Daudé
2022-11-18 9:47 ` [PATCH for-8.0 25/29] tcg/i386: Mark Win64 call-saved vector regs as reserved Richard Henderson
` (4 subsequent siblings)
28 siblings, 1 reply; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
Prepare for TCG_TYPE_I128 by not using a boolean.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
tcg/i386/tcg-target.c.inc | 54 ++++++++++++++++++++++++++-------------
1 file changed, 36 insertions(+), 18 deletions(-)
diff --git a/tcg/i386/tcg-target.c.inc b/tcg/i386/tcg-target.c.inc
index eb93807b5f..e38f08bd12 100644
--- a/tcg/i386/tcg-target.c.inc
+++ b/tcg/i386/tcg-target.c.inc
@@ -1772,7 +1772,7 @@ static void * const qemu_st_helpers[MO_SIZE + 1] = {
* Record the context of a call to the out of line helper code for the slow path
* for a load or store, so that we can later generate the correct helper code
*/
-static void add_qemu_ldst_label(TCGContext *s, bool is_ld, bool is_64,
+static void add_qemu_ldst_label(TCGContext *s, bool is_ld, TCGType type,
MemOpIdx oi,
TCGReg datalo, TCGReg datahi,
TCGReg addrlo, TCGReg addrhi,
@@ -1783,7 +1783,7 @@ static void add_qemu_ldst_label(TCGContext *s, bool is_ld, bool is_64,
label->is_ld = is_ld;
label->oi = oi;
- label->type = is_64 ? TCG_TYPE_I64 : TCG_TYPE_I32;
+ label->type = type;
label->datalo_reg = datalo;
label->datahi_reg = datahi;
label->addrlo_reg = addrlo;
@@ -2124,10 +2124,10 @@ static inline int setup_guest_base_seg(void)
static void tcg_out_qemu_ld_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
TCGReg base, int index, intptr_t ofs,
- int seg, bool is64, MemOp memop)
+ int seg, TCGType type, MemOp memop)
{
bool use_movbe = false;
- int rexw = is64 * P_REXW;
+ int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
int movop = OPC_MOVL_GvEv;
/* Do big-endian loads with movbe. */
@@ -2220,7 +2220,7 @@ static void tcg_out_qemu_ld_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
/* XXX: qemu_ld and qemu_st could be modified to clobber only EDX and
EAX. It will be useful once fixed registers globals are less
common. */
-static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args, bool is64)
+static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args, TCGType type)
{
TCGReg datalo, datahi, addrlo;
TCGReg addrhi __attribute__((unused));
@@ -2232,7 +2232,16 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args, bool is64)
#endif
datalo = *args++;
- datahi = (TCG_TARGET_REG_BITS == 32 && is64 ? *args++ : 0);
+ switch (type) {
+ case TCG_TYPE_I32:
+ datahi = 0;
+ break;
+ case TCG_TYPE_I64:
+ datahi = (TCG_TARGET_REG_BITS == 32 ? *args++ : 0);
+ break;
+ default:
+ g_assert_not_reached();
+ }
addrlo = *args++;
addrhi = (TARGET_LONG_BITS > TCG_TARGET_REG_BITS ? *args++ : 0);
oi = *args++;
@@ -2243,10 +2252,10 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args, bool is64)
label_ptr, offsetof(CPUTLBEntry, addr_read));
/* TLB Hit. */
- tcg_out_qemu_ld_direct(s, datalo, datahi, TCG_REG_L1, -1, 0, 0, is64, opc);
+ tcg_out_qemu_ld_direct(s, datalo, datahi, TCG_REG_L1, -1, 0, 0, type, opc);
/* Record the current context of a load into ldst label */
- add_qemu_ldst_label(s, true, is64, oi, datalo, datahi,
+ add_qemu_ldst_label(s, true, type, oi, datalo, datahi,
TCG_REG_L1, addrhi, s->code_ptr, label_ptr);
#else
a_bits = get_alignment_bits(opc);
@@ -2255,9 +2264,9 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args, bool is64)
}
tcg_out_qemu_ld_direct(s, datalo, datahi, addrlo, x86_guest_base_index,
x86_guest_base_offset, x86_guest_base_seg,
- is64, opc);
+ type, opc);
if (a_bits) {
- add_qemu_ldst_label(s, true, is64, oi, datalo, datahi,
+ add_qemu_ldst_label(s, true, type, oi, datalo, datahi,
addrlo, addrhi, s->code_ptr, label_ptr);
}
#endif
@@ -2315,7 +2324,7 @@ static void tcg_out_qemu_st_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
}
}
-static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, bool is64)
+static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, TCGType type)
{
TCGReg datalo, datahi, addrlo;
TCGReg addrhi __attribute__((unused));
@@ -2327,7 +2336,16 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, bool is64)
#endif
datalo = *args++;
- datahi = (TCG_TARGET_REG_BITS == 32 && is64 ? *args++ : 0);
+ switch (type) {
+ case TCG_TYPE_I32:
+ datahi = 0;
+ break;
+ case TCG_TYPE_I64:
+ datahi = (TCG_TARGET_REG_BITS == 32 ? *args++ : 0);
+ break;
+ default:
+ g_assert_not_reached();
+ }
addrlo = *args++;
addrhi = (TARGET_LONG_BITS > TCG_TARGET_REG_BITS ? *args++ : 0);
oi = *args++;
@@ -2341,7 +2359,7 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, bool is64)
tcg_out_qemu_st_direct(s, datalo, datahi, TCG_REG_L1, -1, 0, 0, opc);
/* Record the current context of a store into ldst label */
- add_qemu_ldst_label(s, false, is64, oi, datalo, datahi,
+ add_qemu_ldst_label(s, false, type, oi, datalo, datahi,
TCG_REG_L1, addrhi, s->code_ptr, label_ptr);
#else
a_bits = get_alignment_bits(opc);
@@ -2351,7 +2369,7 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, bool is64)
tcg_out_qemu_st_direct(s, datalo, datahi, addrlo, x86_guest_base_index,
x86_guest_base_offset, x86_guest_base_seg, opc);
if (a_bits) {
- add_qemu_ldst_label(s, false, is64, oi, datalo, datahi,
+ add_qemu_ldst_label(s, false, type, oi, datalo, datahi,
addrlo, addrhi, s->code_ptr, label_ptr);
}
#endif
@@ -2649,17 +2667,17 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
break;
case INDEX_op_qemu_ld_i32:
- tcg_out_qemu_ld(s, args, 0);
+ tcg_out_qemu_ld(s, args, TCG_TYPE_I32);
break;
case INDEX_op_qemu_ld_i64:
- tcg_out_qemu_ld(s, args, 1);
+ tcg_out_qemu_ld(s, args, TCG_TYPE_I64);
break;
case INDEX_op_qemu_st_i32:
case INDEX_op_qemu_st8_i32:
- tcg_out_qemu_st(s, args, 0);
+ tcg_out_qemu_st(s, args, TCG_TYPE_I32);
break;
case INDEX_op_qemu_st_i64:
- tcg_out_qemu_st(s, args, 1);
+ tcg_out_qemu_st(s, args, TCG_TYPE_I64);
break;
OP_32_64(mulu2):
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* Re: [PATCH for-8.0 24/29] tcg/i386: Replace is64 with type in qemu_ld/st routines
2022-11-18 9:47 ` [PATCH for-8.0 24/29] tcg/i386: Replace is64 with type in qemu_ld/st routines Richard Henderson
@ 2022-11-21 16:27 ` Philippe Mathieu-Daudé
0 siblings, 0 replies; 48+ messages in thread
From: Philippe Mathieu-Daudé @ 2022-11-21 16:27 UTC (permalink / raw)
To: Richard Henderson, qemu-devel
On 18/11/22 10:47, Richard Henderson wrote:
