* [Qemu-devel] [PATCH v2 00/10] tcg vector improvements
@ 2019-01-04 22:31 Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 01/10] tcg: Add logical simplifications during gvec expand Richard Henderson
` (10 more replies)
0 siblings, 11 replies; 18+ messages in thread
From: Richard Henderson @ 2019-01-04 22:31 UTC (permalink / raw)
To: qemu-devel
I've split this out from the target/ppc patch set in which
it was developed.
r~
Richard Henderson (10):
tcg: Add logical simplifications during gvec expand
tcg: Add gvec expanders for nand, nor, eqv
tcg: Add write_aofs to GVecGen4
tcg: Add opcodes for vector saturated arithmetic
tcg: Add opcodes for vector minmax arithmetic
tcg/i386: Split subroutines out of tcg_expand_vec_op
tcg/i386: Implement vector saturating arithmetic
tcg/i386: Implement vector minmax arithmetic
tcg/aarch64: Implement vector saturating arithmetic
tcg/aarch64: Implement vector minmax arithmetic
accel/tcg/tcg-runtime.h | 23 ++
tcg/aarch64/tcg-target.h | 2 +
tcg/i386/tcg-target.h | 2 +
tcg/tcg-op-gvec.h | 18 ++
tcg/tcg-op.h | 11 +
tcg/tcg-opc.h | 8 +
tcg/tcg.h | 2 +
accel/tcg/tcg-runtime-gvec.c | 257 ++++++++++++++++
tcg/aarch64/tcg-target.inc.c | 48 +++
tcg/i386/tcg-target.inc.c | 580 +++++++++++++++++++++--------------
tcg/tcg-op-gvec.c | 305 ++++++++++++++++--
tcg/tcg-op-vec.c | 75 ++++-
tcg/tcg.c | 10 +
13 files changed, 1078 insertions(+), 263 deletions(-)
--
2.17.2
^ permalink raw reply [flat|nested] 18+ messages in thread
* [Qemu-devel] [PATCH v2 01/10] tcg: Add logical simplifications during gvec expand
2019-01-04 22:31 [Qemu-devel] [PATCH v2 00/10] tcg vector improvements Richard Henderson
@ 2019-01-04 22:31 ` Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 02/10] tcg: Add gvec expanders for nand, nor, eqv Richard Henderson
` (9 subsequent siblings)
10 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2019-01-04 22:31 UTC (permalink / raw)
To: qemu-devel
We handle many of these during integer expansion, and the
rest of them during integer optimization.
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
tcg/tcg-op-gvec.c | 35 ++++++++++++++++++++++++++++++-----
1 file changed, 30 insertions(+), 5 deletions(-)
diff --git a/tcg/tcg-op-gvec.c b/tcg/tcg-op-gvec.c
index 61c25f5784..ec231b78fb 100644
--- a/tcg/tcg-op-gvec.c
+++ b/tcg/tcg-op-gvec.c
@@ -1840,7 +1840,12 @@ void tcg_gen_gvec_and(unsigned vece, uint32_t dofs, uint32_t aofs,
.opc = INDEX_op_and_vec,
.prefer_i64 = TCG_TARGET_REG_BITS == 64,
};
- tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
+
+ if (aofs == bofs) {
+ tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
+ } else {
+ tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
+ }
}
void tcg_gen_gvec_or(unsigned vece, uint32_t dofs, uint32_t aofs,
@@ -1853,7 +1858,12 @@ void tcg_gen_gvec_or(unsigned vece, uint32_t dofs, uint32_t aofs,
.opc = INDEX_op_or_vec,
.prefer_i64 = TCG_TARGET_REG_BITS == 64,
};
- tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
+
+ if (aofs == bofs) {
+ tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
+ } else {
+ tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
+ }
}
void tcg_gen_gvec_xor(unsigned vece, uint32_t dofs, uint32_t aofs,
@@ -1866,7 +1876,12 @@ void tcg_gen_gvec_xor(unsigned vece, uint32_t dofs, uint32_t aofs,
.opc = INDEX_op_xor_vec,
.prefer_i64 = TCG_TARGET_REG_BITS == 64,
};
- tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
+
+ if (aofs == bofs) {
+ tcg_gen_gvec_dup8i(dofs, oprsz, maxsz, 0);
+ } else {
+ tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
+ }
}
void tcg_gen_gvec_andc(unsigned vece, uint32_t dofs, uint32_t aofs,
@@ -1879,7 +1894,12 @@ void tcg_gen_gvec_andc(unsigned vece, uint32_t dofs, uint32_t aofs,
.opc = INDEX_op_andc_vec,
.prefer_i64 = TCG_TARGET_REG_BITS == 64,
};
- tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
+
+ if (aofs == bofs) {
+ tcg_gen_gvec_dup8i(dofs, oprsz, maxsz, 0);
+ } else {
+ tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
+ }
}
void tcg_gen_gvec_orc(unsigned vece, uint32_t dofs, uint32_t aofs,
@@ -1892,7 +1912,12 @@ void tcg_gen_gvec_orc(unsigned vece, uint32_t dofs, uint32_t aofs,
.opc = INDEX_op_orc_vec,
.prefer_i64 = TCG_TARGET_REG_BITS == 64,
};
- tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
+
+ if (aofs == bofs) {
+ tcg_gen_gvec_dup8i(dofs, oprsz, maxsz, -1);
+ } else {
+ tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
+ }
}
static const GVecGen2s gop_ands = {
--
2.17.2
^ permalink raw reply related [flat|nested] 18+ messages in thread
* [Qemu-devel] [PATCH v2 02/10] tcg: Add gvec expanders for nand, nor, eqv
2019-01-04 22:31 [Qemu-devel] [PATCH v2 00/10] tcg vector improvements Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 01/10] tcg: Add logical simplifications during gvec expand Richard Henderson
@ 2019-01-04 22:31 ` Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 03/10] tcg: Add write_aofs to GVecGen4 Richard Henderson
` (8 subsequent siblings)
10 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2019-01-04 22:31 UTC (permalink / raw)
To: qemu-devel
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
accel/tcg/tcg-runtime.h | 3 +++
tcg/tcg-op-gvec.h | 6 +++++
tcg/tcg-op.h | 3 +++
accel/tcg/tcg-runtime-gvec.c | 33 +++++++++++++++++++++++
tcg/tcg-op-gvec.c | 51 ++++++++++++++++++++++++++++++++++++
tcg/tcg-op-vec.c | 21 +++++++++++++++
6 files changed, 117 insertions(+)
diff --git a/accel/tcg/tcg-runtime.h b/accel/tcg/tcg-runtime.h
index 1bd39d136d..835ddfebb2 100644
--- a/accel/tcg/tcg-runtime.h
+++ b/accel/tcg/tcg-runtime.h
@@ -211,6 +211,9 @@ DEF_HELPER_FLAGS_4(gvec_or, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_xor, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_andc, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_orc, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_nand, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_nor, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_eqv, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_ands, TCG_CALL_NO_RWG, void, ptr, ptr, i64, i32)
DEF_HELPER_FLAGS_4(gvec_xors, TCG_CALL_NO_RWG, void, ptr, ptr, i64, i32)
diff --git a/tcg/tcg-op-gvec.h b/tcg/tcg-op-gvec.h
index ff43a29a0b..d65b9d9d4c 100644
--- a/tcg/tcg-op-gvec.h
+++ b/tcg/tcg-op-gvec.h
@@ -242,6 +242,12 @@ void tcg_gen_gvec_andc(unsigned vece, uint32_t dofs, uint32_t aofs,
uint32_t bofs, uint32_t oprsz, uint32_t maxsz);
void tcg_gen_gvec_orc(unsigned vece, uint32_t dofs, uint32_t aofs,
uint32_t bofs, uint32_t oprsz, uint32_t maxsz);
+void tcg_gen_gvec_nand(unsigned vece, uint32_t dofs, uint32_t aofs,
+ uint32_t bofs, uint32_t oprsz, uint32_t maxsz);
+void tcg_gen_gvec_nor(unsigned vece, uint32_t dofs, uint32_t aofs,
+ uint32_t bofs, uint32_t oprsz, uint32_t maxsz);
+void tcg_gen_gvec_eqv(unsigned vece, uint32_t dofs, uint32_t aofs,
+ uint32_t bofs, uint32_t oprsz, uint32_t maxsz);
void tcg_gen_gvec_andi(unsigned vece, uint32_t dofs, uint32_t aofs,
int64_t c, uint32_t oprsz, uint32_t maxsz);
diff --git a/tcg/tcg-op.h b/tcg/tcg-op.h
index 7007ec0d4d..f6ef1cd690 100644
--- a/tcg/tcg-op.h
+++ b/tcg/tcg-op.h
@@ -962,6 +962,9 @@ void tcg_gen_or_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b);
void tcg_gen_xor_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b);
void tcg_gen_andc_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b);
void tcg_gen_orc_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b);
+void tcg_gen_nand_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b);
+void tcg_gen_nor_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b);
+void tcg_gen_eqv_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b);
void tcg_gen_not_vec(unsigned vece, TCGv_vec r, TCGv_vec a);
void tcg_gen_neg_vec(unsigned vece, TCGv_vec r, TCGv_vec a);
diff --git a/accel/tcg/tcg-runtime-gvec.c b/accel/tcg/tcg-runtime-gvec.c
index 90340e56e0..d1802467d5 100644
--- a/accel/tcg/tcg-runtime-gvec.c
+++ b/accel/tcg/tcg-runtime-gvec.c
@@ -512,6 +512,39 @@ void HELPER(gvec_orc)(void *d, void *a, void *b, uint32_t desc)
clear_high(d, oprsz, desc);
}
+void HELPER(gvec_nand)(void *d, void *a, void *b, uint32_t desc)
+{
+ intptr_t oprsz = simd_oprsz(desc);
+ intptr_t i;
+
+ for (i = 0; i < oprsz; i += sizeof(vec64)) {
+ *(vec64 *)(d + i) = ~(*(vec64 *)(a + i) & *(vec64 *)(b + i));
+ }
+ clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_nor)(void *d, void *a, void *b, uint32_t desc)
+{
+ intptr_t oprsz = simd_oprsz(desc);
+ intptr_t i;
+
+ for (i = 0; i < oprsz; i += sizeof(vec64)) {
+ *(vec64 *)(d + i) = ~(*(vec64 *)(a + i) | *(vec64 *)(b + i));
+ }
+ clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_eqv)(void *d, void *a, void *b, uint32_t desc)
+{
+ intptr_t oprsz = simd_oprsz(desc);
+ intptr_t i;
+
+ for (i = 0; i < oprsz; i += sizeof(vec64)) {
+ *(vec64 *)(d + i) = ~(*(vec64 *)(a + i) ^ *(vec64 *)(b + i));
+ }
+ clear_high(d, oprsz, desc);
+}
+
void HELPER(gvec_ands)(void *d, void *a, uint64_t b, uint32_t desc)
{
intptr_t oprsz = simd_oprsz(desc);
diff --git a/tcg/tcg-op-gvec.c b/tcg/tcg-op-gvec.c
index ec231b78fb..81689d02f7 100644
--- a/tcg/tcg-op-gvec.c
+++ b/tcg/tcg-op-gvec.c
@@ -1920,6 +1920,57 @@ void tcg_gen_gvec_orc(unsigned vece, uint32_t dofs, uint32_t aofs,
}
}
+void tcg_gen_gvec_nand(unsigned vece, uint32_t dofs, uint32_t aofs,
+ uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
+{
+ static const GVecGen3 g = {
+ .fni8 = tcg_gen_nand_i64,
+ .fniv = tcg_gen_nand_vec,
+ .fno = gen_helper_gvec_nand,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ };
+
+ if (aofs == bofs) {
+ tcg_gen_gvec_not(vece, dofs, aofs, oprsz, maxsz);
+ } else {
+ tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
+ }
+}
+
+void tcg_gen_gvec_nor(unsigned vece, uint32_t dofs, uint32_t aofs,
+ uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
+{
+ static const GVecGen3 g = {
+ .fni8 = tcg_gen_nor_i64,
+ .fniv = tcg_gen_nor_vec,
+ .fno = gen_helper_gvec_nor,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ };
+
+ if (aofs == bofs) {
+ tcg_gen_gvec_not(vece, dofs, aofs, oprsz, maxsz);
+ } else {
+ tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
+ }
+}
+
+void tcg_gen_gvec_eqv(unsigned vece, uint32_t dofs, uint32_t aofs,
+ uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
+{
+ static const GVecGen3 g = {
+ .fni8 = tcg_gen_eqv_i64,
+ .fniv = tcg_gen_eqv_vec,
+ .fno = gen_helper_gvec_eqv,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ };
+
+ if (aofs == bofs) {
+ tcg_gen_gvec_dup8i(dofs, oprsz, maxsz, -1);
+ } else {
+ tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
+ }
+}
+
static const GVecGen2s gop_ands = {
.fni8 = tcg_gen_and_i64,
.fniv = tcg_gen_and_vec,
diff --git a/tcg/tcg-op-vec.c b/tcg/tcg-op-vec.c
index cefba3d185..d77fdf7c1d 100644
--- a/tcg/tcg-op-vec.c
+++ b/tcg/tcg-op-vec.c
@@ -275,6 +275,27 @@ void tcg_gen_orc_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b)
}
}
+void tcg_gen_nand_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b)
+{
+ /* TODO: Add TCG_TARGET_HAS_nand_vec when adding a backend supports it. */
+ tcg_gen_and_vec(0, r, a, b);
+ tcg_gen_not_vec(0, r, r);
+}
+
+void tcg_gen_nor_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b)
+{
+ /* TODO: Add TCG_TARGET_HAS_nor_vec when adding a backend supports it. */
+ tcg_gen_or_vec(0, r, a, b);
+ tcg_gen_not_vec(0, r, r);
+}
+
+void tcg_gen_eqv_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b)
+{
+ /* TODO: Add TCG_TARGET_HAS_eqv_vec when adding a backend supports it. */
+ tcg_gen_xor_vec(0, r, a, b);
+ tcg_gen_not_vec(0, r, r);
+}
+
void tcg_gen_not_vec(unsigned vece, TCGv_vec r, TCGv_vec a)
{
if (TCG_TARGET_HAS_not_vec) {
--
2.17.2
^ permalink raw reply related [flat|nested] 18+ messages in thread
* [Qemu-devel] [PATCH v2 03/10] tcg: Add write_aofs to GVecGen4
2019-01-04 22:31 [Qemu-devel] [PATCH v2 00/10] tcg vector improvements Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 01/10] tcg: Add logical simplifications during gvec expand Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 02/10] tcg: Add gvec expanders for nand, nor, eqv Richard Henderson
@ 2019-01-04 22:31 ` Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 04/10] tcg: Add opcodes for vector saturated arithmetic Richard Henderson
` (7 subsequent siblings)
10 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2019-01-04 22:31 UTC (permalink / raw)
To: qemu-devel
This allows writing 2 output, 3 input operations.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
tcg/tcg-op-gvec.h | 2 ++
tcg/tcg-op-gvec.c | 27 +++++++++++++++++++--------
2 files changed, 21 insertions(+), 8 deletions(-)
diff --git a/tcg/tcg-op-gvec.h b/tcg/tcg-op-gvec.h
index d65b9d9d4c..2cb447112e 100644
--- a/tcg/tcg-op-gvec.h
+++ b/tcg/tcg-op-gvec.h
@@ -181,6 +181,8 @@ typedef struct {
uint8_t vece;
/* Prefer i64 to v64. */
bool prefer_i64;
+ /* Write aofs as a 2nd dest operand. */
+ bool write_aofs;
} GVecGen4;
void tcg_gen_gvec_2(uint32_t dofs, uint32_t aofs,
diff --git a/tcg/tcg-op-gvec.c b/tcg/tcg-op-gvec.c
index 81689d02f7..c10d3d7b26 100644
--- a/tcg/tcg-op-gvec.c
+++ b/tcg/tcg-op-gvec.c
@@ -665,7 +665,7 @@ static void expand_3_i32(uint32_t dofs, uint32_t aofs,
/* Expand OPSZ bytes worth of three-operand operations using i32 elements. */
static void expand_4_i32(uint32_t dofs, uint32_t aofs, uint32_t bofs,
- uint32_t cofs, uint32_t oprsz,
+ uint32_t cofs, uint32_t oprsz, bool write_aofs,
void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32, TCGv_i32))
{
TCGv_i32 t0 = tcg_temp_new_i32();
@@ -680,6 +680,9 @@ static void expand_4_i32(uint32_t dofs, uint32_t aofs, uint32_t bofs,
tcg_gen_ld_i32(t3, cpu_env, cofs + i);
fni(t0, t1, t2, t3);
tcg_gen_st_i32(t0, cpu_env, dofs + i);
+ if (write_aofs) {
+ tcg_gen_st_i32(t1, cpu_env, aofs + i);
+ }
}
tcg_temp_free_i32(t3);
