* [PATCH v2 01/15] target/arm: Create gen_gvec_[us]sra
2020-05-02 22:44 [PATCH v2 00/15] target/arm: partial vector cleanup Richard Henderson
@ 2020-05-02 22:44 ` Richard Henderson
2020-05-02 22:44 ` [PATCH v2 02/15] target/arm: Create gen_gvec_{u,s}{rshr,rsra} Richard Henderson
` (14 subsequent siblings)
15 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2020-05-02 22:44 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
The functions eliminate duplication of the special cases for
this operation. They match up with the GVecGen2iFn typedef.
Add out-of-line helpers. We got away with only having inline
expanders because the neon vector size is only 16 bytes, and
we know that the inline expansion will always succeed.
When we reuse this for SVE, tcg-gvec-op may decide to use an
out-of-line helper due to longer vector lengths.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/helper.h | 10 +++
target/arm/translate.h | 7 +-
target/arm/translate-a64.c | 15 +---
target/arm/translate.c | 161 ++++++++++++++++++++++---------------
target/arm/vec_helper.c | 25 ++++++
5 files changed, 139 insertions(+), 79 deletions(-)
diff --git a/target/arm/helper.h b/target/arm/helper.h
index 5817626b20..9bc162345c 100644
--- a/target/arm/helper.h
+++ b/target/arm/helper.h
@@ -691,6 +691,16 @@ DEF_HELPER_FLAGS_4(gvec_pmull_q, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(neon_pmull_h, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_ssra_b, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_ssra_h, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_ssra_s, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_ssra_d, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+
+DEF_HELPER_FLAGS_3(gvec_usra_b, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_usra_h, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_usra_s, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_usra_d, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+
#ifdef TARGET_AARCH64
#include "helper-a64.h"
#include "helper-sve.h"
diff --git a/target/arm/translate.h b/target/arm/translate.h
index 98b319f3f6..a39cf22666 100644
--- a/target/arm/translate.h
+++ b/target/arm/translate.h
@@ -285,8 +285,6 @@ extern const GVecGen3 mls_op[4];
extern const GVecGen3 cmtst_op[4];
extern const GVecGen3 sshl_op[4];
extern const GVecGen3 ushl_op[4];
-extern const GVecGen2i ssra_op[4];
-extern const GVecGen2i usra_op[4];
extern const GVecGen2i sri_op[4];
extern const GVecGen2i sli_op[4];
extern const GVecGen4 uqadd_op[4];
@@ -299,6 +297,11 @@ void gen_sshl_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b);
void gen_ushl_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b);
void gen_sshl_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b);
+void gen_gvec_ssra(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
+ int64_t shift, uint32_t opr_sz, uint32_t max_sz);
+void gen_gvec_usra(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
+ int64_t shift, uint32_t opr_sz, uint32_t max_sz);
+
/*
* Forward to the isar_feature_* tests given a DisasContext pointer.
*/
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index 010b36633e..03f4dc5805 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -10205,19 +10205,8 @@ static void handle_vec_simd_shri(DisasContext *s, bool is_q, bool is_u,
switch (opcode) {
case 0x02: /* SSRA / USRA (accumulate) */
- if (is_u) {
- /* Shift count same as element size produces zero to add. */
- if (shift == 8 << size) {
- goto done;
- }
- gen_gvec_op2i(s, is_q, rd, rn, shift, &usra_op[size]);
- } else {
- /* Shift count same as element size produces all sign to add. */
- if (shift == 8 << size) {
- shift -= 1;
- }
- gen_gvec_op2i(s, is_q, rd, rn, shift, &ssra_op[size]);
- }
+ gen_gvec_fn2i(s, is_q, rd, rn, shift,
+ is_u ? gen_gvec_usra : gen_gvec_ssra, size);
return;
case 0x08: /* SRI */
/* Shift count same as element size is valid but does nothing. */
diff --git a/target/arm/translate.c b/target/arm/translate.c
index a96899549b..04114906d7 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -4146,33 +4146,51 @@ static void gen_ssra_vec(unsigned vece, TCGv_vec d, TCGv_vec a, int64_t sh)
tcg_gen_add_vec(vece, d, d, a);
}
-static const TCGOpcode vecop_list_ssra[] = {
- INDEX_op_sari_vec, INDEX_op_add_vec, 0
-};
+void gen_gvec_ssra(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
+ int64_t shift, uint32_t opr_sz, uint32_t max_sz)
+{
+ static const TCGOpcode vecop_list[] = {
+ INDEX_op_sari_vec, INDEX_op_add_vec, 0
+ };
+ static const GVecGen2i ops[4] = {
+ { .fni8 = gen_ssra8_i64,
+ .fniv = gen_ssra_vec,
+ .fno = gen_helper_gvec_ssra_b,
+ .load_dest = true,
+ .opt_opc = vecop_list,
+ .vece = MO_8 },
+ { .fni8 = gen_ssra16_i64,
+ .fniv = gen_ssra_vec,
+ .fno = gen_helper_gvec_ssra_h,
+ .load_dest = true,
+ .opt_opc = vecop_list,
+ .vece = MO_16 },
+ { .fni4 = gen_ssra32_i32,
+ .fniv = gen_ssra_vec,
+ .fno = gen_helper_gvec_ssra_s,
+ .load_dest = true,
+ .opt_opc = vecop_list,
+ .vece = MO_32 },
+ { .fni8 = gen_ssra64_i64,
+ .fniv = gen_ssra_vec,
+ .fno = gen_helper_gvec_ssra_b,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ .opt_opc = vecop_list,
+ .load_dest = true,
+ .vece = MO_64 },
+ };
-const GVecGen2i ssra_op[4] = {
- { .fni8 = gen_ssra8_i64,
- .fniv = gen_ssra_vec,
- .load_dest = true,
- .opt_opc = vecop_list_ssra,
- .vece = MO_8 },
- { .fni8 = gen_ssra16_i64,
- .fniv = gen_ssra_vec,
- .load_dest = true,
- .opt_opc = vecop_list_ssra,
- .vece = MO_16 },
- { .fni4 = gen_ssra32_i32,
- .fniv = gen_ssra_vec,
- .load_dest = true,
- .opt_opc = vecop_list_ssra,
- .vece = MO_32 },
- { .fni8 = gen_ssra64_i64,
- .fniv = gen_ssra_vec,
- .prefer_i64 = TCG_TARGET_REG_BITS == 64,
- .opt_opc = vecop_list_ssra,
- .load_dest = true,
- .vece = MO_64 },
-};
+ /* tszimm encoding produces immediates in the range [1..esize]. */
+ tcg_debug_assert(shift > 0);
+ tcg_debug_assert(shift <= (8 << vece));
+
+ /*
+ * Shifts larger than the element size are architecturally valid.
+ * Signed results in all sign bits.
+ */
+ shift = MIN(shift, (8 << vece) - 1);
+ tcg_gen_gvec_2i(rd_ofs, rm_ofs, opr_sz, max_sz, shift, &ops[vece]);
+}
static void gen_usra8_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
{
@@ -4204,33 +4222,55 @@ static void gen_usra_vec(unsigned vece, TCGv_vec d, TCGv_vec a, int64_t sh)
tcg_gen_add_vec(vece, d, d, a);
}
-static const TCGOpcode vecop_list_usra[] = {
- INDEX_op_shri_vec, INDEX_op_add_vec, 0
-};
+void gen_gvec_usra(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
+ int64_t shift, uint32_t opr_sz, uint32_t max_sz)
+{
+ static const TCGOpcode vecop_list[] = {
+ INDEX_op_shri_vec, INDEX_op_add_vec, 0
+ };
+ static const GVecGen2i ops[4] = {
+ { .fni8 = gen_usra8_i64,
+ .fniv = gen_usra_vec,
+ .fno = gen_helper_gvec_usra_b,
+ .load_dest = true,
+ .opt_opc = vecop_list,
+ .vece = MO_8, },
+ { .fni8 = gen_usra16_i64,
+ .fniv = gen_usra_vec,
+ .fno = gen_helper_gvec_usra_h,
+ .load_dest = true,
+ .opt_opc = vecop_list,
+ .vece = MO_16, },
+ { .fni4 = gen_usra32_i32,
+ .fniv = gen_usra_vec,
+ .fno = gen_helper_gvec_usra_s,
+ .load_dest = true,
+ .opt_opc = vecop_list,
+ .vece = MO_32, },
+ { .fni8 = gen_usra64_i64,
+ .fniv = gen_usra_vec,
+ .fno = gen_helper_gvec_usra_d,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ .load_dest = true,
+ .opt_opc = vecop_list,
+ .vece = MO_64, },
+ };
-const GVecGen2i usra_op[4] = {
- { .fni8 = gen_usra8_i64,
- .fniv = gen_usra_vec,
- .load_dest = true,
- .opt_opc = vecop_list_usra,
- .vece = MO_8, },
- { .fni8 = gen_usra16_i64,
- .fniv = gen_usra_vec,
- .load_dest = true,
- .opt_opc = vecop_list_usra,
- .vece = MO_16, },
- { .fni4 = gen_usra32_i32,
- .fniv = gen_usra_vec,
- .load_dest = true,
- .opt_opc = vecop_list_usra,
- .vece = MO_32, },
- { .fni8 = gen_usra64_i64,
- .fniv = gen_usra_vec,
- .prefer_i64 = TCG_TARGET_REG_BITS == 64,
- .load_dest = true,
- .opt_opc = vecop_list_usra,
- .vece = MO_64, },
-};
+ /* tszimm encoding produces immediates in the range [1..esize]. */
+ tcg_debug_assert(shift > 0);
+ tcg_debug_assert(shift <= (8 << vece));
+
+ /*
+ * Shifts larger than the element size are architecturally valid.
+ * Unsigned results in all zeros as input to accumulate: nop.
+ */
+ if (shift < (8 << vece)) {
+ tcg_gen_gvec_2i(rd_ofs, rm_ofs, opr_sz, max_sz, shift, &ops[vece]);
+ } else {
+ /* Nop, but we do need to clear the tail. */
+ tcg_gen_gvec_mov(vece, rd_ofs, rd_ofs, opr_sz, max_sz);
+ }
+}
static void gen_shr8_ins_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
{
@@ -5596,19 +5636,12 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
case 1: /* VSRA */
/* Right shift comes here negative. */
shift = -shift;
- /* Shifts larger than the element size are architecturally
- * valid. Unsigned results in all zeros; signed results
- * in all sign bits.
- */
- if (!u) {
- tcg_gen_gvec_2i(rd_ofs, rm_ofs, vec_size, vec_size,
- MIN(shift, (8 << size) - 1),
- &ssra_op[size]);
- } else if (shift >= 8 << size) {
- /* rd += 0 */
+ if (u) {
+ gen_gvec_usra(size, rd_ofs, rm_ofs, shift,
+ vec_size, vec_size);
} else {
- tcg_gen_gvec_2i(rd_ofs, rm_ofs, vec_size, vec_size,
- shift, &usra_op[size]);
+ gen_gvec_ssra(size, rd_ofs, rm_ofs, shift,
+ vec_size, vec_size);
}
return 0;
diff --git a/target/arm/vec_helper.c b/target/arm/vec_helper.c
index 3d534188a8..230085b35e 100644
--- a/target/arm/vec_helper.c
+++ b/target/arm/vec_helper.c
@@ -899,6 +899,31 @@ void HELPER(gvec_sqsub_d)(void *vd, void *vq, void *vn,
clear_tail(d, oprsz, simd_maxsz(desc));
}
+
+#define DO_SRA(NAME, TYPE) \
+void HELPER(NAME)(void *vd, void *vn, uint32_t desc) \
+{ \
+ intptr_t i, oprsz = simd_oprsz(desc); \
+ int shift = simd_data(desc); \
+ TYPE *d = vd, *n = vn; \
+ for (i = 0; i < oprsz / sizeof(TYPE); i++) { \
+ d[i] += n[i] >> shift; \
+ } \
+ clear_tail(d, oprsz, simd_maxsz(desc)); \
+}
+
+DO_SRA(gvec_ssra_b, int8_t)
+DO_SRA(gvec_ssra_h, int16_t)
+DO_SRA(gvec_ssra_s, int32_t)
+DO_SRA(gvec_ssra_d, int64_t)
+
+DO_SRA(gvec_usra_b, uint8_t)
+DO_SRA(gvec_usra_h, uint16_t)
+DO_SRA(gvec_usra_s, uint32_t)
+DO_SRA(gvec_usra_d, uint64_t)
+
+#undef DO_SRA
+
/*
* Convert float16 to float32, raising no exceptions and
* preserving exceptional values, including SNaN.
--
2.20.1
^ permalink raw reply related [flat|nested] 18+ messages in thread* [PATCH v2 02/15] target/arm: Create gen_gvec_{u,s}{rshr,rsra}
2020-05-02 22:44 [PATCH v2 00/15] target/arm: partial vector cleanup Richard Henderson
2020-05-02 22:44 ` [PATCH v2 01/15] target/arm: Create gen_gvec_[us]sra Richard Henderson
@ 2020-05-02 22:44 ` Richard Henderson
2020-05-02 22:44 ` [PATCH v2 03/15] target/arm: Create gen_gvec_{sri,sli} Richard Henderson
` (13 subsequent siblings)
15 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2020-05-02 22:44 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Create vectorized versions of handle_shri_with_rndacc
for shift+round and shift+round+accumulate. Add out-of-line
helpers in preparation for longer vector lengths from SVE.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/helper.h | 20 ++
target/arm/translate.h | 9 +
target/arm/translate-a64.c | 11 +-
target/arm/translate.c | 461 +++++++++++++++++++++++++++++++++++--
target/arm/vec_helper.c | 50 ++++
5 files changed, 525 insertions(+), 26 deletions(-)
diff --git a/target/arm/helper.h b/target/arm/helper.h
index 9bc162345c..aeb1f52455 100644
--- a/target/arm/helper.h
+++ b/target/arm/helper.h
@@ -701,6 +701,26 @@ DEF_HELPER_FLAGS_3(gvec_usra_h, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
DEF_HELPER_FLAGS_3(gvec_usra_s, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
DEF_HELPER_FLAGS_3(gvec_usra_d, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_srshr_b, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_srshr_h, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_srshr_s, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_srshr_d, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+
+DEF_HELPER_FLAGS_3(gvec_urshr_b, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_urshr_h, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_urshr_s, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_urshr_d, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+
+DEF_HELPER_FLAGS_3(gvec_srsra_b, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_srsra_h, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_srsra_s, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_srsra_d, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+
+DEF_HELPER_FLAGS_3(gvec_ursra_b, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_ursra_h, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_ursra_s, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_ursra_d, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+
#ifdef TARGET_AARCH64
#include "helper-a64.h"
#include "helper-sve.h"
diff --git a/target/arm/translate.h b/target/arm/translate.h
index a39cf22666..823821f82c 100644
--- a/target/arm/translate.h
+++ b/target/arm/translate.h
@@ -302,6 +302,15 @@ void gen_gvec_ssra(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
void gen_gvec_usra(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
int64_t shift, uint32_t opr_sz, uint32_t max_sz);
+void gen_gvec_srshr(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
+ int64_t shift, uint32_t opr_sz, uint32_t max_sz);
+void gen_gvec_urshr(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
+ int64_t shift, uint32_t opr_sz, uint32_t max_sz);
+void gen_gvec_srsra(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
+ int64_t shift, uint32_t opr_sz, uint32_t max_sz);
+void gen_gvec_ursra(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
+ int64_t shift, uint32_t opr_sz, uint32_t max_sz);
+
/*
* Forward to the isar_feature_* tests given a DisasContext pointer.
*/
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index 03f4dc5805..1ef05d5ce1 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -10235,10 +10235,15 @@ static void handle_vec_simd_shri(DisasContext *s, bool is_q, bool is_u,
return;
case 0x04: /* SRSHR / URSHR (rounding) */
- break;
+ gen_gvec_fn2i(s, is_q, rd, rn, shift,
+ is_u ? gen_gvec_urshr : gen_gvec_srshr, size);
+ return;
+
case 0x06: /* SRSRA / URSRA (accum + rounding) */
- accumulate = true;
- break;
+ gen_gvec_fn2i(s, is_q, rd, rn, shift,
+ is_u ? gen_gvec_ursra : gen_gvec_srsra, size);
+ return;
+
default:
g_assert_not_reached();
}
diff --git a/target/arm/translate.c b/target/arm/translate.c
index 04114906d7..d724022cb6 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -4272,6 +4272,422 @@ void gen_gvec_usra(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
}
}
+/*
+ * Shift one less than the requested amount, and the low bit is
+ * the rounding bit. For the 8 and 16-bit operations, because we
+ * mask the low bit, we can perform a normal integer shift instead
+ * of a vector shift.
+ */
+static void gen_srshr8_i64(TCGv_i64 d, TCGv_i64 a, int64_t sh)
+{
+ TCGv_i64 t = tcg_temp_new_i64();
+
+ tcg_gen_shri_i64(t, a, sh - 1);
+ tcg_gen_andi_i64(t, t, dup_const(MO_8, 1));
+ tcg_gen_vec_sar8i_i64(d, a, sh);
+ tcg_gen_vec_add8_i64(d, d, t);
+ tcg_temp_free_i64(t);
+}
+
+static void gen_srshr16_i64(TCGv_i64 d, TCGv_i64 a, int64_t sh)
+{
+ TCGv_i64 t = tcg_temp_new_i64();
+
+ tcg_gen_shri_i64(t, a, sh - 1);
+ tcg_gen_andi_i64(t, t, dup_const(MO_16, 1));
+ tcg_gen_vec_sar16i_i64(d, a, sh);
+ tcg_gen_vec_add16_i64(d, d, t);
+ tcg_temp_free_i64(t);
+}
+
+static void gen_srshr32_i32(TCGv_i32 d, TCGv_i32 a, int32_t sh)
+{
+ TCGv_i32 t = tcg_temp_new_i32();
+
+ tcg_gen_extract_i32(t, a, sh - 1, 1);
+ tcg_gen_sari_i32(d, a, sh);
+ tcg_gen_add_i32(d, d, t);
+ tcg_temp_free_i32(t);
+}
+
+static void gen_srshr64_i64(TCGv_i64 d, TCGv_i64 a, int64_t sh)
+{
+ TCGv_i64 t = tcg_temp_new_i64();
+
+ tcg_gen_extract_i64(t, a, sh - 1, 1);
+ tcg_gen_sari_i64(d, a, sh);
+ tcg_gen_add_i64(d, d, t);
+ tcg_temp_free_i64(t);
+}
+
+static void gen_srshr_vec(unsigned vece, TCGv_vec d, TCGv_vec a, int64_t sh)
+{
+ TCGv_vec t = tcg_temp_new_vec_matching(d);
+ TCGv_vec ones = tcg_temp_new_vec_matching(d);
+
+ tcg_gen_shri_vec(vece, t, a, sh - 1);
+ tcg_gen_dupi_vec(vece, ones, 1);
+ tcg_gen_and_vec(vece, t, t, ones);
+ tcg_gen_sari_vec(vece, d, a, sh);
+ tcg_gen_add_vec(vece, d, d, t);
+
+ tcg_temp_free_vec(t);
+ tcg_temp_free_vec(ones);
+}
+
+void gen_gvec_srshr(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
+ int64_t shift, uint32_t opr_sz, uint32_t max_sz)
+{
+ static const TCGOpcode vecop_list[] = {
+ INDEX_op_shri_vec, INDEX_op_sari_vec, INDEX_op_add_vec, 0
+ };
+ static const GVecGen2i ops[4] = {
+ { .fni8 = gen_srshr8_i64,
+ .fniv = gen_srshr_vec,
+ .fno = gen_helper_gvec_srshr_b,
+ .opt_opc = vecop_list,
+ .vece = MO_8 },
+ { .fni8 = gen_srshr16_i64,
+ .fniv = gen_srshr_vec,
+ .fno = gen_helper_gvec_srshr_h,
+ .opt_opc = vecop_list,
+ .vece = MO_16 },
+ { .fni4 = gen_srshr32_i32,
+ .fniv = gen_srshr_vec,
+ .fno = gen_helper_gvec_srshr_s,
+ .opt_opc = vecop_list,
+ .vece = MO_32 },
+ { .fni8 = gen_srshr64_i64,
+ .fniv = gen_srshr_vec,
+ .fno = gen_helper_gvec_srshr_d,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ .opt_opc = vecop_list,
+ .vece = MO_64 },
+ };
+
+ /* tszimm encoding produces immediates in the range [1..esize] */
+ tcg_debug_assert(shift > 0);
+ tcg_debug_assert(shift <= (8 << vece));
+
+ if (shift == (8 << vece)) {
+ /*
+ * Shifts larger than the element size are architecturally valid.