> Prepare for TCG_TYPE_I128 by not using a boolean.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
> tcg/i386/tcg-target.c.inc | 54 ++++++++++++++++++++++++++-------------
> 1 file changed, 36 insertions(+), 18 deletions(-)
> @@ -2315,7 +2324,7 @@ static void tcg_out_qemu_st_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
> }
> }
>
> -static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, bool is64)
> +static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, TCGType type)
> {
> TCGReg datalo, datahi, addrlo;
> TCGReg addrhi __attribute__((unused));
> @@ -2327,7 +2336,16 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, bool is64)
Confusing git-diff context :)
> #endif
>
> datalo = *args++;
> - datahi = (TCG_TARGET_REG_BITS == 32 && is64 ? *args++ : 0);
> + switch (type) {
> + case TCG_TYPE_I32:
> + datahi = 0;
> + break;
> + case TCG_TYPE_I64:
> + datahi = (TCG_TARGET_REG_BITS == 32 ? *args++ : 0);
> + break;
> + default:
> + g_assert_not_reached();
> + }
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
^ permalink raw reply [flat|nested] 48+ messages in thread
* [PATCH for-8.0 25/29] tcg/i386: Mark Win64 call-saved vector regs as reserved
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (23 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 24/29] tcg/i386: Replace is64 with type in qemu_ld/st routines Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-21 16:28 ` Philippe Mathieu-Daudé
2022-11-18 9:47 ` [PATCH for-8.0 26/29] tcg/i386: Examine MemOp for atomicity and alignment Richard Henderson
` (3 subsequent siblings)
28 siblings, 1 reply; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
While we do not include these in tcg_target_reg_alloc_order,
and therefore they ought never be allocated, it seems safer
to mark them reserved as well.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
tcg/i386/tcg-target.c.inc | 13 +++++++++++++
1 file changed, 13 insertions(+)
diff --git a/tcg/i386/tcg-target.c.inc b/tcg/i386/tcg-target.c.inc
index e38f08bd12..e04818eef6 100644
--- a/tcg/i386/tcg-target.c.inc
+++ b/tcg/i386/tcg-target.c.inc
@@ -4224,6 +4224,19 @@ static void tcg_target_init(TCGContext *s)
s->reserved_regs = 0;
tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
+#ifdef _WIN64
+ /* These are call saved, and not we don't save them, so don't use them. */
+ tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM6);
+ tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM7);
+ tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM8);
+ tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM9);
+ tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM10);
+ tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM11);
+ tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM12);
+ tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM13);
+ tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM14);
+ tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM15);
+#endif
}
typedef struct {
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* Re: [PATCH for-8.0 25/29] tcg/i386: Mark Win64 call-saved vector regs as reserved
2022-11-18 9:47 ` [PATCH for-8.0 25/29] tcg/i386: Mark Win64 call-saved vector regs as reserved Richard Henderson
@ 2022-11-21 16:28 ` Philippe Mathieu-Daudé
0 siblings, 0 replies; 48+ messages in thread
From: Philippe Mathieu-Daudé @ 2022-11-21 16:28 UTC (permalink / raw)
To: Richard Henderson, qemu-devel
On 18/11/22 10:47, Richard Henderson wrote:
> While we do not include these in tcg_target_reg_alloc_order,
> and therefore they ought never be allocated, it seems safer
> to mark them reserved as well.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
> tcg/i386/tcg-target.c.inc | 13 +++++++++++++
> 1 file changed, 13 insertions(+)
>
> diff --git a/tcg/i386/tcg-target.c.inc b/tcg/i386/tcg-target.c.inc
> index e38f08bd12..e04818eef6 100644
> --- a/tcg/i386/tcg-target.c.inc
> +++ b/tcg/i386/tcg-target.c.inc
> @@ -4224,6 +4224,19 @@ static void tcg_target_init(TCGContext *s)
>
> s->reserved_regs = 0;
> tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
> +#ifdef _WIN64
> + /* These are call saved, and not we don't save them, so don't use them. */
s/not//?
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
> + tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM6);
> + tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM7);
> + tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM8);
> + tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM9);
> + tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM10);
> + tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM11);
> + tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM12);
> + tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM13);
> + tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM14);
> + tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM15);
> +#endif
> }
>
> typedef struct {
^ permalink raw reply [flat|nested] 48+ messages in thread
* [PATCH for-8.0 26/29] tcg/i386: Examine MemOp for atomicity and alignment
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (24 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 25/29] tcg/i386: Mark Win64 call-saved vector regs as reserved Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-18 9:47 ` [PATCH for-8.0 27/29] tcg/i386: Support 128-bit load/store with have_atomic16 Richard Henderson
` (2 subsequent siblings)
28 siblings, 0 replies; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
No change to the ultimate load/store routines yet, so some
atomicity conditions not yet honored, but plumbs the change
to alignment through the adjacent functions.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
tcg/i386/tcg-target.c.inc | 128 ++++++++++++++++++++++++++++++--------
1 file changed, 101 insertions(+), 27 deletions(-)
diff --git a/tcg/i386/tcg-target.c.inc b/tcg/i386/tcg-target.c.inc
index e04818eef6..7dc56040d2 100644
--- a/tcg/i386/tcg-target.c.inc
+++ b/tcg/i386/tcg-target.c.inc
@@ -1746,6 +1746,83 @@ tcg_out_testi(TCGContext *s, TCGReg r, uint32_t i)
}
}
+/*
+ * Return the alignment and atomicity to use for the inline fast path
+ * for the given memory operation. The alignment may be larger than
+ * that specified in @opc, and the correct alignment will be diagnosed
+ * by the slow path helper.