tcg_temp_free_i32(t2);
@@ -769,7 +772,7 @@ static void expand_3_i64(uint32_t dofs, uint32_t aofs,
/* Expand OPSZ bytes worth of three-operand operations using i64 elements. */
static void expand_4_i64(uint32_t dofs, uint32_t aofs, uint32_t bofs,
- uint32_t cofs, uint32_t oprsz,
+ uint32_t cofs, uint32_t oprsz, bool write_aofs,
void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64, TCGv_i64))
{
TCGv_i64 t0 = tcg_temp_new_i64();
@@ -784,6 +787,9 @@ static void expand_4_i64(uint32_t dofs, uint32_t aofs, uint32_t bofs,
tcg_gen_ld_i64(t3, cpu_env, cofs + i);
fni(t0, t1, t2, t3);
tcg_gen_st_i64(t0, cpu_env, dofs + i);
+ if (write_aofs) {
+ tcg_gen_st_i64(t1, cpu_env, aofs + i);
+ }
}
tcg_temp_free_i64(t3);
tcg_temp_free_i64(t2);
@@ -880,7 +886,7 @@ static void expand_3_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
/* Expand OPSZ bytes worth of four-operand operations using host vectors. */
static void expand_4_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
uint32_t bofs, uint32_t cofs, uint32_t oprsz,
- uint32_t tysz, TCGType type,
+ uint32_t tysz, TCGType type, bool write_aofs,
void (*fni)(unsigned, TCGv_vec, TCGv_vec,
TCGv_vec, TCGv_vec))
{
@@ -896,6 +902,9 @@ static void expand_4_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
tcg_gen_ld_vec(t3, cpu_env, cofs + i);
fni(vece, t0, t1, t2, t3);
tcg_gen_st_vec(t0, cpu_env, dofs + i);
+ if (write_aofs) {
+ tcg_gen_st_vec(t1, cpu_env, aofs + i);
+ }
}
tcg_temp_free_vec(t3);
tcg_temp_free_vec(t2);
@@ -1187,7 +1196,7 @@ void tcg_gen_gvec_4(uint32_t dofs, uint32_t aofs, uint32_t bofs, uint32_t cofs,
*/
some = QEMU_ALIGN_DOWN(oprsz, 32);
expand_4_vec(g->vece, dofs, aofs, bofs, cofs, some,
- 32, TCG_TYPE_V256, g->fniv);
+ 32, TCG_TYPE_V256, g->write_aofs, g->fniv);
if (some == oprsz) {
break;
}
@@ -1200,18 +1209,20 @@ void tcg_gen_gvec_4(uint32_t dofs, uint32_t aofs, uint32_t bofs, uint32_t cofs,
/* fallthru */
case TCG_TYPE_V128:
expand_4_vec(g->vece, dofs, aofs, bofs, cofs, oprsz,
- 16, TCG_TYPE_V128, g->fniv);
+ 16, TCG_TYPE_V128, g->write_aofs, g->fniv);
break;
case TCG_TYPE_V64:
expand_4_vec(g->vece, dofs, aofs, bofs, cofs, oprsz,
- 8, TCG_TYPE_V64, g->fniv);
+ 8, TCG_TYPE_V64, g->write_aofs, g->fniv);
break;
case 0:
if (g->fni8 && check_size_impl(oprsz, 8)) {
- expand_4_i64(dofs, aofs, bofs, cofs, oprsz, g->fni8);
+ expand_4_i64(dofs, aofs, bofs, cofs, oprsz,
+ g->write_aofs, g->fni8);
} else if (g->fni4 && check_size_impl(oprsz, 4)) {
- expand_4_i32(dofs, aofs, bofs, cofs, oprsz, g->fni4);
+ expand_4_i32(dofs, aofs, bofs, cofs, oprsz,
+ g->write_aofs, g->fni4);
} else {
assert(g->fno != NULL);
tcg_gen_gvec_4_ool(dofs, aofs, bofs, cofs,
--
2.17.2
^ permalink raw reply related [flat|nested] 18+ messages in thread
* [Qemu-devel] [PATCH v2 04/10] tcg: Add opcodes for vector saturated arithmetic
2019-01-04 22:31 [Qemu-devel] [PATCH v2 00/10] tcg vector improvements Richard Henderson
` (2 preceding siblings ...)
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 03/10] tcg: Add write_aofs to GVecGen4 Richard Henderson
@ 2019-01-04 22:31 ` Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 05/10] tcg: Add opcodes for vector minmax arithmetic Richard Henderson
` (6 subsequent siblings)
10 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2019-01-04 22:31 UTC (permalink / raw)
To: qemu-devel
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
tcg/aarch64/tcg-target.h | 1 +
tcg/i386/tcg-target.h | 1 +
tcg/tcg-op.h | 4 ++
tcg/tcg-opc.h | 4 ++
tcg/tcg.h | 1 +
tcg/tcg-op-gvec.c | 84 ++++++++++++++++++++++++++++++----------
tcg/tcg-op-vec.c | 34 ++++++++++++++--
tcg/tcg.c | 5 +++
8 files changed, 110 insertions(+), 24 deletions(-)
diff --git a/tcg/aarch64/tcg-target.h b/tcg/aarch64/tcg-target.h
index f966a4fcb3..98556bcf22 100644
--- a/tcg/aarch64/tcg-target.h
+++ b/tcg/aarch64/tcg-target.h
@@ -135,6 +135,7 @@ typedef enum {
#define TCG_TARGET_HAS_shv_vec 0
#define TCG_TARGET_HAS_cmp_vec 1
#define TCG_TARGET_HAS_mul_vec 1
+#define TCG_TARGET_HAS_sat_vec 0
#define TCG_TARGET_DEFAULT_MO (0)
#define TCG_TARGET_HAS_MEMORY_BSWAP 1
diff --git a/tcg/i386/tcg-target.h b/tcg/i386/tcg-target.h
index f378d29568..44381062e6 100644
--- a/tcg/i386/tcg-target.h
+++ b/tcg/i386/tcg-target.h
@@ -185,6 +185,7 @@ extern bool have_avx2;
#define TCG_TARGET_HAS_shv_vec 0
#define TCG_TARGET_HAS_cmp_vec 1
#define TCG_TARGET_HAS_mul_vec 1
+#define TCG_TARGET_HAS_sat_vec 0
#define TCG_TARGET_deposit_i32_valid(ofs, len) \
(((ofs) == 0 && (len) == 8) || ((ofs) == 8 && (len) == 8) || \
diff --git a/tcg/tcg-op.h b/tcg/tcg-op.h
index f6ef1cd690..4a93d730e8 100644
--- a/tcg/tcg-op.h
+++ b/tcg/tcg-op.h
@@ -967,6 +967,10 @@ void tcg_gen_nor_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b);
void tcg_gen_eqv_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b);
void tcg_gen_not_vec(unsigned vece, TCGv_vec r, TCGv_vec a);
void tcg_gen_neg_vec(unsigned vece, TCGv_vec r, TCGv_vec a);
+void tcg_gen_ssadd_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b);
+void tcg_gen_usadd_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b);
+void tcg_gen_sssub_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b);
+void tcg_gen_ussub_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b);
void tcg_gen_shli_vec(unsigned vece, TCGv_vec r, TCGv_vec a, int64_t i);
void tcg_gen_shri_vec(unsigned vece, TCGv_vec r, TCGv_vec a, int64_t i);
diff --git a/tcg/tcg-opc.h b/tcg/tcg-opc.h
index 7a8a3edb5b..94b2ed80af 100644
--- a/tcg/tcg-opc.h
+++ b/tcg/tcg-opc.h
@@ -222,6 +222,10 @@ DEF(add_vec, 1, 2, 0, IMPLVEC)
DEF(sub_vec, 1, 2, 0, IMPLVEC)
DEF(mul_vec, 1, 2, 0, IMPLVEC | IMPL(TCG_TARGET_HAS_mul_vec))
DEF(neg_vec, 1, 1, 0, IMPLVEC | IMPL(TCG_TARGET_HAS_neg_vec))
+DEF(ssadd_vec, 1, 2, 0, IMPLVEC | IMPL(TCG_TARGET_HAS_sat_vec))
+DEF(usadd_vec, 1, 2, 0, IMPLVEC | IMPL(TCG_TARGET_HAS_sat_vec))
+DEF(sssub_vec, 1, 2, 0, IMPLVEC | IMPL(TCG_TARGET_HAS_sat_vec))
+DEF(ussub_vec, 1, 2, 0, IMPLVEC | IMPL(TCG_TARGET_HAS_sat_vec))
DEF(and_vec, 1, 2, 0, IMPLVEC)
DEF(or_vec, 1, 2, 0, IMPLVEC)
diff --git a/tcg/tcg.h b/tcg/tcg.h
index 3a629991ca..df24afa425 100644
--- a/tcg/tcg.h
+++ b/tcg/tcg.h
@@ -183,6 +183,7 @@ typedef uint64_t TCGRegSet;
#define TCG_TARGET_HAS_shs_vec 0
#define TCG_TARGET_HAS_shv_vec 0
#define TCG_TARGET_HAS_mul_vec 0
+#define TCG_TARGET_HAS_sat_vec 0
#else
#define TCG_TARGET_MAYBE_vec 1
#endif
diff --git a/tcg/tcg-op-gvec.c b/tcg/tcg-op-gvec.c
index c10d3d7b26..0a33f51065 100644
--- a/tcg/tcg-op-gvec.c
+++ b/tcg/tcg-op-gvec.c
@@ -1678,10 +1678,22 @@ void tcg_gen_gvec_ssadd(unsigned vece, uint32_t dofs, uint32_t aofs,
uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
{
static const GVecGen3 g[4] = {
- { .fno = gen_helper_gvec_ssadd8, .vece = MO_8 },
- { .fno = gen_helper_gvec_ssadd16, .vece = MO_16 },
- { .fno = gen_helper_gvec_ssadd32, .vece = MO_32 },
- { .fno = gen_helper_gvec_ssadd64, .vece = MO_64 }
+ { .fniv = tcg_gen_ssadd_vec,
+ .fno = gen_helper_gvec_ssadd8,
+ .opc = INDEX_op_ssadd_vec,
+ .vece = MO_8 },
+ { .fniv = tcg_gen_ssadd_vec,
+ .fno = gen_helper_gvec_ssadd16,
+ .opc = INDEX_op_ssadd_vec,
+ .vece = MO_16 },
+ { .fniv = tcg_gen_ssadd_vec,
+ .fno = gen_helper_gvec_ssadd32,
+ .opc = INDEX_op_ssadd_vec,
+ .vece = MO_32 },
+ { .fniv = tcg_gen_ssadd_vec,
+ .fno = gen_helper_gvec_ssadd64,
+ .opc = INDEX_op_ssadd_vec,
+ .vece = MO_64 },
};
tcg_debug_assert(vece <= MO_64);
tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
@@ -1691,16 +1703,28 @@ void tcg_gen_gvec_sssub(unsigned vece, uint32_t dofs, uint32_t aofs,
uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
{
static const GVecGen3 g[4] = {
- { .fno = gen_helper_gvec_sssub8, .vece = MO_8 },
- { .fno = gen_helper_gvec_sssub16, .vece = MO_16 },
- { .fno = gen_helper_gvec_sssub32, .vece = MO_32 },
- { .fno = gen_helper_gvec_sssub64, .vece = MO_64 }
+ { .fniv = tcg_gen_sssub_vec,
+ .fno = gen_helper_gvec_sssub8,
+ .opc = INDEX_op_sssub_vec,
+ .vece = MO_8 },
+ { .fniv = tcg_gen_sssub_vec,
+ .fno = gen_helper_gvec_sssub16,
+ .opc = INDEX_op_sssub_vec,
+ .vece = MO_16 },
+ { .fniv = tcg_gen_sssub_vec,
+ .fno = gen_helper_gvec_sssub32,
+ .opc = INDEX_op_sssub_vec,
+ .vece = MO_32 },
+ { .fniv = tcg_gen_sssub_vec,
+ .fno = gen_helper_gvec_sssub64,
+ .opc = INDEX_op_sssub_vec,
+ .vece = MO_64 },
};
tcg_debug_assert(vece <= MO_64);
tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
}
-static void tcg_gen_vec_usadd32_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
+static void tcg_gen_usadd_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
{
TCGv_i32 max = tcg_const_i32(-1);
tcg_gen_add_i32(d, a, b);
@@ -1708,7 +1732,7 @@ static void tcg_gen_vec_usadd32_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
tcg_temp_free_i32(max);
}
-static void tcg_gen_vec_usadd32_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
+static void tcg_gen_usadd_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
{
TCGv_i64 max = tcg_const_i64(-1);
tcg_gen_add_i64(d, a, b);
@@ -1720,20 +1744,30 @@ void tcg_gen_gvec_usadd(unsigned vece, uint32_t dofs, uint32_t aofs,
uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
{
static const GVecGen3 g[4] = {
- { .fno = gen_helper_gvec_usadd8, .vece = MO_8 },
- { .fno = gen_helper_gvec_usadd16, .vece = MO_16 },
- { .fni4 = tcg_gen_vec_usadd32_i32,
+ { .fniv = tcg_gen_usadd_vec,
+ .fno = gen_helper_gvec_usadd8,
+ .opc = INDEX_op_usadd_vec,
+ .vece = MO_8 },
+ { .fniv = tcg_gen_usadd_vec,
+ .fno = gen_helper_gvec_usadd16,
+ .opc = INDEX_op_usadd_vec,
+ .vece = MO_16 },
+ { .fni4 = tcg_gen_usadd_i32,
+ .fniv = tcg_gen_usadd_vec,
.fno = gen_helper_gvec_usadd32,
+ .opc = INDEX_op_usadd_vec,
.vece = MO_32 },
- { .fni8 = tcg_gen_vec_usadd32_i64,
+ { .fni8 = tcg_gen_usadd_i64,
+ .fniv = tcg_gen_usadd_vec,
.fno = gen_helper_gvec_usadd64,
+ .opc = INDEX_op_usadd_vec,
.vece = MO_64 }
};
tcg_debug_assert(vece <= MO_64);
tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
}
-static void tcg_gen_vec_ussub32_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
+static void tcg_gen_ussub_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
{
TCGv_i32 min = tcg_const_i32(0);
tcg_gen_sub_i32(d, a, b);
@@ -1741,7 +1775,7 @@ static void tcg_gen_vec_ussub32_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
tcg_temp_free_i32(min);
}
-static void tcg_gen_vec_ussub32_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
+static void tcg_gen_ussub_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
{
TCGv_i64 min = tcg_const_i64(0);
tcg_gen_sub_i64(d, a, b);
@@ -1753,13 +1787,23 @@ void tcg_gen_gvec_ussub(unsigned vece, uint32_t dofs, uint32_t aofs,
uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
{
static const GVecGen3 g[4] = {
- { .fno = gen_helper_gvec_ussub8, .vece = MO_8 },
- { .fno = gen_helper_gvec_ussub16, .vece = MO_16 },
- { .fni4 = tcg_gen_vec_ussub32_i32,
+ { .fniv = tcg_gen_ussub_vec,
+ .fno = gen_helper_gvec_ussub8,
+ .opc = INDEX_op_ussub_vec,
+ .vece = MO_8 },
+ { .fniv = tcg_gen_ussub_vec,
+ .fno = gen_helper_gvec_ussub16,
+ .opc = INDEX_op_ussub_vec,
+ .vece = MO_16 },
+ { .fni4 = tcg_gen_ussub_i32,
+ .fniv = tcg_gen_ussub_vec,
.fno = gen_helper_gvec_ussub32,
+ .opc = INDEX_op_ussub_vec,
.vece = MO_32 },
- { .fni8 = tcg_gen_vec_ussub32_i64,
+ { .fni8 = tcg_gen_ussub_i64,
+ .fniv = tcg_gen_ussub_vec,
.fno = gen_helper_gvec_ussub64,
+ .opc = INDEX_op_ussub_vec,
.vece = MO_64 }
};
tcg_debug_assert(vece <= MO_64);
diff --git a/tcg/tcg-op-vec.c b/tcg/tcg-op-vec.c
index d77fdf7c1d..675aa09258 100644
--- a/tcg/tcg-op-vec.c
+++ b/tcg/tcg-op-vec.c
@@ -386,7 +386,8 @@ void tcg_gen_cmp_vec(TCGCond cond, unsigned vece,
}
}
-void tcg_gen_mul_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b)
+static void do_op3(unsigned vece, TCGv_vec r, TCGv_vec a,
+ TCGv_vec b, TCGOpcode opc)
{
TCGTemp *rt = tcgv_vec_temp(r);
TCGTemp *at = tcgv_vec_temp(a);
@@ -399,11 +400,36 @@ void tcg_gen_mul_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b)
tcg_debug_assert(at->base_type >= type);
tcg_debug_assert(bt->base_type >= type);
- can = tcg_can_emit_vec_op(INDEX_op_mul_vec, type, vece);
+ can = tcg_can_emit_vec_op(opc, type, vece);
if (can > 0) {
- vec_gen_3(INDEX_op_mul_vec, type, vece, ri, ai, bi);
+ vec_gen_3(opc, type, vece, ri, ai, bi);
} else {
tcg_debug_assert(can < 0);
- tcg_expand_vec_op(INDEX_op_mul_vec, type, vece, ri, ai, bi);
+ tcg_expand_vec_op(opc, type, vece, ri, ai, bi);
}
}
+
+void tcg_gen_mul_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b)
+{
+ do_op3(vece, r, a, b, INDEX_op_mul_vec);
+}
+
+void tcg_gen_ssadd_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b)
+{
+ do_op3(vece, r, a, b, INDEX_op_ssadd_vec);
+}
+
+void tcg_gen_usadd_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b)
+{
+ do_op3(vece, r, a, b, INDEX_op_usadd_vec);
+}
+
+void tcg_gen_sssub_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b)
+{
+ do_op3(vece, r, a, b, INDEX_op_sssub_vec);
+}
+
+void tcg_gen_ussub_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b)
+{
+ do_op3(vece, r, a, b, INDEX_op_ussub_vec);
+}
diff --git a/tcg/tcg.c b/tcg/tcg.c
index c54b119020..15ed5af007 100644
--- a/tcg/tcg.c
+++ b/tcg/tcg.c
@@ -1607,6 +1607,11 @@ bool tcg_op_supported(TCGOpcode op)
case INDEX_op_shrv_vec:
case INDEX_op_sarv_vec:
return have_vec && TCG_TARGET_HAS_shv_vec;
+ case INDEX_op_ssadd_vec:
+ case INDEX_op_usadd_vec:
+ case INDEX_op_sssub_vec:
+ case INDEX_op_ussub_vec:
+ return have_vec && TCG_TARGET_HAS_sat_vec;
default:
tcg_debug_assert(op > INDEX_op_last_generic && op < NB_OPS);
--
2.17.2
^ permalink raw reply related [flat|nested] 18+ messages in thread
* [Qemu-devel] [PATCH v2 05/10] tcg: Add opcodes for vector minmax arithmetic
2019-01-04 22:31 [Qemu-devel] [PATCH v2 00/10] tcg vector improvements Richard Henderson
` (3 preceding siblings ...)
