[PATCH v8 00/62] target/riscv: support vector extension v0.7.1

[PATCH v8 00/62] target/riscv: support vector extension v0.7.1

[LIU Zhiwei]

This patchset implements the vector extension for RISC-V on QEMU.

You can also find the patchset and all *test cases* in
my repo(https://github.com/romanheros/qemu.git branch:vector-upstream-v8).
All the test cases are in the directory qemu/tests/riscv/vector/. They are
riscv64 linux user mode programs.

You can test the patchset by the script qemu/tests/riscv/vector/runcase.sh.

Features:
  * support specification riscv-v-spec-0.7.1.(https://github.com/riscv/riscv-v-spec/releases/tag/0.7.1/)
  * support basic vector extension.
  * support Zvlsseg.
  * support Zvamo.
  * not support Zvediv as it is changing.
  * SLEN always equals VLEN.
  * element width support 8bit, 16bit, 32bit, 64bit.

Changelog:
v8
  * support different float rounding modes for vector instructions.
  * use lastest released TCG GVEC DUP IR.
  * set RV_VLEN_MAX to 256 bits, as GVEC IR uses simd_desc.

v7
  * move vl == 0 check to translation time by add a global cpu_vl.
  * implement vector element inline load and store function by TCG IR.
  * based on vec_element_load(store), implement some permutation instructions.
  * implement rsubs GVEC IR.
  * fixup vsmul, vmfne, vfmerge, vslidedown.
  * some other small bugs and indentation errors.

v6
  * use gvec_dup Gvec IR to accellerate move and merge.
  * a better way to implement fixed point instructions.
  * a global check when vl == 0.
  * limit some macros to only one inline function call.
  * fixup sew error when use Gvec IR.
  * fixup bugs for corner cases.

v5
  * fixup a bug in tb flags.

v4
  * no change

v3
  * move check code from execution-time to translation-time
  * use a continous memory block for vector register description.
  * vector registers as direct fields in RISCVCPUState.
  * support VLEN configure from qemu command line.
  * support ELEN configure from qemu command line.
  * support vector specification version configure from qemu command line.
  * probe pages before real load or store access.
  * use probe_page_check for no-fault operations in linux user mode.
  * generation atomic exit exception when in parallel environment.
  * fixup a lot of concrete bugs.

V2
  * use float16_compare{_quiet}
  * only use GETPC() in outer most helper
  * add ctx.ext_v Property

LIU Zhiwei (62):
  target/riscv: add vector extension field in CPURISCVState
  target/riscv: implementation-defined constant parameters
  target/riscv: support vector extension csr
  target/riscv: add vector configure instruction
  target/riscv: add an internals.h header
  target/riscv: add vector stride load and store instructions
  target/riscv: add vector index load and store instructions
  target/riscv: add fault-only-first unit stride load
  target/riscv: add vector amo operations
  target/riscv: vector single-width integer add and subtract
  target/riscv: vector widening integer add and subtract
  target/riscv: vector integer add-with-carry / subtract-with-borrow
    instructions
  target/riscv: vector bitwise logical instructions
  target/riscv: vector single-width bit shift instructions
  target/riscv: vector narrowing integer right shift instructions
  target/riscv: vector integer comparison instructions
  target/riscv: vector integer min/max instructions
  target/riscv: vector single-width integer multiply instructions
  target/riscv: vector integer divide instructions
  target/riscv: vector widening integer multiply instructions
  target/riscv: vector single-width integer multiply-add instructions
  target/riscv: vector widening integer multiply-add instructions
  target/riscv: vector integer merge and move instructions
  target/riscv: vector single-width saturating add and subtract
  target/riscv: vector single-width averaging add and subtract
  target/riscv: vector single-width fractional multiply with rounding
    and saturation
  target/riscv: vector widening saturating scaled multiply-add
  target/riscv: vector single-width scaling shift instructions
  target/riscv: vector narrowing fixed-point clip instructions
  target/riscv: Update fp_status when float rounding mode changes
  target/riscv: vector single-width floating-point add/subtract
    instructions
  target/riscv: vector widening floating-point add/subtract instructions
  target/riscv: vector single-width floating-point multiply/divide
    instructions
  target/riscv: vector widening floating-point multiply
  target/riscv: vector single-width floating-point fused multiply-add
    instructions
  target/riscv: vector widening floating-point fused multiply-add
    instructions
  target/riscv: vector floating-point square-root instruction
  target/riscv: vector floating-point min/max instructions
  target/riscv: vector floating-point sign-injection instructions
  target/riscv: vector floating-point compare instructions
  target/riscv: vector floating-point classify instructions
  target/riscv: vector floating-point merge instructions
  target/riscv: vector floating-point/integer type-convert instructions
  target/riscv: widening floating-point/integer type-convert
    instructions
  target/riscv: narrowing floating-point/integer type-convert
    instructions
  target/riscv: vector single-width integer reduction instructions
  target/riscv: vector wideing integer reduction instructions
  target/riscv: vector single-width floating-point reduction
    instructions
  target/riscv: vector widening floating-point reduction instructions
  target/riscv: vector mask-register logical instructions
  target/riscv: vector mask population count vmpopc
  target/riscv: vmfirst find-first-set mask bit
  target/riscv: set-X-first mask bit
  target/riscv: vector iota instruction
  target/riscv: vector element index instruction
  target/riscv: integer extract instruction
  target/riscv: integer scalar move instruction
  target/riscv: floating-point scalar move instructions
  target/riscv: vector slide instructions
  target/riscv: vector register gather instruction
  target/riscv: vector compress instruction
  target/riscv: configure and turn on vector extension from command line