+ */
+static MemOp atom_and_align_for_opc(TCGContext *s, MemOp opc, MemOp *out_al)
+{
+ MemOp align = get_alignment_bits(opc);
+ MemOp atom, atmax, atsub, size = opc & MO_SIZE;
+
+ /* When serialized, no further atomicity required. */
+ if (s->tb_cflags & CF_PARALLEL) {
+ atom = opc & MO_ATOM_MASK;
+ } else {
+ atom = MO_ATOM_NONE;
+ }
+
+ atmax = opc & MO_ATMAX_MASK;
+ if (atmax == MO_ATMAX_SIZE) {
+ atmax = size;
+ } else {
+ atmax = atmax >> MO_ATMAX_SHIFT;
+ }
+
+ switch (atom) {
+ case MO_ATOM_NONE:
+ /* The operation requires no specific atomicity. */
+ atmax = MO_8;
+ atsub = MO_8;
+ break;
+ case MO_ATOM_IFALIGN:
+ /* If unaligned, the subobjects are bytes. */
+ atsub = MO_8;
+ break;
+ case MO_ATOM_WITHIN16:
+ /* If unaligned, there are subobjects if atmax < size. */
+ atsub = (atmax < size ? atmax : MO_8);
+ atmax = size;
+ break;
+ case MO_ATOM_SUBALIGN:
+ /* If unaligned but not odd, there are subobjects up to atmax - 1. */
+ atsub = (atmax == MO_8 ? MO_8 : atmax - 1);
+ break;
+ default:
+ g_assert_not_reached();
+ }
+
+ /*
+ * Per Intel Architecture SDM, Volume 3 Section 8.1.1,
+ * - Pentium family guarantees atomicity of aligned <= 64-bit.
+ * - P6 family guarantees atomicity of unaligned <= 64-bit
+ * which fit within a cache line.
+ * - AVX guarantees atomicity of aligned 128-bit VMOVDQA (et al).
+ *
+ * There is no language in the Intel manual specifying what happens
+ * with the partial memory operations when crossing a cache line.
+ * When there is required atomicity of subobjects, we must perform
+ * an additional runtime test for alignment and then perform either
+ * the full operation, or two half-sized operations.
+ *
+ * For x86_64, and MO_64, we do not have a scratch register with
+ * which to do this. Only allow splitting for MO_64 on i386,
+ * where the data is already separated, or MO_128.
+ * Otherwise, require full alignment and fall back to the helper
+ * for the misaligned case.
+ */
+ if (align < atmax
+ && atsub != MO_8
+ && size != (TCG_TARGET_REG_BITS == 64 ? MO_128 : MO_64)) {
+ align = size;
+ }
+
+ *out_al = align;
+ return atmax;
+}
+
/*
* helper signature: helper_ld*_mmu(CPUState *env, target_ulong addr,
* int mmu_idx, uintptr_t ra)
@@ -1987,7 +2064,7 @@ static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
* First argument register is clobbered.
*/
static void tcg_out_tlb_load(TCGContext *s, TCGReg addrlo, TCGReg addrhi,
- int mem_index, MemOp opc,
+ int mem_index, MemOp a_bits, MemOp s_bits,
tcg_insn_unit **label_ptr, int which)
{
const TCGReg r0 = TCG_REG_L0;
@@ -1995,8 +2072,6 @@ static void tcg_out_tlb_load(TCGContext *s, TCGReg addrlo, TCGReg addrhi,
TCGType ttype = TCG_TYPE_I32;
TCGType tlbtype = TCG_TYPE_I32;
int trexw = 0, hrexw = 0, tlbrexw = 0;
- unsigned a_bits = get_alignment_bits(opc);
- unsigned s_bits = opc & MO_SIZE;
unsigned a_mask = (1 << a_bits) - 1;
unsigned s_mask = (1 << s_bits) - 1;
target_ulong tlb_mask;
@@ -2124,7 +2199,8 @@ static inline int setup_guest_base_seg(void)
static void tcg_out_qemu_ld_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
TCGReg base, int index, intptr_t ofs,
- int seg, TCGType type, MemOp memop)
+ int seg, TCGType type, MemOp memop,
+ MemOp atom, MemOp align)
{
bool use_movbe = false;
int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
@@ -2225,11 +2301,8 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args, TCGType type)
TCGReg datalo, datahi, addrlo;
TCGReg addrhi __attribute__((unused));
MemOpIdx oi;
- MemOp opc;
+ MemOp opc, atom, align;
tcg_insn_unit *label_ptr[2] = { };
-#ifndef CONFIG_SOFTMMU
- unsigned a_bits;
-#endif
datalo = *args++;
switch (type) {
@@ -2246,26 +2319,27 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args, TCGType type)
addrhi = (TARGET_LONG_BITS > TCG_TARGET_REG_BITS ? *args++ : 0);
oi = *args++;
opc = get_memop(oi);
+ atom = atom_and_align_for_opc(s, opc, &align);
#if defined(CONFIG_SOFTMMU)
- tcg_out_tlb_load(s, addrlo, addrhi, get_mmuidx(oi), opc,
+ tcg_out_tlb_load(s, addrlo, addrhi, get_mmuidx(oi), align, opc & MO_SIZE,
label_ptr, offsetof(CPUTLBEntry, addr_read));
/* TLB Hit. */
- tcg_out_qemu_ld_direct(s, datalo, datahi, TCG_REG_L1, -1, 0, 0, type, opc);
+ tcg_out_qemu_ld_direct(s, datalo, datahi, TCG_REG_L1, -1, 0, 0, type,
+ opc, atom, align);
/* Record the current context of a load into ldst label */
add_qemu_ldst_label(s, true, type, oi, datalo, datahi,
TCG_REG_L1, addrhi, s->code_ptr, label_ptr);
#else
- a_bits = get_alignment_bits(opc);
- if (a_bits) {
- tcg_out_test_alignment(s, addrlo, a_bits, label_ptr);
+ if (align) {
+ tcg_out_test_alignment(s, addrlo, align, label_ptr);
}
tcg_out_qemu_ld_direct(s, datalo, datahi, addrlo, x86_guest_base_index,