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 04/10] tcg: Add opcodes for vector saturated arithmetic Richard Henderson
@ 2019-01-04 22:31 ` Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 06/10] tcg/i386: Split subroutines out of tcg_expand_vec_op Richard Henderson
` (5 subsequent siblings)
10 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2019-01-04 22:31 UTC (permalink / raw)
To: qemu-devel
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
accel/tcg/tcg-runtime.h | 20 ++++
tcg/aarch64/tcg-target.h | 1 +
tcg/i386/tcg-target.h | 1 +
tcg/tcg-op-gvec.h | 10 ++
tcg/tcg-op.h | 4 +
tcg/tcg-opc.h | 4 +
tcg/tcg.h | 1 +
accel/tcg/tcg-runtime-gvec.c | 224 +++++++++++++++++++++++++++++++++++
tcg/tcg-op-gvec.c | 108 +++++++++++++++++
tcg/tcg-op-vec.c | 20 ++++
tcg/tcg.c | 5 +
11 files changed, 398 insertions(+)
diff --git a/accel/tcg/tcg-runtime.h b/accel/tcg/tcg-runtime.h
index 835ddfebb2..dfe325625c 100644
--- a/accel/tcg/tcg-runtime.h
+++ b/accel/tcg/tcg-runtime.h
@@ -200,6 +200,26 @@ DEF_HELPER_FLAGS_4(gvec_ussub16, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_ussub32, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_ussub64, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_smin8, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_smin16, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_smin32, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_smin64, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+
+DEF_HELPER_FLAGS_4(gvec_smax8, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_smax16, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_smax32, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_smax64, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+
+DEF_HELPER_FLAGS_4(gvec_umin8, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_umin16, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_umin32, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_umin64, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+
+DEF_HELPER_FLAGS_4(gvec_umax8, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_umax16, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_umax32, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_umax64, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+
DEF_HELPER_FLAGS_3(gvec_neg8, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
DEF_HELPER_FLAGS_3(gvec_neg16, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
DEF_HELPER_FLAGS_3(gvec_neg32, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
diff --git a/tcg/aarch64/tcg-target.h b/tcg/aarch64/tcg-target.h
index 98556bcf22..545a6eec75 100644
--- a/tcg/aarch64/tcg-target.h
+++ b/tcg/aarch64/tcg-target.h
@@ -136,6 +136,7 @@ typedef enum {
#define TCG_TARGET_HAS_cmp_vec 1
#define TCG_TARGET_HAS_mul_vec 1
#define TCG_TARGET_HAS_sat_vec 0
+#define TCG_TARGET_HAS_minmax_vec 0
#define TCG_TARGET_DEFAULT_MO (0)
#define TCG_TARGET_HAS_MEMORY_BSWAP 1
diff --git a/tcg/i386/tcg-target.h b/tcg/i386/tcg-target.h
index 44381062e6..7bd7eae672 100644
--- a/tcg/i386/tcg-target.h
+++ b/tcg/i386/tcg-target.h
@@ -186,6 +186,7 @@ extern bool have_avx2;
#define TCG_TARGET_HAS_cmp_vec 1
#define TCG_TARGET_HAS_mul_vec 1
#define TCG_TARGET_HAS_sat_vec 0
+#define TCG_TARGET_HAS_minmax_vec 0
#define TCG_TARGET_deposit_i32_valid(ofs, len) \
(((ofs) == 0 && (len) == 8) || ((ofs) == 8 && (len) == 8) || \
diff --git a/tcg/tcg-op-gvec.h b/tcg/tcg-op-gvec.h
index 2cb447112e..4734eef7de 100644
--- a/tcg/tcg-op-gvec.h
+++ b/tcg/tcg-op-gvec.h
@@ -234,6 +234,16 @@ void tcg_gen_gvec_usadd(unsigned vece, uint32_t dofs, uint32_t aofs,
void tcg_gen_gvec_ussub(unsigned vece, uint32_t dofs, uint32_t aofs,
uint32_t bofs, uint32_t oprsz, uint32_t maxsz);
+/* Min/max. */
+void tcg_gen_gvec_smin(unsigned vece, uint32_t dofs, uint32_t aofs,
+ uint32_t bofs, uint32_t oprsz, uint32_t maxsz);
+void tcg_gen_gvec_umin(unsigned vece, uint32_t dofs, uint32_t aofs,
+ uint32_t bofs, uint32_t oprsz, uint32_t maxsz);
+void tcg_gen_gvec_smax(unsigned vece, uint32_t dofs, uint32_t aofs,
+ uint32_t bofs, uint32_t oprsz, uint32_t maxsz);
+void tcg_gen_gvec_umax(unsigned vece, uint32_t dofs, uint32_t aofs,
+ uint32_t bofs, uint32_t oprsz, uint32_t maxsz);
+
void tcg_gen_gvec_and(unsigned vece, uint32_t dofs, uint32_t aofs,
uint32_t bofs, uint32_t oprsz, uint32_t maxsz);
void tcg_gen_gvec_or(unsigned vece, uint32_t dofs, uint32_t aofs,
diff --git a/tcg/tcg-op.h b/tcg/tcg-op.h
index 4a93d730e8..2d98868d8f 100644
--- a/tcg/tcg-op.h
+++ b/tcg/tcg-op.h
@@ -971,6 +971,10 @@ void tcg_gen_ssadd_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b);
void tcg_gen_usadd_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b);
void tcg_gen_sssub_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b);
void tcg_gen_ussub_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b);
+void tcg_gen_smin_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b);
+void tcg_gen_umin_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b);
+void tcg_gen_smax_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b);
+void tcg_gen_umax_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b);
void tcg_gen_shli_vec(unsigned vece, TCGv_vec r, TCGv_vec a, int64_t i);
void tcg_gen_shri_vec(unsigned vece, TCGv_vec r, TCGv_vec a, int64_t i);
diff --git a/tcg/tcg-opc.h b/tcg/tcg-opc.h
index 94b2ed80af..4e0238ad1a 100644
--- a/tcg/tcg-opc.h
+++ b/tcg/tcg-opc.h
@@ -226,6 +226,10 @@ DEF(ssadd_vec, 1, 2, 0, IMPLVEC | IMPL(TCG_TARGET_HAS_sat_vec))
DEF(usadd_vec, 1, 2, 0, IMPLVEC | IMPL(TCG_TARGET_HAS_sat_vec))
DEF(sssub_vec, 1, 2, 0, IMPLVEC | IMPL(TCG_TARGET_HAS_sat_vec))
DEF(ussub_vec, 1, 2, 0, IMPLVEC | IMPL(TCG_TARGET_HAS_sat_vec))
+DEF(smin_vec, 1, 2, 0, IMPLVEC | IMPL(TCG_TARGET_HAS_minmax_vec))
+DEF(umin_vec, 1, 2, 0, IMPLVEC | IMPL(TCG_TARGET_HAS_minmax_vec))
+DEF(smax_vec, 1, 2, 0, IMPLVEC | IMPL(TCG_TARGET_HAS_minmax_vec))
+DEF(umax_vec, 1, 2, 0, IMPLVEC | IMPL(TCG_TARGET_HAS_minmax_vec))
DEF(and_vec, 1, 2, 0, IMPLVEC)
DEF(or_vec, 1, 2, 0, IMPLVEC)
diff --git a/tcg/tcg.h b/tcg/tcg.h
index df24afa425..1c3579077d 100644
--- a/tcg/tcg.h
+++ b/tcg/tcg.h
@@ -184,6 +184,7 @@ typedef uint64_t TCGRegSet;
#define TCG_TARGET_HAS_shv_vec 0
#define TCG_TARGET_HAS_mul_vec 0
#define TCG_TARGET_HAS_sat_vec 0
+#define TCG_TARGET_HAS_minmax_vec 0
#else
#define TCG_TARGET_MAYBE_vec 1
#endif
diff --git a/accel/tcg/tcg-runtime-gvec.c b/accel/tcg/tcg-runtime-gvec.c
index d1802467d5..9358749741 100644
--- a/accel/tcg/tcg-runtime-gvec.c
+++ b/accel/tcg/tcg-runtime-gvec.c
@@ -1028,3 +1028,227 @@ void HELPER(gvec_ussub64)(void *d, void *a, void *b, uint32_t desc)
}
clear_high(d, oprsz, desc);
}
+
+void HELPER(gvec_smin8)(void *d, void *a, void *b, uint32_t desc)
+{
+ intptr_t oprsz = simd_oprsz(desc);
+ intptr_t i;
+
+ for (i = 0; i < oprsz; i += sizeof(int8_t)) {
+ int8_t aa = *(int8_t *)(a + i);
+ int8_t bb = *(int8_t *)(b + i);
+ int8_t dd = aa < bb ? aa : bb;
+ *(int8_t *)(d + i) = dd;
+ }
+ clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_smin16)(void *d, void *a, void *b, uint32_t desc)
+{
+ intptr_t oprsz = simd_oprsz(desc);
+ intptr_t i;
+
+ for (i = 0; i < oprsz; i += sizeof(int16_t)) {
+ int16_t aa = *(int16_t *)(a + i);
+ int16_t bb = *(int16_t *)(b + i);
+ int16_t dd = aa < bb ? aa : bb;
+ *(int16_t *)(d + i) = dd;
+ }
+ clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_smin32)(void *d, void *a, void *b, uint32_t desc)
+{
+ intptr_t oprsz = simd_oprsz(desc);
+ intptr_t i;
+
+ for (i = 0; i < oprsz; i += sizeof(int32_t)) {
+ int32_t aa = *(int32_t *)(a + i);
+ int32_t bb = *(int32_t *)(b + i);
+ int32_t dd = aa < bb ? aa : bb;
+ *(int32_t *)(d + i) = dd;
+ }
+ clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_smin64)(void *d, void *a, void *b, uint32_t desc)
+{
+ intptr_t oprsz = simd_oprsz(desc);
+ intptr_t i;
+
+ for (i = 0; i < oprsz; i += sizeof(int64_t)) {
+ int64_t aa = *(int64_t *)(a + i);
+ int64_t bb = *(int64_t *)(b + i);
+ int64_t dd = aa < bb ? aa : bb;
+ *(int64_t *)(d + i) = dd;
+ }
+ clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_smax8)(void *d, void *a, void *b, uint32_t desc)
+{
+ intptr_t oprsz = simd_oprsz(desc);
+ intptr_t i;
+
+ for (i = 0; i < oprsz; i += sizeof(int8_t)) {
+ int8_t aa = *(int8_t *)(a + i);
+ int8_t bb = *(int8_t *)(b + i);
+ int8_t dd = aa > bb ? aa : bb;
+ *(int8_t *)(d + i) = dd;
+ }
+ clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_smax16)(void *d, void *a, void *b, uint32_t desc)
+{
+ intptr_t oprsz = simd_oprsz(desc);
+ intptr_t i;
+
+ for (i = 0; i < oprsz; i += sizeof(int16_t)) {
+ int16_t aa = *(int16_t *)(a + i);
+ int16_t bb = *(int16_t *)(b + i);
+ int16_t dd = aa > bb ? aa : bb;
+ *(int16_t *)(d + i) = dd;
+ }
+ clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_smax32)(void *d, void *a, void *b, uint32_t desc)
+{
+ intptr_t oprsz = simd_oprsz(desc);
+ intptr_t i;
+
+ for (i = 0; i < oprsz; i += sizeof(int32_t)) {
+ int32_t aa = *(int32_t *)(a + i);
+ int32_t bb = *(int32_t *)(b + i);
+ int32_t dd = aa > bb ? aa : bb;
+ *(int32_t *)(d + i) = dd;
+ }
+ clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_smax64)(void *d, void *a, void *b, uint32_t desc)
+{
+ intptr_t oprsz = simd_oprsz(desc);
+ intptr_t i;
+
+ for (i = 0; i < oprsz; i += sizeof(int64_t)) {
+ int64_t aa = *(int64_t *)(a + i);
+ int64_t bb = *(int64_t *)(b + i);
+ int64_t dd = aa > bb ? aa : bb;
+ *(int64_t *)(d + i) = dd;
+ }
+ clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_umin8)(void *d, void *a, void *b, uint32_t desc)
+{
+ intptr_t oprsz = simd_oprsz(desc);
+ intptr_t i;
+
+ for (i = 0; i < oprsz; i += sizeof(uint8_t)) {
+ uint8_t aa = *(uint8_t *)(a + i);
+ uint8_t bb = *(uint8_t *)(b + i);
+ uint8_t dd = aa < bb ? aa : bb;
+ *(uint8_t *)(d + i) = dd;
+ }
+ clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_umin16)(void *d, void *a, void *b, uint32_t desc)
+{
+ intptr_t oprsz = simd_oprsz(desc);
+ intptr_t i;
+
+ for (i = 0; i < oprsz; i += sizeof(uint16_t)) {
+ uint16_t aa = *(uint16_t *)(a + i);
+ uint16_t bb = *(uint16_t *)(b + i);
+ uint16_t dd = aa < bb ? aa : bb;
+ *(uint16_t *)(d + i) = dd;
+ }
+ clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_umin32)(void *d, void *a, void *b, uint32_t desc)
+{
+ intptr_t oprsz = simd_oprsz(desc);
+ intptr_t i;
+
+ for (i = 0; i < oprsz; i += sizeof(uint32_t)) {
+ uint32_t aa = *(uint32_t *)(a + i);
+ uint32_t bb = *(uint32_t *)(b + i);
+ uint32_t dd = aa < bb ? aa : bb;
