* [Qemu-devel] [PULL 01/11] tcg-s390: Fix off-by-one in wraparound andi
2014-05-15 16:39 [Qemu-devel] [PULL 00/12] tcg s390 updates Richard Henderson
@ 2014-05-15 16:39 ` Richard Henderson
2014-05-15 16:39 ` [Qemu-devel] [PULL 02/11] tcg-s390: Convert to TCGMemOp Richard Henderson
` (10 subsequent siblings)
11 siblings, 0 replies; 15+ messages in thread
From: Richard Henderson @ 2014-05-15 16:39 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Signed-off-by: Richard Henderson <rth@twiddle.net>
---
tcg/s390/tcg-target.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/tcg/s390/tcg-target.c b/tcg/s390/tcg-target.c
index ebdd074..0f972ae 100644
--- a/tcg/s390/tcg-target.c
+++ b/tcg/s390/tcg-target.c
@@ -983,8 +983,8 @@ static void tgen_andi(TCGContext *s, TCGType type, TCGReg dest, uint64_t val)
int msb, lsb;
if ((val & 0x8000000000000001ull) == 0x8000000000000001ull) {
/* Achieve wraparound by swapping msb and lsb. */
- msb = 63 - ctz64(~val);
- lsb = clz64(~val) + 1;
+ msb = 64 - ctz64(~val);
+ lsb = clz64(~val) - 1;
} else {
msb = clz64(val);
lsb = 63 - ctz64(val);
--
1.9.0
^ permalink raw reply related [flat|nested] 15+ messages in thread* [Qemu-devel] [PULL 02/11] tcg-s390: Convert to TCGMemOp
2014-05-15 16:39 [Qemu-devel] [PULL 00/12] tcg s390 updates Richard Henderson
2014-05-15 16:39 ` [Qemu-devel] [PULL 01/11] tcg-s390: Fix off-by-one in wraparound andi Richard Henderson
@ 2014-05-15 16:39 ` Richard Henderson
2014-05-15 16:39 ` [Qemu-devel] [PULL 03/11] tcg-s390: Integrate endianness into TCGMemOp Richard Henderson
` (9 subsequent siblings)
11 siblings, 0 replies; 15+ messages in thread
From: Richard Henderson @ 2014-05-15 16:39 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Signed-off-by: Richard Henderson <rth@twiddle.net>
---
tcg/s390/tcg-target.c | 82 ++++++++++++++++++++++-----------------------------
1 file changed, 36 insertions(+), 46 deletions(-)
diff --git a/tcg/s390/tcg-target.c b/tcg/s390/tcg-target.c
index 0f972ae..e0faacf 100644
--- a/tcg/s390/tcg-target.c
+++ b/tcg/s390/tcg-target.c
@@ -227,16 +227,6 @@ typedef enum S390Opcode {
RX_STH = 0x40,
} S390Opcode;
-#define LD_SIGNED 0x04
-#define LD_UINT8 0x00
-#define LD_INT8 (LD_UINT8 | LD_SIGNED)
-#define LD_UINT16 0x01
-#define LD_INT16 (LD_UINT16 | LD_SIGNED)
-#define LD_UINT32 0x02
-#define LD_INT32 (LD_UINT32 | LD_SIGNED)
-#define LD_UINT64 0x03
-#define LD_INT64 (LD_UINT64 | LD_SIGNED)
-
#ifndef NDEBUG
static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
"%r0", "%r1", "%r2", "%r3", "%r4", "%r5", "%r6", "%r7",
@@ -1280,7 +1270,7 @@ static void tcg_out_call(TCGContext *s, tcg_insn_unit *dest)
}
}
-static void tcg_out_qemu_ld_direct(TCGContext *s, int opc, TCGReg data,
+static void tcg_out_qemu_ld_direct(TCGContext *s, TCGMemOp opc, TCGReg data,
TCGReg base, TCGReg index, int disp)
{
#ifdef TARGET_WORDS_BIGENDIAN
@@ -1289,13 +1279,13 @@ static void tcg_out_qemu_ld_direct(TCGContext *s, int opc, TCGReg data,
const int bswap = 1;
#endif
switch (opc) {
- case LD_UINT8:
+ case MO_UB:
tcg_out_insn(s, RXY, LLGC, data, base, index, disp);
break;
- case LD_INT8:
+ case MO_SB:
tcg_out_insn(s, RXY, LGB, data, base, index, disp);
break;
- case LD_UINT16:
+ case MO_UW:
if (bswap) {
/* swapped unsigned halfword load with upper bits zeroed */
tcg_out_insn(s, RXY, LRVH, data, base, index, disp);
@@ -1304,7 +1294,7 @@ static void tcg_out_qemu_ld_direct(TCGContext *s, int opc, TCGReg data,
tcg_out_insn(s, RXY, LLGH, data, base, index, disp);
}
break;
- case LD_INT16:
+ case MO_SW:
if (bswap) {
/* swapped sign-extended halfword load */
tcg_out_insn(s, RXY, LRVH, data, base, index, disp);
@@ -1313,7 +1303,7 @@ static void tcg_out_qemu_ld_direct(TCGContext *s, int opc, TCGReg data,
tcg_out_insn(s, RXY, LGH, data, base, index, disp);
}
break;
- case LD_UINT32:
+ case MO_UL:
if (bswap) {
/* swapped unsigned int load with upper bits zeroed */
tcg_out_insn(s, RXY, LRV, data, base, index, disp);
@@ -1322,7 +1312,7 @@ static void tcg_out_qemu_ld_direct(TCGContext *s, int opc, TCGReg data,
tcg_out_insn(s, RXY, LLGF, data, base, index, disp);
}
break;
- case LD_INT32:
+ case MO_SL:
if (bswap) {
/* swapped sign-extended int load */
tcg_out_insn(s, RXY, LRV, data, base, index, disp);
@@ -1331,7 +1321,7 @@ static void tcg_out_qemu_ld_direct(TCGContext *s, int opc, TCGReg data,
tcg_out_insn(s, RXY, LGF, data, base, index, disp);
}
break;
- case LD_UINT64:
+ case MO_Q:
if (bswap) {
tcg_out_insn(s, RXY, LRVG, data, base, index, disp);
} else {
@@ -1343,7 +1333,7 @@ static void tcg_out_qemu_ld_direct(TCGContext *s, int opc, TCGReg data,
}
}
-static void tcg_out_qemu_st_direct(TCGContext *s, int opc, TCGReg data,
+static void tcg_out_qemu_st_direct(TCGContext *s, TCGMemOp opc, TCGReg data,
TCGReg base, TCGReg index, int disp)
{
#ifdef TARGET_WORDS_BIGENDIAN
@@ -1352,14 +1342,14 @@ static void tcg_out_qemu_st_direct(TCGContext *s, int opc, TCGReg data,
const int bswap = 1;
#endif
switch (opc) {
- case LD_UINT8:
+ case MO_UB:
if (disp >= 0 && disp < 0x1000) {
tcg_out_insn(s, RX, STC, data, base, index, disp);
} else {
tcg_out_insn(s, RXY, STCY, data, base, index, disp);
}
break;
- case LD_UINT16:
+ case MO_UW:
if (bswap) {
tcg_out_insn(s, RXY, STRVH, data, base, index, disp);
} else if (disp >= 0 && disp < 0x1000) {
@@ -1368,7 +1358,7 @@ static void tcg_out_qemu_st_direct(TCGContext *s, int opc, TCGReg data,
tcg_out_insn(s, RXY, STHY, data, base, index, disp);
}
break;
- case LD_UINT32:
+ case MO_UL:
if (bswap) {
tcg_out_insn(s, RXY, STRV, data, base, index, disp);
} else if (disp >= 0 && disp < 0x1000) {
@@ -1377,7 +1367,7 @@ static void tcg_out_qemu_st_direct(TCGContext *s, int opc, TCGReg data,
tcg_out_insn(s, RXY, STY, data, base, index, disp);
}
break;
- case LD_UINT64:
+ case MO_Q:
if (bswap) {
tcg_out_insn(s, RXY, STRVG, data, base, index, disp);
} else {
@@ -1398,7 +1388,7 @@ static TCGReg tcg_prepare_qemu_ldst(TCGContext* s, TCGReg data_reg,
const TCGReg arg1 = tcg_target_call_iarg_regs[1];
const TCGReg arg2 = tcg_target_call_iarg_regs[2];
const TCGReg arg3 = tcg_target_call_iarg_regs[3];
- int s_bits = opc & 3;
+ TCGMemOp s_bits = opc & MO_SIZE;
tcg_insn_unit *label1_ptr;
tcg_target_long ofs;
@@ -1442,17 +1432,17 @@ static TCGReg tcg_prepare_qemu_ldst(TCGContext* s, TCGReg data_reg,
if (is_store) {
/* Make sure to zero-extend the value to the full register
for the calling convention. */
- switch (opc) {
- case LD_UINT8:
+ switch (s_bits) {
+ case MO_UB:
tgen_ext8u(s, TCG_TYPE_I64, arg2, data_reg);
break;
- case LD_UINT16:
+ case MO_UW:
tgen_ext16u(s, TCG_TYPE_I64, arg2, data_reg);
break;
- case LD_UINT32:
+ case MO_UL:
tgen_ext32u(s, arg2, data_reg);
break;
- case LD_UINT64:
+ case MO_Q:
tcg_out_mov(s, TCG_TYPE_I64, arg2, data_reg);
break;
default:
@@ -1468,13 +1458,13 @@ static TCGReg tcg_prepare_qemu_ldst(TCGContext* s, TCGReg data_reg,
/* sign extension */
switch (opc) {
- case LD_INT8:
+ case MO_SB:
tgen_ext8s(s, TCG_TYPE_I64, data_reg, TCG_REG_R2);
break;
- case LD_INT16:
+ case MO_SW:
tgen_ext16s(s, TCG_TYPE_I64, data_reg, TCG_REG_R2);
break;
- case LD_INT32:
+ case MO_SL:
tgen_ext32s(s, data_reg, TCG_REG_R2);
break;
default:
@@ -1525,7 +1515,7 @@ static void tcg_prepare_user_ldst(TCGContext *s, TCGReg *addr_reg,
/* load data with address translation (if applicable)
and endianness conversion */
-static void tcg_out_qemu_ld(TCGContext* s, const TCGArg* args, int opc)
+static void tcg_out_qemu_ld(TCGContext* s, const TCGArg* args, TCGMemOp opc)
{
TCGReg addr_reg, data_reg;
#if defined(CONFIG_SOFTMMU)
@@ -1554,7 +1544,7 @@ static void tcg_out_qemu_ld(TCGContext* s, const TCGArg* args, int opc)
#endif
}
-static void tcg_out_qemu_st(TCGContext* s, const TCGArg* args, int opc)
+static void tcg_out_qemu_st(TCGContext* s, const TCGArg* args, TCGMemOp opc)
{
TCGReg addr_reg, data_reg;
#if defined(CONFIG_SOFTMMU)
@@ -1812,36 +1802,36 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
break;
case INDEX_op_qemu_ld8u:
- tcg_out_qemu_ld(s, args, LD_UINT8);
+ tcg_out_qemu_ld(s, args, MO_UB);
break;
case INDEX_op_qemu_ld8s:
- tcg_out_qemu_ld(s, args, LD_INT8);
+ tcg_out_qemu_ld(s, args, MO_SB);
break;
case INDEX_op_qemu_ld16u:
- tcg_out_qemu_ld(s, args, LD_UINT16);
+ tcg_out_qemu_ld(s, args, MO_UW);
break;
case INDEX_op_qemu_ld16s:
- tcg_out_qemu_ld(s, args, LD_INT16);
+ tcg_out_qemu_ld(s, args, MO_SW);