x86_guest_base_offset, x86_guest_base_seg,
- type, opc);
- if (a_bits) {
+ type, opc, atom, align);
+ if (align) {
add_qemu_ldst_label(s, true, type, oi, datalo, datahi,
addrlo, addrhi, s->code_ptr, label_ptr);
}
@@ -2274,7 +2348,8 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args, TCGType type)
static void tcg_out_qemu_st_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
TCGReg base, int index, intptr_t ofs,
- int seg, MemOp memop)
+ int seg, MemOp memop,
+ MemOp atom, MemOp align)
{
bool use_movbe = false;
int movop = OPC_MOVL_EvGv;
@@ -2329,11 +2404,8 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, TCGType type)
TCGReg datalo, datahi, addrlo;
TCGReg addrhi __attribute__((unused));
MemOpIdx oi;
- MemOp opc;
+ MemOp opc, atom, align;
tcg_insn_unit *label_ptr[2] = { };
-#ifndef CONFIG_SOFTMMU
- unsigned a_bits;
-#endif
datalo = *args++;
switch (type) {
@@ -2350,25 +2422,27 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, TCGType type)
addrhi = (TARGET_LONG_BITS > TCG_TARGET_REG_BITS ? *args++ : 0);
oi = *args++;
opc = get_memop(oi);
+ atom = atom_and_align_for_opc(s, opc, &align);
#if defined(CONFIG_SOFTMMU)
- tcg_out_tlb_load(s, addrlo, addrhi, get_mmuidx(oi), opc,
+ tcg_out_tlb_load(s, addrlo, addrhi, get_mmuidx(oi), align, opc & MO_SIZE,
label_ptr, offsetof(CPUTLBEntry, addr_write));
/* TLB Hit. */
- tcg_out_qemu_st_direct(s, datalo, datahi, TCG_REG_L1, -1, 0, 0, opc);
+ tcg_out_qemu_st_direct(s, datalo, datahi, TCG_REG_L1, -1, 0, 0,
+ opc, atom, align);
/* Record the current context of a store into ldst label */
add_qemu_ldst_label(s, false, type, oi, datalo, datahi,
TCG_REG_L1, addrhi, s->code_ptr, label_ptr);
#else
- a_bits = get_alignment_bits(opc);
- if (a_bits) {
- tcg_out_test_alignment(s, addrlo, a_bits, label_ptr);
+ if (align) {
+ tcg_out_test_alignment(s, addrlo, align, label_ptr);
}
tcg_out_qemu_st_direct(s, datalo, datahi, addrlo, x86_guest_base_index,
- x86_guest_base_offset, x86_guest_base_seg, opc);
- if (a_bits) {
+ x86_guest_base_offset, x86_guest_base_seg,
+ opc, atom, align);
+ if (align) {
add_qemu_ldst_label(s, false, type, oi, datalo, datahi,
addrlo, addrhi, s->code_ptr, label_ptr);
}
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* [PATCH for-8.0 27/29] tcg/i386: Support 128-bit load/store with have_atomic16
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (25 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 26/29] tcg/i386: Examine MemOp for atomicity and alignment Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-18 9:47 ` [PATCH for-8.0 28/29] tcg/i386: Add vex_v argument to tcg_out_vex_modrm_pool Richard Henderson
2022-11-18 9:47 ` [PATCH for-8.0 29/29] tcg/i386: Honor 64-bit atomicity in 32-bit mode Richard Henderson
28 siblings, 0 replies; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
tcg/i386/tcg-target.h | 3 +-
tcg/i386/tcg-target.c.inc | 325 +++++++++++++++++++++++++++++++++++---
2 files changed, 304 insertions(+), 24 deletions(-)
diff --git a/tcg/i386/tcg-target.h b/tcg/i386/tcg-target.h
index 53d2cb3412..7aafd60d72 100644
--- a/tcg/i386/tcg-target.h
+++ b/tcg/i386/tcg-target.h
@@ -195,7 +195,8 @@ extern bool have_atomic16;
#define TCG_TARGET_HAS_qemu_st8_i32 1
#endif
-#define TCG_TARGET_HAS_qemu_ldst_i128 0
+#define TCG_TARGET_HAS_qemu_ldst_i128 \
+ (TCG_TARGET_REG_BITS == 64 && have_atomic16)
/* We do not support older SSE systems, only beginning with AVX1. */
#define TCG_TARGET_HAS_v64 have_avx1
diff --git a/tcg/i386/tcg-target.c.inc b/tcg/i386/tcg-target.c.inc
index 7dc56040d2..f277085321 100644
--- a/tcg/i386/tcg-target.c.inc
+++ b/tcg/i386/tcg-target.c.inc
@@ -91,6 +91,8 @@ static const int tcg_target_reg_alloc_order[] = {
#endif
};
+#define TCG_TMP_VEC TCG_REG_XMM5
+
static const int tcg_target_call_iarg_regs[] = {
#if TCG_TARGET_REG_BITS == 64
#if defined(_WIN64)
@@ -347,6 +349,8 @@ static bool tcg_target_const_match(int64_t val, TCGType type, int ct)
#define OPC_PCMPGTW (0x65 | P_EXT | P_DATA16)
#define OPC_PCMPGTD (0x66 | P_EXT | P_DATA16)
#define OPC_PCMPGTQ (0x37 | P_EXT38 | P_DATA16)
+#define OPC_PEXTRD (0x16 | P_EXT3A | P_DATA16)
+#define OPC_PINSRD (0x22 | P_EXT3A | P_DATA16)
#define OPC_PMAXSB (0x3c | P_EXT38 | P_DATA16)
#define OPC_PMAXSW (0xee | P_EXT | P_DATA16)
#define OPC_PMAXSD (0x3d | P_EXT38 | P_DATA16)
@@ -1730,8 +1734,7 @@ static void tcg_out_nopn(TCGContext *s, int n)
}
/* Test register R vs immediate bits I, setting Z flag for EQ/NE. */
-static void __attribute__((unused))
-tcg_out_testi(TCGContext *s, TCGReg r, uint32_t i)
+static void tcg_out_testi(TCGContext *s, TCGReg r, uint32_t i)
{
/*
* This is used for testing alignment, so we can usually use testb.