+ *(uint32_t *)(d + i) = dd;
+ }
+ clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_umin64)(void *d, void *a, void *b, uint32_t desc)
+{
+ intptr_t oprsz = simd_oprsz(desc);
+ intptr_t i;
+
+ for (i = 0; i < oprsz; i += sizeof(uint64_t)) {
+ uint64_t aa = *(uint64_t *)(a + i);
+ uint64_t bb = *(uint64_t *)(b + i);
+ uint64_t dd = aa < bb ? aa : bb;
+ *(uint64_t *)(d + i) = dd;
+ }
+ clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_umax8)(void *d, void *a, void *b, uint32_t desc)
+{
+ intptr_t oprsz = simd_oprsz(desc);
+ intptr_t i;
+
+ for (i = 0; i < oprsz; i += sizeof(uint8_t)) {
+ uint8_t aa = *(uint8_t *)(a + i);
+ uint8_t bb = *(uint8_t *)(b + i);
+ uint8_t dd = aa > bb ? aa : bb;
+ *(uint8_t *)(d + i) = dd;
+ }
+ clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_umax16)(void *d, void *a, void *b, uint32_t desc)
+{
+ intptr_t oprsz = simd_oprsz(desc);
+ intptr_t i;
+
+ for (i = 0; i < oprsz; i += sizeof(uint16_t)) {
+ uint16_t aa = *(uint16_t *)(a + i);
+ uint16_t bb = *(uint16_t *)(b + i);
+ uint16_t dd = aa > bb ? aa : bb;
+ *(uint16_t *)(d + i) = dd;
+ }
+ clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_umax32)(void *d, void *a, void *b, uint32_t desc)
+{
+ intptr_t oprsz = simd_oprsz(desc);
+ intptr_t i;
+
+ for (i = 0; i < oprsz; i += sizeof(uint32_t)) {
+ uint32_t aa = *(uint32_t *)(a + i);
+ uint32_t bb = *(uint32_t *)(b + i);
+ uint32_t dd = aa > bb ? aa : bb;
+ *(uint32_t *)(d + i) = dd;
+ }
+ clear_high(d, oprsz, desc);
+}
+
+void HELPER(gvec_umax64)(void *d, void *a, void *b, uint32_t desc)
+{
+ intptr_t oprsz = simd_oprsz(desc);
+ intptr_t i;
+
+ for (i = 0; i < oprsz; i += sizeof(uint64_t)) {
+ uint64_t aa = *(uint64_t *)(a + i);
+ uint64_t bb = *(uint64_t *)(b + i);
+ uint64_t dd = aa > bb ? aa : bb;
+ *(uint64_t *)(d + i) = dd;
+ }
+ clear_high(d, oprsz, desc);
+}
diff --git a/tcg/tcg-op-gvec.c b/tcg/tcg-op-gvec.c
index 0a33f51065..3ee44fcb75 100644
--- a/tcg/tcg-op-gvec.c
+++ b/tcg/tcg-op-gvec.c
@@ -1810,6 +1810,114 @@ void tcg_gen_gvec_ussub(unsigned vece, uint32_t dofs, uint32_t aofs,
tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
}
+void tcg_gen_gvec_smin(unsigned vece, uint32_t dofs, uint32_t aofs,
+ uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
+{
+ static const GVecGen3 g[4] = {
+ { .fniv = tcg_gen_smin_vec,
+ .fno = gen_helper_gvec_smin8,
+ .opc = INDEX_op_smin_vec,
+ .vece = MO_8 },
+ { .fniv = tcg_gen_smin_vec,
+ .fno = gen_helper_gvec_smin16,
+ .opc = INDEX_op_smin_vec,
+ .vece = MO_16 },
+ { .fni4 = tcg_gen_smin_i32,
+ .fniv = tcg_gen_smin_vec,
+ .fno = gen_helper_gvec_smin32,
+ .opc = INDEX_op_smin_vec,
+ .vece = MO_32 },
+ { .fni8 = tcg_gen_smin_i64,
+ .fniv = tcg_gen_smin_vec,
+ .fno = gen_helper_gvec_smin64,
+ .opc = INDEX_op_smin_vec,
+ .vece = MO_64 }
+ };
+ tcg_debug_assert(vece <= MO_64);
+ tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
+}
+
+void tcg_gen_gvec_umin(unsigned vece, uint32_t dofs, uint32_t aofs,
+ uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
+{
+ static const GVecGen3 g[4] = {
+ { .fniv = tcg_gen_umin_vec,
+ .fno = gen_helper_gvec_umin8,
+ .opc = INDEX_op_umin_vec,
+ .vece = MO_8 },
+ { .fniv = tcg_gen_umin_vec,
+ .fno = gen_helper_gvec_umin16,
+ .opc = INDEX_op_umin_vec,
+ .vece = MO_16 },
+ { .fni4 = tcg_gen_umin_i32,
+ .fniv = tcg_gen_umin_vec,
+ .fno = gen_helper_gvec_umin32,
+ .opc = INDEX_op_umin_vec,
+ .vece = MO_32 },
+ { .fni8 = tcg_gen_umin_i64,
+ .fniv = tcg_gen_umin_vec,
+ .fno = gen_helper_gvec_umin64,
+ .opc = INDEX_op_umin_vec,
+ .vece = MO_64 }
+ };
+ tcg_debug_assert(vece <= MO_64);
+ tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
+}
+
+void tcg_gen_gvec_smax(unsigned vece, uint32_t dofs, uint32_t aofs,
+ uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
+{
+ static const GVecGen3 g[4] = {
+ { .fniv = tcg_gen_smax_vec,
+ .fno = gen_helper_gvec_smax8,
+ .opc = INDEX_op_smax_vec,
+ .vece = MO_8 },
+ { .fniv = tcg_gen_smax_vec,
+ .fno = gen_helper_gvec_smax16,
+ .opc = INDEX_op_smax_vec,
+ .vece = MO_16 },
+ { .fni4 = tcg_gen_smax_i32,
+ .fniv = tcg_gen_smax_vec,
+ .fno = gen_helper_gvec_smax32,
+ .opc = INDEX_op_smax_vec,
+ .vece = MO_32 },
+ { .fni8 = tcg_gen_smax_i64,
+ .fniv = tcg_gen_smax_vec,
+ .fno = gen_helper_gvec_smax64,
+ .opc = INDEX_op_smax_vec,
+ .vece = MO_64 }
+ };
+ tcg_debug_assert(vece <= MO_64);
+ tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
+}
+
+void tcg_gen_gvec_umax(unsigned vece, uint32_t dofs, uint32_t aofs,
+ uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
+{
+ static const GVecGen3 g[4] = {
+ { .fniv = tcg_gen_umax_vec,
+ .fno = gen_helper_gvec_umax8,
+ .opc = INDEX_op_umax_vec,
+ .vece = MO_8 },
+ { .fniv = tcg_gen_umax_vec,
+ .fno = gen_helper_gvec_umax16,
+ .opc = INDEX_op_umax_vec,
+ .vece = MO_16 },
+ { .fni4 = tcg_gen_umax_i32,
+ .fniv = tcg_gen_umax_vec,
+ .fno = gen_helper_gvec_umax32,
+ .opc = INDEX_op_umax_vec,
+ .vece = MO_32 },
+ { .fni8 = tcg_gen_umax_i64,
+ .fniv = tcg_gen_umax_vec,
+ .fno = gen_helper_gvec_umax64,
+ .opc = INDEX_op_umax_vec,
+ .vece = MO_64 }
+ };
+ tcg_debug_assert(vece <= MO_64);
+ tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
+}
+
/* Perform a vector negation using normal negation and a mask.
Compare gen_subv_mask above. */
static void gen_negv_mask(TCGv_i64 d, TCGv_i64 b, TCGv_i64 m)
diff --git a/tcg/tcg-op-vec.c b/tcg/tcg-op-vec.c
index 675aa09258..36f35022ac 100644
--- a/tcg/tcg-op-vec.c
+++ b/tcg/tcg-op-vec.c
@@ -433,3 +433,23 @@ void tcg_gen_ussub_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b)
{
do_op3(vece, r, a, b, INDEX_op_ussub_vec);
}
+
+void tcg_gen_smin_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b)
+{
+ do_op3(vece, r, a, b, INDEX_op_smin_vec);
+}
+
+void tcg_gen_umin_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b)
+{
+ do_op3(vece, r, a, b, INDEX_op_umin_vec);
+}
+
+void tcg_gen_smax_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b)
+{
+ do_op3(vece, r, a, b, INDEX_op_smax_vec);
+}
+
+void tcg_gen_umax_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b)
+{
+ do_op3(vece, r, a, b, INDEX_op_umax_vec);
+}
diff --git a/tcg/tcg.c b/tcg/tcg.c
index 15ed5af007..1ae1e788f6 100644
--- a/tcg/tcg.c
+++ b/tcg/tcg.c
@@ -1612,6 +1612,11 @@ bool tcg_op_supported(TCGOpcode op)
case INDEX_op_sssub_vec:
case INDEX_op_ussub_vec:
return have_vec && TCG_TARGET_HAS_sat_vec;
+ case INDEX_op_smin_vec:
+ case INDEX_op_umin_vec:
+ case INDEX_op_smax_vec:
+ case INDEX_op_umax_vec:
+ return have_vec && TCG_TARGET_HAS_minmax_vec;
default:
tcg_debug_assert(op > INDEX_op_last_generic && op < NB_OPS);
--
2.17.2
^ permalink raw reply related [flat|nested] 18+ messages in thread
* [Qemu-devel] [PATCH v2 06/10] tcg/i386: Split subroutines out of tcg_expand_vec_op
2019-01-04 22:31 [Qemu-devel] [PATCH v2 00/10] tcg vector improvements Richard Henderson
` (4 preceding siblings ...)
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 05/10] tcg: Add opcodes for vector minmax arithmetic Richard Henderson
@ 2019-01-04 22:31 ` Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 07/10] tcg/i386: Implement vector saturating arithmetic Richard Henderson
` (4 subsequent siblings)
10 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2019-01-04 22:31 UTC (permalink / raw)
To: qemu-devel
This routine was becoming too large.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
tcg/i386/tcg-target.inc.c | 459 +++++++++++++++++++-------------------
1 file changed, 232 insertions(+), 227 deletions(-)
diff --git a/tcg/i386/tcg-target.inc.c b/tcg/i386/tcg-target.inc.c
index c21c3272f2..ad97386d06 100644
--- a/tcg/i386/tcg-target.inc.c
+++ b/tcg/i386/tcg-target.inc.c
@@ -3079,253 +3079,258 @@ int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
}
}
+static void expand_vec_shi(TCGType type, unsigned vece, bool shr,
+ TCGv_vec v0, TCGv_vec v1, TCGArg imm)
+{
+ TCGv_vec t1, t2;
+
+ tcg_debug_assert(vece == MO_8);
+
+ t1 = tcg_temp_new_vec(type);
+ t2 = tcg_temp_new_vec(type);
+
+ /* Unpack to W, shift, and repack. Tricky bits:
+ (1) Use punpck*bw x,x to produce DDCCBBAA,
+ i.e. duplicate in other half of the 16-bit lane.
+ (2) For right-shift, add 8 so that the high half of
+ the lane becomes zero. For left-shift, we must
+ shift up and down again.
+ (3) Step 2 leaves high half zero such that PACKUSWB
+ (pack with unsigned saturation) does not modify
+ the quantity. */
+ vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
+ tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
+ vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
+ tcgv_vec_arg(t2), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
+
+ if (shr) {
+ tcg_gen_shri_vec(MO_16, t1, t1, imm + 8);
+ tcg_gen_shri_vec(MO_16, t2, t2, imm + 8);
+ } else {
+ tcg_gen_shli_vec(MO_16, t1, t1, imm + 8);
+ tcg_gen_shli_vec(MO_16, t2, t2, imm + 8);
+ tcg_gen_shri_vec(MO_16, t1, t1, 8);
+ tcg_gen_shri_vec(MO_16, t2, t2, 8);
+ }
+
+ vec_gen_3(INDEX_op_x86_packus_vec, type, MO_8,
+ tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t2));
+ tcg_temp_free_vec(t1);
+ tcg_temp_free_vec(t2);
+}
+
+static void expand_vec_sari(TCGType type, unsigned vece,
+ TCGv_vec v0, TCGv_vec v1, TCGArg imm)
+{
+ TCGv_vec t1, t2;
+
+ switch (vece) {
+ case MO_8:
+ /* Unpack to W, shift, and repack, as in expand_vec_shi. */
+ t1 = tcg_temp_new_vec(type);
+ t2 = tcg_temp_new_vec(type);
+ vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
+ tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
+ vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
+ tcgv_vec_arg(t2), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
+ tcg_gen_sari_vec(MO_16, t1, t1, imm + 8);
+ tcg_gen_sari_vec(MO_16, t2, t2, imm + 8);
+ vec_gen_3(INDEX_op_x86_packss_vec, type, MO_8,
+ tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t2));
+ tcg_temp_free_vec(t1);
+ tcg_temp_free_vec(t2);
+ break;
+
+ case MO_64:
+ if (imm <= 32) {
+ /* We can emulate a small sign extend by performing an arithmetic
+ * 32-bit shift and overwriting the high half of a 64-bit logical
+ * shift (note that the ISA says shift of 32 is valid).
+ */
+ t1 = tcg_temp_new_vec(type);
+ tcg_gen_sari_vec(MO_32, t1, v1, imm);
+ tcg_gen_shri_vec(MO_64, v0, v1, imm);
+ vec_gen_4(INDEX_op_x86_blend_vec, type, MO_32,
+ tcgv_vec_arg(v0), tcgv_vec_arg(v0),
+ tcgv_vec_arg(t1), 0xaa);
+ tcg_temp_free_vec(t1);
+ } else {
+ /* Otherwise we will need to use a compare vs 0 to produce
+ * the sign-extend, shift and merge.