break;
case INDEX_op_qemu_ld32:
/* ??? Technically we can use a non-extending instruction. */
- tcg_out_qemu_ld(s, args, LD_UINT32);
+ tcg_out_qemu_ld(s, args, MO_UL);
break;
case INDEX_op_qemu_ld64:
- tcg_out_qemu_ld(s, args, LD_UINT64);
+ tcg_out_qemu_ld(s, args, MO_Q);
break;
case INDEX_op_qemu_st8:
- tcg_out_qemu_st(s, args, LD_UINT8);
+ tcg_out_qemu_st(s, args, MO_UB);
break;
case INDEX_op_qemu_st16:
- tcg_out_qemu_st(s, args, LD_UINT16);
+ tcg_out_qemu_st(s, args, MO_UW);
break;
case INDEX_op_qemu_st32:
- tcg_out_qemu_st(s, args, LD_UINT32);
+ tcg_out_qemu_st(s, args, MO_UL);
break;
case INDEX_op_qemu_st64:
- tcg_out_qemu_st(s, args, LD_UINT64);
+ tcg_out_qemu_st(s, args, MO_Q);
break;
case INDEX_op_ld16s_i64:
@@ -2038,10 +2028,10 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
break;
case INDEX_op_qemu_ld32u:
- tcg_out_qemu_ld(s, args, LD_UINT32);
+ tcg_out_qemu_ld(s, args, MO_UL);
break;
case INDEX_op_qemu_ld32s:
- tcg_out_qemu_ld(s, args, LD_INT32);
+ tcg_out_qemu_ld(s, args, MO_SL);
break;
OP_32_64(deposit):
--
1.9.0
^ permalink raw reply related [flat|nested] 15+ messages in thread* [Qemu-devel] [PULL 03/11] tcg-s390: Integrate endianness into TCGMemOp
2014-05-15 16:39 [Qemu-devel] [PULL 00/12] tcg s390 updates Richard Henderson
2014-05-15 16:39 ` [Qemu-devel] [PULL 01/11] tcg-s390: Fix off-by-one in wraparound andi Richard Henderson
2014-05-15 16:39 ` [Qemu-devel] [PULL 02/11] tcg-s390: Convert to TCGMemOp Richard Henderson
@ 2014-05-15 16:39 ` Richard Henderson
2014-05-15 16:39 ` [Qemu-devel] [PULL 04/11] tcg-s390: Convert to new ldst opcodes Richard Henderson
` (8 subsequent siblings)
11 siblings, 0 replies; 15+ messages in thread
From: Richard Henderson @ 2014-05-15 16:39 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Signed-off-by: Richard Henderson <rth@twiddle.net>
---
tcg/s390/tcg-target.c | 124 ++++++++++++++++++++++++--------------------------
1 file changed, 60 insertions(+), 64 deletions(-)
diff --git a/tcg/s390/tcg-target.c b/tcg/s390/tcg-target.c
index e0faacf..f1d9089 100644
--- a/tcg/s390/tcg-target.c
+++ b/tcg/s390/tcg-target.c
@@ -1273,11 +1273,6 @@ static void tcg_out_call(TCGContext *s, tcg_insn_unit *dest)
static void tcg_out_qemu_ld_direct(TCGContext *s, TCGMemOp opc, TCGReg data,
TCGReg base, TCGReg index, int disp)
{
-#ifdef TARGET_WORDS_BIGENDIAN
- const int bswap = 0;
-#else
- const int bswap = 1;
-#endif
switch (opc) {
case MO_UB:
tcg_out_insn(s, RXY, LLGC, data, base, index, disp);
@@ -1285,49 +1280,50 @@ static void tcg_out_qemu_ld_direct(TCGContext *s, TCGMemOp opc, TCGReg data,
case MO_SB:
tcg_out_insn(s, RXY, LGB, data, base, index, disp);
break;
+
+ case MO_UW | MO_BSWAP:
+ /* swapped unsigned halfword load with upper bits zeroed */
+ tcg_out_insn(s, RXY, LRVH, data, base, index, disp);
+ tgen_ext16u(s, TCG_TYPE_I64, data, data);
+ break;
case MO_UW:
- if (bswap) {
- /* swapped unsigned halfword load with upper bits zeroed */
- tcg_out_insn(s, RXY, LRVH, data, base, index, disp);
- tgen_ext16u(s, TCG_TYPE_I64, data, data);
- } else {
- tcg_out_insn(s, RXY, LLGH, data, base, index, disp);
- }
+ tcg_out_insn(s, RXY, LLGH, data, base, index, disp);
+ break;
+
+ case MO_SW | MO_BSWAP:
+ /* swapped sign-extended halfword load */
+ tcg_out_insn(s, RXY, LRVH, data, base, index, disp);
+ tgen_ext16s(s, TCG_TYPE_I64, data, data);
break;
case MO_SW:
- if (bswap) {
- /* swapped sign-extended halfword load */
- tcg_out_insn(s, RXY, LRVH, data, base, index, disp);
- tgen_ext16s(s, TCG_TYPE_I64, data, data);
- } else {
- tcg_out_insn(s, RXY, LGH, data, base, index, disp);
- }
+ tcg_out_insn(s, RXY, LGH, data, base, index, disp);
+ break;
+
+ case MO_UL | MO_BSWAP:
+ /* swapped unsigned int load with upper bits zeroed */
+ tcg_out_insn(s, RXY, LRV, data, base, index, disp);
+ tgen_ext32u(s, data, data);
break;
case MO_UL:
- if (bswap) {
- /* swapped unsigned int load with upper bits zeroed */
- tcg_out_insn(s, RXY, LRV, data, base, index, disp);
- tgen_ext32u(s, data, data);
- } else {
- tcg_out_insn(s, RXY, LLGF, data, base, index, disp);
- }
+ tcg_out_insn(s, RXY, LLGF, data, base, index, disp);
+ break;
+
+ case MO_SL | MO_BSWAP:
+ /* swapped sign-extended int load */
+ tcg_out_insn(s, RXY, LRV, data, base, index, disp);
+ tgen_ext32s(s, data, data);
break;
case MO_SL:
- if (bswap) {
- /* swapped sign-extended int load */
- tcg_out_insn(s, RXY, LRV, data, base, index, disp);
- tgen_ext32s(s, data, data);
- } else {
- tcg_out_insn(s, RXY, LGF, data, base, index, disp);
- }
+ tcg_out_insn(s, RXY, LGF, data, base, index, disp);
+ break;
+
+ case MO_Q | MO_BSWAP:
+ tcg_out_insn(s, RXY, LRVG, data, base, index, disp);
break;
case MO_Q:
- if (bswap) {
- tcg_out_insn(s, RXY, LRVG, data, base, index, disp);
- } else {
- tcg_out_insn(s, RXY, LG, data, base, index, disp);
- }
+ tcg_out_insn(s, RXY, LG, data, base, index, disp);
break;
+
default:
tcg_abort();
}
@@ -1336,11 +1332,6 @@ static void tcg_out_qemu_ld_direct(TCGContext *s, TCGMemOp opc, TCGReg data,
static void tcg_out_qemu_st_direct(TCGContext *s, TCGMemOp opc, TCGReg data,
TCGReg base, TCGReg index, int disp)
{
-#ifdef TARGET_WORDS_BIGENDIAN
- const int bswap = 0;
-#else
- const int bswap = 1;
-#endif
switch (opc) {
case MO_UB:
if (disp >= 0 && disp < 0x1000) {
@@ -1349,31 +1340,36 @@ static void tcg_out_qemu_st_direct(TCGContext *s, TCGMemOp opc, TCGReg data,
tcg_out_insn(s, RXY, STCY, data, base, index, disp);
}
break;
+
+ case MO_UW | MO_BSWAP:
+ tcg_out_insn(s, RXY, STRVH, data, base, index, disp);
+ break;
case MO_UW:
- if (bswap) {
- tcg_out_insn(s, RXY, STRVH, data, base, index, disp);
- } else if (disp >= 0 && disp < 0x1000) {
+ if (disp >= 0 && disp < 0x1000) {
tcg_out_insn(s, RX, STH, data, base, index, disp);
} else {
tcg_out_insn(s, RXY, STHY, data, base, index, disp);
}
break;
+
+ case MO_UL | MO_BSWAP:
+ tcg_out_insn(s, RXY, STRV, data, base, index, disp);
+ break;
case MO_UL:
- if (bswap) {
- tcg_out_insn(s, RXY, STRV, data, base, index, disp);
- } else if (disp >= 0 && disp < 0x1000) {
+ if (disp >= 0 && disp < 0x1000) {
tcg_out_insn(s, RX, ST, data, base, index, disp);
} else {
tcg_out_insn(s, RXY, STY, data, base, index, disp);
}
break;
+
+ case MO_Q | MO_BSWAP:
+ tcg_out_insn(s, RXY, STRVG, data, base, index, disp);
+ break;
case MO_Q:
- if (bswap) {
- tcg_out_insn(s, RXY, STRVG, data, base, index, disp);
- } else {
- tcg_out_insn(s, RXY, STG, data, base, index, disp);
- }
+ tcg_out_insn(s, RXY, STG, data, base, index, disp);
break;
+
default:
tcg_abort();
}
@@ -1457,7 +1453,7 @@ static TCGReg tcg_prepare_qemu_ldst(TCGContext* s, TCGReg data_reg,
tcg_out_call(s, qemu_ld_helpers[s_bits]);
/* sign extension */
- switch (opc) {
+ switch (opc & MO_SSIZE) {
case MO_SB:
tgen_ext8s(s, TCG_TYPE_I64, data_reg, TCG_REG_R2);
break;
@@ -1808,30 +1804,30 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
tcg_out_qemu_ld(s, args, MO_SB);
break;
case INDEX_op_qemu_ld16u:
- tcg_out_qemu_ld(s, args, MO_UW);
+ tcg_out_qemu_ld(s, args, MO_TEUW);
break;
case INDEX_op_qemu_ld16s:
- tcg_out_qemu_ld(s, args, MO_SW);
+ tcg_out_qemu_ld(s, args, MO_TESW);
break;
case INDEX_op_qemu_ld32:
/* ??? Technically we can use a non-extending instruction. */
- tcg_out_qemu_ld(s, args, MO_UL);
+ tcg_out_qemu_ld(s, args, MO_TEUL);
break;
case INDEX_op_qemu_ld64:
- tcg_out_qemu_ld(s, args, MO_Q);
+ tcg_out_qemu_ld(s, args, MO_TEQ);
break;
case INDEX_op_qemu_st8:
tcg_out_qemu_st(s, args, MO_UB);
break;
case INDEX_op_qemu_st16:
- tcg_out_qemu_st(s, args, MO_UW);
+ tcg_out_qemu_st(s, args, MO_TEUW);
break;
case INDEX_op_qemu_st32:
- tcg_out_qemu_st(s, args, MO_UL);
+ tcg_out_qemu_st(s, args, MO_TEUL);
break;
case INDEX_op_qemu_st64:
- tcg_out_qemu_st(s, args, MO_Q);
+ tcg_out_qemu_st(s, args, MO_TEQ);
break;
case INDEX_op_ld16s_i64:
@@ -2028,10 +2024,10 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
break;
case INDEX_op_qemu_ld32u:
- tcg_out_qemu_ld(s, args, MO_UL);
+ tcg_out_qemu_ld(s, args, MO_TEUL);
break;
case INDEX_op_qemu_ld32s:
- tcg_out_qemu_ld(s, args, MO_SL);
+ tcg_out_qemu_ld(s, args, MO_TESL);
break;
OP_32_64(deposit):
--
1.9.0
^ permalink raw reply related [flat|nested] 15+ messages in thread* [Qemu-devel] [PULL 04/11] tcg-s390: Convert to new ldst opcodes
2014-05-15 16:39 [Qemu-devel] [PULL 00/12] tcg s390 updates Richard Henderson
` (2 preceding siblings ...)