@@ -1828,10 +1831,11 @@ static MemOp atom_and_align_for_opc(TCGContext *s, MemOp opc, MemOp *out_al)
* int mmu_idx, uintptr_t ra)
*/
static void * const qemu_ld_helpers[MO_SIZE + 1] = {
- [MO_UB] = helper_ldub_mmu,
- [MO_UW] = helper_lduw_mmu,
- [MO_UL] = helper_ldul_mmu,
- [MO_UQ] = helper_ldq_mmu,
+ [MO_8] = helper_ldub_mmu,
+ [MO_16] = helper_lduw_mmu,
+ [MO_32] = helper_ldul_mmu,
+ [MO_64] = helper_ldq_mmu,
+ [MO_128] = helper_ld16_mmu,
};
/*
@@ -1839,10 +1843,11 @@ static void * const qemu_ld_helpers[MO_SIZE + 1] = {
* uintxx_t val, int mmu_idx, uintptr_t ra)
*/
static void * const qemu_st_helpers[MO_SIZE + 1] = {
- [MO_UB] = helper_stb_mmu,
- [MO_UW] = helper_stw_mmu,
- [MO_UL] = helper_stl_mmu,
- [MO_UQ] = helper_stq_mmu,
+ [MO_8] = helper_stb_mmu,
+ [MO_16] = helper_stw_mmu,
+ [MO_32] = helper_stl_mmu,
+ [MO_64] = helper_stq_mmu,
+ [MO_128] = helper_st16_mmu,
};
/*
@@ -1870,6 +1875,13 @@ static void add_qemu_ldst_label(TCGContext *s, bool is_ld, TCGType type,
label->label_ptr[1] = label_ptr[1];
}
+static void tcg_out_mov2_xchg(TCGContext *s, TCGType type1, TCGType type2,
+ TCGReg dst1, TCGReg dst2)
+{
+ int w = (type1 == TCG_TYPE_I32 && type2 == TCG_TYPE_I32 ? 0 : P_REXW);
+ tcg_out_modrm(s, OPC_XCHG_EvGv + w, dst1, dst2);
+}
+
/* Move src1 to dst1 and src2 to dst2, minding possible overlap. */
static void tcg_out_mov2(TCGContext *s,
TCGType type1, TCGReg dst1, TCGReg src1,
@@ -1883,11 +1895,69 @@ static void tcg_out_mov2(TCGContext *s,
tcg_out_mov(s, type1, dst1, src1);
} else {
/* dst1 == src2 && dst2 == src1 -> xchg. */
- int w = (type1 == TCG_TYPE_I32 && type2 == TCG_TYPE_I32 ? 0 : P_REXW);
- tcg_out_modrm(s, OPC_XCHG_EvGv + w, dst1, dst2);
+ tcg_out_mov2_xchg(s, type1, type2, dst1, dst2);
}
}
+/* Similarly for 3 pairs. */
+static void tcg_out_mov3(TCGContext *s,
+ TCGType type1, TCGReg dst1, TCGReg src1,
+ TCGType type2, TCGReg dst2, TCGReg src2,
+ TCGType type3, TCGReg dst3, TCGReg src3)
+{
+ if (dst1 != src2 && dst1 != src3) {
+ tcg_out_mov(s, type1, dst1, src1);
+ tcg_out_mov2(s, type2, dst2, src2, type3, dst3, src3);
+ return;
+ }
+ if (dst2 != src2 && dst2 != src3) {
+ tcg_out_mov(s, type2, dst2, src2);
+ tcg_out_mov2(s, type1, dst1, src1, type3, dst3, src3);
+ return;
+ }
+ if (dst3 != src1 && dst3 != src2) {
+ tcg_out_mov(s, type3, dst3, src3);
+ tcg_out_mov2(s, type1, dst1, src1, type2, dst2, src2);
+ return;
+ }
+ /* Three-way overlap present, at least one xchg needed. */
+ if (dst1 == src2) {
+ tcg_out_mov2_xchg(s, type1, type2, src1, src2);
+ tcg_out_mov2(s, type2, dst2, src1, type3, dst3, src3);
+ return;
+ }
+ if (dst1 == src3) {
+ tcg_out_mov2_xchg(s, type1, type3, src1, src3);
+ tcg_out_mov2(s, type2, dst2, src2, type3, dst3, src1);
+ return;
+ }
+ g_assert_not_reached();
+}
+
+static void tcg_out_vec_to_pair(TCGContext *s, TCGType type,
+ TCGReg l, TCGReg h, TCGReg v)
+{
+ int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
+
+ /* vpmov{d,q} %v, %l */
+ tcg_out_vex_modrm(s, OPC_MOVD_EyVy + rexw, v, 0, l);
+ /* vpextr{d,q} $1, %v, %h */
+ tcg_out_vex_modrm(s, OPC_PEXTRD + rexw, v, 0, h);
+ tcg_out8(s, 1);
+}
+
+static void tcg_out_pair_to_vec(TCGContext *s, TCGType type,
+ TCGReg v, TCGReg l, TCGReg h)
+{
+ int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
+
+ /* vmov{d,q} %l, %v */
+ tcg_out_vex_modrm(s, OPC_MOVD_VyEy + rexw, v, 0, l);
+ /* vpinsr{d,q} $1, %h, %v, %v */
+ tcg_out_vex_modrm(s, OPC_PINSRD + rexw, v, v, h);
+ tcg_out8(s, 1);
+}
+
/*
* Generate code for the slow path for a load at the end of block
*/
@@ -1897,7 +1967,7 @@ static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
MemOp opc = get_memop(oi);
TCGReg data_reg;
tcg_insn_unit **label_ptr = &l->label_ptr[0];
- int rexw = (l->type == TCG_TYPE_I64 ? P_REXW : 0);
+ int rexw = (l->type == TCG_TYPE_I32 ? 0 : P_REXW);
/* resolve label address */
tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
@@ -1961,6 +2031,22 @@ static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
TCG_TYPE_I32, l->datahi_reg, TCG_REG_EDX);
}
break;
+ case MO_128:
+ tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
+ switch (TCG_TARGET_CALL_RET_I128) {
+ case TCG_CALL_RET_NORMAL:
+ tcg_out_mov2(s, TCG_TYPE_I64, data_reg, TCG_REG_RAX,
+ TCG_TYPE_I64, l->datahi_reg, TCG_REG_RDX);
+ break;
+ case TCG_CALL_RET_BY_VEC:
+ tcg_out_vec_to_pair(s, TCG_TYPE_I64,
+ data_reg, l->datahi_reg, TCG_REG_XMM0);
+ break;
+ default:
+ qemu_build_not_reached();
+ }
+ break;
+
default:
tcg_abort();
}
@@ -1977,7 +2063,6 @@ static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
{
MemOpIdx oi = l->oi;
MemOp opc = get_memop(oi);
- MemOp s_bits = opc & MO_SIZE;
tcg_insn_unit **label_ptr = &l->label_ptr[0];
TCGReg retaddr;
@@ -2004,9 +2089,15 @@ static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
tcg_out_st(s, TCG_TYPE_I32, l->datalo_reg, TCG_REG_ESP, ofs);
ofs += 4;
- if (s_bits == MO_64) {
+ switch (l->type) {
+ case TCG_TYPE_I32:
+ break;
+ case TCG_TYPE_I64:
tcg_out_st(s, TCG_TYPE_I32, l->datahi_reg, TCG_REG_ESP, ofs);