+ * Signed results in all sign bits. With rounding, this produces
+ * (-1 + 1) >> 1 == 0, or (0 + 1) >> 1 == 0.
+ * I.e. always zero.
+ */
+ tcg_gen_gvec_dup_imm(vece, rd_ofs, opr_sz, max_sz, 0);
+ } else {
+ tcg_gen_gvec_2i(rd_ofs, rm_ofs, opr_sz, max_sz, shift, &ops[vece]);
+ }
+}
+
+static void gen_srsra8_i64(TCGv_i64 d, TCGv_i64 a, int64_t sh)
+{
+ TCGv_i64 t = tcg_temp_new_i64();
+
+ gen_srshr8_i64(t, a, sh);
+ tcg_gen_vec_add8_i64(d, d, t);
+ tcg_temp_free_i64(t);
+}
+
+static void gen_srsra16_i64(TCGv_i64 d, TCGv_i64 a, int64_t sh)
+{
+ TCGv_i64 t = tcg_temp_new_i64();
+
+ gen_srshr16_i64(t, a, sh);
+ tcg_gen_vec_add16_i64(d, d, t);
+ tcg_temp_free_i64(t);
+}
+
+static void gen_srsra32_i32(TCGv_i32 d, TCGv_i32 a, int32_t sh)
+{
+ TCGv_i32 t = tcg_temp_new_i32();
+
+ gen_srshr32_i32(t, a, sh);
+ tcg_gen_add_i32(d, d, t);
+ tcg_temp_free_i32(t);
+}
+
+static void gen_srsra64_i64(TCGv_i64 d, TCGv_i64 a, int64_t sh)
+{
+ TCGv_i64 t = tcg_temp_new_i64();
+
+ gen_srshr64_i64(t, a, sh);
+ tcg_gen_add_i64(d, d, t);
+ tcg_temp_free_i64(t);
+}
+
+static void gen_srsra_vec(unsigned vece, TCGv_vec d, TCGv_vec a, int64_t sh)
+{
+ TCGv_vec t = tcg_temp_new_vec_matching(d);
+
+ gen_srshr_vec(vece, t, a, sh);
+ tcg_gen_add_vec(vece, d, d, t);
+ tcg_temp_free_vec(t);
+}
+
+void gen_gvec_srsra(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
+ int64_t shift, uint32_t opr_sz, uint32_t max_sz)
+{
+ static const TCGOpcode vecop_list[] = {
+ INDEX_op_shri_vec, INDEX_op_sari_vec, INDEX_op_add_vec, 0
+ };
+ static const GVecGen2i ops[4] = {
+ { .fni8 = gen_srsra8_i64,
+ .fniv = gen_srsra_vec,
+ .fno = gen_helper_gvec_srsra_b,
+ .opt_opc = vecop_list,
+ .load_dest = true,
+ .vece = MO_8 },
+ { .fni8 = gen_srsra16_i64,
+ .fniv = gen_srsra_vec,
+ .fno = gen_helper_gvec_srsra_h,
+ .opt_opc = vecop_list,
+ .load_dest = true,
+ .vece = MO_16 },
+ { .fni4 = gen_srsra32_i32,
+ .fniv = gen_srsra_vec,
+ .fno = gen_helper_gvec_srsra_s,
+ .opt_opc = vecop_list,
+ .load_dest = true,
+ .vece = MO_32 },
+ { .fni8 = gen_srsra64_i64,
+ .fniv = gen_srsra_vec,
+ .fno = gen_helper_gvec_srsra_d,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ .opt_opc = vecop_list,
+ .load_dest = true,
+ .vece = MO_64 },
+ };
+
+ /* tszimm encoding produces immediates in the range [1..esize] */
+ tcg_debug_assert(shift > 0);
+ tcg_debug_assert(shift <= (8 << vece));
+
+ /*
+ * Shifts larger than the element size are architecturally valid.
+ * Signed results in all sign bits. With rounding, this produces
+ * (-1 + 1) >> 1 == 0, or (0 + 1) >> 1 == 0.
+ * I.e. always zero. With accumulation, this leaves D unchanged.
+ */
+ if (shift == (8 << vece)) {
+ /* Nop, but we do need to clear the tail. */
+ tcg_gen_gvec_mov(vece, rd_ofs, rd_ofs, opr_sz, max_sz);
+ } else {
+ tcg_gen_gvec_2i(rd_ofs, rm_ofs, opr_sz, max_sz, shift, &ops[vece]);
+ }
+}
+
+static void gen_urshr8_i64(TCGv_i64 d, TCGv_i64 a, int64_t sh)
+{
+ TCGv_i64 t = tcg_temp_new_i64();
+
+ tcg_gen_shri_i64(t, a, sh - 1);
+ tcg_gen_andi_i64(t, t, dup_const(MO_8, 1));
+ tcg_gen_vec_shr8i_i64(d, a, sh);
+ tcg_gen_vec_add8_i64(d, d, t);
+ tcg_temp_free_i64(t);
+}
+
+static void gen_urshr16_i64(TCGv_i64 d, TCGv_i64 a, int64_t sh)
+{
+ TCGv_i64 t = tcg_temp_new_i64();
+
+ tcg_gen_shri_i64(t, a, sh - 1);
+ tcg_gen_andi_i64(t, t, dup_const(MO_16, 1));
+ tcg_gen_vec_shr16i_i64(d, a, sh);
+ tcg_gen_vec_add16_i64(d, d, t);
+ tcg_temp_free_i64(t);
+}
+
+static void gen_urshr32_i32(TCGv_i32 d, TCGv_i32 a, int32_t sh)
+{
+ TCGv_i32 t = tcg_temp_new_i32();
+
+ tcg_gen_extract_i32(t, a, sh - 1, 1);
+ tcg_gen_shri_i32(d, a, sh);
+ tcg_gen_add_i32(d, d, t);
+ tcg_temp_free_i32(t);
+}
+
+static void gen_urshr64_i64(TCGv_i64 d, TCGv_i64 a, int64_t sh)
+{
+ TCGv_i64 t = tcg_temp_new_i64();
+
+ tcg_gen_extract_i64(t, a, sh - 1, 1);
+ tcg_gen_shri_i64(d, a, sh);
+ tcg_gen_add_i64(d, d, t);
+ tcg_temp_free_i64(t);
+}
+
+static void gen_urshr_vec(unsigned vece, TCGv_vec d, TCGv_vec a, int64_t shift)
+{
+ TCGv_vec t = tcg_temp_new_vec_matching(d);
+ TCGv_vec ones = tcg_temp_new_vec_matching(d);
+
+ tcg_gen_shri_vec(vece, t, a, shift - 1);
+ tcg_gen_dupi_vec(vece, ones, 1);
+ tcg_gen_and_vec(vece, t, t, ones);
+ tcg_gen_shri_vec(vece, d, a, shift);
+ tcg_gen_add_vec(vece, d, d, t);
+
+ tcg_temp_free_vec(t);
+ tcg_temp_free_vec(ones);
+}
+
+void gen_gvec_urshr(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
+ int64_t shift, uint32_t opr_sz, uint32_t max_sz)
+{
+ static const TCGOpcode vecop_list[] = {
+ INDEX_op_shri_vec, INDEX_op_add_vec, 0
+ };
+ static const GVecGen2i ops[4] = {
+ { .fni8 = gen_urshr8_i64,
+ .fniv = gen_urshr_vec,
+ .fno = gen_helper_gvec_urshr_b,
+ .opt_opc = vecop_list,
+ .vece = MO_8 },
+ { .fni8 = gen_urshr16_i64,
+ .fniv = gen_urshr_vec,
+ .fno = gen_helper_gvec_urshr_h,
+ .opt_opc = vecop_list,
+ .vece = MO_16 },
+ { .fni4 = gen_urshr32_i32,
+ .fniv = gen_urshr_vec,
+ .fno = gen_helper_gvec_urshr_s,
+ .opt_opc = vecop_list,
+ .vece = MO_32 },
+ { .fni8 = gen_urshr64_i64,
+ .fniv = gen_urshr_vec,
+ .fno = gen_helper_gvec_urshr_d,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ .opt_opc = vecop_list,
+ .vece = MO_64 },
+ };
+
+ /* tszimm encoding produces immediates in the range [1..esize] */
+ tcg_debug_assert(shift > 0);
+ tcg_debug_assert(shift <= (8 << vece));
+
+ if (shift == (8 << vece)) {
+ /*
+ * Shifts larger than the element size are architecturally valid.
+ * Unsigned results in zero. With rounding, this produces a
+ * copy of the most significant bit.
+ */
+ tcg_gen_gvec_shri(vece, rd_ofs, rm_ofs, shift - 1, opr_sz, max_sz);
+ } else {
+ tcg_gen_gvec_2i(rd_ofs, rm_ofs, opr_sz, max_sz, shift, &ops[vece]);
+ }
+}
+
+static void gen_ursra8_i64(TCGv_i64 d, TCGv_i64 a, int64_t sh)
+{
+ TCGv_i64 t = tcg_temp_new_i64();
+
+ if (sh == 8) {
+ tcg_gen_vec_shr8i_i64(t, a, 7);
+ } else {
+ gen_urshr8_i64(t, a, sh);
+ }
+ tcg_gen_vec_add8_i64(d, d, t);
+ tcg_temp_free_i64(t);
+}
+
+static void gen_ursra16_i64(TCGv_i64 d, TCGv_i64 a, int64_t sh)
+{
+ TCGv_i64 t = tcg_temp_new_i64();
+
+ if (sh == 16) {
+ tcg_gen_vec_shr16i_i64(t, a, 15);
+ } else {
+ gen_urshr16_i64(t, a, sh);
+ }
+ tcg_gen_vec_add16_i64(d, d, t);
+ tcg_temp_free_i64(t);
+}
+
+static void gen_ursra32_i32(TCGv_i32 d, TCGv_i32 a, int32_t sh)
+{
+ TCGv_i32 t = tcg_temp_new_i32();
+
+ if (sh == 32) {
+ tcg_gen_shri_i32(t, a, 31);
+ } else {
+ gen_urshr32_i32(t, a, sh);
+ }
+ tcg_gen_add_i32(d, d, t);
+ tcg_temp_free_i32(t);
+}
+
+static void gen_ursra64_i64(TCGv_i64 d, TCGv_i64 a, int64_t sh)
+{
+ TCGv_i64 t = tcg_temp_new_i64();
+
+ if (sh == 64) {
+ tcg_gen_shri_i64(t, a, 63);
+ } else {
+ gen_urshr64_i64(t, a, sh);
+ }
+ tcg_gen_add_i64(d, d, t);
+ tcg_temp_free_i64(t);
+}
+
+static void gen_ursra_vec(unsigned vece, TCGv_vec d, TCGv_vec a, int64_t sh)
+{
+ TCGv_vec t = tcg_temp_new_vec_matching(d);
+
+ if (sh == (8 << vece)) {
+ tcg_gen_shri_vec(vece, t, a, sh - 1);
+ } else {
+ gen_urshr_vec(vece, t, a, sh);
+ }
+ tcg_gen_add_vec(vece, d, d, t);
+ tcg_temp_free_vec(t);
+}
+
+void gen_gvec_ursra(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
+ int64_t shift, uint32_t opr_sz, uint32_t max_sz)
+{
+ static const TCGOpcode vecop_list[] = {
+ INDEX_op_shri_vec, INDEX_op_add_vec, 0
+ };
+ static const GVecGen2i ops[4] = {
+ { .fni8 = gen_ursra8_i64,
+ .fniv = gen_ursra_vec,
+ .fno = gen_helper_gvec_ursra_b,
+ .opt_opc = vecop_list,
+ .load_dest = true,
+ .vece = MO_8 },
+ { .fni8 = gen_ursra16_i64,
+ .fniv = gen_ursra_vec,
+ .fno = gen_helper_gvec_ursra_h,
+ .opt_opc = vecop_list,
+ .load_dest = true,
+ .vece = MO_16 },
+ { .fni4 = gen_ursra32_i32,
+ .fniv = gen_ursra_vec,
+ .fno = gen_helper_gvec_ursra_s,
+ .opt_opc = vecop_list,
+ .load_dest = true,
+ .vece = MO_32 },
+ { .fni8 = gen_ursra64_i64,
+ .fniv = gen_ursra_vec,
+ .fno = gen_helper_gvec_ursra_d,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ .opt_opc = vecop_list,
+ .load_dest = true,
+ .vece = MO_64 },
+ };
+
+ /* tszimm encoding produces immediates in the range [1..esize] */
+ tcg_debug_assert(shift > 0);
+ tcg_debug_assert(shift <= (8 << vece));
+
+ tcg_gen_gvec_2i(rd_ofs, rm_ofs, opr_sz, max_sz, shift, &ops[vece]);
+}
+
static void gen_shr8_ins_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
{
uint64_t mask = dup_const(MO_8, 0xff >> shift);
@@ -5645,6 +6061,28 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
}
return 0;
+ case 2: /* VRSHR */
+ /* Right shift comes here negative. */
+ shift = -shift;
+ if (u) {
+ gen_gvec_urshr(size, rd_ofs, rm_ofs, shift,
+ vec_size, vec_size);
+ } else {
+ gen_gvec_srshr(size, rd_ofs, rm_ofs, shift,
+ vec_size, vec_size);
+ }
+ return 0;
+
+ case 3: /* VRSRA */
+ if (u) {
+ gen_gvec_ursra(size, rd_ofs, rm_ofs, shift,
+ vec_size, vec_size);
+ } else {
+ gen_gvec_srsra(size, rd_ofs, rm_ofs, shift,
+ vec_size, vec_size);
+ }
+ return 0;
+
case 4: /* VSRI */
if (!u) {
return 1;
@@ -5696,13 +6134,6 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
neon_load_reg64(cpu_V0, rm + pass);
tcg_gen_movi_i64(cpu_V1, imm);
switch (op) {
- case 2: /* VRSHR */
- case 3: /* VRSRA */
- if (u)
- gen_helper_neon_rshl_u64(cpu_V0, cpu_V0, cpu_V1);
- else
- gen_helper_neon_rshl_s64(cpu_V0, cpu_V0, cpu_V1);
- break;
case 6: /* VQSHLU */
gen_helper_neon_qshlu_s64(cpu_V0, cpu_env,
cpu_V0, cpu_V1);
@@ -5719,11 +6150,6 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
default:
g_assert_not_reached();
}
- if (op == 3) {
- /* Accumulate. */
- neon_load_reg64(cpu_V1, rd + pass);
- tcg_gen_add_i64(cpu_V0, cpu_V0, cpu_V1);
- }
neon_store_reg64(cpu_V0, rd + pass);
} else { /* size < 3 */
/* Operands in T0 and T1. */
@@ -5731,10 +6157,6 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
tmp2 = tcg_temp_new_i32();
tcg_gen_movi_i32(tmp2, imm);
switch (op) {
- case 2: /* VRSHR */
- case 3: /* VRSRA */
- GEN_NEON_INTEGER_OP(rshl);
- break;
case 6: /* VQSHLU */
switch (size) {
case 0:
@@ -5760,13 +6182,6 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
g_assert_not_reached();
}
tcg_temp_free_i32(tmp2);
-
- if (op == 3) {
- /* Accumulate. */
- tmp2 = neon_load_reg(rd, pass);
- gen_neon_add(size, tmp, tmp2);
- tcg_temp_free_i32(tmp2);
- }
neon_store_reg(rd, pass, tmp);
}
} /* for pass */
diff --git a/target/arm/vec_helper.c b/target/arm/vec_helper.c
index 230085b35e..fd8b2bff49 100644
--- a/target/arm/vec_helper.c
+++ b/target/arm/vec_helper.c
@@ -924,6 +924,56 @@ DO_SRA(gvec_usra_d, uint64_t)
#undef DO_SRA
+#define DO_RSHR(NAME, TYPE) \
+void HELPER(NAME)(void *vd, void *vn, uint32_t desc) \
+{ \
+ intptr_t i, oprsz = simd_oprsz(desc); \
+ int shift = simd_data(desc); \
+ TYPE *d = vd, *n = vn; \
+ for (i = 0; i < oprsz / sizeof(TYPE); i++) { \
+ TYPE tmp = n[i] >> (shift - 1); \
+ d[i] = (tmp >> 1) + (tmp & 1); \
+ } \
+ clear_tail(d, oprsz, simd_maxsz(desc)); \
+}
+
+DO_RSHR(gvec_srshr_b, int8_t)
+DO_RSHR(gvec_srshr_h, int16_t)
+DO_RSHR(gvec_srshr_s, int32_t)
+DO_RSHR(gvec_srshr_d, int64_t)
+
+DO_RSHR(gvec_urshr_b, uint8_t)
+DO_RSHR(gvec_urshr_h, uint16_t)
+DO_RSHR(gvec_urshr_s, uint32_t)
+DO_RSHR(gvec_urshr_d, uint64_t)
+
+#undef DO_RSHR
+
+#define DO_RSRA(NAME, TYPE) \
+void HELPER(NAME)(void *vd, void *vn, uint32_t desc) \
+{ \
+ intptr_t i, oprsz = simd_oprsz(desc); \
+ int shift = simd_data(desc); \
+ TYPE *d = vd, *n = vn; \
+ for (i = 0; i < oprsz / sizeof(TYPE); i++) { \
+ TYPE tmp = n[i] >> (shift - 1); \
+ d[i] += (tmp >> 1) + (tmp & 1); \
+ } \
+ clear_tail(d, oprsz, simd_maxsz(desc)); \
+}
+
+DO_RSRA(gvec_srsra_b, int8_t)
+DO_RSRA(gvec_srsra_h, int16_t)
+DO_RSRA(gvec_srsra_s, int32_t)
+DO_RSRA(gvec_srsra_d, int64_t)
+
+DO_RSRA(gvec_ursra_b, uint8_t)
+DO_RSRA(gvec_ursra_h, uint16_t)
+DO_RSRA(gvec_ursra_s, uint32_t)
+DO_RSRA(gvec_ursra_d, uint64_t)
+
+#undef DO_RSRA
+
/*
* Convert float16 to float32, raising no exceptions and
* preserving exceptional values, including SNaN.