 target/riscv/Makefile.objs              |    2 +-
 target/riscv/cpu.c                      |   49 +
 target/riscv/cpu.h                      |   82 +-
 target/riscv/cpu_bits.h                 |   15 +
 target/riscv/csr.c                      |   82 +-
 target/riscv/fpu_helper.c               |   52 +-
 target/riscv/helper.h                   | 1068 +++++
 target/riscv/insn32-64.decode           |   11 +
 target/riscv/insn32.decode              |  372 ++
 target/riscv/insn_trans/trans_rvv.inc.c | 2884 +++++++++++++
 target/riscv/internals.h                |   44 +
 target/riscv/translate.c                |   27 +-
 target/riscv/vector_helper.c            | 4898 +++++++++++++++++++++++
 13 files changed, 9537 insertions(+), 49 deletions(-)
 create mode 100644 target/riscv/insn_trans/trans_rvv.inc.c
 create mode 100644 target/riscv/internals.h
 create mode 100644 target/riscv/vector_helper.c

-- 
2.23.0

[PATCH v8 01/62] target/riscv: add vector extension field in CPURISCVState

[LIU Zhiwei]

The 32 vector registers will be viewed as a continuous memory block.
It avoids the convension between element index and (regno, offset).
Thus elements can be directly accessed by offset from the first vector
base address.

Signed-off-by: LIU Zhiwei <zhiwei_liu@c-sky.com>
Acked-by: Alistair Francis <alistair.francis@wdc.com>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
---
 target/riscv/cpu.h       | 12 ++++++++++++
 target/riscv/translate.c |  3 ++-
 2 files changed, 14 insertions(+), 1 deletion(-)

diff --git a/target/riscv/cpu.h b/target/riscv/cpu.h
index d0e7f5b9c5..7452c6e118 100644
--- a/target/riscv/cpu.h
+++ b/target/riscv/cpu.h
@@ -65,6 +65,7 @@
 #define RVA RV('A')
 #define RVF RV('F')
 #define RVD RV('D')
+#define RVV RV('V')
 #define RVC RV('C')
 #define RVS RV('S')
 #define RVU RV('U')
@@ -95,9 +96,20 @@ typedef struct CPURISCVState CPURISCVState;
 
 #include "pmp.h"
 
+#define RV_VLEN_MAX 512
+
 struct CPURISCVState {
     target_ulong gpr[32];
     uint64_t fpr[32]; /* assume both F and D extensions */
+
+    /* vector coprocessor state. */
+    uint64_t vreg[32 * RV_VLEN_MAX / 64] QEMU_ALIGNED(16);
+    target_ulong vxrm;
+    target_ulong vxsat;
+    target_ulong vl;
+    target_ulong vstart;
+    target_ulong vtype;
+
     target_ulong pc;
     target_ulong load_res;
     target_ulong load_val;
diff --git a/target/riscv/translate.c b/target/riscv/translate.c
index 43bf7e39a6..b71b7e4bc2 100644
--- a/target/riscv/translate.c
+++ b/target/riscv/translate.c
@@ -32,7 +32,7 @@
 #include "instmap.h"
 