ofs += 4;
+ break;
+ default:
+ g_assert_not_reached();
}
tcg_out_sti(s, TCG_TYPE_I32, oi, TCG_REG_ESP, ofs);
@@ -2016,15 +2107,54 @@ static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
tcg_out_movi(s, TCG_TYPE_PTR, retaddr, (uintptr_t)l->raddr);
tcg_out_st(s, TCG_TYPE_PTR, retaddr, TCG_REG_ESP, ofs);
} else {
- tcg_out_mov2(s, TCG_TYPE_TL,
- tcg_target_call_iarg_regs[1], l->addrlo_reg,
- s_bits == MO_64 ? TCG_TYPE_I64 : TCG_TYPE_I32,
- tcg_target_call_iarg_regs[2], l->datalo_reg);
- tcg_out_mov(s, TCG_TYPE_PTR, tcg_target_call_iarg_regs[0], TCG_AREG0);
- tcg_out_movi(s, TCG_TYPE_I32, tcg_target_call_iarg_regs[3], oi);
+ int slot;
- if (ARRAY_SIZE(tcg_target_call_iarg_regs) > 4) {
- retaddr = tcg_target_call_iarg_regs[4];
+ switch (l->type) {
+ case TCG_TYPE_I32:
+ case TCG_TYPE_I64:
+ tcg_out_mov2(s, TCG_TYPE_TL,
+ tcg_target_call_iarg_regs[1], l->addrlo_reg,
+ l->type, tcg_target_call_iarg_regs[2], l->datalo_reg);
+ slot = 3;
+ break;
+ case TCG_TYPE_I128:
+ switch (TCG_TARGET_CALL_ARG_I128) {
+ case TCG_CALL_ARG_NORMAL:
+ tcg_out_mov3(s, TCG_TYPE_TL,
+ tcg_target_call_iarg_regs[1], l->addrlo_reg,
+ TCG_TYPE_I64,
+ tcg_target_call_iarg_regs[2], l->datalo_reg,
+ TCG_TYPE_I64,
+ tcg_target_call_iarg_regs[3], l->datahi_reg);
+ slot = 4;
+ break;
+ case TCG_CALL_ARG_BY_REF:
+ /* Leave room for retaddr below, take next 16 aligned bytes. */
+ tcg_out_st(s, TCG_TYPE_I64, l->datalo_reg,
+ TCG_REG_ESP, TCG_TARGET_CALL_STACK_OFFSET + 16);
+ tcg_out_st(s, TCG_TYPE_I64, l->datahi_reg,
+ TCG_REG_ESP, TCG_TARGET_CALL_STACK_OFFSET + 24);
+ tcg_out_mov(s, TCG_TYPE_TL,
+ tcg_target_call_iarg_regs[1], l->addrlo_reg);
+ tcg_out_modrm_offset(s, OPC_LEA + P_REXW,
+ tcg_target_call_iarg_regs[2], TCG_REG_ESP,
+ TCG_TARGET_CALL_STACK_OFFSET + 16);
+ slot = 3;
+ break;
+ default:
+ qemu_build_not_reached();
+ }
+ break;
+ default:
+ g_assert_not_reached();
+ }
+
+ tcg_debug_assert(slot < (int)ARRAY_SIZE(tcg_target_call_iarg_regs) - 1);
+ tcg_out_mov(s, TCG_TYPE_PTR, tcg_target_call_iarg_regs[0], TCG_AREG0);
+ tcg_out_movi(s, TCG_TYPE_I32, tcg_target_call_iarg_regs[slot++], oi);
+
+ if (slot < (int)ARRAY_SIZE(tcg_target_call_iarg_regs)) {
+ retaddr = tcg_target_call_iarg_regs[slot];
tcg_out_movi(s, TCG_TYPE_PTR, retaddr, (uintptr_t)l->raddr);
} else {
retaddr = TCG_REG_RAX;
@@ -2288,6 +2418,71 @@ static void tcg_out_qemu_ld_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
}
}
break;
+
+ case MO_128:
+ {
+ TCGLabel *l1 = NULL, *l2 = NULL;
+ bool use_pair = atom < MO_128;
+
+ tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
+
+ if (use_movbe) {
+ TCGReg t = datalo;
+ datalo = datahi;
+ datahi = t;
+ }
+ if (!use_pair) {
+ /*
+ * Atomicity requires that we use use VMOVDQA.
+ * If we've already checked for 16-byte alignment, that's all
+ * we need. If we arrive here with lesser alignment, then we
+ * have determined that less that 16-byte alignment can be
+ * satisfied with two 8-byte loads.
+ */
+ if (align < MO_128) {
+ use_pair = true;
+ l1 = gen_new_label();
+ l2 = gen_new_label();
+
+ tcg_out_testi(s, base, align == MO_64 ? 8 : 15);
+ tcg_out_jxx(s, JCC_JNE, l2, true);
+ }
+
+ tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_VxWx + seg,
+ TCG_TMP_VEC, 0,
+ base, index, 0, ofs);
+ tcg_out_vec_to_pair(s, TCG_TYPE_I64,
+ datalo, datahi, TCG_TMP_VEC);
+
+ if (use_movbe) {
+ tcg_out_bswap64(s, datalo);
+ tcg_out_bswap64(s, datahi);
+ }
+
+ if (use_pair) {
+ tcg_out_jxx(s, JCC_JMP, l1, true);
+ tcg_out_label(s, l2);
+ }
+ }
+ if (use_pair) {
+ if (base != datalo) {
+ tcg_out_modrm_sib_offset(s, movop + P_REXW + seg, datalo,
+ base, index, 0, ofs);
+ tcg_out_modrm_sib_offset(s, movop + P_REXW + seg, datahi,
+ base, index, 0, ofs + 8);
+ } else {
+ tcg_out_modrm_sib_offset(s, movop + P_REXW + seg, datahi,
+ base, index, 0, ofs + 8);
+ tcg_out_modrm_sib_offset(s, movop + P_REXW + seg, datalo,
+ base, index, 0, ofs);
+ }
+ }
+ if (l1) {
+ tcg_out_label(s, l1);
+ }
+ }
+ break;
+
default:
g_assert_not_reached();
}
@@ -2312,6 +2507,10 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args, TCGType type)
case TCG_TYPE_I64:
datahi = (TCG_TARGET_REG_BITS == 32 ? *args++ : 0);
break;
+ case TCG_TYPE_I128:
+ tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
+ datahi = *args++;
+ break;
default:
g_assert_not_reached();
}
@@ -2394,6 +2593,68 @@ static void tcg_out_qemu_st_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
base, index, 0, ofs + 4);
}
break;
+
+ case MO_128:
+ {
+ TCGLabel *l1 = NULL, *l2 = NULL;
+ bool use_pair = atom < MO_128;
+
+ tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
+
+ if (use_movbe) {
+ TCGReg t = datalo;
+ datalo = datahi;
+ datahi = t;
+ }
+ if (!use_pair) {
+ /*
+ * Atomicity requires that we use use VMOVDQA.
+ * If we've already checked for 16-byte alignment, that's all
+ * we need. If we arrive here with lesser alignment, then we
+ * have determined that less that 16-byte alignment can be
+ * satisfied with two 8-byte loads.