2014-05-15 16:39 ` [Qemu-devel] [PULL 03/11] tcg-s390: Integrate endianness into TCGMemOp Richard Henderson
@ 2014-05-15 16:39 ` Richard Henderson
2014-05-15 16:39 ` [Qemu-devel] [PULL 05/11] tcg-s390: Move ldst helpers out of line Richard Henderson
` (7 subsequent siblings)
11 siblings, 0 replies; 15+ messages in thread
From: Richard Henderson @ 2014-05-15 16:39 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Signed-off-by: Richard Henderson <rth@twiddle.net>
---
tcg/s390/tcg-target.c | 170 ++++++++++++++++----------------------------------
tcg/s390/tcg-target.h | 2 +-
2 files changed, 56 insertions(+), 116 deletions(-)
diff --git a/tcg/s390/tcg-target.c b/tcg/s390/tcg-target.c
index f1d9089..4fbadaf 100644
--- a/tcg/s390/tcg-target.c
+++ b/tcg/s390/tcg-target.c
@@ -238,6 +238,7 @@ static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
call-saved registers. Likewise prefer the call-clobbered registers
in reverse order to maximize the chance of avoiding the arguments. */
static const int tcg_target_reg_alloc_order[] = {
+ /* Call saved registers. */
TCG_REG_R13,
TCG_REG_R12,
TCG_REG_R11,
@@ -246,9 +247,11 @@ static const int tcg_target_reg_alloc_order[] = {
TCG_REG_R8,
TCG_REG_R7,
TCG_REG_R6,
+ /* Call clobbered registers. */
TCG_REG_R14,
TCG_REG_R0,
TCG_REG_R1,
+ /* Argument registers, in reverse order of allocation. */
TCG_REG_R5,
TCG_REG_R4,
TCG_REG_R3,
@@ -308,22 +311,29 @@ static const uint8_t tcg_cond_to_ltr_cond[] = {
};
#ifdef CONFIG_SOFTMMU
-/* helper signature: helper_ld_mmu(CPUState *env, target_ulong addr,
- int mmu_idx) */
-static void * const qemu_ld_helpers[4] = {
- helper_ldb_mmu,
- helper_ldw_mmu,
- helper_ldl_mmu,
- helper_ldq_mmu,
+static void * const qemu_ld_helpers[16] = {
+ [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_LEQ] = 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_BEQ] = helper_be_ldq_mmu,
};
-/* helper signature: helper_st_mmu(CPUState *env, target_ulong addr,
- uintxx_t val, int mmu_idx) */
-static void * const qemu_st_helpers[4] = {
- helper_stb_mmu,
- helper_stw_mmu,
- helper_stl_mmu,
- helper_stq_mmu,
+static void * const qemu_st_helpers[16] = {
+ [MO_UB] = helper_ret_stb_mmu,
+ [MO_LEUW] = helper_le_stw_mmu,
+ [MO_LEUL] = helper_le_stl_mmu,
+ [MO_LEQ] = helper_le_stq_mmu,
+ [MO_BEUW] = helper_be_stw_mmu,
+ [MO_BEUL] = helper_be_stl_mmu,
+ [MO_BEQ] = helper_be_stq_mmu,
};
#endif
@@ -1384,6 +1394,7 @@ static TCGReg tcg_prepare_qemu_ldst(TCGContext* s, TCGReg data_reg,
const TCGReg arg1 = tcg_target_call_iarg_regs[1];
const TCGReg arg2 = tcg_target_call_iarg_regs[2];
const TCGReg arg3 = tcg_target_call_iarg_regs[3];
+ const TCGReg arg4 = tcg_target_call_iarg_regs[4];
TCGMemOp s_bits = opc & MO_SIZE;
tcg_insn_unit *label1_ptr;
tcg_target_long ofs;
@@ -1445,29 +1456,15 @@ static TCGReg tcg_prepare_qemu_ldst(TCGContext* s, TCGReg data_reg,
tcg_abort();
}
tcg_out_movi(s, TCG_TYPE_I32, arg3, mem_index);
- tcg_out_mov(s, TCG_TYPE_I64, arg0, TCG_AREG0);
- tcg_out_call(s, qemu_st_helpers[s_bits]);
+ tcg_out_mov(s, TCG_TYPE_PTR, arg0, TCG_AREG0);
+ tcg_out_movi(s, TCG_TYPE_PTR, arg4, (uintptr_t)s->code_ptr);
+ tcg_out_call(s, qemu_st_helpers[opc]);
} else {
tcg_out_movi(s, TCG_TYPE_I32, arg2, mem_index);
- tcg_out_mov(s, TCG_TYPE_I64, arg0, TCG_AREG0);
- tcg_out_call(s, qemu_ld_helpers[s_bits]);
-
- /* sign extension */
- switch (opc & MO_SSIZE) {
- case MO_SB:
- tgen_ext8s(s, TCG_TYPE_I64, data_reg, TCG_REG_R2);
- break;
- case MO_SW:
- tgen_ext16s(s, TCG_TYPE_I64, data_reg, TCG_REG_R2);
- break;
- case MO_SL:
- tgen_ext32s(s, data_reg, TCG_REG_R2);
- break;
- default:
- /* unsigned -> just copy */
- tcg_out_mov(s, TCG_TYPE_I64, data_reg, TCG_REG_R2);
- break;
- }
+ tcg_out_mov(s, TCG_TYPE_PTR, arg0, TCG_AREG0);
+ tcg_out_movi(s, TCG_TYPE_PTR, arg3, (uintptr_t)s->code_ptr);
+ tcg_out_call(s, qemu_ld_helpers[opc]);
+ tcg_out_mov(s, TCG_TYPE_I64, data_reg, TCG_REG_R2);
}
/* jump to label2 (end) */
@@ -1511,59 +1508,39 @@ static void tcg_prepare_user_ldst(TCGContext *s, TCGReg *addr_reg,
/* load data with address translation (if applicable)
and endianness conversion */
-static void tcg_out_qemu_ld(TCGContext* s, const TCGArg* args, TCGMemOp opc)
+static void tcg_out_qemu_ld(TCGContext* s, TCGReg data_reg, TCGReg addr_reg,
+ TCGMemOp opc, int mem_index)
{
- TCGReg addr_reg, data_reg;
#if defined(CONFIG_SOFTMMU)
- int mem_index;
tcg_insn_unit *label2_ptr;
-#else
- TCGReg index_reg;
- tcg_target_long disp;
-#endif
-
- data_reg = *args++;
- addr_reg = *args++;
-
-#if defined(CONFIG_SOFTMMU)
- mem_index = *args;
addr_reg = tcg_prepare_qemu_ldst(s, data_reg, addr_reg, mem_index,
opc, &label2_ptr, 0);
-
tcg_out_qemu_ld_direct(s, opc, data_reg, addr_reg, TCG_REG_NONE, 0);
-
tcg_finish_qemu_ldst(s, label2_ptr);
#else
+ TCGReg index_reg;
+ tcg_target_long disp;
+
tcg_prepare_user_ldst(s, &addr_reg, &index_reg, &disp);
tcg_out_qemu_ld_direct(s, opc, data_reg, addr_reg, index_reg, disp);
#endif
}
-static void tcg_out_qemu_st(TCGContext* s, const TCGArg* args, TCGMemOp opc)
+static void tcg_out_qemu_st(TCGContext* s, TCGReg data_reg, TCGReg addr_reg,
+ TCGMemOp opc, int mem_index)
{
- TCGReg addr_reg, data_reg;
#if defined(CONFIG_SOFTMMU)
- int mem_index;
tcg_insn_unit *label2_ptr;
-#else
- TCGReg index_reg;
- tcg_target_long disp;
-#endif
-
- data_reg = *args++;
- addr_reg = *args++;
-
-#if defined(CONFIG_SOFTMMU)
- mem_index = *args;
addr_reg = tcg_prepare_qemu_ldst(s, data_reg, addr_reg, mem_index,
opc, &label2_ptr, 1);
-
tcg_out_qemu_st_direct(s, opc, data_reg, addr_reg, TCG_REG_NONE, 0);
-
tcg_finish_qemu_ldst(s, label2_ptr);
#else
+ TCGReg index_reg;
+ tcg_target_long disp;
+
tcg_prepare_user_ldst(s, &addr_reg, &index_reg, &disp);
tcg_out_qemu_st_direct(s, opc, data_reg, addr_reg, index_reg, disp);
#endif
@@ -1797,37 +1774,14 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
args[2], const_args[2], args[3]);
break;
- case INDEX_op_qemu_ld8u:
- tcg_out_qemu_ld(s, args, MO_UB);
- break;
- case INDEX_op_qemu_ld8s:
- tcg_out_qemu_ld(s, args, MO_SB);
- break;
- case INDEX_op_qemu_ld16u:
- tcg_out_qemu_ld(s, args, MO_TEUW);
- break;
- case INDEX_op_qemu_ld16s:
- tcg_out_qemu_ld(s, args, MO_TESW);
- break;
- case INDEX_op_qemu_ld32:
+ case INDEX_op_qemu_ld_i32:
/* ??? Technically we can use a non-extending instruction. */
- tcg_out_qemu_ld(s, args, MO_TEUL);
- break;
- case INDEX_op_qemu_ld64:
- tcg_out_qemu_ld(s, args, MO_TEQ);
- break;
-
- case INDEX_op_qemu_st8:
- tcg_out_qemu_st(s, args, MO_UB);
- break;
- case INDEX_op_qemu_st16:
- tcg_out_qemu_st(s, args, MO_TEUW);
- break;
- case INDEX_op_qemu_st32:
- tcg_out_qemu_st(s, args, MO_TEUL);
+ case INDEX_op_qemu_ld_i64:
+ tcg_out_qemu_ld(s, args[0], args[1], args[2], args[3]);
break;
- case INDEX_op_qemu_st64:
- tcg_out_qemu_st(s, args, MO_TEQ);
+ case INDEX_op_qemu_st_i32:
+ case INDEX_op_qemu_st_i64:
+ tcg_out_qemu_st(s, args[0], args[1], args[2], args[3]);
break;
case INDEX_op_ld16s_i64:
@@ -2023,13 +1977,6 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
args[2], const_args[2], args[3]);
break;
- case INDEX_op_qemu_ld32u:
- tcg_out_qemu_ld(s, args, MO_TEUL);
- break;
- case INDEX_op_qemu_ld32s:
- tcg_out_qemu_ld(s, args, MO_TESL);
- break;
-
OP_32_64(deposit):
tgen_deposit(s, args[0], args[2], args[3], args[4]);
break;
@@ -2094,17 +2041,10 @@ static const TCGTargetOpDef s390_op_defs[] = {
{ INDEX_op_movcond_i32, { "r", "r", "rC", "r", "0" } },
{ INDEX_op_deposit_i32, { "r", "0", "r" } },
- { INDEX_op_qemu_ld8u, { "r", "L" } },
- { INDEX_op_qemu_ld8s, { "r", "L" } },
- { INDEX_op_qemu_ld16u, { "r", "L" } },
- { INDEX_op_qemu_ld16s, { "r", "L" } },
- { INDEX_op_qemu_ld32, { "r", "L" } },
- { INDEX_op_qemu_ld64, { "r", "L" } },
-
- { INDEX_op_qemu_st8, { "L", "L" } },
- { INDEX_op_qemu_st16, { "L", "L" } },
- { INDEX_op_qemu_st32, { "L", "L" } },
- { INDEX_op_qemu_st64, { "L", "L" } },
+ { INDEX_op_qemu_ld_i32, { "r", "L" } },
+ { INDEX_op_qemu_ld_i64, { "r", "L" } },
+ { INDEX_op_qemu_st_i32, { "L", "L" } },
+ { INDEX_op_qemu_st_i64, { "L", "L" } },
{ INDEX_op_ld8u_i64, { "r", "r" } },
{ INDEX_op_ld8s_i64, { "r", "r" } },
@@ -2159,9 +2099,6 @@ static const TCGTargetOpDef s390_op_defs[] = {
{ INDEX_op_movcond_i64, { "r", "r", "rC", "r", "0" } },
{ INDEX_op_deposit_i64, { "r", "0", "r" } },
- { INDEX_op_qemu_ld32u, { "r", "L" } },
- { INDEX_op_qemu_ld32s, { "r", "L" } },
-
{ -1 },
};
@@ -2196,6 +2133,9 @@ static void tcg_target_init(TCGContext *s)
tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R3);
tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R4);
tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R5);
+ /* The r6 register is technically call-saved, but it's also a parameter
+ register, so it can get killed by setup for the qemu_st helper. */
+ tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R6);
/* The return register can be considered call-clobbered. */
tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R14);
diff --git a/tcg/s390/tcg-target.h b/tcg/s390/tcg-target.h
index 5bf733e..ad2c6dd 100644
--- a/tcg/s390/tcg-target.h
+++ b/tcg/s390/tcg-target.h
@@ -100,7 +100,7 @@ typedef enum TCGReg {
#define TCG_TARGET_HAS_muluh_i64 0
#define TCG_TARGET_HAS_mulsh_i64 0
-#define TCG_TARGET_HAS_new_ldst 0
+#define TCG_TARGET_HAS_new_ldst 1
extern bool tcg_target_deposit_valid(int ofs, int len);
#define TCG_TARGET_deposit_i32_valid tcg_target_deposit_valid
--
1.9.0
^ permalink raw reply related [flat|nested] 15+ messages in thread* [Qemu-devel] [PULL 05/11] tcg-s390: Move ldst helpers out of line
2014-05-15 16:39 [Qemu-devel] [PULL 00/12] tcg s390 updates Richard Henderson
` (3 preceding siblings ...)