--
2.20.1
^ permalink raw reply related [flat|nested] 18+ messages in thread* [PATCH v2 03/15] target/arm: Create gen_gvec_{sri,sli}
2020-05-02 22:44 [PATCH v2 00/15] target/arm: partial vector cleanup Richard Henderson
2020-05-02 22:44 ` [PATCH v2 01/15] target/arm: Create gen_gvec_[us]sra Richard Henderson
2020-05-02 22:44 ` [PATCH v2 02/15] target/arm: Create gen_gvec_{u,s}{rshr,rsra} Richard Henderson
@ 2020-05-02 22:44 ` Richard Henderson
2020-05-02 22:44 ` [PATCH v2 04/15] target/arm: Remove unnecessary range check for VSHL Richard Henderson
` (12 subsequent siblings)
15 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2020-05-02 22:44 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
The functions eliminate duplication of the special cases for
this operation. They match up with the GVecGen2iFn typedef.
Add out-of-line helpers. We got away with only having inline
expanders because the neon vector size is only 16 bytes, and
we know that the inline expansion will always succeed.
When we reuse this for SVE, tcg-gvec-op may decide to use an
out-of-line helper due to longer vector lengths.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/helper.h | 10 ++
target/arm/translate.h | 7 +-
target/arm/translate-a64.c | 20 +---
target/arm/translate.c | 186 +++++++++++++++++++++----------------
target/arm/vec_helper.c | 38 ++++++++
5 files changed, 160 insertions(+), 101 deletions(-)
diff --git a/target/arm/helper.h b/target/arm/helper.h
index aeb1f52455..33c76192d2 100644
--- a/target/arm/helper.h
+++ b/target/arm/helper.h
@@ -721,6 +721,16 @@ DEF_HELPER_FLAGS_3(gvec_ursra_h, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
DEF_HELPER_FLAGS_3(gvec_ursra_s, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
DEF_HELPER_FLAGS_3(gvec_ursra_d, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_sri_b, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_sri_h, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_sri_s, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_sri_d, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+
+DEF_HELPER_FLAGS_3(gvec_sli_b, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_sli_h, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_sli_s, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_3(gvec_sli_d, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+
#ifdef TARGET_AARCH64
#include "helper-a64.h"
#include "helper-sve.h"
diff --git a/target/arm/translate.h b/target/arm/translate.h
index 823821f82c..7a2008f0dd 100644
--- a/target/arm/translate.h
+++ b/target/arm/translate.h
@@ -285,8 +285,6 @@ extern const GVecGen3 mls_op[4];
extern const GVecGen3 cmtst_op[4];
extern const GVecGen3 sshl_op[4];
extern const GVecGen3 ushl_op[4];
-extern const GVecGen2i sri_op[4];
-extern const GVecGen2i sli_op[4];
extern const GVecGen4 uqadd_op[4];
extern const GVecGen4 sqadd_op[4];
extern const GVecGen4 uqsub_op[4];
@@ -311,6 +309,11 @@ void gen_gvec_srsra(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
void gen_gvec_ursra(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
int64_t shift, uint32_t opr_sz, uint32_t max_sz);
+void gen_gvec_sri(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
+ int64_t shift, uint32_t opr_sz, uint32_t max_sz);
+void gen_gvec_sli(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
+ int64_t shift, uint32_t opr_sz, uint32_t max_sz);
+
/*
* Forward to the isar_feature_* tests given a DisasContext pointer.
*/
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index 1ef05d5ce1..bc326dadda 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -602,16 +602,6 @@ static void gen_gvec_op2(DisasContext *s, bool is_q, int rd,
is_q ? 16 : 8, vec_full_reg_size(s), gvec_op);
}
-/* Expand a 2-operand + immediate AdvSIMD vector operation using
- * an op descriptor.
- */
-static void gen_gvec_op2i(DisasContext *s, bool is_q, int rd,
- int rn, int64_t imm, const GVecGen2i *gvec_op)
-{
- tcg_gen_gvec_2i(vec_full_reg_offset(s, rd), vec_full_reg_offset(s, rn),
- is_q ? 16 : 8, vec_full_reg_size(s), imm, gvec_op);
-}
-
/* Expand a 3-operand AdvSIMD vector operation using an op descriptor. */
static void gen_gvec_op3(DisasContext *s, bool is_q, int rd,
int rn, int rm, const GVecGen3 *gvec_op)
@@ -10208,12 +10198,9 @@ static void handle_vec_simd_shri(DisasContext *s, bool is_q, bool is_u,
gen_gvec_fn2i(s, is_q, rd, rn, shift,
is_u ? gen_gvec_usra : gen_gvec_ssra, size);
return;
+
case 0x08: /* SRI */
- /* Shift count same as element size is valid but does nothing. */
- if (shift == 8 << size) {
- goto done;
- }
- gen_gvec_op2i(s, is_q, rd, rn, shift, &sri_op[size]);
+ gen_gvec_fn2i(s, is_q, rd, rn, shift, gen_gvec_sri, size);
return;
case 0x00: /* SSHR / USHR */
@@ -10264,7 +10251,6 @@ static void handle_vec_simd_shri(DisasContext *s, bool is_q, bool is_u,
}
tcg_temp_free_i64(tcg_round);
- done:
clear_vec_high(s, is_q, rd);
}
@@ -10289,7 +10275,7 @@ static void handle_vec_simd_shli(DisasContext *s, bool is_q, bool insert,
}
if (insert) {
- gen_gvec_op2i(s, is_q, rd, rn, shift, &sli_op[size]);
+ gen_gvec_fn2i(s, is_q, rd, rn, shift, gen_gvec_sli, size);
} else {
gen_gvec_fn2i(s, is_q, rd, rn, shift, tcg_gen_gvec_shli, size);
}
diff --git a/target/arm/translate.c b/target/arm/translate.c
index d724022cb6..f730eb5b75 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -4726,47 +4726,62 @@ static void gen_shr64_ins_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
static void gen_shr_ins_vec(unsigned vece, TCGv_vec d, TCGv_vec a, int64_t sh)
{
- if (sh == 0) {
- tcg_gen_mov_vec(d, a);
- } else {
- TCGv_vec t = tcg_temp_new_vec_matching(d);
- TCGv_vec m = tcg_temp_new_vec_matching(d);
+ TCGv_vec t = tcg_temp_new_vec_matching(d);
+ TCGv_vec m = tcg_temp_new_vec_matching(d);
- tcg_gen_dupi_vec(vece, m, MAKE_64BIT_MASK((8 << vece) - sh, sh));
- tcg_gen_shri_vec(vece, t, a, sh);
- tcg_gen_and_vec(vece, d, d, m);
- tcg_gen_or_vec(vece, d, d, t);
+ tcg_gen_dupi_vec(vece, m, MAKE_64BIT_MASK((8 << vece) - sh, sh));
+ tcg_gen_shri_vec(vece, t, a, sh);
+ tcg_gen_and_vec(vece, d, d, m);
+ tcg_gen_or_vec(vece, d, d, t);
- tcg_temp_free_vec(t);
- tcg_temp_free_vec(m);
- }
+ tcg_temp_free_vec(t);
+ tcg_temp_free_vec(m);
}
-static const TCGOpcode vecop_list_sri[] = { INDEX_op_shri_vec, 0 };
+void gen_gvec_sri(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
+ int64_t shift, uint32_t opr_sz, uint32_t max_sz)
+{
+ static const TCGOpcode vecop_list[] = { INDEX_op_shri_vec, 0 };
+ const GVecGen2i ops[4] = {
+ { .fni8 = gen_shr8_ins_i64,
+ .fniv = gen_shr_ins_vec,
+ .fno = gen_helper_gvec_sri_b,
+ .load_dest = true,
+ .opt_opc = vecop_list,
+ .vece = MO_8 },
+ { .fni8 = gen_shr16_ins_i64,
+ .fniv = gen_shr_ins_vec,
+ .fno = gen_helper_gvec_sri_h,
+ .load_dest = true,
+ .opt_opc = vecop_list,
+ .vece = MO_16 },
+ { .fni4 = gen_shr32_ins_i32,
+ .fniv = gen_shr_ins_vec,
+ .fno = gen_helper_gvec_sri_s,
+ .load_dest = true,
+ .opt_opc = vecop_list,
+ .vece = MO_32 },
+ { .fni8 = gen_shr64_ins_i64,
+ .fniv = gen_shr_ins_vec,
+ .fno = gen_helper_gvec_sri_d,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ .load_dest = true,
+ .opt_opc = vecop_list,
+ .vece = MO_64 },
+ };
-const GVecGen2i sri_op[4] = {
- { .fni8 = gen_shr8_ins_i64,
- .fniv = gen_shr_ins_vec,
- .load_dest = true,
- .opt_opc = vecop_list_sri,
- .vece = MO_8 },
- { .fni8 = gen_shr16_ins_i64,
- .fniv = gen_shr_ins_vec,
- .load_dest = true,
- .opt_opc = vecop_list_sri,
- .vece = MO_16 },
- { .fni4 = gen_shr32_ins_i32,
- .fniv = gen_shr_ins_vec,
- .load_dest = true,
- .opt_opc = vecop_list_sri,
- .vece = MO_32 },
- { .fni8 = gen_shr64_ins_i64,
- .fniv = gen_shr_ins_vec,
- .prefer_i64 = TCG_TARGET_REG_BITS == 64,
- .load_dest = true,
- .opt_opc = vecop_list_sri,
- .vece = MO_64 },
-};
+ /* tszimm encoding produces immediates in the range [1..esize]. */
+ tcg_debug_assert(shift > 0);
+ tcg_debug_assert(shift <= (8 << vece));
+
+ /* Shift of esize leaves destination unchanged. */
+ if (shift < (8 << vece)) {
+ tcg_gen_gvec_2i(rd_ofs, rm_ofs, opr_sz, max_sz, shift, &ops[vece]);
+ } else {
+ /* Nop, but we do need to clear the tail. */
+ tcg_gen_gvec_mov(vece, rd_ofs, rd_ofs, opr_sz, max_sz);
+ }
+}
static void gen_shl8_ins_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
{
@@ -4804,47 +4819,60 @@ static void gen_shl64_ins_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
static void gen_shl_ins_vec(unsigned vece, TCGv_vec d, TCGv_vec a, int64_t sh)
{
- if (sh == 0) {
- tcg_gen_mov_vec(d, a);
- } else {
- TCGv_vec t = tcg_temp_new_vec_matching(d);
- TCGv_vec m = tcg_temp_new_vec_matching(d);
+ TCGv_vec t = tcg_temp_new_vec_matching(d);
+ TCGv_vec m = tcg_temp_new_vec_matching(d);
- tcg_gen_dupi_vec(vece, m, MAKE_64BIT_MASK(0, sh));
- tcg_gen_shli_vec(vece, t, a, sh);
- tcg_gen_and_vec(vece, d, d, m);
- tcg_gen_or_vec(vece, d, d, t);
+ tcg_gen_shli_vec(vece, t, a, sh);
+ tcg_gen_dupi_vec(vece, m, MAKE_64BIT_MASK(0, sh));
+ tcg_gen_and_vec(vece, d, d, m);
+ tcg_gen_or_vec(vece, d, d, t);
- tcg_temp_free_vec(t);
- tcg_temp_free_vec(m);
- }
+ tcg_temp_free_vec(t);
+ tcg_temp_free_vec(m);
}
-static const TCGOpcode vecop_list_sli[] = { INDEX_op_shli_vec, 0 };
+void gen_gvec_sli(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
+ int64_t shift, uint32_t opr_sz, uint32_t max_sz)
+{
+ static const TCGOpcode vecop_list[] = { INDEX_op_shli_vec, 0 };
+ const GVecGen2i ops[4] = {
+ { .fni8 = gen_shl8_ins_i64,
+ .fniv = gen_shl_ins_vec,
+ .fno = gen_helper_gvec_sli_b,
+ .load_dest = true,
+ .opt_opc = vecop_list,
+ .vece = MO_8 },
+ { .fni8 = gen_shl16_ins_i64,
+ .fniv = gen_shl_ins_vec,
+ .fno = gen_helper_gvec_sli_h,
+ .load_dest = true,
+ .opt_opc = vecop_list,
+ .vece = MO_16 },
+ { .fni4 = gen_shl32_ins_i32,
+ .fniv = gen_shl_ins_vec,
+ .fno = gen_helper_gvec_sli_s,
+ .load_dest = true,
+ .opt_opc = vecop_list,
+ .vece = MO_32 },
+ { .fni8 = gen_shl64_ins_i64,
+ .fniv = gen_shl_ins_vec,
+ .fno = gen_helper_gvec_sli_d,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ .load_dest = true,
+ .opt_opc = vecop_list,
+ .vece = MO_64 },
+ };
-const GVecGen2i sli_op[4] = {
- { .fni8 = gen_shl8_ins_i64,
- .fniv = gen_shl_ins_vec,
- .load_dest = true,
- .opt_opc = vecop_list_sli,
- .vece = MO_8 },
- { .fni8 = gen_shl16_ins_i64,
- .fniv = gen_shl_ins_vec,
- .load_dest = true,
- .opt_opc = vecop_list_sli,
- .vece = MO_16 },
- { .fni4 = gen_shl32_ins_i32,
- .fniv = gen_shl_ins_vec,
- .load_dest = true,
- .opt_opc = vecop_list_sli,
- .vece = MO_32 },
- { .fni8 = gen_shl64_ins_i64,
- .fniv = gen_shl_ins_vec,
- .prefer_i64 = TCG_TARGET_REG_BITS == 64,
- .load_dest = true,
- .opt_opc = vecop_list_sli,
- .vece = MO_64 },
-};
+ /* tszimm encoding produces immediates in the range [0..esize-1]. */
+ tcg_debug_assert(shift >= 0);
+ tcg_debug_assert(shift < (8 << vece));
+
+ if (shift == 0) {
+ tcg_gen_gvec_mov(vece, rd_ofs, rm_ofs, opr_sz, max_sz);
+ } else {
+ tcg_gen_gvec_2i(rd_ofs, rm_ofs, opr_sz, max_sz, shift, &ops[vece]);
+ }
+}
static void gen_mla8_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
{
@@ -6089,20 +6117,14 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
}
/* Right shift comes here negative. */
shift = -shift;
- /* Shift out of range leaves destination unchanged. */
- if (shift < 8 << size) {
- tcg_gen_gvec_2i(rd_ofs, rm_ofs, vec_size, vec_size,
- shift, &sri_op[size]);
- }
+ gen_gvec_sri(size, rd_ofs, rm_ofs, shift,
+ vec_size, vec_size);
return 0;
case 5: /* VSHL, VSLI */
if (u) { /* VSLI */
- /* Shift out of range leaves destination unchanged. */
- if (shift < 8 << size) {
- tcg_gen_gvec_2i(rd_ofs, rm_ofs, vec_size,
- vec_size, shift, &sli_op[size]);
- }
+ gen_gvec_sli(size, rd_ofs, rm_ofs, shift,
+ vec_size, vec_size);
} else { /* VSHL */
/* Shifts larger than the element size are
* architecturally valid and results in zero.
diff --git a/target/arm/vec_helper.c b/target/arm/vec_helper.c
index fd8b2bff49..096fea67ef 100644
--- a/target/arm/vec_helper.c
+++ b/target/arm/vec_helper.c
@@ -974,6 +974,44 @@ DO_RSRA(gvec_ursra_d, uint64_t)
#undef DO_RSRA
+#define DO_SRI(NAME, TYPE) \
+void HELPER(NAME)(void *vd, void *vn, uint32_t desc) \
+{ \
+ intptr_t i, oprsz = simd_oprsz(desc); \
+ int shift = simd_data(desc); \
+ TYPE *d = vd, *n = vn; \
+ for (i = 0; i < oprsz / sizeof(TYPE); i++) { \
+ d[i] = deposit64(d[i], 0, sizeof(TYPE) * 8 - shift, n[i] >> shift); \
+ } \
+ clear_tail(d, oprsz, simd_maxsz(desc)); \
+}
+
+DO_SRI(gvec_sri_b, uint8_t)
+DO_SRI(gvec_sri_h, uint16_t)
+DO_SRI(gvec_sri_s, uint32_t)
+DO_SRI(gvec_sri_d, uint64_t)
+
+#undef DO_SRI
+
+#define DO_SLI(NAME, TYPE) \
+void HELPER(NAME)(void *vd, void *vn, uint32_t desc) \
+{ \
+ intptr_t i, oprsz = simd_oprsz(desc); \
+ int shift = simd_data(desc); \
+ TYPE *d = vd, *n = vn; \
+ for (i = 0; i < oprsz / sizeof(TYPE); i++) { \
+ d[i] = deposit64(d[i], shift, sizeof(TYPE) * 8 - shift, n[i]); \
+ } \
+ clear_tail(d, oprsz, simd_maxsz(desc)); \
+}
+
+DO_SLI(gvec_sli_b, uint8_t)
+DO_SLI(gvec_sli_h, uint16_t)
+DO_SLI(gvec_sli_s, uint32_t)
+DO_SLI(gvec_sli_d, uint64_t)
+
+#undef DO_SLI
+
/*
* Convert float16 to float32, raising no exceptions and
* preserving exceptional values, including SNaN.
--
2.20.1
^ permalink raw reply related [flat|nested] 18+ messages in thread* [PATCH v2 04/15] target/arm: Remove unnecessary range check for VSHL
2020-05-02 22:44 [PATCH v2 00/15] target/arm: partial vector cleanup Richard Henderson
` (2 preceding siblings ...)
2020-05-02 22:44 ` [PATCH v2 03/15] target/arm: Create gen_gvec_{sri,sli} Richard Henderson
@ 2020-05-02 22:44 ` Richard Henderson
2020-05-02 22:44 ` [PATCH v2 05/15] target/arm: Tidy handle_vec_simd_shri Richard Henderson
` (11 subsequent siblings)
15 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2020-05-02 22:44 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
In 1dc8425e551, while converting to gvec, I added an extra range check
against the shift count. This was unnecessary because the encoding of
the shift count produces 0 to the element size - 1.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate.c | 12 ++----------
1 file changed, 2 insertions(+), 10 deletions(-)
diff --git a/target/arm/translate.c b/target/arm/translate.c
index f730eb5b75..f082384117 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -6126,16 +6126,8 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
gen_gvec_sli(size, rd_ofs, rm_ofs, shift,
vec_size, vec_size);
} else { /* VSHL */
- /* Shifts larger than the element size are
- * architecturally valid and results in zero.
- */
- if (shift >= 8 << size) {
- tcg_gen_gvec_dup_imm(size, rd_ofs,
- vec_size, vec_size, 0);
- } else {
- tcg_gen_gvec_shli(size, rd_ofs, rm_ofs, shift,
- vec_size, vec_size);
- }
+ tcg_gen_gvec_shli(size, rd_ofs, rm_ofs, shift,
+ vec_size, vec_size);
}
return 0;
}
--
2.20.1
^ permalink raw reply related [flat|nested] 18+ messages in thread* [PATCH v2 05/15] target/arm: Tidy handle_vec_simd_shri
2020-05-02 22:44 [PATCH v2 00/15] target/arm: partial vector cleanup Richard Henderson
` (3 preceding siblings ...)
2020-05-02 22:44 ` [PATCH v2 04/15] target/arm: Remove unnecessary range check for VSHL Richard Henderson
@ 2020-05-02 22:44 ` Richard Henderson
2020-05-02 22:44 ` [PATCH v2 06/15] target/arm: Wrap vector compare zero GVecGen2 in GVecGen2Fn Richard Henderson
` (10 subsequent siblings)
15 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2020-05-02 22:44 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Now that we've converted all cases to gvec, there is quite a bit
of dead code at the end of the function. Remove it.