+ */
+ if (align < MO_128) {
+ use_pair = true;
+ l1 = gen_new_label();
+ l2 = gen_new_label();
+
+ tcg_out_testi(s, base, align == MO_64 ? 8 : 15);
+ tcg_out_jxx(s, JCC_JNE, l2, true);
+ }
+
+ if (use_movbe) {
+ /* Byte swap while storing to the stack. */
+ tcg_out_modrm_offset(s, movop + P_REXW + seg, datalo,
+ TCG_REG_ESP, 0);
+ tcg_out_modrm_offset(s, movop + P_REXW + seg, datahi,
+ TCG_REG_ESP, 8);
+ tcg_out_ld(s, TCG_TYPE_V128, TCG_TMP_VEC, TCG_REG_ESP, 0);
+ } else {
+ tcg_out_pair_to_vec(s, TCG_TYPE_I64,
+ TCG_TMP_VEC, datalo, datahi);
+ }
+ tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + seg,
+ TCG_TMP_VEC, 0,
+ base, index, 0, ofs);
+
+ if (use_pair) {
+ tcg_out_jxx(s, JCC_JMP, l1, true);
+ tcg_out_label(s, l2);
+ }
+ }
+ if (use_pair) {
+ tcg_out_modrm_sib_offset(s, movop + P_REXW + seg, datalo,
+ base, index, 0, ofs);
+ tcg_out_modrm_sib_offset(s, movop + P_REXW + seg, datahi,
+ base, index, 0, ofs + 8);
+ }
+ if (l1) {
+ tcg_out_label(s, l1);
+ }
+ }
+ break;
+
default:
g_assert_not_reached();
}
@@ -2415,6 +2676,10 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, TCGType type)
case TCG_TYPE_I64:
datahi = (TCG_TARGET_REG_BITS == 32 ? *args++ : 0);
break;
+ case TCG_TYPE_I128:
+ tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
+ datahi = *args++;
+ break;
default:
g_assert_not_reached();
}
@@ -2746,6 +3011,9 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
case INDEX_op_qemu_ld_i64:
tcg_out_qemu_ld(s, args, TCG_TYPE_I64);
break;
+ case INDEX_op_qemu_ld_i128:
+ tcg_out_qemu_ld(s, args, TCG_TYPE_I128);
+ break;
case INDEX_op_qemu_st_i32:
case INDEX_op_qemu_st8_i32:
tcg_out_qemu_st(s, args, TCG_TYPE_I32);
@@ -2753,6 +3021,9 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
case INDEX_op_qemu_st_i64:
tcg_out_qemu_st(s, args, TCG_TYPE_I64);
break;
+ case INDEX_op_qemu_st_i128:
+ tcg_out_qemu_st(s, args, TCG_TYPE_I128);
+ break;
OP_32_64(mulu2):
tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_MUL, args[3]);
@@ -3441,6 +3712,13 @@ static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
: TARGET_LONG_BITS <= TCG_TARGET_REG_BITS ? C_O0_I3(L, L, L)
: C_O0_I4(L, L, L, L));
+ case INDEX_op_qemu_ld_i128:
+ tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
+ return C_O2_I1(r, r, L);
+ case INDEX_op_qemu_st_i128:
+ tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
+ return C_O0_I3(L, L, L);
+
case INDEX_op_brcond2_i32:
return C_O0_I4(r, r, ri, ri);
@@ -4298,6 +4576,7 @@ static void tcg_target_init(TCGContext *s)
s->reserved_regs = 0;
tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
+ tcg_regset_set_reg(s->reserved_regs, TCG_TMP_VEC);
#ifdef _WIN64
/* These are call saved, and not we don't save them, so don't use them. */
tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM6);
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* [PATCH for-8.0 28/29] tcg/i386: Add vex_v argument to tcg_out_vex_modrm_pool
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (26 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 27/29] tcg/i386: Support 128-bit load/store with have_atomic16 Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
2022-11-18 9:47 ` [PATCH for-8.0 29/29] tcg/i386: Honor 64-bit atomicity in 32-bit mode Richard Henderson
28 siblings, 0 replies; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
tcg/i386/tcg-target.c.inc | 16 ++++++++--------
1 file changed, 8 insertions(+), 8 deletions(-)
diff --git a/tcg/i386/tcg-target.c.inc b/tcg/i386/tcg-target.c.inc
index f277085321..3f0cb4bc66 100644
--- a/tcg/i386/tcg-target.c.inc
+++ b/tcg/i386/tcg-target.c.inc
@@ -841,9 +841,9 @@ static inline void tcg_out_modrm_pool(TCGContext *s, int opc, int r)
}
/* Output an opcode with an expected reference to the constant pool. */
-static inline void tcg_out_vex_modrm_pool(TCGContext *s, int opc, int r)
+static inline void tcg_out_vex_modrm_pool(TCGContext *s, int opc, int r, int v)
{
- tcg_out_vex_opc(s, opc, r, 0, 0, 0);
+ tcg_out_vex_opc(s, opc, r, v, 0, 0);
/* Absolute for 32-bit, pc-relative for 64-bit. */
tcg_out8(s, LOWREGMASK(r) << 3 | 5);
tcg_out32(s, 0);
@@ -990,18 +990,18 @@ static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
if (TCG_TARGET_REG_BITS == 32 && vece < MO_64) {
if (have_avx2) {
- tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTD + vex_l, ret);
+ tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTD + vex_l, ret, 0);
} else {
- tcg_out_vex_modrm_pool(s, OPC_VBROADCASTSS, ret);
+ tcg_out_vex_modrm_pool(s, OPC_VBROADCASTSS, ret, 0);
}
new_pool_label(s, arg, R_386_32, s->code_ptr - 4, 0);
} else {
if (type == TCG_TYPE_V64) {
- tcg_out_vex_modrm_pool(s, OPC_MOVQ_VqWq, ret);
+ tcg_out_vex_modrm_pool(s, OPC_MOVQ_VqWq, ret, 0);
} else if (have_avx2) {
- tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTQ + vex_l, ret);
+ tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTQ + vex_l, ret, 0);
} else {
- tcg_out_vex_modrm_pool(s, OPC_MOVDDUP, ret);
+ tcg_out_vex_modrm_pool(s, OPC_MOVDDUP, ret, 0);
}
if (TCG_TARGET_REG_BITS == 64) {
new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
@@ -1024,7 +1024,7 @@ static void tcg_out_movi_vec(TCGContext *s, TCGType type,
}
int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
- tcg_out_vex_modrm_pool(s, OPC_MOVD_VyEy + rexw, ret);
+ tcg_out_vex_modrm_pool(s, OPC_MOVD_VyEy + rexw, ret, 0);
if (TCG_TARGET_REG_BITS == 64) {
new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
} else {
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread* [PATCH for-8.0 29/29] tcg/i386: Honor 64-bit atomicity in 32-bit mode
2022-11-18 9:47 [PATCH for-8.0 00/29] tcg: Improve atomicity support Richard Henderson
` (27 preceding siblings ...)