2014-05-15 16:39 ` [Qemu-devel] [PULL 04/11] tcg-s390: Convert to new ldst opcodes Richard Henderson
@ 2014-05-15 16:39 ` Richard Henderson
2015-01-29 16:01 ` Alexander Graf
2014-05-15 16:39 ` [Qemu-devel] [PULL 06/11] tcg-s390: Use more risbg in the tlb sequence Richard Henderson
` (6 subsequent siblings)
11 siblings, 1 reply; 15+ messages in thread
From: Richard Henderson @ 2014-05-15 16:39 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
That is, the old LDST_OPTIMIZATION.
Signed-off-by: Richard Henderson <rth@twiddle.net>
---
tcg/s390/tcg-target.c | 210 ++++++++++++++++++++++++++++----------------------
1 file changed, 118 insertions(+), 92 deletions(-)
diff --git a/tcg/s390/tcg-target.c b/tcg/s390/tcg-target.c
index 4fbadaf..1c955db 100644
--- a/tcg/s390/tcg-target.c
+++ b/tcg/s390/tcg-target.c
@@ -24,7 +24,7 @@
* THE SOFTWARE.
*/
-#include "tcg-be-null.h"
+#include "tcg-be-ldst.h"
/* We only support generating code for 64-bit mode. */
#if TCG_TARGET_REG_BITS != 64
@@ -1386,107 +1386,123 @@ static void tcg_out_qemu_st_direct(TCGContext *s, TCGMemOp opc, TCGReg data,
}
#if defined(CONFIG_SOFTMMU)
-static TCGReg tcg_prepare_qemu_ldst(TCGContext* s, TCGReg data_reg,
- TCGReg addr_reg, int mem_index, int opc,
- tcg_insn_unit **label2_ptr_p, int is_store)
+/* We're expecting to use a 20-bit signed offset on the tlb memory ops.
+ Using the offset of the second entry in the last tlb table ensures
+ that we can index all of the elements of the first entry. */
+QEMU_BUILD_BUG_ON(offsetof(CPUArchState, tlb_table[NB_MMU_MODES - 1][1])
+ > 0x7ffff);
+
+/* Load and compare a TLB entry, leaving the flags set. Loads the TLB
+ addend into R2. Returns a register with the santitized guest address. */
+static TCGReg tcg_out_tlb_read(TCGContext* s, TCGReg addr_reg, TCGMemOp opc,
+ int mem_index, bool is_ld)
{
- const TCGReg arg0 = tcg_target_call_iarg_regs[0];
- const TCGReg arg1 = tcg_target_call_iarg_regs[1];
- const TCGReg arg2 = tcg_target_call_iarg_regs[2];
- const TCGReg arg3 = tcg_target_call_iarg_regs[3];
- const TCGReg arg4 = tcg_target_call_iarg_regs[4];
TCGMemOp s_bits = opc & MO_SIZE;
- tcg_insn_unit *label1_ptr;
- tcg_target_long ofs;
+ int ofs;
+
+ tcg_out_sh64(s, RSY_SRLG, TCG_REG_R2, addr_reg, TCG_REG_NONE,
+ TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS);
if (TARGET_LONG_BITS == 32) {
- tgen_ext32u(s, arg1, addr_reg);
+ tgen_ext32u(s, TCG_REG_R3, addr_reg);
} else {
- tcg_out_mov(s, TCG_TYPE_I64, arg1, addr_reg);
+ tcg_out_mov(s, TCG_TYPE_I64, TCG_REG_R3, addr_reg);
}
- tcg_out_sh64(s, RSY_SRLG, arg2, addr_reg, TCG_REG_NONE,
- TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS);
-
- tgen_andi(s, TCG_TYPE_I64, arg1, TARGET_PAGE_MASK | ((1 << s_bits) - 1));
- tgen_andi(s, TCG_TYPE_I64, arg2, (CPU_TLB_SIZE - 1) << CPU_TLB_ENTRY_BITS);
+ tgen_andi(s, TCG_TYPE_I64, TCG_REG_R2,
+ (CPU_TLB_SIZE - 1) << CPU_TLB_ENTRY_BITS);
+ tgen_andi(s, TCG_TYPE_I64, TCG_REG_R3,
+ TARGET_PAGE_MASK | ((1 << s_bits) - 1));
- if (is_store) {
- ofs = offsetof(CPUArchState, tlb_table[mem_index][0].addr_write);
- } else {
+ if (is_ld) {
ofs = offsetof(CPUArchState, tlb_table[mem_index][0].addr_read);
+ } else {
+ ofs = offsetof(CPUArchState, tlb_table[mem_index][0].addr_write);
}
- assert(ofs < 0x80000);
-
if (TARGET_LONG_BITS == 32) {
- tcg_out_mem(s, RX_C, RXY_CY, arg1, arg2, TCG_AREG0, ofs);
+ tcg_out_mem(s, RX_C, RXY_CY, TCG_REG_R3, TCG_REG_R2, TCG_AREG0, ofs);
} else {
- tcg_out_mem(s, 0, RXY_CG, arg1, arg2, TCG_AREG0, ofs);
+ tcg_out_mem(s, 0, RXY_CG, TCG_REG_R3, TCG_REG_R2, TCG_AREG0, ofs);
}
+ ofs = offsetof(CPUArchState, tlb_table[mem_index][0].addend);
+ tcg_out_mem(s, 0, RXY_LG, TCG_REG_R2, TCG_REG_R2, TCG_AREG0, ofs);
+
if (TARGET_LONG_BITS == 32) {
- tgen_ext32u(s, arg1, addr_reg);
- } else {
- tcg_out_mov(s, TCG_TYPE_I64, arg1, addr_reg);
- }
-
- label1_ptr = s->code_ptr;
-
- /* je label1 (offset will be patched in later) */
- tcg_out_insn(s, RI, BRC, S390_CC_EQ, 0);
-
- /* call load/store helper */
- if (is_store) {
- /* Make sure to zero-extend the value to the full register
- for the calling convention. */
- switch (s_bits) {
- case MO_UB:
- tgen_ext8u(s, TCG_TYPE_I64, arg2, data_reg);
- break;
- case MO_UW:
- tgen_ext16u(s, TCG_TYPE_I64, arg2, data_reg);
- break;
- case MO_UL:
- tgen_ext32u(s, arg2, data_reg);
- break;
- case MO_Q:
- tcg_out_mov(s, TCG_TYPE_I64, arg2, data_reg);
- break;
- default:
- tcg_abort();
- }
- tcg_out_movi(s, TCG_TYPE_I32, arg3, mem_index);
- tcg_out_mov(s, TCG_TYPE_PTR, arg0, TCG_AREG0);
- tcg_out_movi(s, TCG_TYPE_PTR, arg4, (uintptr_t)s->code_ptr);
- tcg_out_call(s, qemu_st_helpers[opc]);
- } else {
- tcg_out_movi(s, TCG_TYPE_I32, arg2, mem_index);
- tcg_out_mov(s, TCG_TYPE_PTR, arg0, TCG_AREG0);
- tcg_out_movi(s, TCG_TYPE_PTR, arg3, (uintptr_t)s->code_ptr);
- tcg_out_call(s, qemu_ld_helpers[opc]);
- tcg_out_mov(s, TCG_TYPE_I64, data_reg, TCG_REG_R2);
+ tgen_ext32u(s, TCG_REG_R3, addr_reg);
+ return TCG_REG_R3;
}
+ return addr_reg;
+}
- /* jump to label2 (end) */
- *label2_ptr_p = s->code_ptr;
-
- tcg_out_insn(s, RI, BRC, S390_CC_ALWAYS, 0);
+static void add_qemu_ldst_label(TCGContext *s, bool is_ld, TCGMemOp opc,
+ TCGReg data, TCGReg addr, int mem_index,
+ tcg_insn_unit *raddr, tcg_insn_unit *label_ptr)
+{
+ TCGLabelQemuLdst *label = new_ldst_label(s);
+
+ label->is_ld = is_ld;
+ label->opc = opc;
+ label->datalo_reg = data;
+ label->addrlo_reg = addr;
+ label->mem_index = mem_index;
+ label->raddr = raddr;
+ label->label_ptr[0] = label_ptr;
+}
- /* this is label1, patch branch */
- label1_ptr[1] = s->code_ptr - label1_ptr;
+static void tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
+{
+ TCGReg addr_reg = lb->addrlo_reg;
+ TCGReg data_reg = lb->datalo_reg;
+ TCGMemOp opc = lb->opc;
- ofs = offsetof(CPUArchState, tlb_table[mem_index][0].addend);
- assert(ofs < 0x80000);
+ patch_reloc(lb->label_ptr[0], R_390_PC16DBL, (intptr_t)s->code_ptr, -2);
- tcg_out_mem(s, 0, RXY_AG, arg1, arg2, TCG_AREG0, ofs);
+ tcg_out_mov(s, TCG_TYPE_PTR, TCG_REG_R2, TCG_AREG0);
+ if (TARGET_LONG_BITS == 64) {
+ tcg_out_mov(s, TCG_TYPE_I64, TCG_REG_R3, addr_reg);
+ }
+ tcg_out_movi(s, TCG_TYPE_I32, TCG_REG_R4, lb->mem_index);
+ tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R5, (uintptr_t)lb->raddr);
+ tcg_out_call(s, qemu_ld_helpers[opc]);
+ tcg_out_mov(s, TCG_TYPE_I64, data_reg, TCG_REG_R2);
- return arg1;
+ tgen_gotoi(s, S390_CC_ALWAYS, lb->raddr);
}
-static void tcg_finish_qemu_ldst(TCGContext* s, tcg_insn_unit *label2_ptr)
+static void tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
{
- /* patch branch */
- label2_ptr[1] = s->code_ptr - label2_ptr;
+ TCGReg addr_reg = lb->addrlo_reg;
+ TCGReg data_reg = lb->datalo_reg;
+ TCGMemOp opc = lb->opc;
+
+ patch_reloc(lb->label_ptr[0], R_390_PC16DBL, (intptr_t)s->code_ptr, -2);
+
+ tcg_out_mov(s, TCG_TYPE_PTR, TCG_REG_R2, TCG_AREG0);
+ if (TARGET_LONG_BITS == 64) {
+ tcg_out_mov(s, TCG_TYPE_I64, TCG_REG_R3, addr_reg);
+ }
+ switch (opc & MO_SIZE) {
+ case MO_UB:
+ tgen_ext8u(s, TCG_TYPE_I64, TCG_REG_R4, data_reg);
+ break;
+ case MO_UW:
+ tgen_ext16u(s, TCG_TYPE_I64, TCG_REG_R4, data_reg);
+ break;
+ case MO_UL:
+ tgen_ext32u(s, TCG_REG_R4, data_reg);
+ break;
+ case MO_Q:
+ tcg_out_mov(s, TCG_TYPE_I64, TCG_REG_R4, data_reg);
+ break;
+ default:
+ tcg_abort();
+ }
+ tcg_out_movi(s, TCG_TYPE_I32, TCG_REG_R5, lb->mem_index);
+ tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R6, (uintptr_t)lb->raddr);
+ tcg_out_call(s, qemu_st_helpers[opc]);
+
+ tgen_gotoi(s, S390_CC_ALWAYS, lb->raddr);
}
#else
static void tcg_prepare_user_ldst(TCGContext *s, TCGReg *addr_reg,
@@ -1506,18 +1522,22 @@ static void tcg_prepare_user_ldst(TCGContext *s, TCGReg *addr_reg,
}
#endif /* CONFIG_SOFTMMU */
-/* load data with address translation (if applicable)
- and endianness conversion */
static void tcg_out_qemu_ld(TCGContext* s, TCGReg data_reg, TCGReg addr_reg,
TCGMemOp opc, int mem_index)
{
-#if defined(CONFIG_SOFTMMU)
- tcg_insn_unit *label2_ptr;
+#ifdef CONFIG_SOFTMMU
+ tcg_insn_unit *label_ptr;
+ TCGReg base_reg;
+
+ base_reg = tcg_out_tlb_read(s, addr_reg, opc, mem_index, 1);
+
+ label_ptr = s->code_ptr + 1;
+ tcg_out_insn(s, RI, BRC, S390_CC_NE, 0);
+
+ tcg_out_qemu_ld_direct(s, opc, data_reg, base_reg, TCG_REG_R2, 0);
- addr_reg = tcg_prepare_qemu_ldst(s, data_reg, addr_reg, mem_index,
- opc, &label2_ptr, 0);
- tcg_out_qemu_ld_direct(s, opc, data_reg, addr_reg, TCG_REG_NONE, 0);
- tcg_finish_qemu_ldst(s, label2_ptr);
+ add_qemu_ldst_label(s, 1, opc, data_reg, addr_reg, mem_index,
+ s->code_ptr, label_ptr);
#else