+ */
+ t1 = tcg_const_zeros_vec(type);
+ tcg_gen_cmp_vec(TCG_COND_GT, MO_64, t1, t1, v1);
+ tcg_gen_shri_vec(MO_64, v0, v1, imm);
+ tcg_gen_shli_vec(MO_64, t1, t1, 64 - imm);
+ tcg_gen_or_vec(MO_64, v0, v0, t1);
+ tcg_temp_free_vec(t1);
+ }
+ break;
+
+ default:
+ g_assert_not_reached();
+ }
+}
+
+static void expand_vec_mul(TCGType type, unsigned vece,
+ TCGv_vec v0, TCGv_vec v1, TCGv_vec v2)
+{
+ TCGv_vec t1, t2, t3, t4;
+
+ tcg_debug_assert(vece == MO_8);
+
+ /*
+ * Unpack v1 bytes to words, 0 | x.
+ * Unpack v2 bytes to words, y | 0.
+ * This leaves the 8-bit result, x * y, with 8 bits of right padding.
+ * Shift logical right by 8 bits to clear the high 8 bytes before
+ * using an unsigned saturated pack.
+ *
+ * The difference between the V64, V128 and V256 cases is merely how
+ * we distribute the expansion between temporaries.
+ */
+ switch (type) {
+ case TCG_TYPE_V64:
+ t1 = tcg_temp_new_vec(TCG_TYPE_V128);
+ t2 = tcg_temp_new_vec(TCG_TYPE_V128);
+ tcg_gen_dup16i_vec(t2, 0);
+ vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
+ tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(t2));
+ vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
+ tcgv_vec_arg(t2), tcgv_vec_arg(t2), tcgv_vec_arg(v2));
+ tcg_gen_mul_vec(MO_16, t1, t1, t2);
+ tcg_gen_shri_vec(MO_16, t1, t1, 8);
+ vec_gen_3(INDEX_op_x86_packus_vec, TCG_TYPE_V128, MO_8,
+ tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t1));
+ tcg_temp_free_vec(t1);
+ tcg_temp_free_vec(t2);
+ break;
+
+ case TCG_TYPE_V128:
+ case TCG_TYPE_V256:
+ t1 = tcg_temp_new_vec(type);
+ t2 = tcg_temp_new_vec(type);
+ t3 = tcg_temp_new_vec(type);
+ t4 = tcg_temp_new_vec(type);
+ tcg_gen_dup16i_vec(t4, 0);
+ vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
+ tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(t4));
+ vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
+ tcgv_vec_arg(t2), tcgv_vec_arg(t4), tcgv_vec_arg(v2));
+ vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
+ tcgv_vec_arg(t3), tcgv_vec_arg(v1), tcgv_vec_arg(t4));
+ vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
+ tcgv_vec_arg(t4), tcgv_vec_arg(t4), tcgv_vec_arg(v2));
+ tcg_gen_mul_vec(MO_16, t1, t1, t2);
+ tcg_gen_mul_vec(MO_16, t3, t3, t4);
+ tcg_gen_shri_vec(MO_16, t1, t1, 8);
+ tcg_gen_shri_vec(MO_16, t3, t3, 8);
+ vec_gen_3(INDEX_op_x86_packus_vec, type, MO_8,
+ tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t3));
+ tcg_temp_free_vec(t1);
+ tcg_temp_free_vec(t2);
+ tcg_temp_free_vec(t3);
+ tcg_temp_free_vec(t4);
+ break;
+
+ default:
+ g_assert_not_reached();
+ }
+}
+
+static void expand_vec_cmp(TCGType type, unsigned vece, TCGv_vec v0,
+ TCGv_vec v1, TCGv_vec v2, TCGCond cond)
+{
+ enum {
+ NEED_SWAP = 1,
+ NEED_INV = 2,
+ NEED_BIAS = 4
+ };
+ static const uint8_t fixups[16] = {
+ [0 ... 15] = -1,
+ [TCG_COND_EQ] = 0,
+ [TCG_COND_NE] = NEED_INV,
+ [TCG_COND_GT] = 0,
+ [TCG_COND_LT] = NEED_SWAP,
+ [TCG_COND_LE] = NEED_INV,
+ [TCG_COND_GE] = NEED_SWAP | NEED_INV,
+ [TCG_COND_GTU] = NEED_BIAS,
+ [TCG_COND_LTU] = NEED_BIAS | NEED_SWAP,
+ [TCG_COND_LEU] = NEED_BIAS | NEED_INV,
+ [TCG_COND_GEU] = NEED_BIAS | NEED_SWAP | NEED_INV,
+ };
+ TCGv_vec t1, t2;
+ uint8_t fixup;
+
+ fixup = fixups[cond & 15];
+ tcg_debug_assert(fixup != 0xff);
+
+ if (fixup & NEED_INV) {
+ cond = tcg_invert_cond(cond);
+ }
+ if (fixup & NEED_SWAP) {
+ t1 = v1, v1 = v2, v2 = t1;
+ cond = tcg_swap_cond(cond);
+ }
+
+ t1 = t2 = NULL;
+ if (fixup & NEED_BIAS) {
+ t1 = tcg_temp_new_vec(type);
+ t2 = tcg_temp_new_vec(type);
+ tcg_gen_dupi_vec(vece, t2, 1ull << ((8 << vece) - 1));
+ tcg_gen_sub_vec(vece, t1, v1, t2);
+ tcg_gen_sub_vec(vece, t2, v2, t2);
+ v1 = t1;
+ v2 = t2;
+ cond = tcg_signed_cond(cond);
+ }
+
+ tcg_debug_assert(cond == TCG_COND_EQ || cond == TCG_COND_GT);
+ /* Expand directly; do not recurse. */
+ vec_gen_4(INDEX_op_cmp_vec, type, vece,
+ tcgv_vec_arg(v0), tcgv_vec_arg(v1), tcgv_vec_arg(v2), cond);
+
+ if (t1) {
+ tcg_temp_free_vec(t1);
+ if (t2) {
+ tcg_temp_free_vec(t2);
+ }
+ }
+ if (fixup & NEED_INV) {
+ tcg_gen_not_vec(vece, v0, v0);
+ }
+}
+
void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
TCGArg a0, ...)
{
va_list va;
- TCGArg a1, a2;
- TCGv_vec v0, t1, t2, t3, t4;
+ TCGArg a2;
+ TCGv_vec v0, v1, v2;
va_start(va, a0);
v0 = temp_tcgv_vec(arg_temp(a0));
+ v1 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
+ a2 = va_arg(va, TCGArg);
switch (opc) {
case INDEX_op_shli_vec:
case INDEX_op_shri_vec:
- tcg_debug_assert(vece == MO_8);
- a1 = va_arg(va, TCGArg);
- a2 = va_arg(va, TCGArg);
- /* Unpack to W, shift, and repack. Tricky bits:
- (1) Use punpck*bw x,x to produce DDCCBBAA,
- i.e. duplicate in other half of the 16-bit lane.
- (2) For right-shift, add 8 so that the high half of
- the lane becomes zero. For left-shift, we must
- shift up and down again.
- (3) Step 2 leaves high half zero such that PACKUSWB
- (pack with unsigned saturation) does not modify
- the quantity. */
- t1 = tcg_temp_new_vec(type);
- t2 = tcg_temp_new_vec(type);
- vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
- tcgv_vec_arg(t1), a1, a1);
- vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
- tcgv_vec_arg(t2), a1, a1);
- if (opc == INDEX_op_shri_vec) {
- vec_gen_3(INDEX_op_shri_vec, type, MO_16,
- tcgv_vec_arg(t1), tcgv_vec_arg(t1), a2 + 8);
- vec_gen_3(INDEX_op_shri_vec, type, MO_16,
- tcgv_vec_arg(t2), tcgv_vec_arg(t2), a2 + 8);
- } else {
- vec_gen_3(INDEX_op_shli_vec, type, MO_16,
- tcgv_vec_arg(t1), tcgv_vec_arg(t1), a2 + 8);
- vec_gen_3(INDEX_op_shli_vec, type, MO_16,
- tcgv_vec_arg(t2), tcgv_vec_arg(t2), a2 + 8);
- vec_gen_3(INDEX_op_shri_vec, type, MO_16,
- tcgv_vec_arg(t1), tcgv_vec_arg(t1), 8);
- vec_gen_3(INDEX_op_shri_vec, type, MO_16,
- tcgv_vec_arg(t2), tcgv_vec_arg(t2), 8);
- }
- vec_gen_3(INDEX_op_x86_packus_vec, type, MO_8,
- a0, tcgv_vec_arg(t1), tcgv_vec_arg(t2));
- tcg_temp_free_vec(t1);
- tcg_temp_free_vec(t2);
+ expand_vec_shi(type, vece, opc == INDEX_op_shri_vec, v0, v1, a2);
break;
case INDEX_op_sari_vec:
- a1 = va_arg(va, TCGArg);
- a2 = va_arg(va, TCGArg);
- if (vece == MO_8) {
- /* Unpack to W, shift, and repack, as above. */
- t1 = tcg_temp_new_vec(type);
- t2 = tcg_temp_new_vec(type);
- vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
- tcgv_vec_arg(t1), a1, a1);
- vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
- tcgv_vec_arg(t2), a1, a1);
- vec_gen_3(INDEX_op_sari_vec, type, MO_16,
- tcgv_vec_arg(t1), tcgv_vec_arg(t1), a2 + 8);
- vec_gen_3(INDEX_op_sari_vec, type, MO_16,
- tcgv_vec_arg(t2), tcgv_vec_arg(t2), a2 + 8);
- vec_gen_3(INDEX_op_x86_packss_vec, type, MO_8,
- a0, tcgv_vec_arg(t1), tcgv_vec_arg(t2));
- tcg_temp_free_vec(t1);
- tcg_temp_free_vec(t2);
- break;
- }
- tcg_debug_assert(vece == MO_64);
- /* MO_64: If the shift is <= 32, we can emulate the sign extend by
- performing an arithmetic 32-bit shift and overwriting the high
- half of the result (note that the ISA says shift of 32 is valid). */
- if (a2 <= 32) {
- t1 = tcg_temp_new_vec(type);
- vec_gen_3(INDEX_op_sari_vec, type, MO_32, tcgv_vec_arg(t1), a1, a2);
- vec_gen_3(INDEX_op_shri_vec, type, MO_64, a0, a1, a2);
- vec_gen_4(INDEX_op_x86_blend_vec, type, MO_32,
- a0, a0, tcgv_vec_arg(t1), 0xaa);
- tcg_temp_free_vec(t1);
- break;
- }
- /* Otherwise we will need to use a compare vs 0 to produce the
- sign-extend, shift and merge. */
- t1 = tcg_temp_new_vec(type);
- t2 = tcg_const_zeros_vec(type);
- vec_gen_4(INDEX_op_cmp_vec, type, MO_64,
- tcgv_vec_arg(t1), tcgv_vec_arg(t2), a1, TCG_COND_GT);
- tcg_temp_free_vec(t2);
- vec_gen_3(INDEX_op_shri_vec, type, MO_64, a0, a1, a2);
- vec_gen_3(INDEX_op_shli_vec, type, MO_64,
- tcgv_vec_arg(t1), tcgv_vec_arg(t1), 64 - a2);
- vec_gen_3(INDEX_op_or_vec, type, MO_64, a0, a0, tcgv_vec_arg(t1));
- tcg_temp_free_vec(t1);
+ expand_vec_sari(type, vece, v0, v1, a2);
break;
case INDEX_op_mul_vec:
- tcg_debug_assert(vece == MO_8);
- a1 = va_arg(va, TCGArg);
- a2 = va_arg(va, TCGArg);
- switch (type) {
- case TCG_TYPE_V64:
- t1 = tcg_temp_new_vec(TCG_TYPE_V128);
- t2 = tcg_temp_new_vec(TCG_TYPE_V128);
- tcg_gen_dup16i_vec(t2, 0);
- vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
- tcgv_vec_arg(t1), a1, tcgv_vec_arg(t2));
- vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
- tcgv_vec_arg(t2), tcgv_vec_arg(t2), a2);
- tcg_gen_mul_vec(MO_16, t1, t1, t2);
- tcg_gen_shri_vec(MO_16, t1, t1, 8);
- vec_gen_3(INDEX_op_x86_packus_vec, TCG_TYPE_V128, MO_8,
- a0, tcgv_vec_arg(t1), tcgv_vec_arg(t1));
- tcg_temp_free_vec(t1);
- tcg_temp_free_vec(t2);
- break;
-
- case TCG_TYPE_V128:
- t1 = tcg_temp_new_vec(TCG_TYPE_V128);
- t2 = tcg_temp_new_vec(TCG_TYPE_V128);
- t3 = tcg_temp_new_vec(TCG_TYPE_V128);
- t4 = tcg_temp_new_vec(TCG_TYPE_V128);
- tcg_gen_dup16i_vec(t4, 0);
- vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
- tcgv_vec_arg(t1), a1, tcgv_vec_arg(t4));
- vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
- tcgv_vec_arg(t2), tcgv_vec_arg(t4), a2);
- vec_gen_3(INDEX_op_x86_punpckh_vec, TCG_TYPE_V128, MO_8,
- tcgv_vec_arg(t3), a1, tcgv_vec_arg(t4));
- vec_gen_3(INDEX_op_x86_punpckh_vec, TCG_TYPE_V128, MO_8,
- tcgv_vec_arg(t4), tcgv_vec_arg(t4), a2);
- tcg_gen_mul_vec(MO_16, t1, t1, t2);
- tcg_gen_mul_vec(MO_16, t3, t3, t4);
- tcg_gen_shri_vec(MO_16, t1, t1, 8);
- tcg_gen_shri_vec(MO_16, t3, t3, 8);
- vec_gen_3(INDEX_op_x86_packus_vec, TCG_TYPE_V128, MO_8,
- a0, tcgv_vec_arg(t1), tcgv_vec_arg(t3));
- tcg_temp_free_vec(t1);
- tcg_temp_free_vec(t2);
- tcg_temp_free_vec(t3);
- tcg_temp_free_vec(t4);
- break;
-
- case TCG_TYPE_V256:
- t1 = tcg_temp_new_vec(TCG_TYPE_V256);
- t2 = tcg_temp_new_vec(TCG_TYPE_V256);
- t3 = tcg_temp_new_vec(TCG_TYPE_V256);
- t4 = tcg_temp_new_vec(TCG_TYPE_V256);
- tcg_gen_dup16i_vec(t4, 0);
- /* a1: A[0-7] ... D[0-7]; a2: W[0-7] ... Z[0-7]
- t1: extends of B[0-7], D[0-7]
- t2: extends of X[0-7], Z[0-7]
- t3: extends of A[0-7], C[0-7]
- t4: extends of W[0-7], Y[0-7]. */
- vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V256, MO_8,
- tcgv_vec_arg(t1), a1, tcgv_vec_arg(t4));
- vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V256, MO_8,
- tcgv_vec_arg(t2), tcgv_vec_arg(t4), a2);
- vec_gen_3(INDEX_op_x86_punpckh_vec, TCG_TYPE_V256, MO_8,
- tcgv_vec_arg(t3), a1, tcgv_vec_arg(t4));
- vec_gen_3(INDEX_op_x86_punpckh_vec, TCG_TYPE_V256, MO_8,
- tcgv_vec_arg(t4), tcgv_vec_arg(t4), a2);
- /* t1: BX DZ; t2: AW CY. */
- tcg_gen_mul_vec(MO_16, t1, t1, t2);
- tcg_gen_mul_vec(MO_16, t3, t3, t4);
- tcg_gen_shri_vec(MO_16, t1, t1, 8);
- tcg_gen_shri_vec(MO_16, t3, t3, 8);
- /* a0: AW BX CY DZ. */
- vec_gen_3(INDEX_op_x86_packus_vec, TCG_TYPE_V256, MO_8,
- a0, tcgv_vec_arg(t1), tcgv_vec_arg(t3));
- tcg_temp_free_vec(t1);
- tcg_temp_free_vec(t2);
- tcg_temp_free_vec(t3);
- tcg_temp_free_vec(t4);
- break;
-
- default:
- g_assert_not_reached();
- }
+ v2 = temp_tcgv_vec(arg_temp(a2));
+ expand_vec_mul(type, vece, v0, v1, v2);
break;
case INDEX_op_cmp_vec:
- {
- enum {
- NEED_SWAP = 1,
- NEED_INV = 2,
- NEED_BIAS = 4
- };
- static const uint8_t fixups[16] = {
- [0 ... 15] = -1,
- [TCG_COND_EQ] = 0,
- [TCG_COND_NE] = NEED_INV,
- [TCG_COND_GT] = 0,
- [TCG_COND_LT] = NEED_SWAP,
- [TCG_COND_LE] = NEED_INV,
- [TCG_COND_GE] = NEED_SWAP | NEED_INV,
- [TCG_COND_GTU] = NEED_BIAS,
- [TCG_COND_LTU] = NEED_BIAS | NEED_SWAP,
- [TCG_COND_LEU] = NEED_BIAS | NEED_INV,
- [TCG_COND_GEU] = NEED_BIAS | NEED_SWAP | NEED_INV,
- };
-
- TCGCond cond;
- uint8_t fixup;
-
- a1 = va_arg(va, TCGArg);
- a2 = va_arg(va, TCGArg);
- cond = va_arg(va, TCGArg);
- fixup = fixups[cond & 15];
- tcg_debug_assert(fixup != 0xff);
-
- if (fixup & NEED_INV) {
- cond = tcg_invert_cond(cond);
- }
- if (fixup & NEED_SWAP) {
- TCGArg t;
- t = a1, a1 = a2, a2 = t;
- cond = tcg_swap_cond(cond);
- }
-
- t1 = t2 = NULL;
- if (fixup & NEED_BIAS) {
- t1 = tcg_temp_new_vec(type);
- t2 = tcg_temp_new_vec(type);
- tcg_gen_dupi_vec(vece, t2, 1ull << ((8 << vece) - 1));
- tcg_gen_sub_vec(vece, t1, temp_tcgv_vec(arg_temp(a1)), t2);
- tcg_gen_sub_vec(vece, t2, temp_tcgv_vec(arg_temp(a2)), t2);
- a1 = tcgv_vec_arg(t1);
- a2 = tcgv_vec_arg(t2);
- cond = tcg_signed_cond(cond);
- }
-
- tcg_debug_assert(cond == TCG_COND_EQ || cond == TCG_COND_GT);
- vec_gen_4(INDEX_op_cmp_vec, type, vece, a0, a1, a2, cond);
-
- if (fixup & NEED_BIAS) {
- tcg_temp_free_vec(t1);
- tcg_temp_free_vec(t2);
- }
- if (fixup & NEED_INV) {
- tcg_gen_not_vec(vece, v0, v0);
- }
- }
+ v2 = temp_tcgv_vec(arg_temp(a2));
+ expand_vec_cmp(type, vece, v0, v1, v2, va_arg(va, TCGArg));
break;
default:
--
2.17.2
^ permalink raw reply related [flat|nested] 18+ messages in thread
* [Qemu-devel] [PATCH v2 07/10] tcg/i386: Implement vector saturating arithmetic
2019-01-04 22:31 [Qemu-devel] [PATCH v2 00/10] tcg vector improvements Richard Henderson
` (5 preceding siblings ...)