 /* global register indices */
-static TCGv cpu_gpr[32], cpu_pc;
+static TCGv cpu_gpr[32], cpu_pc, cpu_vl;
 static TCGv_i64 cpu_fpr[32]; /* assume F and D extensions */
 static TCGv load_res;
 static TCGv load_val;
@@ -886,6 +886,7 @@ void riscv_translate_init(void)
     }
 
     cpu_pc = tcg_global_mem_new(cpu_env, offsetof(CPURISCVState, pc), "pc");
+    cpu_vl = tcg_global_mem_new(cpu_env, offsetof(CPURISCVState, vl), "vl");
     load_res = tcg_global_mem_new(cpu_env, offsetof(CPURISCVState, load_res),
                              "load_res");
     load_val = tcg_global_mem_new(cpu_env, offsetof(CPURISCVState, load_val),
-- 
2.23.0

[PATCH v8 01/62] target/riscv: add vector extension field in CPURISCVState

[LIU Zhiwei]

The 32 vector registers will be viewed as a continuous memory block.
It avoids the convension between element index and (regno, offset).
Thus elements can be directly accessed by offset from the first vector
base address.

Signed-off-by: LIU Zhiwei <zhiwei_liu@c-sky.com>
Acked-by: Alistair Francis <alistair.francis@wdc.com>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
---
 target/riscv/cpu.h       | 12 ++++++++++++
 target/riscv/translate.c |  3 ++-
 2 files changed, 14 insertions(+), 1 deletion(-)

diff --git a/target/riscv/cpu.h b/target/riscv/cpu.h
index d0e7f5b9c5..7452c6e118 100644
--- a/target/riscv/cpu.h
+++ b/target/riscv/cpu.h
@@ -65,6 +65,7 @@
 #define RVA RV('A')
 #define RVF RV('F')
 #define RVD RV('D')
+#define RVV RV('V')
 #define RVC RV('C')
 #define RVS RV('S')
 #define RVU RV('U')
@@ -95,9 +96,20 @@ typedef struct CPURISCVState CPURISCVState;
 
 #include "pmp.h"
 
+#define RV_VLEN_MAX 512
+
 struct CPURISCVState {
     target_ulong gpr[32];
     uint64_t fpr[32]; /* assume both F and D extensions */
+
+    /* vector coprocessor state. */
+    uint64_t vreg[32 * RV_VLEN_MAX / 64] QEMU_ALIGNED(16);
+    target_ulong vxrm;
+    target_ulong vxsat;
+    target_ulong vl;
+    target_ulong vstart;
+    target_ulong vtype;
+
     target_ulong pc;
     target_ulong load_res;
     target_ulong load_val;
diff --git a/target/riscv/translate.c b/target/riscv/translate.c
index 43bf7e39a6..b71b7e4bc2 100644
--- a/target/riscv/translate.c
+++ b/target/riscv/translate.c
@@ -32,7 +32,7 @@
 #include "instmap.h"
 
 /* global register indices */
-static TCGv cpu_gpr[32], cpu_pc;
+static TCGv cpu_gpr[32], cpu_pc, cpu_vl;
 static TCGv_i64 cpu_fpr[32]; /* assume F and D extensions */
 static TCGv load_res;
 static TCGv load_val;
@@ -886,6 +886,7 @@ void riscv_translate_init(void)
     }
 
     cpu_pc = tcg_global_mem_new(cpu_env, offsetof(CPURISCVState, pc), "pc");
+    cpu_vl = tcg_global_mem_new(cpu_env, offsetof(CPURISCVState, vl), "vl");
     load_res = tcg_global_mem_new(cpu_env, offsetof(CPURISCVState, load_res),
                              "load_res");
     load_val = tcg_global_mem_new(cpu_env, offsetof(CPURISCVState, load_val),
-- 
2.23.0