TCGReg index_reg;
tcg_target_long disp;
@@ -1530,13 +1550,19 @@ static void tcg_out_qemu_ld(TCGContext* s, TCGReg data_reg, TCGReg addr_reg,
static void tcg_out_qemu_st(TCGContext* s, TCGReg data_reg, TCGReg addr_reg,
TCGMemOp opc, int mem_index)
{
-#if defined(CONFIG_SOFTMMU)
- tcg_insn_unit *label2_ptr;
+#ifdef CONFIG_SOFTMMU
+ tcg_insn_unit *label_ptr;
+ TCGReg base_reg;
+
+ base_reg = tcg_out_tlb_read(s, addr_reg, opc, mem_index, 0);
+
+ label_ptr = s->code_ptr + 1;
+ tcg_out_insn(s, RI, BRC, S390_CC_NE, 0);
+
+ tcg_out_qemu_st_direct(s, opc, data_reg, base_reg, TCG_REG_R2, 0);
- addr_reg = tcg_prepare_qemu_ldst(s, data_reg, addr_reg, mem_index,
- opc, &label2_ptr, 1);
- tcg_out_qemu_st_direct(s, opc, data_reg, addr_reg, TCG_REG_NONE, 0);
- tcg_finish_qemu_ldst(s, label2_ptr);
+ add_qemu_ldst_label(s, 0, opc, data_reg, addr_reg, mem_index,
+ s->code_ptr, label_ptr);
#else
TCGReg index_reg;
tcg_target_long disp;
--
1.9.0
^ permalink raw reply related [flat|nested] 15+ messages in thread* Re: [Qemu-devel] [PULL 05/11] tcg-s390: Move ldst helpers out of line
2014-05-15 16:39 ` [Qemu-devel] [PULL 05/11] tcg-s390: Move ldst helpers out of line Richard Henderson
@ 2015-01-29 16:01 ` Alexander Graf
2015-01-29 17:12 ` Richard Henderson
0 siblings, 1 reply; 15+ messages in thread
From: Alexander Graf @ 2015-01-29 16:01 UTC (permalink / raw)
To: Richard Henderson; +Cc: peter.maydell, qemu-devel
> On 15.05.2014, at 18:39, Richard Henderson <rth@twiddle.net> wrote:
>
> That is, the old LDST_OPTIMIZATION.
>
> Signed-off-by: Richard Henderson <rth@twiddle.net>
> ---
> tcg/s390/tcg-target.c | 210 ++++++++++++++++++++++++++++----------------------
> 1 file changed, 118 insertions(+), 92 deletions(-)
>
> diff --git a/tcg/s390/tcg-target.c b/tcg/s390/tcg-target.c
> index 4fbadaf..1c955db 100644
> --- a/tcg/s390/tcg-target.c
> +++ b/tcg/s390/tcg-target.c
> @@ -24,7 +24,7 @@
> * THE SOFTWARE.
> */
>
> -#include "tcg-be-null.h"
> +#include "tcg-be-ldst.h"
>
> /* We only support generating code for 64-bit mode. */
> #if TCG_TARGET_REG_BITS != 64
> @@ -1386,107 +1386,123 @@ static void tcg_out_qemu_st_direct(TCGContext *s, TCGMemOp opc, TCGReg data,
> }
>
> #if defined(CONFIG_SOFTMMU)
> -static TCGReg tcg_prepare_qemu_ldst(TCGContext* s, TCGReg data_reg,
> - TCGReg addr_reg, int mem_index, int opc,
> - tcg_insn_unit **label2_ptr_p, int is_store)
> +/* We're expecting to use a 20-bit signed offset on the tlb memory ops.
> + Using the offset of the second entry in the last tlb table ensures
> + that we can index all of the elements of the first entry. */
> +QEMU_BUILD_BUG_ON(offsetof(CPUArchState, tlb_table[NB_MMU_MODES - 1][1])
> + > 0x7ffff);
> +
> +/* Load and compare a TLB entry, leaving the flags set. Loads the TLB
> + addend into R2. Returns a register with the santitized guest address. */
> +static TCGReg tcg_out_tlb_read(TCGContext* s, TCGReg addr_reg, TCGMemOp opc,
> + int mem_index, bool is_ld)
> {
> - const TCGReg arg0 = tcg_target_call_iarg_regs[0];
> - const TCGReg arg1 = tcg_target_call_iarg_regs[1];
> - const TCGReg arg2 = tcg_target_call_iarg_regs[2];
> - const TCGReg arg3 = tcg_target_call_iarg_regs[3];
> - const TCGReg arg4 = tcg_target_call_iarg_regs[4];
> TCGMemOp s_bits = opc & MO_SIZE;
> - tcg_insn_unit *label1_ptr;
> - tcg_target_long ofs;
> + int ofs;
> +
> + tcg_out_sh64(s, RSY_SRLG, TCG_REG_R2, addr_reg, TCG_REG_NONE,
> + TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS);
>
> if (TARGET_LONG_BITS == 32) {
> - tgen_ext32u(s, arg1, addr_reg);
> + tgen_ext32u(s, TCG_REG_R3, addr_reg);
> } else {
> - tcg_out_mov(s, TCG_TYPE_I64, arg1, addr_reg);
> + tcg_out_mov(s, TCG_TYPE_I64, TCG_REG_R3, addr_reg);
> }
>
> - tcg_out_sh64(s, RSY_SRLG, arg2, addr_reg, TCG_REG_NONE,
> - TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS);
> -
> - tgen_andi(s, TCG_TYPE_I64, arg1, TARGET_PAGE_MASK | ((1 << s_bits) - 1));
> - tgen_andi(s, TCG_TYPE_I64, arg2, (CPU_TLB_SIZE - 1) << CPU_TLB_ENTRY_BITS);
> + tgen_andi(s, TCG_TYPE_I64, TCG_REG_R2,
> + (CPU_TLB_SIZE - 1) << CPU_TLB_ENTRY_BITS);
> + tgen_andi(s, TCG_TYPE_I64, TCG_REG_R3,
> + TARGET_PAGE_MASK | ((1 << s_bits) - 1));
>
> - if (is_store) {
> - ofs = offsetof(CPUArchState, tlb_table[mem_index][0].addr_write);
> - } else {
> + if (is_ld) {
> ofs = offsetof(CPUArchState, tlb_table[mem_index][0].addr_read);
> + } else {
> + ofs = offsetof(CPUArchState, tlb_table[mem_index][0].addr_write);
> }
> - assert(ofs < 0x80000);
> -
> if (TARGET_LONG_BITS == 32) {
> - tcg_out_mem(s, RX_C, RXY_CY, arg1, arg2, TCG_AREG0, ofs);
> + tcg_out_mem(s, RX_C, RXY_CY, TCG_REG_R3, TCG_REG_R2, TCG_AREG0, ofs);
> } else {
> - tcg_out_mem(s, 0, RXY_CG, arg1, arg2, TCG_AREG0, ofs);
> + tcg_out_mem(s, 0, RXY_CG, TCG_REG_R3, TCG_REG_R2, TCG_AREG0, ofs);
> }
>
> + ofs = offsetof(CPUArchState, tlb_table[mem_index][0].addend);
> + tcg_out_mem(s, 0, RXY_LG, TCG_REG_R2, TCG_REG_R2, TCG_AREG0, ofs);
> +
> if (TARGET_LONG_BITS == 32) {
> - tgen_ext32u(s, arg1, addr_reg);
> - } else {
> - tcg_out_mov(s, TCG_TYPE_I64, arg1, addr_reg);
> - }
> -
> - label1_ptr = s->code_ptr;
> -
> - /* je label1 (offset will be patched in later) */
> - tcg_out_insn(s, RI, BRC, S390_CC_EQ, 0);
> -
> - /* call load/store helper */
> - if (is_store) {
> - /* Make sure to zero-extend the value to the full register
> - for the calling convention. */
> - switch (s_bits) {
> - case MO_UB:
> - tgen_ext8u(s, TCG_TYPE_I64, arg2, data_reg);
> - break;
> - case MO_UW:
> - tgen_ext16u(s, TCG_TYPE_I64, arg2, data_reg);
> - break;
> - case MO_UL:
> - tgen_ext32u(s, arg2, data_reg);
> - break;
> - case MO_Q:
> - tcg_out_mov(s, TCG_TYPE_I64, arg2, data_reg);
> - break;
> - default:
> - tcg_abort();
> - }
> - tcg_out_movi(s, TCG_TYPE_I32, arg3, mem_index);
> - tcg_out_mov(s, TCG_TYPE_PTR, arg0, TCG_AREG0);
> - tcg_out_movi(s, TCG_TYPE_PTR, arg4, (uintptr_t)s->code_ptr);
> - tcg_out_call(s, qemu_st_helpers[opc]);
> - } else {
> - tcg_out_movi(s, TCG_TYPE_I32, arg2, mem_index);
> - tcg_out_mov(s, TCG_TYPE_PTR, arg0, TCG_AREG0);
> - tcg_out_movi(s, TCG_TYPE_PTR, arg3, (uintptr_t)s->code_ptr);
> - tcg_out_call(s, qemu_ld_helpers[opc]);
> - tcg_out_mov(s, TCG_TYPE_I64, data_reg, TCG_REG_R2);
> + tgen_ext32u(s, TCG_REG_R3, addr_reg);
> + return TCG_REG_R3;
> }
> + return addr_reg;
> +}
>
> - /* jump to label2 (end) */
> - *label2_ptr_p = s->code_ptr;
> -
> - tcg_out_insn(s, RI, BRC, S390_CC_ALWAYS, 0);
> +static void add_qemu_ldst_label(TCGContext *s, bool is_ld, TCGMemOp opc,
> + TCGReg data, TCGReg addr, int mem_index,
> + tcg_insn_unit *raddr, tcg_insn_unit *label_ptr)
> +{
> + TCGLabelQemuLdst *label = new_ldst_label(s);
> +
> + label->is_ld = is_ld;
> + label->opc = opc;
> + label->datalo_reg = data;
> + label->addrlo_reg = addr;
> + label->mem_index = mem_index;
> + label->raddr = raddr;
> + label->label_ptr[0] = label_ptr;
> +}
>
> - /* this is label1, patch branch */
> - label1_ptr[1] = s->code_ptr - label1_ptr;
> +static void tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
> +{
> + TCGReg addr_reg = lb->addrlo_reg;
> + TCGReg data_reg = lb->datalo_reg;
> + TCGMemOp opc = lb->opc;
>
> - ofs = offsetof(CPUArchState, tlb_table[mem_index][0].addend);
> - assert(ofs < 0x80000);
> + patch_reloc(lb->label_ptr[0], R_390_PC16DBL, (intptr_t)s->code_ptr, -2);
>
> - tcg_out_mem(s, 0, RXY_AG, arg1, arg2, TCG_AREG0, ofs);
> + tcg_out_mov(s, TCG_TYPE_PTR, TCG_REG_R2, TCG_AREG0);
> + if (TARGET_LONG_BITS == 64) {
> + tcg_out_mov(s, TCG_TYPE_I64, TCG_REG_R3, addr_reg);
> + }
> + tcg_out_movi(s, TCG_TYPE_I32, TCG_REG_R4, lb->mem_index);
> + tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R5, (uintptr_t)lb->raddr);
> + tcg_out_call(s, qemu_ld_helpers[opc]);
> + tcg_out_mov(s, TCG_TYPE_I64, data_reg, TCG_REG_R2);
>
> - return arg1;
> + tgen_gotoi(s, S390_CC_ALWAYS, lb->raddr);
> }
>
> -static void tcg_finish_qemu_ldst(TCGContext* s, tcg_insn_unit *label2_ptr)
> +static void tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
> {
> - /* patch branch */
> - label2_ptr[1] = s->code_ptr - label2_ptr;
> + TCGReg addr_reg = lb->addrlo_reg;
> + TCGReg data_reg = lb->datalo_reg;
> + TCGMemOp opc = lb->opc;
> +
> + patch_reloc(lb->label_ptr[0], R_390_PC16DBL, (intptr_t)s->code_ptr, -2);
Under certain situations this seems to create a jump to itself:
(gdb)
0x000000009082a3f0 in code_gen_buffer ()
1: x/i $pc
=> 0x9082a3f0 <code_gen_buffer+8561648>:
jlh 0x9082a3f0 <code_gen_buffer+8561648>
Any idea why exactly this would happen? Or do we endless loop on purpose? ;)
Reverting this patch makes the guest work fine btw.