Sink the call to gen_gvec_fn2i to the end, loading a function
pointer within the switch statement.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate-a64.c | 56 ++++++++++----------------------------
1 file changed, 14 insertions(+), 42 deletions(-)
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index bc326dadda..5937069992 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -10172,16 +10172,7 @@ static void handle_vec_simd_shri(DisasContext *s, bool is_q, bool is_u,
int size = 32 - clz32(immh) - 1;
int immhb = immh << 3 | immb;
int shift = 2 * (8 << size) - immhb;
- bool accumulate = false;
- int dsize = is_q ? 128 : 64;
- int esize = 8 << size;
- int elements = dsize/esize;
- MemOp memop = size | (is_u ? 0 : MO_SIGN);
- TCGv_i64 tcg_rn = new_tmp_a64(s);
- TCGv_i64 tcg_rd = new_tmp_a64(s);
- TCGv_i64 tcg_round;
- uint64_t round_const;
- int i;
+ GVecGen2iFn *gvec_fn;
if (extract32(immh, 3, 1) && !is_q) {
unallocated_encoding(s);
@@ -10195,13 +10186,12 @@ static void handle_vec_simd_shri(DisasContext *s, bool is_q, bool is_u,
switch (opcode) {
case 0x02: /* SSRA / USRA (accumulate) */
- gen_gvec_fn2i(s, is_q, rd, rn, shift,
- is_u ? gen_gvec_usra : gen_gvec_ssra, size);
- return;
+ gvec_fn = is_u ? gen_gvec_usra : gen_gvec_ssra;
+ break;
case 0x08: /* SRI */
- gen_gvec_fn2i(s, is_q, rd, rn, shift, gen_gvec_sri, size);
- return;
+ gvec_fn = gen_gvec_sri;
+ break;
case 0x00: /* SSHR / USHR */
if (is_u) {
@@ -10209,49 +10199,31 @@ static void handle_vec_simd_shri(DisasContext *s, bool is_q, bool is_u,
/* Shift count the same size as element size produces zero. */
tcg_gen_gvec_dup_imm(size, vec_full_reg_offset(s, rd),
is_q ? 16 : 8, vec_full_reg_size(s), 0);
- } else {
- gen_gvec_fn2i(s, is_q, rd, rn, shift, tcg_gen_gvec_shri, size);
+ return;
}
+ gvec_fn = tcg_gen_gvec_shri;
} else {
/* Shift count the same size as element size produces all sign. */
if (shift == 8 << size) {
shift -= 1;
}
- gen_gvec_fn2i(s, is_q, rd, rn, shift, tcg_gen_gvec_sari, size);
+ gvec_fn = tcg_gen_gvec_sari;
}
- return;
+ break;
case 0x04: /* SRSHR / URSHR (rounding) */
- gen_gvec_fn2i(s, is_q, rd, rn, shift,
- is_u ? gen_gvec_urshr : gen_gvec_srshr, size);
- return;
+ gvec_fn = is_u ? gen_gvec_urshr : gen_gvec_srshr;
+ break;
case 0x06: /* SRSRA / URSRA (accum + rounding) */
- gen_gvec_fn2i(s, is_q, rd, rn, shift,
- is_u ? gen_gvec_ursra : gen_gvec_srsra, size);
- return;
+ gvec_fn = is_u ? gen_gvec_ursra : gen_gvec_srsra;
+ break;
default:
g_assert_not_reached();
}
- round_const = 1ULL << (shift - 1);
- tcg_round = tcg_const_i64(round_const);
-
- for (i = 0; i < elements; i++) {
- read_vec_element(s, tcg_rn, rn, i, memop);
- if (accumulate) {
- read_vec_element(s, tcg_rd, rd, i, memop);
- }
-
- handle_shri_with_rndacc(tcg_rd, tcg_rn, tcg_round,
- accumulate, is_u, size, shift);
-
- write_vec_element(s, tcg_rd, rd, i, size);
- }
- tcg_temp_free_i64(tcg_round);
-
- clear_vec_high(s, is_q, rd);
+ gen_gvec_fn2i(s, is_q, rd, rn, shift, gvec_fn, size);
}
/* SHL/SLI - Vector shift left */
--
2.20.1
^ permalink raw reply related [flat|nested] 18+ messages in thread* [PATCH v2 06/15] target/arm: Wrap vector compare zero GVecGen2 in GVecGen2Fn
2020-05-02 22:44 [PATCH v2 00/15] target/arm: partial vector cleanup Richard Henderson
` (4 preceding siblings ...)
2020-05-02 22:44 ` [PATCH v2 05/15] target/arm: Tidy handle_vec_simd_shri Richard Henderson
@ 2020-05-02 22:44 ` Richard Henderson
2020-05-02 22:44 ` [PATCH v2 07/15] target/arm: Wrap vector mla/mls GVecGen3 in GVecGen3Fn Richard Henderson
` (9 subsequent siblings)
15 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2020-05-02 22:44 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Provide a functional interface for the vector expansion.
This fits better with the existing set of helpers that
we provide for other operations.
Macro-ize the 5 nearly identical comparisons.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate.h | 16 ++-
target/arm/translate-a64.c | 22 ++--
target/arm/translate.c | 254 ++++++++-----------------------------
3 files changed, 74 insertions(+), 218 deletions(-)
diff --git a/target/arm/translate.h b/target/arm/translate.h
index 7a2008f0dd..20ec9cedd7 100644
--- a/target/arm/translate.h
+++ b/target/arm/translate.h
@@ -275,11 +275,17 @@ static inline void gen_swstep_exception(DisasContext *s, int isv, int ex)
uint64_t vfp_expand_imm(int size, uint8_t imm8);
/* Vector operations shared between ARM and AArch64. */
-extern const GVecGen2 ceq0_op[4];
-extern const GVecGen2 clt0_op[4];
-extern const GVecGen2 cgt0_op[4];
-extern const GVecGen2 cle0_op[4];
-extern const GVecGen2 cge0_op[4];
+void gen_gvec_ceq0(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
+ uint32_t opr_sz, uint32_t max_sz);
+void gen_gvec_clt0(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
+ uint32_t opr_sz, uint32_t max_sz);
+void gen_gvec_cgt0(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
+ uint32_t opr_sz, uint32_t max_sz);
+void gen_gvec_cle0(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
+ uint32_t opr_sz, uint32_t max_sz);
+void gen_gvec_cge0(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
+ uint32_t opr_sz, uint32_t max_sz);
+
extern const GVecGen3 mla_op[4];
extern const GVecGen3 mls_op[4];
extern const GVecGen3 cmtst_op[4];
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index 5937069992..8208651394 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -594,14 +594,6 @@ static void gen_gvec_fn4(DisasContext *s, bool is_q, int rd, int rn, int rm,
is_q ? 16 : 8, vec_full_reg_size(s));
}
-/* Expand a 2-operand AdvSIMD vector operation using an op descriptor. */
-static void gen_gvec_op2(DisasContext *s, bool is_q, int rd,
- int rn, const GVecGen2 *gvec_op)
-{
- tcg_gen_gvec_2(vec_full_reg_offset(s, rd), vec_full_reg_offset(s, rn),
- is_q ? 16 : 8, vec_full_reg_size(s), gvec_op);
-}
-
/* Expand a 3-operand AdvSIMD vector operation using an op descriptor. */
static void gen_gvec_op3(DisasContext *s, bool is_q, int rd,
int rn, int rm, const GVecGen3 *gvec_op)
@@ -12327,13 +12319,21 @@ static void disas_simd_two_reg_misc(DisasContext *s, uint32_t insn)
}
break;
case 0x8: /* CMGT, CMGE */
- gen_gvec_op2(s, is_q, rd, rn, u ? &cge0_op[size] : &cgt0_op[size]);
+ if (u) {
+ gen_gvec_fn2(s, is_q, rd, rn, gen_gvec_cge0, size);
+ } else {
+ gen_gvec_fn2(s, is_q, rd, rn, gen_gvec_cgt0, size);
+ }
return;
case 0x9: /* CMEQ, CMLE */
- gen_gvec_op2(s, is_q, rd, rn, u ? &cle0_op[size] : &ceq0_op[size]);
+ if (u) {
+ gen_gvec_fn2(s, is_q, rd, rn, gen_gvec_cle0, size);
+ } else {
+ gen_gvec_fn2(s, is_q, rd, rn, gen_gvec_ceq0, size);
+ }
return;
case 0xa: /* CMLT */
- gen_gvec_op2(s, is_q, rd, rn, &clt0_op[size]);
+ gen_gvec_fn2(s, is_q, rd, rn, gen_gvec_clt0, size);
return;
case 0xb:
if (u) { /* ABS, NEG */
diff --git a/target/arm/translate.c b/target/arm/translate.c
index f082384117..b08c4a2527 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -3917,204 +3917,59 @@ static int do_v81_helper(DisasContext *s, gen_helper_gvec_3_ptr *fn,
return 1;
}
-static void gen_ceq0_i32(TCGv_i32 d, TCGv_i32 a)
-{
- tcg_gen_setcondi_i32(TCG_COND_EQ, d, a, 0);
- tcg_gen_neg_i32(d, d);
-}
-
-static void gen_ceq0_i64(TCGv_i64 d, TCGv_i64 a)
-{
- tcg_gen_setcondi_i64(TCG_COND_EQ, d, a, 0);
- tcg_gen_neg_i64(d, d);
-}
-
-static void gen_ceq0_vec(unsigned vece, TCGv_vec d, TCGv_vec a)
-{
- TCGv_vec zero = tcg_const_zeros_vec_matching(d);
- tcg_gen_cmp_vec(TCG_COND_EQ, vece, d, a, zero);
- tcg_temp_free_vec(zero);
-}
+#define GEN_CMP0(NAME, COND) \
+ static void gen_##NAME##0_i32(TCGv_i32 d, TCGv_i32 a) \
+ { \
+ tcg_gen_setcondi_i32(COND, d, a, 0); \
+ tcg_gen_neg_i32(d, d); \
+ } \
+ static void gen_##NAME##0_i64(TCGv_i64 d, TCGv_i64 a) \
+ { \
+ tcg_gen_setcondi_i64(COND, d, a, 0); \
+ tcg_gen_neg_i64(d, d); \
+ } \
+ static void gen_##NAME##0_vec(unsigned vece, TCGv_vec d, TCGv_vec a) \
+ { \
+ TCGv_vec zero = tcg_const_zeros_vec_matching(d); \
+ tcg_gen_cmp_vec(COND, vece, d, a, zero); \
+ tcg_temp_free_vec(zero); \
+ } \
+ void gen_gvec_##NAME##0(unsigned vece, uint32_t d, uint32_t m, \
+ uint32_t opr_sz, uint32_t max_sz) \
+ { \
+ const GVecGen2 op[4] = { \
+ { .fno = gen_helper_gvec_##NAME##0_b, \
+ .fniv = gen_##NAME##0_vec, \
+ .opt_opc = vecop_list_cmp, \
+ .vece = MO_8 }, \
+ { .fno = gen_helper_gvec_##NAME##0_h, \
+ .fniv = gen_##NAME##0_vec, \
+ .opt_opc = vecop_list_cmp, \
+ .vece = MO_16 }, \
+ { .fni4 = gen_##NAME##0_i32, \
+ .fniv = gen_##NAME##0_vec, \
+ .opt_opc = vecop_list_cmp, \
+ .vece = MO_32 }, \
+ { .fni8 = gen_##NAME##0_i64, \
+ .fniv = gen_##NAME##0_vec, \
+ .opt_opc = vecop_list_cmp, \
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64, \
+ .vece = MO_64 }, \
+ }; \
+ tcg_gen_gvec_2(d, m, opr_sz, max_sz, &op[vece]); \
+ }
static const TCGOpcode vecop_list_cmp[] = {
INDEX_op_cmp_vec, 0
};
-const GVecGen2 ceq0_op[4] = {
- { .fno = gen_helper_gvec_ceq0_b,
- .fniv = gen_ceq0_vec,
- .opt_opc = vecop_list_cmp,
- .vece = MO_8 },
- { .fno = gen_helper_gvec_ceq0_h,
- .fniv = gen_ceq0_vec,
- .opt_opc = vecop_list_cmp,
- .vece = MO_16 },
- { .fni4 = gen_ceq0_i32,
- .fniv = gen_ceq0_vec,
- .opt_opc = vecop_list_cmp,
- .vece = MO_32 },
- { .fni8 = gen_ceq0_i64,
- .fniv = gen_ceq0_vec,
- .opt_opc = vecop_list_cmp,
- .prefer_i64 = TCG_TARGET_REG_BITS == 64,
- .vece = MO_64 },
-};
+GEN_CMP0(ceq, TCG_COND_EQ)
+GEN_CMP0(cle, TCG_COND_LE)
+GEN_CMP0(cge, TCG_COND_GE)
+GEN_CMP0(clt, TCG_COND_LT)
+GEN_CMP0(cgt, TCG_COND_GT)
-static void gen_cle0_i32(TCGv_i32 d, TCGv_i32 a)
-{
- tcg_gen_setcondi_i32(TCG_COND_LE, d, a, 0);
- tcg_gen_neg_i32(d, d);
-}
-
-static void gen_cle0_i64(TCGv_i64 d, TCGv_i64 a)
-{
- tcg_gen_setcondi_i64(TCG_COND_LE, d, a, 0);
- tcg_gen_neg_i64(d, d);
-}
-
-static void gen_cle0_vec(unsigned vece, TCGv_vec d, TCGv_vec a)
-{
- TCGv_vec zero = tcg_const_zeros_vec_matching(d);
- tcg_gen_cmp_vec(TCG_COND_LE, vece, d, a, zero);
- tcg_temp_free_vec(zero);
-}
-
-const GVecGen2 cle0_op[4] = {
- { .fno = gen_helper_gvec_cle0_b,
- .fniv = gen_cle0_vec,
- .opt_opc = vecop_list_cmp,
- .vece = MO_8 },
- { .fno = gen_helper_gvec_cle0_h,
- .fniv = gen_cle0_vec,
- .opt_opc = vecop_list_cmp,
- .vece = MO_16 },
- { .fni4 = gen_cle0_i32,
- .fniv = gen_cle0_vec,
- .opt_opc = vecop_list_cmp,
- .vece = MO_32 },
- { .fni8 = gen_cle0_i64,
- .fniv = gen_cle0_vec,
- .opt_opc = vecop_list_cmp,
- .prefer_i64 = TCG_TARGET_REG_BITS == 64,
- .vece = MO_64 },
-};
-
-static void gen_cge0_i32(TCGv_i32 d, TCGv_i32 a)
-{
- tcg_gen_setcondi_i32(TCG_COND_GE, d, a, 0);
- tcg_gen_neg_i32(d, d);
-}
-
-static void gen_cge0_i64(TCGv_i64 d, TCGv_i64 a)
-{
- tcg_gen_setcondi_i64(TCG_COND_GE, d, a, 0);
- tcg_gen_neg_i64(d, d);
-}
-
-static void gen_cge0_vec(unsigned vece, TCGv_vec d, TCGv_vec a)
-{
- TCGv_vec zero = tcg_const_zeros_vec_matching(d);
- tcg_gen_cmp_vec(TCG_COND_GE, vece, d, a, zero);
- tcg_temp_free_vec(zero);
-}
-
-const GVecGen2 cge0_op[4] = {
- { .fno = gen_helper_gvec_cge0_b,
- .fniv = gen_cge0_vec,
- .opt_opc = vecop_list_cmp,
- .vece = MO_8 },
- { .fno = gen_helper_gvec_cge0_h,
- .fniv = gen_cge0_vec,
- .opt_opc = vecop_list_cmp,
- .vece = MO_16 },
- { .fni4 = gen_cge0_i32,
- .fniv = gen_cge0_vec,
- .opt_opc = vecop_list_cmp,
- .vece = MO_32 },
- { .fni8 = gen_cge0_i64,
- .fniv = gen_cge0_vec,
- .opt_opc = vecop_list_cmp,
- .prefer_i64 = TCG_TARGET_REG_BITS == 64,
- .vece = MO_64 },
-};
-
-static void gen_clt0_i32(TCGv_i32 d, TCGv_i32 a)
-{
- tcg_gen_setcondi_i32(TCG_COND_LT, d, a, 0);
- tcg_gen_neg_i32(d, d);
-}
-
-static void gen_clt0_i64(TCGv_i64 d, TCGv_i64 a)
-{
- tcg_gen_setcondi_i64(TCG_COND_LT, d, a, 0);
- tcg_gen_neg_i64(d, d);
-}
-
-static void gen_clt0_vec(unsigned vece, TCGv_vec d, TCGv_vec a)
-{
- TCGv_vec zero = tcg_const_zeros_vec_matching(d);
- tcg_gen_cmp_vec(TCG_COND_LT, vece, d, a, zero);
- tcg_temp_free_vec(zero);
-}
-
-const GVecGen2 clt0_op[4] = {
- { .fno = gen_helper_gvec_clt0_b,
- .fniv = gen_clt0_vec,
- .opt_opc = vecop_list_cmp,
- .vece = MO_8 },
- { .fno = gen_helper_gvec_clt0_h,
- .fniv = gen_clt0_vec,
- .opt_opc = vecop_list_cmp,
- .vece = MO_16 },
- { .fni4 = gen_clt0_i32,
- .fniv = gen_clt0_vec,
- .opt_opc = vecop_list_cmp,
- .vece = MO_32 },
- { .fni8 = gen_clt0_i64,
- .fniv = gen_clt0_vec,
- .opt_opc = vecop_list_cmp,
- .prefer_i64 = TCG_TARGET_REG_BITS == 64,
- .vece = MO_64 },
-};
-
-static void gen_cgt0_i32(TCGv_i32 d, TCGv_i32 a)
-{
- tcg_gen_setcondi_i32(TCG_COND_GT, d, a, 0);
- tcg_gen_neg_i32(d, d);
-}
-
-static void gen_cgt0_i64(TCGv_i64 d, TCGv_i64 a)
-{
- tcg_gen_setcondi_i64(TCG_COND_GT, d, a, 0);
- tcg_gen_neg_i64(d, d);
-}
-
-static void gen_cgt0_vec(unsigned vece, TCGv_vec d, TCGv_vec a)
-{
- TCGv_vec zero = tcg_const_zeros_vec_matching(d);
- tcg_gen_cmp_vec(TCG_COND_GT, vece, d, a, zero);
- tcg_temp_free_vec(zero);
-}
-
-const GVecGen2 cgt0_op[4] = {
- { .fno = gen_helper_gvec_cgt0_b,
- .fniv = gen_cgt0_vec,
- .opt_opc = vecop_list_cmp,
- .vece = MO_8 },
- { .fno = gen_helper_gvec_cgt0_h,
- .fniv = gen_cgt0_vec,
- .opt_opc = vecop_list_cmp,
- .vece = MO_16 },
- { .fni4 = gen_cgt0_i32,
- .fniv = gen_cgt0_vec,
- .opt_opc = vecop_list_cmp,
- .vece = MO_32 },
- { .fni8 = gen_cgt0_i64,
- .fniv = gen_cgt0_vec,
- .opt_opc = vecop_list_cmp,
- .prefer_i64 = TCG_TARGET_REG_BITS == 64,
- .vece = MO_64 },
-};
+#undef GEN_CMP0
static void gen_ssra8_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
{
@@ -7146,24 +7001,19 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
break;
case NEON_2RM_VCEQ0:
- tcg_gen_gvec_2(rd_ofs, rm_ofs, vec_size,
- vec_size, &ceq0_op[size]);
+ gen_gvec_ceq0(size, rd_ofs, rm_ofs, vec_size, vec_size);
break;
case NEON_2RM_VCGT0:
- tcg_gen_gvec_2(rd_ofs, rm_ofs, vec_size,
- vec_size, &cgt0_op[size]);
+ gen_gvec_cgt0(size, rd_ofs, rm_ofs, vec_size, vec_size);
break;
case NEON_2RM_VCLE0:
- tcg_gen_gvec_2(rd_ofs, rm_ofs, vec_size,
- vec_size, &cle0_op[size]);
+ gen_gvec_cle0(size, rd_ofs, rm_ofs, vec_size, vec_size);
break;
case NEON_2RM_VCGE0:
- tcg_gen_gvec_2(rd_ofs, rm_ofs, vec_size,
- vec_size, &cge0_op[size]);
+ gen_gvec_cge0(size, rd_ofs, rm_ofs, vec_size, vec_size);
break;
case NEON_2RM_VCLT0:
- tcg_gen_gvec_2(rd_ofs, rm_ofs, vec_size,
- vec_size, &clt0_op[size]);
+ gen_gvec_clt0(size, rd_ofs, rm_ofs, vec_size, vec_size);
break;
default:
--
2.20.1
^ permalink raw reply related [flat|nested] 18+ messages in thread* [PATCH v2 07/15] target/arm: Wrap vector mla/mls GVecGen3 in GVecGen3Fn
2020-05-02 22:44 [PATCH v2 00/15] target/arm: partial vector cleanup Richard Henderson
` (5 preceding siblings ...)