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 06/10] tcg/i386: Split subroutines out of tcg_expand_vec_op Richard Henderson
@ 2019-01-04 22:31 ` Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 08/10] tcg/i386: Implement vector minmax arithmetic Richard Henderson
` (3 subsequent siblings)
10 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2019-01-04 22:31 UTC (permalink / raw)
To: qemu-devel
Only MO_8 and MO_16 are implemented, since that's all the
instruction set provides.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
tcg/i386/tcg-target.h | 2 +-
tcg/i386/tcg-target.inc.c | 42 +++++++++++++++++++++++++++++++++++++++
2 files changed, 43 insertions(+), 1 deletion(-)
diff --git a/tcg/i386/tcg-target.h b/tcg/i386/tcg-target.h
index 7bd7eae672..efbd5a6fc9 100644
--- a/tcg/i386/tcg-target.h
+++ b/tcg/i386/tcg-target.h
@@ -185,7 +185,7 @@ extern bool have_avx2;
#define TCG_TARGET_HAS_shv_vec 0
#define TCG_TARGET_HAS_cmp_vec 1
#define TCG_TARGET_HAS_mul_vec 1
-#define TCG_TARGET_HAS_sat_vec 0
+#define TCG_TARGET_HAS_sat_vec 1
#define TCG_TARGET_HAS_minmax_vec 0
#define TCG_TARGET_deposit_i32_valid(ofs, len) \
diff --git a/tcg/i386/tcg-target.inc.c b/tcg/i386/tcg-target.inc.c
index ad97386d06..feec40a412 100644
--- a/tcg/i386/tcg-target.inc.c
+++ b/tcg/i386/tcg-target.inc.c
@@ -377,6 +377,10 @@ static inline int tcg_target_const_match(tcg_target_long val, TCGType type,
#define OPC_PADDW (0xfd | P_EXT | P_DATA16)
#define OPC_PADDD (0xfe | P_EXT | P_DATA16)
#define OPC_PADDQ (0xd4 | P_EXT | P_DATA16)
+#define OPC_PADDSB (0xec | P_EXT | P_DATA16)
+#define OPC_PADDSW (0xed | P_EXT | P_DATA16)
+#define OPC_PADDUB (0xdc | P_EXT | P_DATA16)
+#define OPC_PADDUW (0xdd | P_EXT | P_DATA16)
#define OPC_PAND (0xdb | P_EXT | P_DATA16)
#define OPC_PANDN (0xdf | P_EXT | P_DATA16)
#define OPC_PBLENDW (0x0e | P_EXT3A | P_DATA16)
@@ -408,6 +412,10 @@ static inline int tcg_target_const_match(tcg_target_long val, TCGType type,
#define OPC_PSUBW (0xf9 | P_EXT | P_DATA16)
#define OPC_PSUBD (0xfa | P_EXT | P_DATA16)
#define OPC_PSUBQ (0xfb | P_EXT | P_DATA16)
+#define OPC_PSUBSB (0xe8 | P_EXT | P_DATA16)
+#define OPC_PSUBSW (0xe9 | P_EXT | P_DATA16)
+#define OPC_PSUBUB (0xd8 | P_EXT | P_DATA16)
+#define OPC_PSUBUW (0xd9 | P_EXT | P_DATA16)
#define OPC_PUNPCKLBW (0x60 | P_EXT | P_DATA16)
#define OPC_PUNPCKLWD (0x61 | P_EXT | P_DATA16)
#define OPC_PUNPCKLDQ (0x62 | P_EXT | P_DATA16)
@@ -2591,9 +2599,21 @@ static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
static int const add_insn[4] = {
OPC_PADDB, OPC_PADDW, OPC_PADDD, OPC_PADDQ
};
+ static int const ssadd_insn[4] = {
+ OPC_PADDSB, OPC_PADDSW, OPC_UD2, OPC_UD2
+ };
+ static int const usadd_insn[4] = {
+ OPC_PADDSB, OPC_PADDSW, OPC_UD2, OPC_UD2
+ };
static int const sub_insn[4] = {
OPC_PSUBB, OPC_PSUBW, OPC_PSUBD, OPC_PSUBQ
};
+ static int const sssub_insn[4] = {
+ OPC_PSUBSB, OPC_PSUBSW, OPC_UD2, OPC_UD2
+ };
+ static int const ussub_insn[4] = {
+ OPC_PSUBSB, OPC_PSUBSW, OPC_UD2, OPC_UD2
+ };
static int const mul_insn[4] = {
OPC_UD2, OPC_PMULLW, OPC_PMULLD, OPC_UD2
};
@@ -2631,9 +2651,21 @@ static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
case INDEX_op_add_vec:
insn = add_insn[vece];
goto gen_simd;
+ case INDEX_op_ssadd_vec:
+ insn = ssadd_insn[vece];
+ goto gen_simd;
+ case INDEX_op_usadd_vec:
+ insn = usadd_insn[vece];
+ goto gen_simd;
case INDEX_op_sub_vec:
insn = sub_insn[vece];
goto gen_simd;
+ case INDEX_op_sssub_vec:
+ insn = sssub_insn[vece];
+ goto gen_simd;
+ case INDEX_op_ussub_vec:
+ insn = ussub_insn[vece];
+ goto gen_simd;
case INDEX_op_mul_vec:
insn = mul_insn[vece];
goto gen_simd;
@@ -3007,6 +3039,10 @@ static const TCGTargetOpDef *tcg_target_op_def(TCGOpcode op)
case INDEX_op_or_vec:
case INDEX_op_xor_vec:
case INDEX_op_andc_vec:
+ case INDEX_op_ssadd_vec:
+ case INDEX_op_usadd_vec:
+ case INDEX_op_sssub_vec:
+ case INDEX_op_ussub_vec:
case INDEX_op_cmp_vec:
case INDEX_op_x86_shufps_vec:
case INDEX_op_x86_blend_vec:
@@ -3074,6 +3110,12 @@ int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
}
return 1;
+ case INDEX_op_ssadd_vec:
+ case INDEX_op_usadd_vec:
+ case INDEX_op_sssub_vec:
+ case INDEX_op_ussub_vec:
+ return vece <= MO_16;
+
default:
return 0;
}
--
2.17.2
^ permalink raw reply related [flat|nested] 18+ messages in thread
* [Qemu-devel] [PATCH v2 08/10] tcg/i386: Implement vector minmax arithmetic
2019-01-04 22:31 [Qemu-devel] [PATCH v2 00/10] tcg vector improvements Richard Henderson
` (6 preceding siblings ...)
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 07/10] tcg/i386: Implement vector saturating arithmetic Richard Henderson
@ 2019-01-04 22:31 ` Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 09/10] tcg/aarch64: Implement vector saturating arithmetic Richard Henderson
` (2 subsequent siblings)
10 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2019-01-04 22:31 UTC (permalink / raw)
To: qemu-devel
The avx instruction set does not directly provide MO_64.
We can still implement 64-bit with comparison and vpblendvb.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
tcg/i386/tcg-target.h | 2 +-
tcg/i386/tcg-target.inc.c | 81 +++++++++++++++++++++++++++++++++++++++
2 files changed, 82 insertions(+), 1 deletion(-)
diff --git a/tcg/i386/tcg-target.h b/tcg/i386/tcg-target.h
index efbd5a6fc9..7995fe3eab 100644
--- a/tcg/i386/tcg-target.h
+++ b/tcg/i386/tcg-target.h
@@ -186,7 +186,7 @@ extern bool have_avx2;
#define TCG_TARGET_HAS_cmp_vec 1
#define TCG_TARGET_HAS_mul_vec 1
#define TCG_TARGET_HAS_sat_vec 1
-#define TCG_TARGET_HAS_minmax_vec 0
+#define TCG_TARGET_HAS_minmax_vec 1
#define TCG_TARGET_deposit_i32_valid(ofs, len) \
(((ofs) == 0 && (len) == 8) || ((ofs) == 8 && (len) == 8) || \
diff --git a/tcg/i386/tcg-target.inc.c b/tcg/i386/tcg-target.inc.c
index feec40a412..94007c7aa5 100644
--- a/tcg/i386/tcg-target.inc.c
+++ b/tcg/i386/tcg-target.inc.c
@@ -392,6 +392,18 @@ static inline int tcg_target_const_match(tcg_target_long val, TCGType type,
#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_PMAXSB (0x3c | P_EXT38 | P_DATA16)
+#define OPC_PMAXSW (0xee | P_EXT | P_DATA16)
+#define OPC_PMAXSD (0x3d | P_EXT38 | P_DATA16)
+#define OPC_PMAXUB (0xde | P_EXT | P_DATA16)
+#define OPC_PMAXUW (0x3e | P_EXT38 | P_DATA16)
+#define OPC_PMAXUD (0x3f | P_EXT38 | P_DATA16)
+#define OPC_PMINSB (0x38 | P_EXT38 | P_DATA16)
+#define OPC_PMINSW (0xea | P_EXT | P_DATA16)
+#define OPC_PMINSD (0x39 | P_EXT38 | P_DATA16)
+#define OPC_PMINUB (0xda | P_EXT | P_DATA16)
+#define OPC_PMINUW (0x3a | P_EXT38 | P_DATA16)
+#define OPC_PMINUD (0x3b | P_EXT38 | P_DATA16)
#define OPC_PMOVSXBW (0x20 | P_EXT38 | P_DATA16)
#define OPC_PMOVSXWD (0x23 | P_EXT38 | P_DATA16)
#define OPC_PMOVSXDQ (0x25 | P_EXT38 | P_DATA16)
@@ -2638,6 +2650,18 @@ static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
static int const packus_insn[4] = {
OPC_PACKUSWB, OPC_PACKUSDW, OPC_UD2, OPC_UD2
};
+ static int const smin_insn[4] = {
+ OPC_PMINSB, OPC_PMINSW, OPC_PMINSD, OPC_UD2
+ };
+ static int const smax_insn[4] = {
+ OPC_PMAXSB, OPC_PMAXSW, OPC_PMAXSD, OPC_UD2
+ };
+ static int const umin_insn[4] = {
+ OPC_PMINUB, OPC_PMINUW, OPC_PMINUD, OPC_UD2
+ };
+ static int const umax_insn[4] = {
+ OPC_PMAXUB, OPC_PMAXUW, OPC_PMAXUD, OPC_UD2
+ };
TCGType type = vecl + TCG_TYPE_V64;
int insn, sub;
@@ -2678,6 +2702,18 @@ static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
case INDEX_op_xor_vec:
insn = OPC_PXOR;
goto gen_simd;
+ case INDEX_op_smin_vec:
+ insn = smin_insn[vece];
+ goto gen_simd;
+ case INDEX_op_umin_vec:
+ insn = umin_insn[vece];
+ goto gen_simd;
+ case INDEX_op_smax_vec:
+ insn = smax_insn[vece];
+ goto gen_simd;
+ case INDEX_op_umax_vec:
+ insn = umax_insn[vece];
+ goto gen_simd;
case INDEX_op_x86_punpckl_vec:
insn = punpckl_insn[vece];
goto gen_simd;
@@ -3043,6 +3079,10 @@ static const TCGTargetOpDef *tcg_target_op_def(TCGOpcode op)
case INDEX_op_usadd_vec:
case INDEX_op_sssub_vec:
case INDEX_op_ussub_vec:
+ case INDEX_op_smin_vec:
+ case INDEX_op_umin_vec:
+ case INDEX_op_smax_vec:
+ case INDEX_op_umax_vec:
case INDEX_op_cmp_vec:
case INDEX_op_x86_shufps_vec:
case INDEX_op_x86_blend_vec:
@@ -3115,6 +3155,11 @@ int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
case INDEX_op_sssub_vec:
case INDEX_op_ussub_vec:
return vece <= MO_16;
+ case INDEX_op_smin_vec:
+ case INDEX_op_smax_vec:
+ case INDEX_op_umin_vec:
+ case INDEX_op_umax_vec:
+ return vece <= MO_32 ? 1 : -1;
default:
return 0;
@@ -3343,6 +3388,25 @@ static void expand_vec_cmp(TCGType type, unsigned vece, TCGv_vec v0,
}
}
+static void expand_vec_minmax(TCGType type, unsigned vece,
+ TCGCond cond, bool min,
+ TCGv_vec v0, TCGv_vec v1, TCGv_vec v2)
+{
+ TCGv_vec t1 = tcg_temp_new_vec(type);
+
+ tcg_debug_assert(vece == MO_64);
+
+ tcg_gen_cmp_vec(cond, vece, t1, v1, v2);
+ if (min) {
+ TCGv_vec t2;
+ t2 = v1, v1 = v2, v2 = t2;
+ }
+ vec_gen_4(INDEX_op_x86_vpblendvb_vec, type, vece,
+ tcgv_vec_arg(v0), tcgv_vec_arg(v1),
+ tcgv_vec_arg(v2), tcgv_vec_arg(t1));
+ tcg_temp_free_vec(t1);
+}
+
void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
TCGArg a0, ...)