Alex
^ permalink raw reply [flat|nested] 15+ messages in thread* Re: [Qemu-devel] [PULL 05/11] tcg-s390: Move ldst helpers out of line
2015-01-29 16:01 ` Alexander Graf
@ 2015-01-29 17:12 ` Richard Henderson
0 siblings, 0 replies; 15+ messages in thread
From: Richard Henderson @ 2015-01-29 17:12 UTC (permalink / raw)
To: Alexander Graf; +Cc: peter.maydell, qemu-devel
On 01/29/2015 08:01 AM, Alexander Graf wrote:
>> -static void tcg_finish_qemu_ldst(TCGContext* s, tcg_insn_unit *label2_ptr)
>> +static void tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
>> {
>> - /* patch branch */
>> - label2_ptr[1] = s->code_ptr - label2_ptr;
>> + TCGReg addr_reg = lb->addrlo_reg;
>> + TCGReg data_reg = lb->datalo_reg;
>> + TCGMemOp opc = lb->opc;
>> +
>> + patch_reloc(lb->label_ptr[0], R_390_PC16DBL, (intptr_t)s->code_ptr, -2);
>
> Under certain situations this seems to create a jump to itself:
>
> (gdb)
> 0x000000009082a3f0 in code_gen_buffer ()
> 1: x/i $pc
> => 0x9082a3f0 <code_gen_buffer+8561648>:
> jlh 0x9082a3f0 <code_gen_buffer+8561648>
>
> Any idea why exactly this would happen? Or do we endless loop on purpose? ;)
You mean we actually arrive at this patch_reloc in that case?
Jump-to-self is, if I'm not mistaken, the uninitialized state
before we arrive here to output the slow path. I would have
expected this symptom to indicate that we *hadn't* patched a
jump rather than the reverse.
My first guess is that there's some sort of data corruption going on.
r~
^ permalink raw reply [flat|nested] 15+ messages in thread
* [Qemu-devel] [PULL 06/11] tcg-s390: Use more risbg in the tlb sequence
2014-05-15 16:39 [Qemu-devel] [PULL 00/12] tcg s390 updates Richard Henderson
` (4 preceding siblings ...)
2014-05-15 16:39 ` [Qemu-devel] [PULL 05/11] tcg-s390: Move ldst helpers out of line Richard Henderson
@ 2014-05-15 16:39 ` Richard Henderson
2014-05-15 16:39 ` [Qemu-devel] [PULL 07/11] tcg-s390: Implement tcg_register_jit Richard Henderson
` (5 subsequent siblings)
11 siblings, 0 replies; 15+ messages in thread
From: Richard Henderson @ 2014-05-15 16:39 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Elides two insns from the sequence. The resulting tlb compare
sequence is satisfyingly minimal:
risbg %r2,%r8,51,186,56
risbg %r3,%r8,61,178,0
cg %r3,904(%r10,%r2)
lg %r2,920(%r10,%r2)
jlh tlb_miss
Signed-off-by: Richard Henderson <rth@twiddle.net>
---
tcg/s390/tcg-target.c | 49 ++++++++++++++++++++++++++++---------------------
1 file changed, 28 insertions(+), 21 deletions(-)
diff --git a/tcg/s390/tcg-target.c b/tcg/s390/tcg-target.c
index 1c955db..cd60c01 100644
--- a/tcg/s390/tcg-target.c
+++ b/tcg/s390/tcg-target.c
@@ -933,6 +933,20 @@ static inline bool risbg_mask(uint64_t c)
return c == -lsb;
}
+static void tgen_andi_risbg(TCGContext *s, TCGReg out, TCGReg in, uint64_t val)
+{
+ int msb, lsb;
+ if ((val & 0x8000000000000001ull) == 0x8000000000000001ull) {
+ /* Achieve wraparound by swapping msb and lsb. */
+ msb = 64 - ctz64(~val);
+ lsb = clz64(~val) - 1;
+ } else {
+ msb = clz64(val);
+ lsb = 63 - ctz64(val);
+ }
+ tcg_out_risbg(s, out, in, msb, lsb, 0, 1);
+}
+
static void tgen_andi(TCGContext *s, TCGType type, TCGReg dest, uint64_t val)
{
static const S390Opcode ni_insns[4] = {
@@ -980,16 +994,7 @@ static void tgen_andi(TCGContext *s, TCGType type, TCGReg dest, uint64_t val)
}
}
if ((facilities & FACILITY_GEN_INST_EXT) && risbg_mask(val)) {
- int msb, lsb;
- if ((val & 0x8000000000000001ull) == 0x8000000000000001ull) {
- /* Achieve wraparound by swapping msb and lsb. */
- msb = 64 - ctz64(~val);
- lsb = clz64(~val) - 1;
- } else {
- msb = clz64(val);
- lsb = 63 - ctz64(val);
- }
- tcg_out_risbg(s, dest, dest, msb, lsb, 0, 1);
+ tgen_andi_risbg(s, dest, dest, val);
return;
}
@@ -1398,22 +1403,24 @@ static TCGReg tcg_out_tlb_read(TCGContext* s, TCGReg addr_reg, TCGMemOp opc,
int mem_index, bool is_ld)
{
TCGMemOp s_bits = opc & MO_SIZE;
+ uint64_t tlb_mask = TARGET_PAGE_MASK | ((1 << s_bits) - 1);
int ofs;
- tcg_out_sh64(s, RSY_SRLG, TCG_REG_R2, addr_reg, TCG_REG_NONE,
- TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS);
-
- if (TARGET_LONG_BITS == 32) {
- tgen_ext32u(s, TCG_REG_R3, addr_reg);
+ if (facilities & FACILITY_GEN_INST_EXT) {
+ tcg_out_risbg(s, TCG_REG_R2, addr_reg,
+ 64 - CPU_TLB_BITS - CPU_TLB_ENTRY_BITS,
+ 63 - CPU_TLB_ENTRY_BITS,
+ 64 + CPU_TLB_ENTRY_BITS - TARGET_PAGE_BITS, 1);
+ tgen_andi_risbg(s, TCG_REG_R3, addr_reg, tlb_mask);
} else {
- tcg_out_mov(s, TCG_TYPE_I64, TCG_REG_R3, addr_reg);
+ tcg_out_sh64(s, RSY_SRLG, TCG_REG_R2, addr_reg, TCG_REG_NONE,
+ TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS);
+ tcg_out_movi(s, TCG_TYPE_TL, TCG_REG_R3, addr_reg);
+ tgen_andi(s, TCG_TYPE_I64, TCG_REG_R2,
+ (CPU_TLB_SIZE - 1) << CPU_TLB_ENTRY_BITS);
+ tgen_andi(s, TCG_TYPE_TL, TCG_REG_R3, tlb_mask);
}
- tgen_andi(s, TCG_TYPE_I64, TCG_REG_R2,
- (CPU_TLB_SIZE - 1) << CPU_TLB_ENTRY_BITS);
- tgen_andi(s, TCG_TYPE_I64, TCG_REG_R3,
- TARGET_PAGE_MASK | ((1 << s_bits) - 1));
-
if (is_ld) {
ofs = offsetof(CPUArchState, tlb_table[mem_index][0].addr_read);
} else {
--
1.9.0
^ permalink raw reply related [flat|nested] 15+ messages in thread* [Qemu-devel] [PULL 07/11] tcg-s390: Implement tcg_register_jit
2014-05-15 16:39 [Qemu-devel] [PULL 00/12] tcg s390 updates Richard Henderson
` (5 preceding siblings ...)