2020-05-02 22:44 ` [PATCH v2 06/15] target/arm: Wrap vector compare zero GVecGen2 in GVecGen2Fn Richard Henderson
@ 2020-05-02 22:44 ` Richard Henderson
2020-05-02 22:44 ` [PATCH v2 08/15] target/arm: Wrap vector cmtst/ushl/sshl " Richard Henderson
` (8 subsequent siblings)
15 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2020-05-02 22:44 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Provide a functional interface for the vector expansion.
This fits better with the existing set of helpers that
we provide for other operations.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate.h | 7 ++-
target/arm/translate-a64.c | 4 +-
target/arm/translate.c | 124 ++++++++++++++++++++-----------------
3 files changed, 74 insertions(+), 61 deletions(-)
diff --git a/target/arm/translate.h b/target/arm/translate.h
index 20ec9cedd7..4fbcdf1294 100644
--- a/target/arm/translate.h
+++ b/target/arm/translate.h
@@ -286,8 +286,11 @@ void gen_gvec_cle0(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
void gen_gvec_cge0(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
uint32_t opr_sz, uint32_t max_sz);
-extern const GVecGen3 mla_op[4];
-extern const GVecGen3 mls_op[4];
+void gen_gvec_mla(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz);
+void gen_gvec_mls(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz);
+
extern const GVecGen3 cmtst_op[4];
extern const GVecGen3 sshl_op[4];
extern const GVecGen3 ushl_op[4];
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index 8208651394..2b5ae4d43a 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -11243,9 +11243,9 @@ static void disas_simd_3same_int(DisasContext *s, uint32_t insn)
return;
case 0x12: /* MLA, MLS */
if (u) {
- gen_gvec_op3(s, is_q, rd, rn, rm, &mls_op[size]);
+ gen_gvec_fn3(s, is_q, rd, rn, rm, gen_gvec_mls, size);
} else {
- gen_gvec_op3(s, is_q, rd, rn, rm, &mla_op[size]);
+ gen_gvec_fn3(s, is_q, rd, rn, rm, gen_gvec_mla, size);
}
return;
case 0x11:
diff --git a/target/arm/translate.c b/target/arm/translate.c
index b08c4a2527..da807242ff 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -4792,62 +4792,69 @@ static void gen_mls_vec(unsigned vece, TCGv_vec d, TCGv_vec a, TCGv_vec b)
/* Note that while NEON does not support VMLA and VMLS as 64-bit ops,
* these tables are shared with AArch64 which does support them.
*/
+void gen_gvec_mla(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz)
+{
+ static const TCGOpcode vecop_list[] = {
+ INDEX_op_mul_vec, INDEX_op_add_vec, 0
+ };
+ static const GVecGen3 ops[4] = {
+ { .fni4 = gen_mla8_i32,
+ .fniv = gen_mla_vec,
+ .load_dest = true,
+ .opt_opc = vecop_list,
+ .vece = MO_8 },
+ { .fni4 = gen_mla16_i32,
+ .fniv = gen_mla_vec,
+ .load_dest = true,
+ .opt_opc = vecop_list,
+ .vece = MO_16 },
+ { .fni4 = gen_mla32_i32,
+ .fniv = gen_mla_vec,
+ .load_dest = true,
+ .opt_opc = vecop_list,
+ .vece = MO_32 },
+ { .fni8 = gen_mla64_i64,
+ .fniv = gen_mla_vec,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ .load_dest = true,
+ .opt_opc = vecop_list,
+ .vece = MO_64 },
+ };
+ tcg_gen_gvec_3(rd_ofs, rn_ofs, rm_ofs, opr_sz, max_sz, &ops[vece]);
+}
-static const TCGOpcode vecop_list_mla[] = {
- INDEX_op_mul_vec, INDEX_op_add_vec, 0
-};
-
-static const TCGOpcode vecop_list_mls[] = {
- INDEX_op_mul_vec, INDEX_op_sub_vec, 0
-};
-
-const GVecGen3 mla_op[4] = {
- { .fni4 = gen_mla8_i32,
- .fniv = gen_mla_vec,
- .load_dest = true,
- .opt_opc = vecop_list_mla,
- .vece = MO_8 },
- { .fni4 = gen_mla16_i32,
- .fniv = gen_mla_vec,
- .load_dest = true,
- .opt_opc = vecop_list_mla,
- .vece = MO_16 },
- { .fni4 = gen_mla32_i32,
- .fniv = gen_mla_vec,
- .load_dest = true,
- .opt_opc = vecop_list_mla,
- .vece = MO_32 },
- { .fni8 = gen_mla64_i64,
- .fniv = gen_mla_vec,
- .prefer_i64 = TCG_TARGET_REG_BITS == 64,
- .load_dest = true,
- .opt_opc = vecop_list_mla,
- .vece = MO_64 },
-};
-
-const GVecGen3 mls_op[4] = {
- { .fni4 = gen_mls8_i32,
- .fniv = gen_mls_vec,
- .load_dest = true,
- .opt_opc = vecop_list_mls,
- .vece = MO_8 },
- { .fni4 = gen_mls16_i32,
- .fniv = gen_mls_vec,
- .load_dest = true,
- .opt_opc = vecop_list_mls,
- .vece = MO_16 },
- { .fni4 = gen_mls32_i32,
- .fniv = gen_mls_vec,
- .load_dest = true,
- .opt_opc = vecop_list_mls,
- .vece = MO_32 },
- { .fni8 = gen_mls64_i64,
- .fniv = gen_mls_vec,
- .prefer_i64 = TCG_TARGET_REG_BITS == 64,
- .load_dest = true,
- .opt_opc = vecop_list_mls,
- .vece = MO_64 },
-};
+void gen_gvec_mls(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz)
+{
+ static const TCGOpcode vecop_list[] = {
+ INDEX_op_mul_vec, INDEX_op_sub_vec, 0
+ };
+ static const GVecGen3 ops[4] = {
+ { .fni4 = gen_mls8_i32,
+ .fniv = gen_mls_vec,
+ .load_dest = true,
+ .opt_opc = vecop_list,
+ .vece = MO_8 },
+ { .fni4 = gen_mls16_i32,
+ .fniv = gen_mls_vec,
+ .load_dest = true,
+ .opt_opc = vecop_list,
+ .vece = MO_16 },
+ { .fni4 = gen_mls32_i32,
+ .fniv = gen_mls_vec,
+ .load_dest = true,
+ .opt_opc = vecop_list,
+ .vece = MO_32 },
+ { .fni8 = gen_mls64_i64,
+ .fniv = gen_mls_vec,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ .load_dest = true,
+ .opt_opc = vecop_list,
+ .vece = MO_64 },
+ };
+ tcg_gen_gvec_3(rd_ofs, rn_ofs, rm_ofs, opr_sz, max_sz, &ops[vece]);
+}
/* CMTST : test is "if (X & Y != 0)". */
static void gen_cmtst_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
@@ -5529,8 +5536,11 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
return 0;
case NEON_3R_VML: /* VMLA, VMLS */
- tcg_gen_gvec_3(rd_ofs, rn_ofs, rm_ofs, vec_size, vec_size,
- u ? &mls_op[size] : &mla_op[size]);
+ if (u) {
+ gen_gvec_mls(size, rd_ofs, rn_ofs, rm_ofs, vec_size, vec_size);
+ } else {
+ gen_gvec_mla(size, rd_ofs, rn_ofs, rm_ofs, vec_size, vec_size);
+ }
return 0;
case NEON_3R_VTST_VCEQ:
--
2.20.1
^ permalink raw reply related [flat|nested] 18+ messages in thread* [PATCH v2 08/15] target/arm: Wrap vector cmtst/ushl/sshl GVecGen3 in GVecGen3Fn
2020-05-02 22:44 [PATCH v2 00/15] target/arm: partial vector cleanup Richard Henderson
` (6 preceding siblings ...)
2020-05-02 22:44 ` [PATCH v2 07/15] target/arm: Wrap vector mla/mls GVecGen3 in GVecGen3Fn Richard Henderson
@ 2020-05-02 22:44 ` Richard Henderson
2020-05-02 22:44 ` [PATCH v2 09/15] target/arm: Wrap vector uqadd/sqadd/uqsub/sqsub GVecGen4 " Richard Henderson
` (7 subsequent siblings)
15 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2020-05-02 22:44 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Provide a functional interface for the vector expansion.
This fits better with the existing set of helpers that
we provide for other operations.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate.h | 10 ++-
target/arm/translate-a64.c | 18 ++---
target/arm/translate.c | 159 ++++++++++++++++++++-----------------
3 files changed, 101 insertions(+), 86 deletions(-)
diff --git a/target/arm/translate.h b/target/arm/translate.h
index 4fbcdf1294..b3e47e7a7f 100644
--- a/target/arm/translate.h
+++ b/target/arm/translate.h
@@ -291,9 +291,13 @@ void gen_gvec_mla(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
void gen_gvec_mls(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz);
-extern const GVecGen3 cmtst_op[4];
-extern const GVecGen3 sshl_op[4];
-extern const GVecGen3 ushl_op[4];
+void gen_gvec_cmtst(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz);
+void gen_gvec_sshl(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz);
+void gen_gvec_ushl(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz);
+
extern const GVecGen4 uqadd_op[4];
extern const GVecGen4 sqadd_op[4];
extern const GVecGen4 uqsub_op[4];
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index 2b5ae4d43a..2be6ab541e 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -594,15 +594,6 @@ static void gen_gvec_fn4(DisasContext *s, bool is_q, int rd, int rn, int rm,
is_q ? 16 : 8, vec_full_reg_size(s));
}
-/* Expand a 3-operand AdvSIMD vector operation using an op descriptor. */
-static void gen_gvec_op3(DisasContext *s, bool is_q, int rd,
- int rn, int rm, const GVecGen3 *gvec_op)
-{
- tcg_gen_gvec_3(vec_full_reg_offset(s, rd), vec_full_reg_offset(s, rn),
- vec_full_reg_offset(s, rm), is_q ? 16 : 8,
- vec_full_reg_size(s), gvec_op);
-}
-
/* Expand a 3-operand operation using an out-of-line helper. */
static void gen_gvec_op3_ool(DisasContext *s, bool is_q, int rd,
int rn, int rm, int data, gen_helper_gvec_3 *fn)
@@ -11210,8 +11201,11 @@ static void disas_simd_3same_int(DisasContext *s, uint32_t insn)
(u ? uqsub_op : sqsub_op) + size);
return;
case 0x08: /* SSHL, USHL */
- gen_gvec_op3(s, is_q, rd, rn, rm,
- u ? &ushl_op[size] : &sshl_op[size]);
+ if (u) {
+ gen_gvec_fn3(s, is_q, rd, rn, rm, gen_gvec_ushl, size);
+ } else {
+ gen_gvec_fn3(s, is_q, rd, rn, rm, gen_gvec_sshl, size);
+ }
return;
case 0x0c: /* SMAX, UMAX */
if (u) {
@@ -11250,7 +11244,7 @@ static void disas_simd_3same_int(DisasContext *s, uint32_t insn)
return;
case 0x11:
if (!u) { /* CMTST */
- gen_gvec_op3(s, is_q, rd, rn, rm, &cmtst_op[size]);
+ gen_gvec_fn3(s, is_q, rd, rn, rm, gen_gvec_cmtst, size);
return;
}
/* else CMEQ */
diff --git a/target/arm/translate.c b/target/arm/translate.c
index da807242ff..e5aa78c88a 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -4878,27 +4878,31 @@ static void gen_cmtst_vec(unsigned vece, TCGv_vec d, TCGv_vec a, TCGv_vec b)
tcg_gen_cmp_vec(TCG_COND_NE, vece, d, d, a);
}
-static const TCGOpcode vecop_list_cmtst[] = { INDEX_op_cmp_vec, 0 };
-
-const GVecGen3 cmtst_op[4] = {
- { .fni4 = gen_helper_neon_tst_u8,
- .fniv = gen_cmtst_vec,
- .opt_opc = vecop_list_cmtst,
- .vece = MO_8 },
- { .fni4 = gen_helper_neon_tst_u16,
- .fniv = gen_cmtst_vec,
- .opt_opc = vecop_list_cmtst,
- .vece = MO_16 },
- { .fni4 = gen_cmtst_i32,
- .fniv = gen_cmtst_vec,
- .opt_opc = vecop_list_cmtst,
- .vece = MO_32 },
- { .fni8 = gen_cmtst_i64,
- .fniv = gen_cmtst_vec,
- .prefer_i64 = TCG_TARGET_REG_BITS == 64,
- .opt_opc = vecop_list_cmtst,
- .vece = MO_64 },
-};
+void gen_gvec_cmtst(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz)
+{
+ static const TCGOpcode vecop_list[] = { INDEX_op_cmp_vec, 0 };
+ static const GVecGen3 ops[4] = {
+ { .fni4 = gen_helper_neon_tst_u8,
+ .fniv = gen_cmtst_vec,
+ .opt_opc = vecop_list,
+ .vece = MO_8 },
+ { .fni4 = gen_helper_neon_tst_u16,
+ .fniv = gen_cmtst_vec,
+ .opt_opc = vecop_list,
+ .vece = MO_16 },
+ { .fni4 = gen_cmtst_i32,
+ .fniv = gen_cmtst_vec,
+ .opt_opc = vecop_list,
+ .vece = MO_32 },
+ { .fni8 = gen_cmtst_i64,
+ .fniv = gen_cmtst_vec,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ .opt_opc = vecop_list,
+ .vece = MO_64 },
+ };
+ tcg_gen_gvec_3(rd_ofs, rn_ofs, rm_ofs, opr_sz, max_sz, &ops[vece]);
+}
void gen_ushl_i32(TCGv_i32 dst, TCGv_i32 src, TCGv_i32 shift)
{
@@ -5016,29 +5020,33 @@ static void gen_ushl_vec(unsigned vece, TCGv_vec dst,
tcg_temp_free_vec(rsh);
}
-static const TCGOpcode ushl_list[] = {
- INDEX_op_neg_vec, INDEX_op_shlv_vec,
- INDEX_op_shrv_vec, INDEX_op_cmp_vec, 0
-};
-
-const GVecGen3 ushl_op[4] = {
- { .fniv = gen_ushl_vec,
- .fno = gen_helper_gvec_ushl_b,
- .opt_opc = ushl_list,
- .vece = MO_8 },
- { .fniv = gen_ushl_vec,
- .fno = gen_helper_gvec_ushl_h,
- .opt_opc = ushl_list,
- .vece = MO_16 },
- { .fni4 = gen_ushl_i32,
- .fniv = gen_ushl_vec,
- .opt_opc = ushl_list,
- .vece = MO_32 },
- { .fni8 = gen_ushl_i64,
- .fniv = gen_ushl_vec,
- .opt_opc = ushl_list,
- .vece = MO_64 },
-};
+void gen_gvec_ushl(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz)
+{
+ static const TCGOpcode vecop_list[] = {
+ INDEX_op_neg_vec, INDEX_op_shlv_vec,
+ INDEX_op_shrv_vec, INDEX_op_cmp_vec, 0
+ };
+ static const GVecGen3 ops[4] = {
+ { .fniv = gen_ushl_vec,
+ .fno = gen_helper_gvec_ushl_b,
+ .opt_opc = vecop_list,
+ .vece = MO_8 },
+ { .fniv = gen_ushl_vec,
+ .fno = gen_helper_gvec_ushl_h,
+ .opt_opc = vecop_list,
+ .vece = MO_16 },
+ { .fni4 = gen_ushl_i32,
+ .fniv = gen_ushl_vec,
+ .opt_opc = vecop_list,
+ .vece = MO_32 },
+ { .fni8 = gen_ushl_i64,
+ .fniv = gen_ushl_vec,
+ .opt_opc = vecop_list,
+ .vece = MO_64 },
+ };
+ tcg_gen_gvec_3(rd_ofs, rn_ofs, rm_ofs, opr_sz, max_sz, &ops[vece]);
+}
void gen_sshl_i32(TCGv_i32 dst, TCGv_i32 src, TCGv_i32 shift)
{
@@ -5150,29 +5158,33 @@ static void gen_sshl_vec(unsigned vece, TCGv_vec dst,
tcg_temp_free_vec(tmp);
}
-static const TCGOpcode sshl_list[] = {
- INDEX_op_neg_vec, INDEX_op_umin_vec, INDEX_op_shlv_vec,
- INDEX_op_sarv_vec, INDEX_op_cmp_vec, INDEX_op_cmpsel_vec, 0
-};
-
-const GVecGen3 sshl_op[4] = {
- { .fniv = gen_sshl_vec,
- .fno = gen_helper_gvec_sshl_b,
- .opt_opc = sshl_list,
- .vece = MO_8 },
- { .fniv = gen_sshl_vec,
- .fno = gen_helper_gvec_sshl_h,
- .opt_opc = sshl_list,
- .vece = MO_16 },
- { .fni4 = gen_sshl_i32,
- .fniv = gen_sshl_vec,
- .opt_opc = sshl_list,
- .vece = MO_32 },
- { .fni8 = gen_sshl_i64,
- .fniv = gen_sshl_vec,
- .opt_opc = sshl_list,
- .vece = MO_64 },
-};
+void gen_gvec_sshl(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz)
+{
+ static const TCGOpcode vecop_list[] = {
+ INDEX_op_neg_vec, INDEX_op_umin_vec, INDEX_op_shlv_vec,
+ INDEX_op_sarv_vec, INDEX_op_cmp_vec, INDEX_op_cmpsel_vec, 0
+ };
+ static const GVecGen3 ops[4] = {
+ { .fniv = gen_sshl_vec,
+ .fno = gen_helper_gvec_sshl_b,
+ .opt_opc = vecop_list,
+ .vece = MO_8 },
+ { .fniv = gen_sshl_vec,
+ .fno = gen_helper_gvec_sshl_h,
+ .opt_opc = vecop_list,
+ .vece = MO_16 },
+ { .fni4 = gen_sshl_i32,
+ .fniv = gen_sshl_vec,
+ .opt_opc = vecop_list,
+ .vece = MO_32 },
+ { .fni8 = gen_sshl_i64,
+ .fniv = gen_sshl_vec,
+ .opt_opc = vecop_list,
+ .vece = MO_64 },
+ };
+ tcg_gen_gvec_3(rd_ofs, rn_ofs, rm_ofs, opr_sz, max_sz, &ops[vece]);
+}
static void gen_uqadd_vec(unsigned vece, TCGv_vec t, TCGv_vec sat,
TCGv_vec a, TCGv_vec b)
@@ -5548,8 +5560,8 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
tcg_gen_gvec_cmp(TCG_COND_EQ, size, rd_ofs, rn_ofs, rm_ofs,
vec_size, vec_size);
} else { /* VTST */
- tcg_gen_gvec_3(rd_ofs, rn_ofs, rm_ofs,
- vec_size, vec_size, &cmtst_op[size]);
+ gen_gvec_cmtst(size, rd_ofs, rn_ofs, rm_ofs,
+ vec_size, vec_size);
}
return 0;
@@ -5584,8 +5596,13 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
case NEON_3R_VSHL:
/* Note the operation is vshl vd,vm,vn */
- tcg_gen_gvec_3(rd_ofs, rm_ofs, rn_ofs, vec_size, vec_size,
- u ? &ushl_op[size] : &sshl_op[size]);
+ if (u) {
+ gen_gvec_ushl(size, rd_ofs, rm_ofs, rn_ofs,
+ vec_size, vec_size);
+ } else {
+ gen_gvec_sshl(size, rd_ofs, rm_ofs, rn_ofs,
+ vec_size, vec_size);
+ }
return 0;
}
--
2.20.1
^ permalink raw reply related [flat|nested] 18+ messages in thread* [PATCH v2 09/15] target/arm: Wrap vector uqadd/sqadd/uqsub/sqsub GVecGen4 in GVecGen3Fn
2020-05-02 22:44 [PATCH v2 00/15] target/arm: partial vector cleanup Richard Henderson
` (7 preceding siblings ...)