{
@@ -3375,6 +3439,23 @@ void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
expand_vec_cmp(type, vece, v0, v1, v2, va_arg(va, TCGArg));
break;
+ case INDEX_op_smin_vec:
+ v2 = temp_tcgv_vec(arg_temp(a2));
+ expand_vec_minmax(type, vece, TCG_COND_GT, true, v0, v1, v2);
+ break;
+ case INDEX_op_smax_vec:
+ v2 = temp_tcgv_vec(arg_temp(a2));
+ expand_vec_minmax(type, vece, TCG_COND_GT, false, v0, v1, v2);
+ break;
+ case INDEX_op_umin_vec:
+ v2 = temp_tcgv_vec(arg_temp(a2));
+ expand_vec_minmax(type, vece, TCG_COND_GTU, true, v0, v1, v2);
+ break;
+ case INDEX_op_umax_vec:
+ v2 = temp_tcgv_vec(arg_temp(a2));
+ expand_vec_minmax(type, vece, TCG_COND_GTU, false, v0, v1, v2);
+ break;
+
default:
break;
}
--
2.17.2
^ permalink raw reply related [flat|nested] 18+ messages in thread
* [Qemu-devel] [PATCH v2 09/10] tcg/aarch64: Implement vector saturating arithmetic
2019-01-04 22:31 [Qemu-devel] [PATCH v2 00/10] tcg vector improvements Richard Henderson
` (7 preceding siblings ...)
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 08/10] tcg/i386: Implement vector minmax arithmetic Richard Henderson
@ 2019-01-04 22:31 ` Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 10/10] tcg/aarch64: Implement vector minmax arithmetic Richard Henderson
2019-01-07 13:11 ` [Qemu-devel] [PATCH v2 00/10] tcg vector improvements Mark Cave-Ayland
10 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2019-01-04 22:31 UTC (permalink / raw)
To: qemu-devel
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
tcg/aarch64/tcg-target.h | 2 +-
tcg/aarch64/tcg-target.inc.c | 24 ++++++++++++++++++++++++
2 files changed, 25 insertions(+), 1 deletion(-)
diff --git a/tcg/aarch64/tcg-target.h b/tcg/aarch64/tcg-target.h
index 545a6eec75..a1884543d0 100644
--- a/tcg/aarch64/tcg-target.h
+++ b/tcg/aarch64/tcg-target.h
@@ -135,7 +135,7 @@ typedef enum {
#define TCG_TARGET_HAS_shv_vec 0
#define TCG_TARGET_HAS_cmp_vec 1
#define TCG_TARGET_HAS_mul_vec 1
-#define TCG_TARGET_HAS_sat_vec 0
+#define TCG_TARGET_HAS_sat_vec 1
#define TCG_TARGET_HAS_minmax_vec 0
#define TCG_TARGET_DEFAULT_MO (0)
diff --git a/tcg/aarch64/tcg-target.inc.c b/tcg/aarch64/tcg-target.inc.c
index 0562e0aa40..b2b011f130 100644
--- a/tcg/aarch64/tcg-target.inc.c
+++ b/tcg/aarch64/tcg-target.inc.c
@@ -528,6 +528,10 @@ typedef enum {
I3616_CMHI = 0x2e203400,
I3616_CMHS = 0x2e203c00,
I3616_CMEQ = 0x2e208c00,
+ I3616_SQADD = 0x0e200c00,
+ I3616_SQSUB = 0x0e202c00,
+ I3616_UQADD = 0x2e200c00,
+ I3616_UQSUB = 0x2e202c00,
/* AdvSIMD two-reg misc. */
I3617_CMGT0 = 0x0e208800,
@@ -2137,6 +2141,18 @@ static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
case INDEX_op_orc_vec:
tcg_out_insn(s, 3616, ORN, is_q, 0, a0, a1, a2);
break;
+ case INDEX_op_ssadd_vec:
+ tcg_out_insn(s, 3616, SQADD, is_q, vece, a0, a1, a2);
+ break;
+ case INDEX_op_sssub_vec:
+ tcg_out_insn(s, 3616, SQSUB, is_q, vece, a0, a1, a2);
+ break;
+ case INDEX_op_usadd_vec:
+ tcg_out_insn(s, 3616, UQADD, is_q, vece, a0, a1, a2);
+ break;
+ case INDEX_op_ussub_vec:
+ tcg_out_insn(s, 3616, UQSUB, is_q, vece, a0, a1, a2);
+ break;
case INDEX_op_not_vec:
tcg_out_insn(s, 3617, NOT, is_q, 0, a0, a1);
break;
@@ -2207,6 +2223,10 @@ int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
case INDEX_op_shli_vec:
case INDEX_op_shri_vec:
case INDEX_op_sari_vec:
+ case INDEX_op_ssadd_vec:
+ case INDEX_op_sssub_vec:
+ case INDEX_op_usadd_vec:
+ case INDEX_op_ussub_vec:
return 1;
case INDEX_op_mul_vec:
return vece < MO_64;
@@ -2386,6 +2406,10 @@ static const TCGTargetOpDef *tcg_target_op_def(TCGOpcode op)
case INDEX_op_xor_vec:
case INDEX_op_andc_vec:
case INDEX_op_orc_vec:
+ case INDEX_op_ssadd_vec:
+ case INDEX_op_sssub_vec:
+ case INDEX_op_usadd_vec:
+ case INDEX_op_ussub_vec:
return &w_w_w;
case INDEX_op_not_vec:
case INDEX_op_neg_vec:
--
2.17.2
^ permalink raw reply related [flat|nested] 18+ messages in thread
* [Qemu-devel] [PATCH v2 10/10] tcg/aarch64: Implement vector minmax arithmetic
2019-01-04 22:31 [Qemu-devel] [PATCH v2 00/10] tcg vector improvements Richard Henderson
` (8 preceding siblings ...)
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 09/10] tcg/aarch64: Implement vector saturating arithmetic Richard Henderson
@ 2019-01-04 22:31 ` Richard Henderson
2019-01-07 13:11 ` [Qemu-devel] [PATCH v2 00/10] tcg vector improvements Mark Cave-Ayland
10 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2019-01-04 22:31 UTC (permalink / raw)
To: qemu-devel
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
tcg/aarch64/tcg-target.h | 2 +-
tcg/aarch64/tcg-target.inc.c | 24 ++++++++++++++++++++++++
2 files changed, 25 insertions(+), 1 deletion(-)
diff --git a/tcg/aarch64/tcg-target.h b/tcg/aarch64/tcg-target.h
index a1884543d0..2d93cf404e 100644
--- a/tcg/aarch64/tcg-target.h
+++ b/tcg/aarch64/tcg-target.h
@@ -136,7 +136,7 @@ typedef enum {
#define TCG_TARGET_HAS_cmp_vec 1
#define TCG_TARGET_HAS_mul_vec 1
#define TCG_TARGET_HAS_sat_vec 1
-#define TCG_TARGET_HAS_minmax_vec 0
+#define TCG_TARGET_HAS_minmax_vec 1
#define TCG_TARGET_DEFAULT_MO (0)
#define TCG_TARGET_HAS_MEMORY_BSWAP 1
diff --git a/tcg/aarch64/tcg-target.inc.c b/tcg/aarch64/tcg-target.inc.c
index b2b011f130..ee0d5819af 100644
--- a/tcg/aarch64/tcg-target.inc.c
+++ b/tcg/aarch64/tcg-target.inc.c
@@ -528,8 +528,12 @@ typedef enum {
I3616_CMHI = 0x2e203400,
I3616_CMHS = 0x2e203c00,
I3616_CMEQ = 0x2e208c00,
+ I3616_SMAX = 0x0e206400,
+ I3616_SMIN = 0x0e206c00,
I3616_SQADD = 0x0e200c00,
I3616_SQSUB = 0x0e202c00,
+ I3616_UMAX = 0x2e206400,
+ I3616_UMIN = 0x2e206c00,
I3616_UQADD = 0x2e200c00,
I3616_UQSUB = 0x2e202c00,
@@ -2153,6 +2157,18 @@ static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
case INDEX_op_ussub_vec:
tcg_out_insn(s, 3616, UQSUB, is_q, vece, a0, a1, a2);
break;
+ case INDEX_op_smax_vec:
+ tcg_out_insn(s, 3616, SMAX, is_q, vece, a0, a1, a2);
+ break;
+ case INDEX_op_smin_vec:
+ tcg_out_insn(s, 3616, SMIN, is_q, vece, a0, a1, a2);
+ break;
+ case INDEX_op_umax_vec:
+ tcg_out_insn(s, 3616, UMAX, is_q, vece, a0, a1, a2);
+ break;
+ case INDEX_op_umin_vec:
+ tcg_out_insn(s, 3616, UMIN, is_q, vece, a0, a1, a2);
+ break;
case INDEX_op_not_vec:
tcg_out_insn(s, 3617, NOT, is_q, 0, a0, a1);
break;
@@ -2227,6 +2243,10 @@ int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
case INDEX_op_sssub_vec:
case INDEX_op_usadd_vec:
case INDEX_op_ussub_vec:
+ case INDEX_op_smax_vec:
+ case INDEX_op_smin_vec:
+ case INDEX_op_umax_vec:
+ case INDEX_op_umin_vec:
return 1;
case INDEX_op_mul_vec:
return vece < MO_64;
@@ -2410,6 +2430,10 @@ static const TCGTargetOpDef *tcg_target_op_def(TCGOpcode op)
case INDEX_op_sssub_vec:
case INDEX_op_usadd_vec:
case INDEX_op_ussub_vec:
+ case INDEX_op_smax_vec:
+ case INDEX_op_smin_vec:
+ case INDEX_op_umax_vec:
+ case INDEX_op_umin_vec:
return &w_w_w;
case INDEX_op_not_vec:
case INDEX_op_neg_vec:
--
2.17.2
^ permalink raw reply related [flat|nested] 18+ messages in thread
* Re: [Qemu-devel] [PATCH v2 00/10] tcg vector improvements
2019-01-04 22:31 [Qemu-devel] [PATCH v2 00/10] tcg vector improvements Richard Henderson
` (9 preceding siblings ...)
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 10/10] tcg/aarch64: Implement vector minmax arithmetic Richard Henderson
@ 2019-01-07 13:11 ` Mark Cave-Ayland
2019-01-23 5:09 ` Richard Henderson
10 siblings, 1 reply; 18+ messages in thread
From: Mark Cave-Ayland @ 2019-01-07 13:11 UTC (permalink / raw)
To: Richard Henderson, qemu-devel; +Cc: Howard Spoelstra
On 04/01/2019 22:31, Richard Henderson wrote:
> I've split this out from the target/ppc patch set in which
> it was developed.
>
>
> r~
>
>
> Richard Henderson (10):
> tcg: Add logical simplifications during gvec expand
> tcg: Add gvec expanders for nand, nor, eqv
> tcg: Add write_aofs to GVecGen4
> tcg: Add opcodes for vector saturated arithmetic
> tcg: Add opcodes for vector minmax arithmetic
> tcg/i386: Split subroutines out of tcg_expand_vec_op
> tcg/i386: Implement vector saturating arithmetic
> tcg/i386: Implement vector minmax arithmetic
> tcg/aarch64: Implement vector saturating arithmetic
> tcg/aarch64: Implement vector minmax arithmetic
>
> accel/tcg/tcg-runtime.h | 23 ++
> tcg/aarch64/tcg-target.h | 2 +
> tcg/i386/tcg-target.h | 2 +
> tcg/tcg-op-gvec.h | 18 ++
> tcg/tcg-op.h | 11 +
> tcg/tcg-opc.h | 8 +
> tcg/tcg.h | 2 +
> accel/tcg/tcg-runtime-gvec.c | 257 ++++++++++++++++
> tcg/aarch64/tcg-target.inc.c | 48 +++
> tcg/i386/tcg-target.inc.c | 580 +++++++++++++++++++++--------------
> tcg/tcg-op-gvec.c | 305 ++++++++++++++++--
> tcg/tcg-op-vec.c | 75 ++++-
> tcg/tcg.c | 10 +
> 13 files changed, 1078 insertions(+), 263 deletions(-)
Not sure that I'm particularly qualified to give this an R-B, however should there be
a corresponding update to tcg/README for the new instructions?
One other thing is that Howard sent me off-list a backtrace from trying my combined
branch at https://github.com/mcayland/qemu/tree/ppc-altivec-v5.5-rth booting MacOS
10.5 in the guest and ended up hitting an assert in an --enable-debug build:
Thread 5 (Thread 0x7fffe3fff700 (LWP 10627)):
#0 0x00007ffff698d53f in raise () at /lib64/libc.so.6
#1 0x00007ffff6977895 in abort () at /lib64/libc.so.6
#2 0x00007ffff6977769 in _nl_load_domain.cold.0 () at /lib64/libc.so.6
#3 0x00007ffff69859f6 in .annobin_assert.c_end () at /lib64/libc.so.6
#4 0x000055555584bb67 in do_op3 (vece=2, r=0x1848, a=0x18b8,
b=0x18f0, opc=INDEX_op_ssadd_vec) at
/home/hsp/src/qemu-altivec-55/tcg/tcg-op-vec.c:407
rt = 0x7fffdc002368
at = 0x7fffdc0023d8
bt = 0x7fffdc002410
ri = 140736884384616
ai = 140736884384728
bi = 140736884384784
type = TCG_TYPE_V128
can = 0
__PRETTY_FUNCTION__ = "do_op3"
#5 0x000055555584bbfa in tcg_gen_ssadd_vec (vece=2, r=0x1848,
a=0x18b8, b=0x18f0) at
/home/hsp/src/qemu-altivec-55/tcg/tcg-op-vec.c:419
#6 0x0000555555991887 in gen_vaddsws_vec (vece=2, t=0x1848,
sat=0x1880, a=0x18b8, b=0x18f0) at
/home/hsp/src/qemu-altivec-55/target/ppc/translate/vmx-impl.inc.c:597
x = 0x1928
#7 0x0000555555852e53 in expand_4_vec (vece=2, dofs=197872,
aofs=198288, bofs=197776, cofs=197792, oprsz=16, tysz=16,
type=TCG_TYPE_V128, write_aofs=true, fni=0x55555599182a
<gen_vaddsws_vec>) at
/home/hsp/src/qemu-altivec-55/tcg/tcg-op-gvec.c:903
t0 = 0x1848
t1 = 0x1880
t2 = 0x18b8
t3 = 0x18f0
i = 0
#8 0x0000555555853cc4 in tcg_gen_gvec_4 (dofs=197872, aofs=198288,
bofs=197776, cofs=197792, oprsz=16, maxsz=16, g=0x5555562d33c0 <g>) at
/home/hsp/src/qemu-altivec-55/tcg/tcg-op-gvec.c:1211
type = TCG_TYPE_V128
some = 21845
__PRETTY_FUNCTION__ = "tcg_gen_gvec_4"
__func__ = "tcg_gen_gvec_4"
#9 0x0000555555991987 in gen_vaddsws (ctx=0x7fffe3ffe5f0) at
/home/hsp/src/qemu-altivec-55/target/ppc/translate/vmx-impl.inc.c:597
g = {fni8 = 0x0, fni4 = 0x0, fniv = 0x55555599182a
<gen_vaddsws_vec>, fno = 0x5555559601a1 <gen_helper_vaddsws>, opc =
INDEX_op_add_vec, data = 0, vece = 2 '\002', prefer_i64 = false,
write_aofs = true}
Certainly according to patch 7 of the series only 8-bit and 16-bit accesses are
supported on i386 hosts, but shouldn't we be falling back to the previous
implementations rather than hitting an assert()?
ATB,
Mark.