2014-05-15 16:39 ` [Qemu-devel] [PULL 06/11] tcg-s390: Use more risbg in the tlb sequence Richard Henderson
@ 2014-05-15 16:39 ` Richard Henderson
2014-05-15 16:39 ` [Qemu-devel] [PULL 08/11] tcg-s390: Allow immediate operands to add2 and sub2 Richard Henderson
` (4 subsequent siblings)
11 siblings, 0 replies; 15+ messages in thread
From: Richard Henderson @ 2014-05-15 16:39 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Signed-off-by: Richard Henderson <rth@twiddle.net>
---
tcg/s390/tcg-target.c | 62 +++++++++++++++++++++++++++++++++++++++++++++------
1 file changed, 55 insertions(+), 7 deletions(-)
diff --git a/tcg/s390/tcg-target.c b/tcg/s390/tcg-target.c
index cd60c01..2f0cdf8 100644
--- a/tcg/s390/tcg-target.c
+++ b/tcg/s390/tcg-target.c
@@ -2181,18 +2181,17 @@ static void tcg_target_init(TCGContext *s)
tcg_add_target_add_op_defs(s390_op_defs);
}
+#define FRAME_SIZE ((int)(TCG_TARGET_CALL_STACK_OFFSET \
+ + TCG_STATIC_CALL_ARGS_SIZE \
+ + CPU_TEMP_BUF_NLONGS * sizeof(long)))
+
static void tcg_target_qemu_prologue(TCGContext *s)
{
- tcg_target_long frame_size;
-
/* stmg %r6,%r15,48(%r15) (save registers) */
tcg_out_insn(s, RXY, STMG, TCG_REG_R6, TCG_REG_R15, TCG_REG_R15, 48);
/* aghi %r15,-frame_size */
- frame_size = TCG_TARGET_CALL_STACK_OFFSET;
- frame_size += TCG_STATIC_CALL_ARGS_SIZE;
- frame_size += CPU_TEMP_BUF_NLONGS * sizeof(long);
- tcg_out_insn(s, RI, AGHI, TCG_REG_R15, -frame_size);
+ tcg_out_insn(s, RI, AGHI, TCG_REG_R15, -FRAME_SIZE);
tcg_set_frame(s, TCG_REG_CALL_STACK,
TCG_STATIC_CALL_ARGS_SIZE + TCG_TARGET_CALL_STACK_OFFSET,
@@ -2211,8 +2210,57 @@ static void tcg_target_qemu_prologue(TCGContext *s)
/* lmg %r6,%r15,fs+48(%r15) (restore registers) */
tcg_out_insn(s, RXY, LMG, TCG_REG_R6, TCG_REG_R15, TCG_REG_R15,
- frame_size + 48);
+ FRAME_SIZE + 48);
/* br %r14 (return) */
tcg_out_insn(s, RR, BCR, S390_CC_ALWAYS, TCG_REG_R14);
}
+
+typedef struct {
+ DebugFrameCIE cie;
+ DebugFrameFDEHeader fde;
+ uint8_t fde_def_cfa[4];
+ uint8_t fde_reg_ofs[18];
+} DebugFrame;
+
+/* We're expecting a 2 byte uleb128 encoded value. */
+QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
+
+#define ELF_HOST_MACHINE EM_S390
+
+static DebugFrame debug_frame = {
+ .cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
+ .cie.id = -1,
+ .cie.version = 1,
+ .cie.code_align = 1,
+ .cie.data_align = 8, /* sleb128 8 */
+ .cie.return_column = TCG_REG_R14,
+
+ /* Total FDE size does not include the "len" member. */
+ .fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, fde.cie_offset),
+
+ .fde_def_cfa = {
+ 12, TCG_REG_CALL_STACK, /* DW_CFA_def_cfa %r15, ... */
+ (FRAME_SIZE & 0x7f) | 0x80, /* ... uleb128 FRAME_SIZE */
+ (FRAME_SIZE >> 7)
+ },
+ .fde_reg_ofs = {
+ 0x86, 6, /* DW_CFA_offset, %r6, 48 */
+ 0x87, 7, /* DW_CFA_offset, %r7, 56 */
+ 0x88, 8, /* DW_CFA_offset, %r8, 64 */
+ 0x89, 9, /* DW_CFA_offset, %r92, 72 */
+ 0x8a, 10, /* DW_CFA_offset, %r10, 80 */
+ 0x8b, 11, /* DW_CFA_offset, %r11, 88 */
+ 0x8c, 12, /* DW_CFA_offset, %r12, 96 */
+ 0x8d, 13, /* DW_CFA_offset, %r13, 104 */
+ 0x8e, 14, /* DW_CFA_offset, %r14, 112 */
+ }
+};
+
+void tcg_register_jit(void *buf, size_t buf_size)
+{
+ debug_frame.fde.func_start = (uintptr_t)buf;
+ debug_frame.fde.func_len = buf_size;
+
+ tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
+}
--
1.9.0
^ permalink raw reply related [flat|nested] 15+ messages in thread* [Qemu-devel] [PULL 08/11] tcg-s390: Allow immediate operands to add2 and sub2
2014-05-15 16:39 [Qemu-devel] [PULL 00/12] tcg s390 updates Richard Henderson
` (6 preceding siblings ...)
2014-05-15 16:39 ` [Qemu-devel] [PULL 07/11] tcg-s390: Implement tcg_register_jit Richard Henderson
@ 2014-05-15 16:39 ` Richard Henderson
2014-05-15 16:39 ` [Qemu-devel] [PULL 09/11] tcg-s390: Improve setcond Richard Henderson
` (3 subsequent siblings)
11 siblings, 0 replies; 15+ messages in thread
From: Richard Henderson @ 2014-05-15 16:39 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Signed-off-by: Richard Henderson <rth@twiddle.net>
---
tcg/s390/tcg-target.c | 64 +++++++++++++++++++++++++++++++++++++++++----------
1 file changed, 52 insertions(+), 12 deletions(-)
diff --git a/tcg/s390/tcg-target.c b/tcg/s390/tcg-target.c
index 2f0cdf8..e912f43 100644
--- a/tcg/s390/tcg-target.c
+++ b/tcg/s390/tcg-target.c
@@ -42,6 +42,7 @@
#define TCG_CT_CONST_ORI 0x200
#define TCG_CT_CONST_XORI 0x400
#define TCG_CT_CONST_CMPI 0x800
+#define TCG_CT_CONST_ADLI 0x1000
/* Several places within the instruction set 0 means "no register"
rather than TCG_REG_R0. */
@@ -403,6 +404,9 @@ static int target_parse_constraint(TCGArgConstraint *ct, const char **pct_str)
tcg_regset_clear(ct->u.regs);
tcg_regset_set_reg(ct->u.regs, TCG_REG_R3);
break;
+ case 'A':
+ ct->ct |= TCG_CT_CONST_ADLI;
+ break;
case 'K':
ct->ct |= TCG_CT_CONST_MULI;
break;
@@ -507,6 +511,20 @@ static int tcg_match_cmpi(TCGType type, tcg_target_long val)
}
}
+/* Immediates to be used with add2/sub2. */
+
+static int tcg_match_add2i(TCGType type, tcg_target_long val)
+{
+ if (facilities & FACILITY_EXT_IMM) {
+ if (type == TCG_TYPE_I32) {
+ return 1;
+ } else if (val >= -0xffffffffll && val <= 0xffffffffll) {
+ return 1;
+ }
+ }
+ return 0;
+}
+
/* Test if a constant matches the constraint. */
static int tcg_target_const_match(tcg_target_long val, TCGType type,
const TCGArgConstraint *arg_ct)
@@ -532,6 +550,8 @@ static int tcg_target_const_match(tcg_target_long val, TCGType type,
} else {
return val == (int16_t)val;
}
+ } else if (ct & TCG_CT_CONST_ADLI) {
+ return tcg_match_add2i(type, val);
} else if (ct & TCG_CT_CONST_ORI) {
return tcg_match_ori(type, val);
} else if (ct & TCG_CT_CONST_XORI) {
@@ -1780,13 +1800,19 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
break;
case INDEX_op_add2_i32:
- /* ??? Make use of ALFI. */
- tcg_out_insn(s, RR, ALR, args[0], args[4]);
+ if (const_args[4]) {
+ tcg_out_insn(s, RIL, ALFI, args[0], args[4]);
+ } else {
+ tcg_out_insn(s, RR, ALR, args[0], args[4]);
+ }
tcg_out_insn(s, RRE, ALCR, args[1], args[5]);
break;
case INDEX_op_sub2_i32:
- /* ??? Make use of SLFI. */
- tcg_out_insn(s, RR, SLR, args[0], args[4]);
+ if (const_args[4]) {
+ tcg_out_insn(s, RIL, SLFI, args[0], args[4]);
+ } else {
+ tcg_out_insn(s, RR, SLR, args[0], args[4]);
+ }
tcg_out_insn(s, RRE, SLBR, args[1], args[5]);
break;
@@ -1987,13 +2013,27 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
break;
case INDEX_op_add2_i64:
- /* ??? Make use of ALGFI and SLGFI. */
- tcg_out_insn(s, RRE, ALGR, args[0], args[4]);
+ if (const_args[4]) {
+ if ((int64_t)args[4] >= 0) {
+ tcg_out_insn(s, RIL, ALGFI, args[0], args[4]);
+ } else {
+ tcg_out_insn(s, RIL, SLGFI, args[0], -args[4]);
+ }
+ } else {
+ tcg_out_insn(s, RRE, ALGR, args[0], args[4]);
+ }
tcg_out_insn(s, RRE, ALCGR, args[1], args[5]);
break;
case INDEX_op_sub2_i64:
- /* ??? Make use of ALGFI and SLGFI. */
- tcg_out_insn(s, RRE, SLGR, args[0], args[4]);
+ if (const_args[4]) {
+ if ((int64_t)args[4] >= 0) {
+ tcg_out_insn(s, RIL, SLGFI, args[0], args[4]);
+ } else {
+ tcg_out_insn(s, RIL, ALGFI, args[0], -args[4]);
+ }
+ } else {
+ tcg_out_insn(s, RRE, SLGR, args[0], args[4]);
+ }
tcg_out_insn(s, RRE, SLBGR, args[1], args[5]);
break;
@@ -2066,8 +2106,8 @@ static const TCGTargetOpDef s390_op_defs[] = {
{ INDEX_op_bswap16_i32, { "r", "r" } },
{ INDEX_op_bswap32_i32, { "r", "r" } },
- { INDEX_op_add2_i32, { "r", "r", "0", "1", "r", "r" } },
- { INDEX_op_sub2_i32, { "r", "r", "0", "1", "r", "r" } },
+ { INDEX_op_add2_i32, { "r", "r", "0", "1", "rA", "r" } },
+ { INDEX_op_sub2_i32, { "r", "r", "0", "1", "rA", "r" } },
{ INDEX_op_brcond_i32, { "r", "rC" } },
{ INDEX_op_setcond_i32, { "r", "r", "rC" } },
@@ -2124,8 +2164,8 @@ static const TCGTargetOpDef s390_op_defs[] = {
{ INDEX_op_bswap32_i64, { "r", "r" } },
{ INDEX_op_bswap64_i64, { "r", "r" } },
- { INDEX_op_add2_i64, { "r", "r", "0", "1", "r", "r" } },
- { INDEX_op_sub2_i64, { "r", "r", "0", "1", "r", "r" } },
+ { INDEX_op_add2_i64, { "r", "r", "0", "1", "rA", "r" } },
+ { INDEX_op_sub2_i64, { "r", "r", "0", "1", "rA", "r" } },
{ INDEX_op_brcond_i64, { "r", "rC" } },
{ INDEX_op_setcond_i64, { "r", "r", "rC" } },
--
1.9.0
^ permalink raw reply related [flat|nested] 15+ messages in thread* [Qemu-devel] [PULL 09/11] tcg-s390: Improve setcond
2014-05-15 16:39 [Qemu-devel] [PULL 00/12] tcg s390 updates Richard Henderson
` (7 preceding siblings ...)
2014-05-15 16:39 ` [Qemu-devel] [PULL 08/11] tcg-s390: Allow immediate operands to add2 and sub2 Richard Henderson
@ 2014-05-15 16:39 ` Richard Henderson
2014-05-15 16:39 ` [Qemu-devel] [PULL 10/11] tcg-s390: Don't force -march=z990 Richard Henderson
` (2 subsequent siblings)
11 siblings, 0 replies; 15+ messages in thread
From: Richard Henderson @ 2014-05-15 16:39 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
There are a variety of common cases for which we can use carry tricks to
avoid a conditional branch. On very new hardware, use LOAD ON CONDITION
instead of a conditional branch.