2020-05-02 22:44 ` [PATCH v2 08/15] target/arm: Wrap vector cmtst/ushl/sshl " Richard Henderson
@ 2020-05-02 22:44 ` Richard Henderson
2020-05-02 22:44 ` [PATCH v2 10/15] target/arm: Remove fp_status from helper_{recpe, rsqrte}_u32 Richard Henderson
` (6 subsequent siblings)
15 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2020-05-02 22:44 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Provide a functional interface for the vector expansion.
This fits better with the existing set of helpers that
we provide for other operations.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate.h | 13 +-
target/arm/translate-a64.c | 22 ++--
target/arm/translate.c | 248 +++++++++++++++++++++----------------
3 files changed, 157 insertions(+), 126 deletions(-)
diff --git a/target/arm/translate.h b/target/arm/translate.h
index b3e47e7a7f..ada84d411d 100644
--- a/target/arm/translate.h
+++ b/target/arm/translate.h
@@ -298,16 +298,21 @@ void gen_gvec_sshl(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
void gen_gvec_ushl(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz);
-extern const GVecGen4 uqadd_op[4];
-extern const GVecGen4 sqadd_op[4];
-extern const GVecGen4 uqsub_op[4];
-extern const GVecGen4 sqsub_op[4];
void gen_cmtst_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b);
void gen_ushl_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b);
void gen_sshl_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b);
void gen_ushl_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b);
void gen_sshl_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b);
+void gen_gvec_uqadd_qc(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz);
+void gen_gvec_sqadd_qc(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz);
+void gen_gvec_uqsub_qc(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz);
+void gen_gvec_sqsub_qc(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz);
+
void gen_gvec_ssra(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
int64_t shift, uint32_t opr_sz, uint32_t max_sz);
void gen_gvec_usra(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index 2be6ab541e..eeaa92b9f1 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -11185,20 +11185,18 @@ static void disas_simd_3same_int(DisasContext *s, uint32_t insn)
switch (opcode) {
case 0x01: /* SQADD, UQADD */
- tcg_gen_gvec_4(vec_full_reg_offset(s, rd),
- offsetof(CPUARMState, vfp.qc),
- vec_full_reg_offset(s, rn),
- vec_full_reg_offset(s, rm),
- is_q ? 16 : 8, vec_full_reg_size(s),
- (u ? uqadd_op : sqadd_op) + size);
+ if (u) {
+ gen_gvec_fn3(s, is_q, rd, rn, rm, gen_gvec_uqadd_qc, size);
+ } else {
+ gen_gvec_fn3(s, is_q, rd, rn, rm, gen_gvec_sqadd_qc, size);
+ }
return;
case 0x05: /* SQSUB, UQSUB */
- tcg_gen_gvec_4(vec_full_reg_offset(s, rd),
- offsetof(CPUARMState, vfp.qc),
- vec_full_reg_offset(s, rn),
- vec_full_reg_offset(s, rm),
- is_q ? 16 : 8, vec_full_reg_size(s),
- (u ? uqsub_op : sqsub_op) + size);
+ if (u) {
+ gen_gvec_fn3(s, is_q, rd, rn, rm, gen_gvec_uqsub_qc, size);
+ } else {
+ gen_gvec_fn3(s, is_q, rd, rn, rm, gen_gvec_sqsub_qc, size);
+ }
return;
case 0x08: /* SSHL, USHL */
if (u) {
diff --git a/target/arm/translate.c b/target/arm/translate.c
index e5aa78c88a..8e6c6f7b00 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -5197,32 +5197,37 @@ static void gen_uqadd_vec(unsigned vece, TCGv_vec t, TCGv_vec sat,
tcg_temp_free_vec(x);
}
-static const TCGOpcode vecop_list_uqadd[] = {
- INDEX_op_usadd_vec, INDEX_op_cmp_vec, INDEX_op_add_vec, 0
-};
-
-const GVecGen4 uqadd_op[4] = {
- { .fniv = gen_uqadd_vec,
- .fno = gen_helper_gvec_uqadd_b,
- .write_aofs = true,
- .opt_opc = vecop_list_uqadd,
- .vece = MO_8 },
- { .fniv = gen_uqadd_vec,
- .fno = gen_helper_gvec_uqadd_h,
- .write_aofs = true,
- .opt_opc = vecop_list_uqadd,
- .vece = MO_16 },
- { .fniv = gen_uqadd_vec,
- .fno = gen_helper_gvec_uqadd_s,
- .write_aofs = true,
- .opt_opc = vecop_list_uqadd,
- .vece = MO_32 },
- { .fniv = gen_uqadd_vec,
- .fno = gen_helper_gvec_uqadd_d,
- .write_aofs = true,
- .opt_opc = vecop_list_uqadd,
- .vece = MO_64 },
-};
+void gen_gvec_uqadd_qc(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz)
+{
+ static const TCGOpcode vecop_list[] = {
+ INDEX_op_usadd_vec, INDEX_op_cmp_vec, INDEX_op_add_vec, 0
+ };
+ static const GVecGen4 ops[4] = {
+ { .fniv = gen_uqadd_vec,
+ .fno = gen_helper_gvec_uqadd_b,
+ .write_aofs = true,
+ .opt_opc = vecop_list,
+ .vece = MO_8 },
+ { .fniv = gen_uqadd_vec,
+ .fno = gen_helper_gvec_uqadd_h,
+ .write_aofs = true,
+ .opt_opc = vecop_list,
+ .vece = MO_16 },
+ { .fniv = gen_uqadd_vec,
+ .fno = gen_helper_gvec_uqadd_s,
+ .write_aofs = true,
+ .opt_opc = vecop_list,
+ .vece = MO_32 },
+ { .fniv = gen_uqadd_vec,
+ .fno = gen_helper_gvec_uqadd_d,
+ .write_aofs = true,
+ .opt_opc = vecop_list,
+ .vece = MO_64 },
+ };
+ tcg_gen_gvec_4(rd_ofs, offsetof(CPUARMState, vfp.qc),
+ rn_ofs, rm_ofs, opr_sz, max_sz, &ops[vece]);
+}
static void gen_sqadd_vec(unsigned vece, TCGv_vec t, TCGv_vec sat,
TCGv_vec a, TCGv_vec b)
@@ -5235,32 +5240,37 @@ static void gen_sqadd_vec(unsigned vece, TCGv_vec t, TCGv_vec sat,
tcg_temp_free_vec(x);
}
-static const TCGOpcode vecop_list_sqadd[] = {
- INDEX_op_ssadd_vec, INDEX_op_cmp_vec, INDEX_op_add_vec, 0
-};
-
-const GVecGen4 sqadd_op[4] = {
- { .fniv = gen_sqadd_vec,
- .fno = gen_helper_gvec_sqadd_b,
- .opt_opc = vecop_list_sqadd,
- .write_aofs = true,
- .vece = MO_8 },
- { .fniv = gen_sqadd_vec,
- .fno = gen_helper_gvec_sqadd_h,
- .opt_opc = vecop_list_sqadd,
- .write_aofs = true,
- .vece = MO_16 },
- { .fniv = gen_sqadd_vec,
- .fno = gen_helper_gvec_sqadd_s,
- .opt_opc = vecop_list_sqadd,
- .write_aofs = true,
- .vece = MO_32 },
- { .fniv = gen_sqadd_vec,
- .fno = gen_helper_gvec_sqadd_d,
- .opt_opc = vecop_list_sqadd,
- .write_aofs = true,
- .vece = MO_64 },
-};
+void gen_gvec_sqadd_qc(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz)
+{
+ static const TCGOpcode vecop_list[] = {
+ INDEX_op_ssadd_vec, INDEX_op_cmp_vec, INDEX_op_add_vec, 0
+ };
+ static const GVecGen4 ops[4] = {
+ { .fniv = gen_sqadd_vec,
+ .fno = gen_helper_gvec_sqadd_b,
+ .opt_opc = vecop_list,
+ .write_aofs = true,
+ .vece = MO_8 },
+ { .fniv = gen_sqadd_vec,
+ .fno = gen_helper_gvec_sqadd_h,
+ .opt_opc = vecop_list,
+ .write_aofs = true,
+ .vece = MO_16 },
+ { .fniv = gen_sqadd_vec,
+ .fno = gen_helper_gvec_sqadd_s,
+ .opt_opc = vecop_list,
+ .write_aofs = true,
+ .vece = MO_32 },
+ { .fniv = gen_sqadd_vec,
+ .fno = gen_helper_gvec_sqadd_d,
+ .opt_opc = vecop_list,
+ .write_aofs = true,
+ .vece = MO_64 },
+ };
+ tcg_gen_gvec_4(rd_ofs, offsetof(CPUARMState, vfp.qc),
+ rn_ofs, rm_ofs, opr_sz, max_sz, &ops[vece]);
+}
static void gen_uqsub_vec(unsigned vece, TCGv_vec t, TCGv_vec sat,
TCGv_vec a, TCGv_vec b)
@@ -5273,32 +5283,37 @@ static void gen_uqsub_vec(unsigned vece, TCGv_vec t, TCGv_vec sat,
tcg_temp_free_vec(x);
}
-static const TCGOpcode vecop_list_uqsub[] = {
- INDEX_op_ussub_vec, INDEX_op_cmp_vec, INDEX_op_sub_vec, 0
-};
-
-const GVecGen4 uqsub_op[4] = {
- { .fniv = gen_uqsub_vec,
- .fno = gen_helper_gvec_uqsub_b,
- .opt_opc = vecop_list_uqsub,
- .write_aofs = true,
- .vece = MO_8 },
- { .fniv = gen_uqsub_vec,
- .fno = gen_helper_gvec_uqsub_h,
- .opt_opc = vecop_list_uqsub,
- .write_aofs = true,
- .vece = MO_16 },
- { .fniv = gen_uqsub_vec,
- .fno = gen_helper_gvec_uqsub_s,
- .opt_opc = vecop_list_uqsub,
- .write_aofs = true,
- .vece = MO_32 },
- { .fniv = gen_uqsub_vec,
- .fno = gen_helper_gvec_uqsub_d,
- .opt_opc = vecop_list_uqsub,
- .write_aofs = true,
- .vece = MO_64 },
-};
+void gen_gvec_uqsub_qc(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz)
+{
+ static const TCGOpcode vecop_list[] = {
+ INDEX_op_ussub_vec, INDEX_op_cmp_vec, INDEX_op_sub_vec, 0
+ };
+ static const GVecGen4 ops[4] = {
+ { .fniv = gen_uqsub_vec,
+ .fno = gen_helper_gvec_uqsub_b,
+ .opt_opc = vecop_list,
+ .write_aofs = true,
+ .vece = MO_8 },
+ { .fniv = gen_uqsub_vec,
+ .fno = gen_helper_gvec_uqsub_h,
+ .opt_opc = vecop_list,
+ .write_aofs = true,
+ .vece = MO_16 },
+ { .fniv = gen_uqsub_vec,
+ .fno = gen_helper_gvec_uqsub_s,
+ .opt_opc = vecop_list,
+ .write_aofs = true,
+ .vece = MO_32 },
+ { .fniv = gen_uqsub_vec,
+ .fno = gen_helper_gvec_uqsub_d,
+ .opt_opc = vecop_list,
+ .write_aofs = true,
+ .vece = MO_64 },
+ };
+ tcg_gen_gvec_4(rd_ofs, offsetof(CPUARMState, vfp.qc),
+ rn_ofs, rm_ofs, opr_sz, max_sz, &ops[vece]);
+}
static void gen_sqsub_vec(unsigned vece, TCGv_vec t, TCGv_vec sat,
TCGv_vec a, TCGv_vec b)
@@ -5311,32 +5326,37 @@ static void gen_sqsub_vec(unsigned vece, TCGv_vec t, TCGv_vec sat,
tcg_temp_free_vec(x);
}
-static const TCGOpcode vecop_list_sqsub[] = {
- INDEX_op_sssub_vec, INDEX_op_cmp_vec, INDEX_op_sub_vec, 0
-};
-
-const GVecGen4 sqsub_op[4] = {
- { .fniv = gen_sqsub_vec,
- .fno = gen_helper_gvec_sqsub_b,
- .opt_opc = vecop_list_sqsub,
- .write_aofs = true,
- .vece = MO_8 },
- { .fniv = gen_sqsub_vec,
- .fno = gen_helper_gvec_sqsub_h,
- .opt_opc = vecop_list_sqsub,
- .write_aofs = true,
- .vece = MO_16 },
- { .fniv = gen_sqsub_vec,
- .fno = gen_helper_gvec_sqsub_s,
- .opt_opc = vecop_list_sqsub,
- .write_aofs = true,
- .vece = MO_32 },
- { .fniv = gen_sqsub_vec,
- .fno = gen_helper_gvec_sqsub_d,
- .opt_opc = vecop_list_sqsub,
- .write_aofs = true,
- .vece = MO_64 },
-};
+void gen_gvec_sqsub_qc(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz)
+{
+ static const TCGOpcode vecop_list[] = {
+ INDEX_op_sssub_vec, INDEX_op_cmp_vec, INDEX_op_sub_vec, 0
+ };
+ static const GVecGen4 ops[4] = {
+ { .fniv = gen_sqsub_vec,
+ .fno = gen_helper_gvec_sqsub_b,
+ .opt_opc = vecop_list,
+ .write_aofs = true,
+ .vece = MO_8 },
+ { .fniv = gen_sqsub_vec,
+ .fno = gen_helper_gvec_sqsub_h,
+ .opt_opc = vecop_list,
+ .write_aofs = true,
+ .vece = MO_16 },
+ { .fniv = gen_sqsub_vec,
+ .fno = gen_helper_gvec_sqsub_s,
+ .opt_opc = vecop_list,
+ .write_aofs = true,
+ .vece = MO_32 },
+ { .fniv = gen_sqsub_vec,
+ .fno = gen_helper_gvec_sqsub_d,
+ .opt_opc = vecop_list,
+ .write_aofs = true,
+ .vece = MO_64 },
+ };
+ tcg_gen_gvec_4(rd_ofs, offsetof(CPUARMState, vfp.qc),
+ rn_ofs, rm_ofs, opr_sz, max_sz, &ops[vece]);
+}
/* Translate a NEON data processing instruction. Return nonzero if the
instruction is invalid.
@@ -5522,15 +5542,23 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
return 0;
case NEON_3R_VQADD:
- tcg_gen_gvec_4(rd_ofs, offsetof(CPUARMState, vfp.qc),
- rn_ofs, rm_ofs, vec_size, vec_size,
- (u ? uqadd_op : sqadd_op) + size);
+ if (u) {
+ gen_gvec_uqadd_qc(size, rd_ofs, rn_ofs, rm_ofs,
+ vec_size, vec_size);
+ } else {
+ gen_gvec_sqadd_qc(size, rd_ofs, rn_ofs, rm_ofs,
+ vec_size, vec_size);
+ }
return 0;
case NEON_3R_VQSUB:
- tcg_gen_gvec_4(rd_ofs, offsetof(CPUARMState, vfp.qc),
- rn_ofs, rm_ofs, vec_size, vec_size,
- (u ? uqsub_op : sqsub_op) + size);
+ if (u) {
+ gen_gvec_uqsub_qc(size, rd_ofs, rn_ofs, rm_ofs,
+ vec_size, vec_size);
+ } else {
+ gen_gvec_sqsub_qc(size, rd_ofs, rn_ofs, rm_ofs,
+ vec_size, vec_size);
+ }
return 0;
case NEON_3R_VMUL: /* VMUL */
--
2.20.1
^ permalink raw reply related [flat|nested] 18+ messages in thread* [PATCH v2 10/15] target/arm: Remove fp_status from helper_{recpe, rsqrte}_u32
2020-05-02 22:44 [PATCH v2 00/15] target/arm: partial vector cleanup Richard Henderson
` (8 preceding siblings ...)