2022-11-18 9:47 ` [PATCH for-8.0 28/29] tcg/i386: Add vex_v argument to tcg_out_vex_modrm_pool Richard Henderson
@ 2022-11-18 9:47 ` Richard Henderson
28 siblings, 0 replies; 48+ messages in thread
From: Richard Henderson @ 2022-11-18 9:47 UTC (permalink / raw)
To: qemu-devel
Use one of the coprocessors to perform 64-bit stores.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
tcg/i386/tcg-target.c.inc | 119 +++++++++++++++++++++++++++++++++-----
1 file changed, 106 insertions(+), 13 deletions(-)
diff --git a/tcg/i386/tcg-target.c.inc b/tcg/i386/tcg-target.c.inc
index 3f0cb4bc66..3d3ee4b20a 100644
--- a/tcg/i386/tcg-target.c.inc
+++ b/tcg/i386/tcg-target.c.inc
@@ -472,6 +472,10 @@ static bool tcg_target_const_match(int64_t val, TCGType type, int ct)
#define OPC_GRP5 (0xff)
#define OPC_GRP14 (0x73 | P_EXT | P_DATA16)
+#define OPC_ESCDF (0xdf)
+#define ESCDF_FILD_m64 5
+#define ESCDF_FISTP_m64 7
+
/* Group 1 opcode extensions for 0x80-0x83.
These are also used as modifiers for OPC_ARITH. */
#define ARITH_ADD 0
@@ -2400,21 +2404,65 @@ static void tcg_out_qemu_ld_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
tcg_out_modrm_sib_offset(s, movop + P_REXW + seg, datalo,
base, index, 0, ofs);
} else {
+ TCGLabel *l1 = NULL, *l2 = NULL;
+ bool use_pair = atom < MO_64;
+
if (use_movbe) {
TCGReg t = datalo;
datalo = datahi;
datahi = t;
}
- if (base != datalo) {
- tcg_out_modrm_sib_offset(s, movop + seg, datalo,
- base, index, 0, ofs);
- tcg_out_modrm_sib_offset(s, movop + seg, datahi,
- base, index, 0, ofs + 4);
- } else {
- tcg_out_modrm_sib_offset(s, movop + seg, datahi,
- base, index, 0, ofs + 4);
- tcg_out_modrm_sib_offset(s, movop + seg, datalo,
+
+ if (!use_pair) {
+ /*
+ * Atomicity requires that we use use a single 8-byte load.
+ * For simplicity, and code size, always use the FPU for this.
+ * Similar insns using SSE/AVX are merely larger.
+ * Load from memory in one go, then store back to the stack,
+ * from whence we can load into the correct integer regs.
+ *
+ * If we've already checked for 8-byte alignment, or not
+ * checked for alignment at all, that's all we need.
+ * If we arrive here with lesser but non-zero alignment,
+ * then we have determined that subalignment can be
+ * satisfied with two 4-byte loads.
+ */
+ if (align > MO_8 && align < MO_64) {
+ use_pair = true;
+ l1 = gen_new_label();
+ l2 = gen_new_label();
+
+ tcg_out_testi(s, base, align == MO_32 ? 4 : 7);
+ tcg_out_jxx(s, JCC_JNE, l2, true);
+ }
+
+ tcg_out_modrm_sib_offset(s, OPC_ESCDF + seg, ESCDF_FILD_m64,
base, index, 0, ofs);
+ tcg_out_modrm_offset(s, OPC_ESCDF, ESCDF_FISTP_m64,
+ TCG_REG_ESP, 0);
+ tcg_out_modrm_offset(s, movop, datalo, TCG_REG_ESP, 0);
+ tcg_out_modrm_offset(s, movop, datahi, TCG_REG_ESP, 4);
+
+ if (use_pair) {
+ tcg_out_jxx(s, JCC_JMP, l1, true);
+ tcg_out_label(s, l2);
+ }
+ }
+ if (use_pair) {
+ if (base != datalo) {
+ tcg_out_modrm_sib_offset(s, movop + seg, datalo,
+ base, index, 0, ofs);
+ tcg_out_modrm_sib_offset(s, movop + seg, datahi,
+ base, index, 0, ofs + 4);
+ } else {
+ tcg_out_modrm_sib_offset(s, movop + seg, datahi,
+ base, index, 0, ofs + 4);
+ tcg_out_modrm_sib_offset(s, movop + seg, datalo,
+ base, index, 0, ofs);
+ }
+ }
+ if (l1) {
+ tcg_out_label(s, l1);
}
}
break;
@@ -2577,20 +2625,65 @@ static void tcg_out_qemu_st_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
case MO_32:
tcg_out_modrm_sib_offset(s, movop + seg, datalo, base, index, 0, ofs);
break;
+
case MO_64:
if (TCG_TARGET_REG_BITS == 64) {
tcg_out_modrm_sib_offset(s, movop + P_REXW + seg, datalo,
base, index, 0, ofs);
} else {
+ TCGLabel *l1 = NULL, *l2 = NULL;
+ bool use_pair = atom < MO_64;
+
if (use_movbe) {
TCGReg t = datalo;
datalo = datahi;
datahi = t;
}
- tcg_out_modrm_sib_offset(s, movop + seg, datalo,
- base, index, 0, ofs);
- tcg_out_modrm_sib_offset(s, movop + seg, datahi,
- base, index, 0, ofs + 4);
+
+ if (!use_pair) {
+ /*
+ * Atomicity requires that we use use one 8-byte store.
+ * For simplicity, and code size, always use the FPU for this.
+ * Similar insns using SSE/AVX are merely larger.
+ * Assemble the 8-byte quantity in required endianness
+ * on the stack, load to coproc unit, and store.
+ *
+ * If we've already checked for 8-byte alignment, or not
+ * checked for alignment at all, that's all we need.
+ * If we arrive here with lesser but non-zero alignment,
+ * then we have determined that subalignment can be
+ * satisfied with two 4-byte stores.
+ */
+ if (align > MO_8 && align < MO_64) {
+ use_pair = true;
+ l1 = gen_new_label();
+ l2 = gen_new_label();
+
+ tcg_out_testi(s, base, align == MO_32 ? 4 : 7);
+ tcg_out_jxx(s, JCC_JNE, l2, true);
+ }
+
+ tcg_out_modrm_offset(s, movop, datalo, TCG_REG_ESP, 0);
+ tcg_out_modrm_offset(s, movop, datahi, TCG_REG_ESP, 4);
+ tcg_out_modrm_offset(s, OPC_ESCDF, ESCDF_FILD_m64,
+ TCG_REG_ESP, 0);
+ tcg_out_modrm_sib_offset(s, OPC_ESCDF + seg, ESCDF_FISTP_m64,
+ base, index, 0, ofs);
+
+ if (use_pair) {
+ tcg_out_jxx(s, JCC_JMP, l1, true);
+ tcg_out_label(s, l2);
+ }
+ }
+ if (use_pair) {
+ tcg_out_modrm_sib_offset(s, movop + seg, datalo,
+ base, index, 0, ofs);
+ tcg_out_modrm_sib_offset(s, movop + seg, datahi,
+ base, index, 0, ofs + 4);
+ }
+ if (l1) {
+ tcg_out_label(s, l1);
+ }
}
break;
--
2.34.1
^ permalink raw reply related [flat|nested] 48+ messages in thread