Signed-off-by: Richard Henderson <rth@twiddle.net>
---
tcg/s390/tcg-target.c | 97 +++++++++++++++++++++++++++++++++++++++++++++++----
1 file changed, 91 insertions(+), 6 deletions(-)
diff --git a/tcg/s390/tcg-target.c b/tcg/s390/tcg-target.c
index e912f43..e235130 100644
--- a/tcg/s390/tcg-target.c
+++ b/tcg/s390/tcg-target.c
@@ -1135,15 +1135,100 @@ static int tgen_cmp(TCGContext *s, TCGType type, TCGCond c, TCGReg r1,
return tcg_cond_to_s390_cond[c];
}
-static void tgen_setcond(TCGContext *s, TCGType type, TCGCond c,
+static void tgen_setcond(TCGContext *s, TCGType type, TCGCond cond,
TCGReg dest, TCGReg c1, TCGArg c2, int c2const)
{
- int cc = tgen_cmp(s, type, c, c1, c2, c2const);
+ int cc;
+
+ switch (cond) {
+ case TCG_COND_GTU:
+ case TCG_COND_GT:
+ do_greater:
+ /* The result of a compare has CC=2 for GT and CC=3 unused.
+ ADD LOGICAL WITH CARRY considers (CC & 2) the carry bit. */
+ tgen_cmp(s, type, cond, c1, c2, c2const);
+ tcg_out_movi(s, type, dest, 0);
+ tcg_out_insn(s, RRE, ALCGR, dest, dest);
+ return;
+
+ case TCG_COND_GEU:
+ do_geu:
+ /* We need "real" carry semantics, so use SUBTRACT LOGICAL
+ instead of COMPARE LOGICAL. This needs an extra move. */
+ tcg_out_mov(s, type, TCG_TMP0, c1);
+ if (c2const) {
+ tcg_out_movi(s, TCG_TYPE_I64, dest, 0);
+ if (type == TCG_TYPE_I32) {
+ tcg_out_insn(s, RIL, SLFI, TCG_TMP0, c2);
+ } else {
+ tcg_out_insn(s, RIL, SLGFI, TCG_TMP0, c2);
+ }
+ } else {
+ if (type == TCG_TYPE_I32) {
+ tcg_out_insn(s, RR, SLR, TCG_TMP0, c2);
+ } else {
+ tcg_out_insn(s, RRE, SLGR, TCG_TMP0, c2);
+ }
+ tcg_out_movi(s, TCG_TYPE_I64, dest, 0);
+ }
+ tcg_out_insn(s, RRE, ALCGR, dest, dest);
+ return;
+
+ case TCG_COND_LEU:
+ case TCG_COND_LTU:
+ case TCG_COND_LT:
+ /* Swap operands so that we can use GEU/GTU/GT. */
+ if (c2const) {
+ tcg_out_movi(s, type, TCG_TMP0, c2);
+ c2 = c1;
+ c2const = 0;
+ c1 = TCG_TMP0;
+ } else {
+ TCGReg t = c1;
+ c1 = c2;
+ c2 = t;
+ }
+ if (cond == TCG_COND_LEU) {
+ goto do_geu;
+ }
+ cond = tcg_swap_cond(cond);
+ goto do_greater;
+
+ case TCG_COND_NE:
+ /* X != 0 is X > 0. */
+ if (c2const && c2 == 0) {
+ cond = TCG_COND_GTU;
+ goto do_greater;
+ }
+ break;
- /* Emit: r1 = 1; if (cc) goto over; r1 = 0; over: */
- tcg_out_movi(s, type, dest, 1);
- tcg_out_insn(s, RI, BRC, cc, (4 + 4) >> 1);
- tcg_out_movi(s, type, dest, 0);
+ case TCG_COND_EQ:
+ /* X == 0 is X <= 0 is 0 >= X. */
+ if (c2const && c2 == 0) {
+ tcg_out_movi(s, TCG_TYPE_I64, TCG_TMP0, 0);
+ c2 = c1;
+ c2const = 0;
+ c1 = TCG_TMP0;
+ goto do_geu;
+ }
+ break;
+
+ default:
+ break;
+ }
+
+ cc = tgen_cmp(s, type, cond, c1, c2, c2const);
+ if (facilities & FACILITY_LOAD_ON_COND) {
+ /* Emit: d = 0, t = 1, d = (cc ? t : d). */
+ tcg_out_movi(s, TCG_TYPE_I64, dest, 0);
+ tcg_out_movi(s, TCG_TYPE_I64, TCG_TMP0, 1);
+ tcg_out_insn(s, RRF, LOCGR, dest, TCG_TMP0, cc);
+ } else {
+ /* Emit: d = 1; if (cc) goto over; d = 0; over: */
+ tcg_out_movi(s, type, dest, 1);
+ tcg_out_insn(s, RI, BRC, cc, (4 + 4) >> 1);
+ tcg_out_movi(s, type, dest, 0);
+ }
}
static void tgen_movcond(TCGContext *s, TCGType type, TCGCond c, TCGReg dest,
--
1.9.0
^ permalink raw reply related [flat|nested] 15+ messages in thread* [Qemu-devel] [PULL 10/11] tcg-s390: Don't force -march=z990
2014-05-15 16:39 [Qemu-devel] [PULL 00/12] tcg s390 updates Richard Henderson
` (8 preceding siblings ...)
2014-05-15 16:39 ` [Qemu-devel] [PULL 09/11] tcg-s390: Improve setcond Richard Henderson
@ 2014-05-15 16:39 ` Richard Henderson
2014-05-15 16:39 ` [Qemu-devel] [PULL 11/11] tcg-s390: Implement direct chaining of TBs Richard Henderson
2014-05-15 18:45 ` [Qemu-devel] [PULL 00/12] tcg s390 updates Peter Maydell
11 siblings, 0 replies; 15+ messages in thread
From: Richard Henderson @ 2014-05-15 16:39 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
While we still require the LONG DISPLACEMENT facility, defaults
have moved on since then. Don't override the system compiler,
whose default may be set to z9-109 or later.
Acked-by: Christian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: Richard Henderson <rth@twiddle.net>
---
configure | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/configure b/configure
index 6adfa72..332eea0 100755
--- a/configure
+++ b/configure
@@ -1137,11 +1137,11 @@ case "$cpu" in
CPU_CFLAGS="-m64 -mcpu=ultrasparc"
;;
s390)
- CPU_CFLAGS="-m31 -march=z990"
+ CPU_CFLAGS="-m31"
LDFLAGS="-m31 $LDFLAGS"
;;
s390x)
- CPU_CFLAGS="-m64 -march=z990"
+ CPU_CFLAGS="-m64"
LDFLAGS="-m64 $LDFLAGS"
;;
i386)
--
1.9.0
^ permalink raw reply related [flat|nested] 15+ messages in thread* [Qemu-devel] [PULL 11/11] tcg-s390: Implement direct chaining of TBs
2014-05-15 16:39 [Qemu-devel] [PULL 00/12] tcg s390 updates Richard Henderson
` (9 preceding siblings ...)
2014-05-15 16:39 ` [Qemu-devel] [PULL 10/11] tcg-s390: Don't force -march=z990 Richard Henderson
@ 2014-05-15 16:39 ` Richard Henderson
2014-05-15 18:45 ` [Qemu-devel] [PULL 00/12] tcg s390 updates Peter Maydell
11 siblings, 0 replies; 15+ messages in thread
From: Richard Henderson @ 2014-05-15 16:39 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Signed-off-by: Richard Henderson <rth@twiddle.net>
---
include/exec/exec-all.h | 9 +++++++++
tcg/s390/tcg-target.c | 4 +++-
2 files changed, 12 insertions(+), 1 deletion(-)
diff --git a/include/exec/exec-all.h b/include/exec/exec-all.h
index 444b4d9..8bc2eb6 100644
--- a/include/exec/exec-all.h
+++ b/include/exec/exec-all.h
@@ -131,6 +131,7 @@ static inline void tlb_flush(CPUState *cpu, int flush_global)
#if defined(__arm__) || defined(_ARCH_PPC) \
|| defined(__x86_64__) || defined(__i386__) \
|| defined(__sparc__) || defined(__aarch64__) \
+ || defined(__s390x__) \
|| defined(CONFIG_TCG_INTERPRETER)
#define USE_DIRECT_JUMP
#endif
@@ -232,6 +233,14 @@ static inline void tb_set_jmp_target1(uintptr_t jmp_addr, uintptr_t addr)
stl_le_p((void*)jmp_addr, addr - (jmp_addr + 4));
/* no need to flush icache explicitly */
}
+#elif defined(__s390x__)
+static inline void tb_set_jmp_target1(uintptr_t jmp_addr, uintptr_t addr)
+{
+ /* patch the branch destination */
+ intptr_t disp = addr - (jmp_addr - 2);
+ stl_be_p((void*)jmp_addr, disp / 2);
+ /* no need to flush icache explicitly */
+}
#elif defined(__aarch64__)
void aarch64_tb_set_jmp_target(uintptr_t jmp_addr, uintptr_t addr);
#define tb_set_jmp_target1 aarch64_tb_set_jmp_target
diff --git a/tcg/s390/tcg-target.c b/tcg/s390/tcg-target.c
index e235130..07164e5 100644
--- a/tcg/s390/tcg-target.c
+++ b/tcg/s390/tcg-target.c
@@ -1703,7 +1703,9 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
case INDEX_op_goto_tb:
if (s->tb_jmp_offset) {
- tcg_abort();
+ tcg_out16(s, RIL_BRCL | (S390_CC_ALWAYS << 4));
+ s->tb_jmp_offset[args[0]] = tcg_current_code_size(s);
+ s->code_ptr += 2;
} else {
/* load address stored at s->tb_next + args[0] */
tcg_out_ld_abs(s, TCG_TYPE_PTR, TCG_TMP0, s->tb_next + args[0]);
--
1.9.0
^ permalink raw reply related [flat|nested] 15+ messages in thread* Re: [Qemu-devel] [PULL 00/12] tcg s390 updates
2014-05-15 16:39 [Qemu-devel] [PULL 00/12] tcg s390 updates Richard Henderson
` (10 preceding siblings ...)
2014-05-15 16:39 ` [Qemu-devel] [PULL 11/11] tcg-s390: Implement direct chaining of TBs Richard Henderson
@ 2014-05-15 18:45 ` Peter Maydell
11 siblings, 0 replies; 15+ messages in thread
From: Peter Maydell @ 2014-05-15 18:45 UTC (permalink / raw)
To: Richard Henderson; +Cc: QEMU Developers, Thomas Huth
On 15 May 2014 17:39, Richard Henderson <rth@twiddle.net> wrote:
> This version differs from the patch set I posted 2 weeks ago in that
> the andi bugfix is split out separately, a typo in the add2/sub2 patch
> is fixed, and a long forgotten patch to implement direct branches
> between TBs was merged.
>
> In addition to my own testing, Thomas Huth was kind enough to do a quick
> check against the bug he originally reported to me.
>
>
> r~
>
>
> The following changes since commit 450445d543a33fdb1d18a18ab7669a33dc1ca7ae:
>
> tcg: Fix tcg_reg_alloc_mov vs no-op truncation (2014-05-14 09:56:13 -0700)
>
> are available in the git repository at:
>
> git://github.com/rth7680/qemu.git tags/pull-tcg-s390-20140515
>
> for you to fetch changes up to a10c64e0df3acbefdee1b2b837f76f05b0231aae:
>
> tcg-s390: Implement direct chaining of TBs (2014-05-15 09:22:32 -0700)
>
> ----------------------------------------------------------------
> tcg/s390 updates
>
Applied, thanks.
-- PMM
^ permalink raw reply [flat|nested] 15+ messages in thread