2020-05-02 22:44 ` [PATCH v2 09/15] target/arm: Wrap vector uqadd/sqadd/uqsub/sqsub GVecGen4 " Richard Henderson
@ 2020-05-02 22:44 ` Richard Henderson
2020-05-02 22:44 ` [PATCH v2 11/15] target/arm: Wrap vector qrdmla/qrdmls in GVecGen3Fn Richard Henderson
` (5 subsequent siblings)
15 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2020-05-02 22:44 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
These operations do not touch fp_status.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/helper.h | 4 ++--
target/arm/translate-a64.c | 5 ++---
target/arm/translate.c | 12 ++----------
target/arm/vfp_helper.c | 4 ++--
4 files changed, 8 insertions(+), 17 deletions(-)
diff --git a/target/arm/helper.h b/target/arm/helper.h
index 33c76192d2..aed3050965 100644
--- a/target/arm/helper.h
+++ b/target/arm/helper.h
@@ -211,8 +211,8 @@ DEF_HELPER_FLAGS_2(recpe_f64, TCG_CALL_NO_RWG, f64, f64, ptr)
DEF_HELPER_FLAGS_2(rsqrte_f16, TCG_CALL_NO_RWG, f16, f16, ptr)
DEF_HELPER_FLAGS_2(rsqrte_f32, TCG_CALL_NO_RWG, f32, f32, ptr)
DEF_HELPER_FLAGS_2(rsqrte_f64, TCG_CALL_NO_RWG, f64, f64, ptr)
-DEF_HELPER_2(recpe_u32, i32, i32, ptr)
-DEF_HELPER_FLAGS_2(rsqrte_u32, TCG_CALL_NO_RWG, i32, i32, ptr)
+DEF_HELPER_FLAGS_1(recpe_u32, TCG_CALL_NO_RWG, i32, i32)
+DEF_HELPER_FLAGS_1(rsqrte_u32, TCG_CALL_NO_RWG, i32, i32)
DEF_HELPER_FLAGS_4(neon_tbl, TCG_CALL_NO_RWG, i32, i32, i32, ptr, i32)
DEF_HELPER_3(shl_cc, i32, env, i32, i32)
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index eeaa92b9f1..4d5cdcef2f 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -9716,7 +9716,7 @@ static void handle_2misc_reciprocal(DisasContext *s, int opcode,
switch (opcode) {
case 0x3c: /* URECPE */
- gen_helper_recpe_u32(tcg_res, tcg_op, fpst);
+ gen_helper_recpe_u32(tcg_res, tcg_op);
break;
case 0x3d: /* FRECPE */
gen_helper_recpe_f32(tcg_res, tcg_op, fpst);
@@ -12261,7 +12261,6 @@ static void disas_simd_two_reg_misc(DisasContext *s, uint32_t insn)
unallocated_encoding(s);
return;
}
- need_fpstatus = true;
break;
case 0x1e: /* FRINT32Z */
case 0x1f: /* FRINT64Z */
@@ -12429,7 +12428,7 @@ static void disas_simd_two_reg_misc(DisasContext *s, uint32_t insn)
gen_helper_rints_exact(tcg_res, tcg_op, tcg_fpstatus);
break;
case 0x7c: /* URSQRTE */
- gen_helper_rsqrte_u32(tcg_res, tcg_op, tcg_fpstatus);
+ gen_helper_rsqrte_u32(tcg_res, tcg_op);
break;
case 0x1e: /* FRINT32Z */
case 0x5e: /* FRINT32X */
diff --git a/target/arm/translate.c b/target/arm/translate.c
index 8e6c6f7b00..a5bb4b0040 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -7265,19 +7265,11 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
break;
}
case NEON_2RM_VRECPE:
- {
- TCGv_ptr fpstatus = get_fpstatus_ptr(1);
- gen_helper_recpe_u32(tmp, tmp, fpstatus);
- tcg_temp_free_ptr(fpstatus);
+ gen_helper_recpe_u32(tmp, tmp);
break;
- }
case NEON_2RM_VRSQRTE:
- {
- TCGv_ptr fpstatus = get_fpstatus_ptr(1);
- gen_helper_rsqrte_u32(tmp, tmp, fpstatus);
- tcg_temp_free_ptr(fpstatus);
+ gen_helper_rsqrte_u32(tmp, tmp);
break;
- }
case NEON_2RM_VRECPE_F:
{
TCGv_ptr fpstatus = get_fpstatus_ptr(1);
diff --git a/target/arm/vfp_helper.c b/target/arm/vfp_helper.c
index 930d6e747f..a792661166 100644
--- a/target/arm/vfp_helper.c
+++ b/target/arm/vfp_helper.c
@@ -1023,7 +1023,7 @@ float64 HELPER(rsqrte_f64)(float64 input, void *fpstp)
return make_float64(val);
}
-uint32_t HELPER(recpe_u32)(uint32_t a, void *fpstp)
+uint32_t HELPER(recpe_u32)(uint32_t a)
{
/* float_status *s = fpstp; */
int input, estimate;
@@ -1038,7 +1038,7 @@ uint32_t HELPER(recpe_u32)(uint32_t a, void *fpstp)
return deposit32(0, (32 - 9), 9, estimate);
}
-uint32_t HELPER(rsqrte_u32)(uint32_t a, void *fpstp)
+uint32_t HELPER(rsqrte_u32)(uint32_t a)
{
int estimate;
--
2.20.1
^ permalink raw reply related [flat|nested] 18+ messages in thread* [PATCH v2 11/15] target/arm: Wrap vector qrdmla/qrdmls in GVecGen3Fn
2020-05-02 22:44 [PATCH v2 00/15] target/arm: partial vector cleanup Richard Henderson
` (9 preceding siblings ...)
2020-05-02 22:44 ` [PATCH v2 10/15] target/arm: Remove fp_status from helper_{recpe, rsqrte}_u32 Richard Henderson
@ 2020-05-02 22:44 ` Richard Henderson
2020-05-02 22:45 ` [PATCH v2 12/15] target/arm: Pass pointer to qc to qrdmla/qrdmls Richard Henderson
` (4 subsequent siblings)
15 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2020-05-02 22:44 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Provide a functional interface for the vector expansion.
This fits better with the existing set of helpers that
we provide for other operations.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate.h | 5 ++++
target/arm/translate-a64.c | 34 ++----------------------
target/arm/translate.c | 54 +++++++++++++++++++-------------------
3 files changed, 34 insertions(+), 59 deletions(-)
diff --git a/target/arm/translate.h b/target/arm/translate.h
index ada84d411d..76cd3d31c7 100644
--- a/target/arm/translate.h
+++ b/target/arm/translate.h
@@ -332,6 +332,11 @@ void gen_gvec_sri(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
void gen_gvec_sli(unsigned vece, uint32_t rd_ofs, uint32_t rm_ofs,
int64_t shift, uint32_t opr_sz, uint32_t max_sz);
+void gen_gvec_sqrdmlah_qc(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz);
+void gen_gvec_sqrdmlsh_qc(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz);
+
/*
* Forward to the isar_feature_* tests given a DisasContext pointer.
*/
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index 4d5cdcef2f..1821a8e09d 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -604,18 +604,6 @@ static void gen_gvec_op3_ool(DisasContext *s, bool is_q, int rd,
is_q ? 16 : 8, vec_full_reg_size(s), data, fn);
}
-/* Expand a 3-operand + env pointer operation using
- * an out-of-line helper.
- */
-static void gen_gvec_op3_env(DisasContext *s, bool is_q, int rd,
- int rn, int rm, gen_helper_gvec_3_ptr *fn)
-{
- tcg_gen_gvec_3_ptr(vec_full_reg_offset(s, rd),
- vec_full_reg_offset(s, rn),
- vec_full_reg_offset(s, rm), cpu_env,
- is_q ? 16 : 8, vec_full_reg_size(s), 0, fn);
-}
-
/* Expand a 3-operand + fpstatus pointer + simd data value operation using
* an out-of-line helper.
*/
@@ -11710,29 +11698,11 @@ static void disas_simd_three_reg_same_extra(DisasContext *s, uint32_t insn)
switch (opcode) {
case 0x0: /* SQRDMLAH (vector) */
- switch (size) {
- case 1:
- gen_gvec_op3_env(s, is_q, rd, rn, rm, gen_helper_gvec_qrdmlah_s16);
- break;
- case 2:
- gen_gvec_op3_env(s, is_q, rd, rn, rm, gen_helper_gvec_qrdmlah_s32);
- break;
- default:
- g_assert_not_reached();
- }
+ gen_gvec_fn3(s, is_q, rd, rn, rm, gen_gvec_sqrdmlah_qc, size);
return;
case 0x1: /* SQRDMLSH (vector) */
- switch (size) {
- case 1:
- gen_gvec_op3_env(s, is_q, rd, rn, rm, gen_helper_gvec_qrdmlsh_s16);
- break;
- case 2:
- gen_gvec_op3_env(s, is_q, rd, rn, rm, gen_helper_gvec_qrdmlsh_s32);
- break;
- default:
- g_assert_not_reached();
- }
+ gen_gvec_fn3(s, is_q, rd, rn, rm, gen_gvec_sqrdmlsh_qc, size);
return;
case 0x2: /* SDOT / UDOT */
diff --git a/target/arm/translate.c b/target/arm/translate.c
index a5bb4b0040..8d20726dcf 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -3901,20 +3901,26 @@ static const uint8_t neon_2rm_sizes[] = {
[NEON_2RM_VCVT_UF] = 0x4,
};
-
-/* Expand v8.1 simd helper. */
-static int do_v81_helper(DisasContext *s, gen_helper_gvec_3_ptr *fn,
- int q, int rd, int rn, int rm)
+void gen_gvec_sqrdmlah_qc(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz)
{
- if (dc_isar_feature(aa32_rdm, s)) {
- int opr_sz = (1 + q) * 8;
- tcg_gen_gvec_3_ptr(vfp_reg_offset(1, rd),
- vfp_reg_offset(1, rn),
- vfp_reg_offset(1, rm), cpu_env,
- opr_sz, opr_sz, 0, fn);
- return 0;
- }
- return 1;
+ static gen_helper_gvec_3_ptr * const fns[2] = {
+ gen_helper_gvec_qrdmlah_s16, gen_helper_gvec_qrdmlah_s32
+ };
+ tcg_debug_assert(vece >= 1 && vece <= 2);
+ tcg_gen_gvec_3_ptr(rd_ofs, rn_ofs, rm_ofs, cpu_env,
+ opr_sz, max_sz, 0, fns[vece - 1]);
+}
+
+void gen_gvec_sqrdmlsh_qc(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz)
+{
+ static gen_helper_gvec_3_ptr * const fns[2] = {
+ gen_helper_gvec_qrdmlsh_s16, gen_helper_gvec_qrdmlsh_s32
+ };
+ tcg_debug_assert(vece >= 1 && vece <= 2);
+ tcg_gen_gvec_3_ptr(rd_ofs, rn_ofs, rm_ofs, cpu_env,
+ opr_sz, max_sz, 0, fns[vece - 1]);
}
#define GEN_CMP0(NAME, COND) \
@@ -5465,13 +5471,10 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
break; /* VPADD */
}
/* VQRDMLAH */
- switch (size) {
- case 1:
- return do_v81_helper(s, gen_helper_gvec_qrdmlah_s16,
- q, rd, rn, rm);
- case 2:
- return do_v81_helper(s, gen_helper_gvec_qrdmlah_s32,
- q, rd, rn, rm);
+ if (dc_isar_feature(aa32_rdm, s) && (size == 1 || size == 2)) {
+ gen_gvec_sqrdmlah_qc(size, rd_ofs, rn_ofs, rm_ofs,
+ vec_size, vec_size);
+ return 0;
}
return 1;
@@ -5484,13 +5487,10 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
break;
}
/* VQRDMLSH */
- switch (size) {
- case 1:
- return do_v81_helper(s, gen_helper_gvec_qrdmlsh_s16,
- q, rd, rn, rm);
- case 2:
- return do_v81_helper(s, gen_helper_gvec_qrdmlsh_s32,
- q, rd, rn, rm);
+ if (dc_isar_feature(aa32_rdm, s) && (size == 1 || size == 2)) {
+ gen_gvec_sqrdmlsh_qc(size, rd_ofs, rn_ofs, rm_ofs,
+ vec_size, vec_size);
+ return 0;
}
return 1;
--
2.20.1
^ permalink raw reply related [flat|nested] 18+ messages in thread* [PATCH v2 12/15] target/arm: Pass pointer to qc to qrdmla/qrdmls
2020-05-02 22:44 [PATCH v2 00/15] target/arm: partial vector cleanup Richard Henderson
` (10 preceding siblings ...)
2020-05-02 22:44 ` [PATCH v2 11/15] target/arm: Wrap vector qrdmla/qrdmls in GVecGen3Fn Richard Henderson
@ 2020-05-02 22:45 ` Richard Henderson
2020-05-02 22:45 ` [PATCH v2 13/15] target/arm: Clear tail in gvec_fmul_idx_*, gvec_fmla_idx_* Richard Henderson
` (3 subsequent siblings)
15 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2020-05-02 22:45 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Pass a pointer directly to env->vfp.qc[0], rather than env.
This will allow SVE2, which does not modify QC, to pass a
pointer to dummy storage.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate.c | 18 ++++++++---
target/arm/vec_helper.c | 70 +++++++++++++++++++++++------------------
2 files changed, 54 insertions(+), 34 deletions(-)
diff --git a/target/arm/translate.c b/target/arm/translate.c
index 8d20726dcf..532768d65f 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -3901,6 +3901,18 @@ static const uint8_t neon_2rm_sizes[] = {
[NEON_2RM_VCVT_UF] = 0x4,
};
+static void gen_gvec_fn3_qc(uint32_t rd_ofs, uint32_t rn_ofs, uint32_t rm_ofs,
+ uint32_t opr_sz, uint32_t max_sz,
+ gen_helper_gvec_3_ptr *fn)
+{
+ TCGv_ptr qc_ptr = tcg_temp_new_ptr();
+
+ tcg_gen_addi_ptr(qc_ptr, cpu_env, offsetof(CPUARMState, vfp.qc));
+ tcg_gen_gvec_3_ptr(rd_ofs, rn_ofs, rm_ofs, qc_ptr,
+ opr_sz, max_sz, 0, fn);
+ tcg_temp_free_ptr(qc_ptr);
+}
+
void gen_gvec_sqrdmlah_qc(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz)
{
@@ -3908,8 +3920,7 @@ void gen_gvec_sqrdmlah_qc(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
gen_helper_gvec_qrdmlah_s16, gen_helper_gvec_qrdmlah_s32
};
tcg_debug_assert(vece >= 1 && vece <= 2);
- tcg_gen_gvec_3_ptr(rd_ofs, rn_ofs, rm_ofs, cpu_env,
- opr_sz, max_sz, 0, fns[vece - 1]);
+ gen_gvec_fn3_qc(rd_ofs, rn_ofs, rm_ofs, opr_sz, max_sz, fns[vece - 1]);
}
void gen_gvec_sqrdmlsh_qc(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
@@ -3919,8 +3930,7 @@ void gen_gvec_sqrdmlsh_qc(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
gen_helper_gvec_qrdmlsh_s16, gen_helper_gvec_qrdmlsh_s32
};
tcg_debug_assert(vece >= 1 && vece <= 2);
- tcg_gen_gvec_3_ptr(rd_ofs, rn_ofs, rm_ofs, cpu_env,
- opr_sz, max_sz, 0, fns[vece - 1]);
+ gen_gvec_fn3_qc(rd_ofs, rn_ofs, rm_ofs, opr_sz, max_sz, fns[vece - 1]);
}
#define GEN_CMP0(NAME, COND) \
diff --git a/target/arm/vec_helper.c b/target/arm/vec_helper.c
index 096fea67ef..6aa2ca0827 100644
--- a/target/arm/vec_helper.c
+++ b/target/arm/vec_helper.c
@@ -36,8 +36,6 @@
#define H4(x) (x)
#endif
-#define SET_QC() env->vfp.qc[0] = 1
-
static void clear_tail(void *vd, uintptr_t opr_sz, uintptr_t max_sz)
{
uint64_t *d = vd + opr_sz;
@@ -49,8 +47,8 @@ static void clear_tail(void *vd, uintptr_t opr_sz, uintptr_t max_sz)
}
/* Signed saturating rounding doubling multiply-accumulate high half, 16-bit */
-static uint16_t inl_qrdmlah_s16(CPUARMState *env, int16_t src1,
- int16_t src2, int16_t src3)
+static int16_t inl_qrdmlah_s16(int16_t src1, int16_t src2,
+ int16_t src3, uint32_t *sat)
{
/* Simplify:
* = ((a3 << 16) + ((e1 * e2) << 1) + (1 << 15)) >> 16
@@ -60,7 +58,7 @@ static uint16_t inl_qrdmlah_s16(CPUARMState *env, int16_t src1,
ret = ((int32_t)src3 << 15) + ret + (1 << 14);
ret >>= 15;
if (ret != (int16_t)ret) {
- SET_QC();
+ *sat = 1;
ret = (ret < 0 ? -0x8000 : 0x7fff);
}
return ret;
@@ -69,30 +67,30 @@ static uint16_t inl_qrdmlah_s16(CPUARMState *env, int16_t src1,
uint32_t HELPER(neon_qrdmlah_s16)(CPUARMState *env, uint32_t src1,
uint32_t src2, uint32_t src3)
{
- uint16_t e1 = inl_qrdmlah_s16(env, src1, src2, src3);
- uint16_t e2 = inl_qrdmlah_s16(env, src1 >> 16, src2 >> 16, src3 >> 16);
+ uint32_t *sat = &env->vfp.qc[0];
+ uint16_t e1 = inl_qrdmlah_s16(src1, src2, src3, sat);
+ uint16_t e2 = inl_qrdmlah_s16(src1 >> 16, src2 >> 16, src3 >> 16, sat);
return deposit32(e1, 16, 16, e2);
}
void HELPER(gvec_qrdmlah_s16)(void *vd, void *vn, void *vm,
- void *ve, uint32_t desc)
+ void *vq, uint32_t desc)
{
uintptr_t opr_sz = simd_oprsz(desc);
int16_t *d = vd;
int16_t *n = vn;
int16_t *m = vm;
- CPUARMState *env = ve;
uintptr_t i;
for (i = 0; i < opr_sz / 2; ++i) {
- d[i] = inl_qrdmlah_s16(env, n[i], m[i], d[i]);
+ d[i] = inl_qrdmlah_s16(n[i], m[i], d[i], vq);
}
clear_tail(d, opr_sz, simd_maxsz(desc));
}
/* Signed saturating rounding doubling multiply-subtract high half, 16-bit */
-static uint16_t inl_qrdmlsh_s16(CPUARMState *env, int16_t src1,
- int16_t src2, int16_t src3)
+static int16_t inl_qrdmlsh_s16(int16_t src1, int16_t src2,
+ int16_t src3, uint32_t *sat)
{
/* Similarly, using subtraction:
* = ((a3 << 16) - ((e1 * e2) << 1) + (1 << 15)) >> 16
@@ -102,7 +100,7 @@ static uint16_t inl_qrdmlsh_s16(CPUARMState *env, int16_t src1,
ret = ((int32_t)src3 << 15) - ret + (1 << 14);
ret >>= 15;
if (ret != (int16_t)ret) {
- SET_QC();
+ *sat = 1;
ret = (ret < 0 ? -0x8000 : 0x7fff);
}
return ret;
@@ -111,85 +109,97 @@ static uint16_t inl_qrdmlsh_s16(CPUARMState *env, int16_t src1,
uint32_t HELPER(neon_qrdmlsh_s16)(CPUARMState *env, uint32_t src1,
uint32_t src2, uint32_t src3)
{
- uint16_t e1 = inl_qrdmlsh_s16(env, src1, src2, src3);
- uint16_t e2 = inl_qrdmlsh_s16(env, src1 >> 16, src2 >> 16, src3 >> 16);
+ uint32_t *sat = &env->vfp.qc[0];
+ uint16_t e1 = inl_qrdmlsh_s16(src1, src2, src3, sat);
+ uint16_t e2 = inl_qrdmlsh_s16(src1 >> 16, src2 >> 16, src3 >> 16, sat);
return deposit32(e1, 16, 16, e2);
}
void HELPER(gvec_qrdmlsh_s16)(void *vd, void *vn, void *vm,
- void *ve, uint32_t desc)
+ void *vq, uint32_t desc)
{
uintptr_t opr_sz = simd_oprsz(desc);
int16_t *d = vd;
int16_t *n = vn;
int16_t *m = vm;
- CPUARMState *env = ve;
uintptr_t i;
for (i = 0; i < opr_sz / 2; ++i) {
- d[i] = inl_qrdmlsh_s16(env, n[i], m[i], d[i]);
+ d[i] = inl_qrdmlsh_s16(n[i], m[i], d[i], vq);
}
clear_tail(d, opr_sz, simd_maxsz(desc));
}
/* Signed saturating rounding doubling multiply-accumulate high half, 32-bit */
-uint32_t HELPER(neon_qrdmlah_s32)(CPUARMState *env, int32_t src1,
- int32_t src2, int32_t src3)
+static int32_t inl_qrdmlah_s32(int32_t src1, int32_t src2,
+ int32_t src3, uint32_t *sat)
{
/* Simplify similarly to int_qrdmlah_s16 above. */
int64_t ret = (int64_t)src1 * src2;
ret = ((int64_t)src3 << 31) + ret + (1 << 30);
ret >>= 31;
if (ret != (int32_t)ret) {
- SET_QC();
+ *sat = 1;
ret = (ret < 0 ? INT32_MIN : INT32_MAX);
}
return ret;
}
+uint32_t HELPER(neon_qrdmlah_s32)(CPUARMState *env, int32_t src1,
+ int32_t src2, int32_t src3)
+{
+ uint32_t *sat = &env->vfp.qc[0];
+ return inl_qrdmlah_s32(src1, src2, src3, sat);
+}
+
void HELPER(gvec_qrdmlah_s32)(void *vd, void *vn, void *vm,
- void *ve, uint32_t desc)
+ void *vq, uint32_t desc)
{
uintptr_t opr_sz = simd_oprsz(desc);
int32_t *d = vd;
int32_t *n = vn;
int32_t *m = vm;
- CPUARMState *env = ve;
uintptr_t i;
for (i = 0; i < opr_sz / 4; ++i) {
- d[i] = helper_neon_qrdmlah_s32(env, n[i], m[i], d[i]);
+ d[i] = inl_qrdmlah_s32(n[i], m[i], d[i], vq);
}
clear_tail(d, opr_sz, simd_maxsz(desc));
}
/* Signed saturating rounding doubling multiply-subtract high half, 32-bit */
-uint32_t HELPER(neon_qrdmlsh_s32)(CPUARMState *env, int32_t src1,
- int32_t src2, int32_t src3)
+static int32_t inl_qrdmlsh_s32(int32_t src1, int32_t src2,
+ int32_t src3, uint32_t *sat)
{
/* Simplify similarly to int_qrdmlsh_s16 above. */
int64_t ret = (int64_t)src1 * src2;
ret = ((int64_t)src3 << 31) - ret + (1 << 30);
ret >>= 31;
if (ret != (int32_t)ret) {
- SET_QC();
+ *sat = 1;
ret = (ret < 0 ? INT32_MIN : INT32_MAX);
}
return ret;
}
+uint32_t HELPER(neon_qrdmlsh_s32)(CPUARMState *env, int32_t src1,
+ int32_t src2, int32_t src3)
+{
+ uint32_t *sat = &env->vfp.qc[0];
+ return inl_qrdmlsh_s32(src1, src2, src3, sat);
+}
+
void HELPER(gvec_qrdmlsh_s32)(void *vd, void *vn, void *vm,
- void *ve, uint32_t desc)
+ void *vq, uint32_t desc)
{
uintptr_t opr_sz = simd_oprsz(desc);
int32_t *d = vd;
int32_t *n = vn;
int32_t *m = vm;
- CPUARMState *env = ve;
uintptr_t i;
for (i = 0; i < opr_sz / 4; ++i) {
- d[i] = helper_neon_qrdmlsh_s32(env, n[i], m[i], d[i]);
+ d[i] = inl_qrdmlsh_s32(n[i], m[i], d[i], vq);
}
clear_tail(d, opr_sz, simd_maxsz(desc));
}
--
2.20.1
^ permalink raw reply related [flat|nested] 18+ messages in thread* [PATCH v2 13/15] target/arm: Clear tail in gvec_fmul_idx_*, gvec_fmla_idx_*
2020-05-02 22:44 [PATCH v2 00/15] target/arm: partial vector cleanup Richard Henderson
` (11 preceding siblings ...)