^ permalink raw reply [flat|nested] 18+ messages in thread
* Re: [Qemu-devel] [PATCH v2 00/10] tcg vector improvements
2019-01-07 13:11 ` [Qemu-devel] [PATCH v2 00/10] tcg vector improvements Mark Cave-Ayland
@ 2019-01-23 5:09 ` Richard Henderson
2019-02-05 21:29 ` Mark Cave-Ayland
0 siblings, 1 reply; 18+ messages in thread
From: Richard Henderson @ 2019-01-23 5:09 UTC (permalink / raw)
To: Mark Cave-Ayland, qemu-devel; +Cc: Howard Spoelstra
On 1/7/19 5:11 AM, Mark Cave-Ayland wrote:
> #7 0x0000555555852e53 in expand_4_vec (vece=2, dofs=197872,
> aofs=198288, bofs=197776, cofs=197792, oprsz=16, tysz=16,
> type=TCG_TYPE_V128, write_aofs=true, fni=0x55555599182a
> <gen_vaddsws_vec>) at
> /home/hsp/src/qemu-altivec-55/tcg/tcg-op-gvec.c:903
> t0 = 0x1848
> t1 = 0x1880
> t2 = 0x18b8
> t3 = 0x18f0
> i = 0
> #8 0x0000555555853cc4 in tcg_gen_gvec_4 (dofs=197872, aofs=198288,
> bofs=197776, cofs=197792, oprsz=16, maxsz=16, g=0x5555562d33c0 <g>) at
> /home/hsp/src/qemu-altivec-55/tcg/tcg-op-gvec.c:1211
> type = TCG_TYPE_V128
> some = 21845
> __PRETTY_FUNCTION__ = "tcg_gen_gvec_4"
> __func__ = "tcg_gen_gvec_4"
> #9 0x0000555555991987 in gen_vaddsws (ctx=0x7fffe3ffe5f0) at
> /home/hsp/src/qemu-altivec-55/target/ppc/translate/vmx-impl.inc.c:597
> g = {fni8 = 0x0, fni4 = 0x0, fniv = 0x55555599182a
> <gen_vaddsws_vec>, fno = 0x5555559601a1 <gen_helper_vaddsws>, opc =
> INDEX_op_add_vec, data = 0, vece = 2 '\002', prefer_i64 = false,
> write_aofs = true}
>
>
> Certainly according to patch 7 of the series only 8-bit and 16-bit accesses are
> supported on i386 hosts, but shouldn't we be falling back to the previous
> implementations rather than hitting an assert()?
In here:
#define GEN_VXFORM_SAT(NAME, VECE, NORM, SAT, OPC2, OPC3) \
static void glue(glue(gen_, NAME), _vec)(unsigned vece, TCGv_vec t, \
TCGv_vec sat, TCGv_vec a, \
TCGv_vec b) \
{ \
TCGv_vec x = tcg_temp_new_vec_matching(t); \
glue(glue(tcg_gen_, NORM), _vec)(VECE, x, a, b); \
glue(glue(tcg_gen_, SAT), _vec)(VECE, t, a, b); \
tcg_gen_cmp_vec(TCG_COND_NE, VECE, x, x, t); \
tcg_gen_or_vec(VECE, sat, sat, x); \
tcg_temp_free_vec(x); \
} \
static void glue(gen_, NAME)(DisasContext *ctx) \
{ \
static const GVecGen4 g = { \
.fniv = glue(glue(gen_, NAME), _vec), \
.fno = glue(gen_helper_, NAME), \
.opc = glue(glue(INDEX_op_, NORM), _vec), \
s/NORM/SAT/, so that we query whether the saturated opcode is supported. The
normal arithmetic, cmp, and or opcodes are mandatory; we don't need to do
anything with those.
r~
^ permalink raw reply [flat|nested] 18+ messages in thread
* Re: [Qemu-devel] [PATCH v2 00/10] tcg vector improvements
2019-01-23 5:09 ` Richard Henderson
@ 2019-02-05 21:29 ` Mark Cave-Ayland
2019-02-06 3:37 ` Richard Henderson
2019-02-07 22:31 ` Mark Cave-Ayland
0 siblings, 2 replies; 18+ messages in thread
From: Mark Cave-Ayland @ 2019-02-05 21:29 UTC (permalink / raw)
To: Richard Henderson, qemu-devel; +Cc: Howard Spoelstra
On 23/01/2019 05:09, Richard Henderson wrote:
> On 1/7/19 5:11 AM, Mark Cave-Ayland wrote:
>> #7 0x0000555555852e53 in expand_4_vec (vece=2, dofs=197872,
>> aofs=198288, bofs=197776, cofs=197792, oprsz=16, tysz=16,
>> type=TCG_TYPE_V128, write_aofs=true, fni=0x55555599182a
>> <gen_vaddsws_vec>) at
>> /home/hsp/src/qemu-altivec-55/tcg/tcg-op-gvec.c:903
>> t0 = 0x1848
>> t1 = 0x1880
>> t2 = 0x18b8
>> t3 = 0x18f0
>> i = 0
>> #8 0x0000555555853cc4 in tcg_gen_gvec_4 (dofs=197872, aofs=198288,
>> bofs=197776, cofs=197792, oprsz=16, maxsz=16, g=0x5555562d33c0 <g>) at
>> /home/hsp/src/qemu-altivec-55/tcg/tcg-op-gvec.c:1211
>> type = TCG_TYPE_V128
>> some = 21845
>> __PRETTY_FUNCTION__ = "tcg_gen_gvec_4"
>> __func__ = "tcg_gen_gvec_4"
>> #9 0x0000555555991987 in gen_vaddsws (ctx=0x7fffe3ffe5f0) at
>> /home/hsp/src/qemu-altivec-55/target/ppc/translate/vmx-impl.inc.c:597
>> g = {fni8 = 0x0, fni4 = 0x0, fniv = 0x55555599182a
>> <gen_vaddsws_vec>, fno = 0x5555559601a1 <gen_helper_vaddsws>, opc =
>> INDEX_op_add_vec, data = 0, vece = 2 '\002', prefer_i64 = false,
>> write_aofs = true}
>>
>>
>> Certainly according to patch 7 of the series only 8-bit and 16-bit accesses are
>> supported on i386 hosts, but shouldn't we be falling back to the previous
>> implementations rather than hitting an assert()?
>
> In here:
>
> #define GEN_VXFORM_SAT(NAME, VECE, NORM, SAT, OPC2, OPC3) \
> static void glue(glue(gen_, NAME), _vec)(unsigned vece, TCGv_vec t, \
> TCGv_vec sat, TCGv_vec a, \
> TCGv_vec b) \
> { \
> TCGv_vec x = tcg_temp_new_vec_matching(t); \
> glue(glue(tcg_gen_, NORM), _vec)(VECE, x, a, b); \
> glue(glue(tcg_gen_, SAT), _vec)(VECE, t, a, b); \
> tcg_gen_cmp_vec(TCG_COND_NE, VECE, x, x, t); \
> tcg_gen_or_vec(VECE, sat, sat, x); \
> tcg_temp_free_vec(x); \
> } \
> static void glue(gen_, NAME)(DisasContext *ctx) \
> { \
> static const GVecGen4 g = { \
> .fniv = glue(glue(gen_, NAME), _vec), \
> .fno = glue(gen_helper_, NAME), \
> .opc = glue(glue(INDEX_op_, NORM), _vec), \
>
> s/NORM/SAT/, so that we query whether the saturated opcode is supported. The
> normal arithmetic, cmp, and or opcodes are mandatory; we don't need to do
> anything with those.
Now that this and the other pre-requisite patches have been merged into master, I've
rebased the outstanding PPC parts of your "tcg, target/ppc vector improvements" on
master including the above fix and pushed the result to
https://github.com/mcayland/qemu/commits/ppc-altivec-v6.
The good news is that the graphics corruption I originally noticed caused by the
patch introducing the saturating add/sub vector ops has now gone, and with my
little-endian vsplt fix included then both OS X and MacOS 9 appear to run without any
obvious issues on an x86 host, and certainly feel smoother compared to before.
The only minor question I had with the patchset in its current form is whether to use
the new VsrD() macro for vscr_sat, or whether we don't really care enough?
ATB,
Mark.
^ permalink raw reply [flat|nested] 18+ messages in thread
* Re: [Qemu-devel] [PATCH v2 00/10] tcg vector improvements
2019-02-05 21:29 ` Mark Cave-Ayland
@ 2019-02-06 3:37 ` Richard Henderson
2019-02-06 7:15 ` Mark Cave-Ayland
2019-02-07 22:31 ` Mark Cave-Ayland
1 sibling, 1 reply; 18+ messages in thread
From: Richard Henderson @ 2019-02-06 3:37 UTC (permalink / raw)
To: Mark Cave-Ayland, qemu-devel; +Cc: Howard Spoelstra
On 2/5/19 9:29 PM, Mark Cave-Ayland wrote:
> The only minor question I had with the patchset in its current form is whether to use
> the new VsrD() macro for vscr_sat, or whether we don't really care enough?
Given the comment
/* Which bit we set is completely arbitrary, but clear the rest. */
I don't think VsrD is helpful.
In "target/ppc: Split out VSCR_SAT to a vector field":
ppc_vsr_t vsr[64] QEMU_ALIGNED(16);
+ /* Non-zero if and only if VSCR_SAT should be set. */
+ ppc_vsr_t vscr_sat;
Better to add the QEMU_ALIGNED(16) to vscr_sat as well. Yes, it will already
happen to be aligned by placement, but this is also a bit documentation.
In "target/ppc: convert vadd*s and vsub*s to vector operations":
if (sat) { \
- set_vscr_sat(env); \
+ vscr_sat->u32[0] = 1; \
} \
Changed in error?
r~
^ permalink raw reply [flat|nested] 18+ messages in thread
* Re: [Qemu-devel] [PATCH v2 00/10] tcg vector improvements
2019-02-06 3:37 ` Richard Henderson
@ 2019-02-06 7:15 ` Mark Cave-Ayland
2019-02-06 15:44 ` BALATON Zoltan
0 siblings, 1 reply; 18+ messages in thread
From: Mark Cave-Ayland @ 2019-02-06 7:15 UTC (permalink / raw)
To: Richard Henderson, qemu-devel; +Cc: Howard Spoelstra
On 06/02/2019 03:37, Richard Henderson wrote:
> On 2/5/19 9:29 PM, Mark Cave-Ayland wrote:
>> The only minor question I had with the patchset in its current form is whether to use
>> the new VsrD() macro for vscr_sat, or whether we don't really care enough?
>
> Given the comment
>
> /* Which bit we set is completely arbitrary, but clear the rest. */
>
> I don't think VsrD is helpful.
Okay, I can leave that for now.
> In "target/ppc: Split out VSCR_SAT to a vector field":
>
> ppc_vsr_t vsr[64] QEMU_ALIGNED(16);
> + /* Non-zero if and only if VSCR_SAT should be set. */
> + ppc_vsr_t vscr_sat;
>
> Better to add the QEMU_ALIGNED(16) to vscr_sat as well. Yes, it will already
> happen to be aligned by placement, but this is also a bit documentation.
I've added this to my latest branch.
> In "target/ppc: convert vadd*s and vsub*s to vector operations":
>
> if (sat) { \
> - set_vscr_sat(env); \
> + vscr_sat->u32[0] = 1; \
> } \
>
> Changed in error?
It looks like this was in the original patch, presumably because GEN_VXFORM_SAT
doesn't include the env parameter which was present in GEN_VXFORM_ENV. Should the env
parameter be added to GEN_VXFORM_SAT?
Howard has also pointed out that he's still spotted some corruption in his tests, so
I will do a bit more investigation and report back.
ATB,
Mark.
^ permalink raw reply [flat|nested] 18+ messages in thread
* Re: [Qemu-devel] [PATCH v2 00/10] tcg vector improvements
2019-02-06 7:15 ` Mark Cave-Ayland
@ 2019-02-06 15:44 ` BALATON Zoltan
0 siblings, 0 replies; 18+ messages in thread
From: BALATON Zoltan @ 2019-02-06 15:44 UTC (permalink / raw)
To: Mark Cave-Ayland; +Cc: Richard Henderson, qemu-devel, Howard Spoelstra
On Wed, 6 Feb 2019, Mark Cave-Ayland wrote:
> Howard has also pointed out that he's still spotted some corruption in his tests, so
> I will do a bit more investigation and report back.
I wonder if instead of "visually testing" shouldn't it be verified by some
more exact tests? Could any of these be useful:
https://gcc.gnu.org/ml/gcc-patches/2002-11/msg01391.html
https://gcc.gnu.org/git/?p=gcc.git;a=tree;f=gcc/testsuite/gcc.dg/vmx;h=4765c35dabb4e3891c98084af077c2549fb45dfc;hb=HEAD
https://github.com/llvm-mirror/test-suite/tree/master/SingleSource/UnitTests/Vector/Altivec
or there could be other similar ones around that already have test cases
for these instructions that could provide some more confidence in the
correctness of the implementation.
Regards,
BALATON Zoltan
^ permalink raw reply [flat|nested] 18+ messages in thread
* Re: [Qemu-devel] [PATCH v2 00/10] tcg vector improvements
2019-02-05 21:29 ` Mark Cave-Ayland
2019-02-06 3:37 ` Richard Henderson
@ 2019-02-07 22:31 ` Mark Cave-Ayland
1 sibling, 0 replies; 18+ messages in thread
From: Mark Cave-Ayland @ 2019-02-07 22:31 UTC (permalink / raw)
To: Richard Henderson, qemu-devel; +Cc: Howard Spoelstra
On 05/02/2019 21:29, Mark Cave-Ayland wrote:
> Now that this and the other pre-requisite patches have been merged into master, I've
> rebased the outstanding PPC parts of your "tcg, target/ppc vector improvements" on
> master including the above fix and pushed the result to
> https://github.com/mcayland/qemu/commits/ppc-altivec-v6.
>
> The good news is that the graphics corruption I originally noticed caused by the
> patch introducing the saturating add/sub vector ops has now gone, and with my
> little-endian vsplt fix included then both OS X and MacOS 9 appear to run without any
> obvious issues on an x86 host, and certainly feel smoother compared to before.
I started to follow up Howard's report that he could still see graphical corruption
with these patches, and I was surprised to notice that it had reappeared locally once
again :/
After working out that this was because the relevant instructions weren't being
generated for all Mac machines, I was able to figure it out after a few hours of
digging. Patch to follow shortly.
ATB,
Mark.
^ permalink raw reply [flat|nested] 18+ messages in thread
end of thread, other threads:[~2019-02-07 22:31 UTC | newest]
Thread overview: 18+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2019-01-04 22:31 [Qemu-devel] [PATCH v2 00/10] tcg vector improvements Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 01/10] tcg: Add logical simplifications during gvec expand Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 02/10] tcg: Add gvec expanders for nand, nor, eqv Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 03/10] tcg: Add write_aofs to GVecGen4 Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 04/10] tcg: Add opcodes for vector saturated arithmetic Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 05/10] tcg: Add opcodes for vector minmax arithmetic Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 06/10] tcg/i386: Split subroutines out of tcg_expand_vec_op Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 07/10] tcg/i386: Implement vector saturating arithmetic Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 08/10] tcg/i386: Implement vector minmax arithmetic Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 09/10] tcg/aarch64: Implement vector saturating arithmetic Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 10/10] tcg/aarch64: Implement vector minmax arithmetic Richard Henderson
2019-01-07 13:11 ` [Qemu-devel] [PATCH v2 00/10] tcg vector improvements Mark Cave-Ayland
2019-01-23 5:09 ` Richard Henderson
2019-02-05 21:29 ` Mark Cave-Ayland
2019-02-06 3:37 ` Richard Henderson
2019-02-06 7:15 ` Mark Cave-Ayland
2019-02-06 15:44 ` BALATON Zoltan
2019-02-07 22:31 ` Mark Cave-Ayland
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.