2020-05-02 22:45 ` [PATCH v2 12/15] target/arm: Pass pointer to qc to qrdmla/qrdmls Richard Henderson
@ 2020-05-02 22:45 ` Richard Henderson
2020-05-02 22:45 ` [PATCH v2 14/15] target/arm: Vectorize SABD/UABD Richard Henderson
` (2 subsequent siblings)
15 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2020-05-02 22:45 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Must clear the tail for AdvSIMD when SVE is enabled.
Fixes: ca40a6e6e39
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/vec_helper.c | 2 ++
1 file changed, 2 insertions(+)
diff --git a/target/arm/vec_helper.c b/target/arm/vec_helper.c
index 6aa2ca0827..a483841add 100644
--- a/target/arm/vec_helper.c
+++ b/target/arm/vec_helper.c
@@ -747,6 +747,7 @@ void HELPER(NAME)(void *vd, void *vn, void *vm, void *stat, uint32_t desc) \
d[i + j] = TYPE##_mul(n[i + j], mm, stat); \
} \
} \
+ clear_tail(d, oprsz, simd_maxsz(desc)); \
}
DO_MUL_IDX(gvec_fmul_idx_h, float16, H2)
@@ -771,6 +772,7 @@ void HELPER(NAME)(void *vd, void *vn, void *vm, void *va, \
mm, a[i + j], 0, stat); \
} \
} \
+ clear_tail(d, oprsz, simd_maxsz(desc)); \
}
DO_FMLA_IDX(gvec_fmla_idx_h, float16, H2)
--
2.20.1
^ permalink raw reply related [flat|nested] 18+ messages in thread* [PATCH v2 14/15] target/arm: Vectorize SABD/UABD
2020-05-02 22:44 [PATCH v2 00/15] target/arm: partial vector cleanup Richard Henderson
` (12 preceding siblings ...)
2020-05-02 22:45 ` [PATCH v2 13/15] target/arm: Clear tail in gvec_fmul_idx_*, gvec_fmla_idx_* Richard Henderson
@ 2020-05-02 22:45 ` Richard Henderson
2020-05-02 22:45 ` [PATCH v2 15/15] target/arm: Vectorize SABA/UABA Richard Henderson
2020-05-05 10:38 ` [PATCH v2 00/15] target/arm: partial vector cleanup Peter Maydell
15 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2020-05-02 22:45 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Include 64-bit element size in preparation for SVE2.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/helper.h | 10 +++
target/arm/translate.h | 5 ++
target/arm/translate-a64.c | 8 ++-
target/arm/translate.c | 133 ++++++++++++++++++++++++++++++++++++-
target/arm/vec_helper.c | 24 +++++++
5 files changed, 176 insertions(+), 4 deletions(-)
diff --git a/target/arm/helper.h b/target/arm/helper.h
index aed3050965..4678d3a6f4 100644
--- a/target/arm/helper.h
+++ b/target/arm/helper.h
@@ -731,6 +731,16 @@ DEF_HELPER_FLAGS_3(gvec_sli_h, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
DEF_HELPER_FLAGS_3(gvec_sli_s, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
DEF_HELPER_FLAGS_3(gvec_sli_d, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_sabd_b, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_sabd_h, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_sabd_s, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_sabd_d, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+
+DEF_HELPER_FLAGS_4(gvec_uabd_b, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_uabd_h, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_uabd_s, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_uabd_d, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+
#ifdef TARGET_AARCH64
#include "helper-a64.h"
#include "helper-sve.h"
diff --git a/target/arm/translate.h b/target/arm/translate.h
index 76cd3d31c7..d4c4111a5c 100644
--- a/target/arm/translate.h
+++ b/target/arm/translate.h
@@ -337,6 +337,11 @@ void gen_gvec_sqrdmlah_qc(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
void gen_gvec_sqrdmlsh_qc(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz);
+void gen_gvec_sabd(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz);
+void gen_gvec_uabd(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz);
+
/*
* Forward to the isar_feature_* tests given a DisasContext pointer.
*/
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index 1821a8e09d..38f72bf550 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -11207,6 +11207,13 @@ static void disas_simd_3same_int(DisasContext *s, uint32_t insn)
gen_gvec_fn3(s, is_q, rd, rn, rm, tcg_gen_gvec_smin, size);
}
return;
+ case 0xe: /* SABD, UABD */
+ if (u) {
+ gen_gvec_fn3(s, is_q, rd, rn, rm, gen_gvec_uabd, size);
+ } else {
+ gen_gvec_fn3(s, is_q, rd, rn, rm, gen_gvec_sabd, size);
+ }
+ return;
case 0x10: /* ADD, SUB */
if (u) {
gen_gvec_fn3(s, is_q, rd, rn, rm, tcg_gen_gvec_sub, size);
@@ -11339,7 +11346,6 @@ static void disas_simd_3same_int(DisasContext *s, uint32_t insn)
genenvfn = fns[size][u];
break;
}
- case 0xe: /* SABD, UABD */
case 0xf: /* SABA, UABA */
{
static NeonGenTwoOpFn * const fns[3][2] = {
diff --git a/target/arm/translate.c b/target/arm/translate.c
index 532768d65f..e0c4de2898 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -5374,6 +5374,126 @@ void gen_gvec_sqsub_qc(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
rn_ofs, rm_ofs, opr_sz, max_sz, &ops[vece]);
}
+static void gen_sabd_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
+{
+ TCGv_i32 t = tcg_temp_new_i32();
+
+ tcg_gen_sub_i32(t, a, b);
+ tcg_gen_sub_i32(d, b, a);
+ tcg_gen_movcond_i32(TCG_COND_LT, d, a, b, d, t);
+ tcg_temp_free_i32(t);
+}
+
+static void gen_sabd_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
+{
+ TCGv_i64 t = tcg_temp_new_i64();
+
+ tcg_gen_sub_i64(t, a, b);
+ tcg_gen_sub_i64(d, b, a);
+ tcg_gen_movcond_i64(TCG_COND_LT, d, a, b, d, t);
+ tcg_temp_free_i64(t);
+}
+
+static void gen_sabd_vec(unsigned vece, TCGv_vec d, TCGv_vec a, TCGv_vec b)
+{
+ TCGv_vec t = tcg_temp_new_vec_matching(d);
+
+ tcg_gen_smin_vec(vece, t, a, b);
+ tcg_gen_smax_vec(vece, d, a, b);
+ tcg_gen_sub_vec(vece, d, d, t);
+ tcg_temp_free_vec(t);
+}
+
+void gen_gvec_sabd(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz)
+{
+ static const TCGOpcode vecop_list[] = {
+ INDEX_op_sub_vec, INDEX_op_smin_vec, INDEX_op_smax_vec, 0
+ };
+ static const GVecGen3 ops[4] = {
+ { .fniv = gen_sabd_vec,
+ .fno = gen_helper_gvec_sabd_b,
+ .opt_opc = vecop_list,
+ .vece = MO_8 },
+ { .fniv = gen_sabd_vec,
+ .fno = gen_helper_gvec_sabd_h,
+ .opt_opc = vecop_list,
+ .vece = MO_16 },
+ { .fni4 = gen_sabd_i32,
+ .fniv = gen_sabd_vec,
+ .fno = gen_helper_gvec_sabd_s,
+ .opt_opc = vecop_list,
+ .vece = MO_32 },
+ { .fni8 = gen_sabd_i64,
+ .fniv = gen_sabd_vec,
+ .fno = gen_helper_gvec_sabd_d,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ .opt_opc = vecop_list,
+ .vece = MO_64 },
+ };
+ tcg_gen_gvec_3(rd_ofs, rn_ofs, rm_ofs, opr_sz, max_sz, &ops[vece]);
+}
+
+static void gen_uabd_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
+{
+ TCGv_i32 t = tcg_temp_new_i32();
+
+ tcg_gen_sub_i32(t, a, b);
+ tcg_gen_sub_i32(d, b, a);
+ tcg_gen_movcond_i32(TCG_COND_LTU, d, a, b, d, t);
+ tcg_temp_free_i32(t);
+}
+
+static void gen_uabd_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
+{
+ TCGv_i64 t = tcg_temp_new_i64();
+
+ tcg_gen_sub_i64(t, a, b);
+ tcg_gen_sub_i64(d, b, a);
+ tcg_gen_movcond_i64(TCG_COND_LTU, d, a, b, d, t);
+ tcg_temp_free_i64(t);
+}
+
+static void gen_uabd_vec(unsigned vece, TCGv_vec d, TCGv_vec a, TCGv_vec b)
+{
+ TCGv_vec t = tcg_temp_new_vec_matching(d);
+
+ tcg_gen_umin_vec(vece, t, a, b);
+ tcg_gen_umax_vec(vece, d, a, b);
+ tcg_gen_sub_vec(vece, d, d, t);
+ tcg_temp_free_vec(t);
+}
+
+void gen_gvec_uabd(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz)
+{
+ static const TCGOpcode vecop_list[] = {
+ INDEX_op_sub_vec, INDEX_op_umin_vec, INDEX_op_umax_vec, 0
+ };
+ static const GVecGen3 ops[4] = {
+ { .fniv = gen_uabd_vec,
+ .fno = gen_helper_gvec_uabd_b,
+ .opt_opc = vecop_list,
+ .vece = MO_8 },
+ { .fniv = gen_uabd_vec,
+ .fno = gen_helper_gvec_uabd_h,
+ .opt_opc = vecop_list,
+ .vece = MO_16 },
+ { .fni4 = gen_uabd_i32,
+ .fniv = gen_uabd_vec,
+ .fno = gen_helper_gvec_uabd_s,
+ .opt_opc = vecop_list,
+ .vece = MO_32 },
+ { .fni8 = gen_uabd_i64,
+ .fniv = gen_uabd_vec,
+ .fno = gen_helper_gvec_uabd_d,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ .opt_opc = vecop_list,
+ .vece = MO_64 },
+ };
+ tcg_gen_gvec_3(rd_ofs, rn_ofs, rm_ofs, opr_sz, max_sz, &ops[vece]);
+}
+
/* Translate a NEON data processing instruction. Return nonzero if the
instruction is invalid.
We process data in a mixture of 32-bit and 64-bit chunks.
@@ -5642,6 +5762,16 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
vec_size, vec_size);
}
return 0;
+
+ case NEON_3R_VABD:
+ if (u) {
+ gen_gvec_uabd(size, rd_ofs, rn_ofs, rm_ofs,
+ vec_size, vec_size);
+ } else {
+ gen_gvec_sabd(size, rd_ofs, rn_ofs, rm_ofs,
+ vec_size, vec_size);
+ }
+ return 0;
}
if (size == 3) {
@@ -5772,9 +5902,6 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
case NEON_3R_VQRSHL:
GEN_NEON_INTEGER_OP_ENV(qrshl);
break;
- case NEON_3R_VABD:
- GEN_NEON_INTEGER_OP(abd);
- break;
case NEON_3R_VABA:
GEN_NEON_INTEGER_OP(abd);
tcg_temp_free_i32(tmp2);
diff --git a/target/arm/vec_helper.c b/target/arm/vec_helper.c
index a483841add..a4972d02fc 100644
--- a/target/arm/vec_helper.c
+++ b/target/arm/vec_helper.c
@@ -1407,3 +1407,27 @@ DO_CMP0(gvec_cgt0_h, int16_t, >)
DO_CMP0(gvec_cge0_h, int16_t, >=)
#undef DO_CMP0
+
+#define DO_ABD(NAME, TYPE) \
+void HELPER(NAME)(void *vd, void *vn, void *vm, uint32_t desc) \
+{ \
+ intptr_t i, opr_sz = simd_oprsz(desc); \
+ TYPE *d = vd, *n = vn, *m = vm; \
+ \
+ for (i = 0; i < opr_sz / sizeof(TYPE); ++i) { \
+ d[i] = n[i] < m[i] ? m[i] - n[i] : n[i] - m[i]; \
+ } \
+ clear_tail(d, opr_sz, simd_maxsz(desc)); \
+}
+
+DO_ABD(gvec_sabd_b, int8_t)
+DO_ABD(gvec_sabd_h, int16_t)
+DO_ABD(gvec_sabd_s, int32_t)
+DO_ABD(gvec_sabd_d, int64_t)
+
+DO_ABD(gvec_uabd_b, uint8_t)
+DO_ABD(gvec_uabd_h, uint16_t)
+DO_ABD(gvec_uabd_s, uint32_t)
+DO_ABD(gvec_uabd_d, uint64_t)
+
+#undef DO_ABD
--
2.20.1
^ permalink raw reply related [flat|nested] 18+ messages in thread* [PATCH v2 15/15] target/arm: Vectorize SABA/UABA
2020-05-02 22:44 [PATCH v2 00/15] target/arm: partial vector cleanup Richard Henderson
` (13 preceding siblings ...)
2020-05-02 22:45 ` [PATCH v2 14/15] target/arm: Vectorize SABD/UABD Richard Henderson
@ 2020-05-02 22:45 ` Richard Henderson
2020-05-05 10:38 ` [PATCH v2 00/15] target/arm: partial vector cleanup Peter Maydell
15 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2020-05-02 22:45 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Include 64-bit element size in preparation for SVE2.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/helper.h | 17 +++--
target/arm/translate.h | 5 ++
target/arm/neon_helper.c | 10 ---
target/arm/translate-a64.c | 17 ++---
target/arm/translate.c | 134 +++++++++++++++++++++++++++++++++++--
target/arm/vec_helper.c | 24 +++++++
6 files changed, 174 insertions(+), 33 deletions(-)
diff --git a/target/arm/helper.h b/target/arm/helper.h
index 4678d3a6f4..1857f4ee46 100644
--- a/target/arm/helper.h
+++ b/target/arm/helper.h
@@ -284,13 +284,6 @@ DEF_HELPER_2(neon_pmax_s8, i32, i32, i32)
DEF_HELPER_2(neon_pmax_u16, i32, i32, i32)
DEF_HELPER_2(neon_pmax_s16, i32, i32, i32)
-DEF_HELPER_2(neon_abd_u8, i32, i32, i32)
-DEF_HELPER_2(neon_abd_s8, i32, i32, i32)
-DEF_HELPER_2(neon_abd_u16, i32, i32, i32)
-DEF_HELPER_2(neon_abd_s16, i32, i32, i32)
-DEF_HELPER_2(neon_abd_u32, i32, i32, i32)
-DEF_HELPER_2(neon_abd_s32, i32, i32, i32)
-
DEF_HELPER_2(neon_shl_u16, i32, i32, i32)
DEF_HELPER_2(neon_shl_s16, i32, i32, i32)
DEF_HELPER_2(neon_rshl_u8, i32, i32, i32)
@@ -741,6 +734,16 @@ DEF_HELPER_FLAGS_4(gvec_uabd_h, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_uabd_s, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_uabd_d, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_saba_b, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_saba_h, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_saba_s, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_saba_d, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+
+DEF_HELPER_FLAGS_4(gvec_uaba_b, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_uaba_h, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_uaba_s, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(gvec_uaba_d, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
+
#ifdef TARGET_AARCH64
#include "helper-a64.h"
#include "helper-sve.h"
diff --git a/target/arm/translate.h b/target/arm/translate.h
index d4c4111a5c..70139efcee 100644
--- a/target/arm/translate.h
+++ b/target/arm/translate.h
@@ -342,6 +342,11 @@ void gen_gvec_sabd(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
void gen_gvec_uabd(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz);
+void gen_gvec_saba(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz);
+void gen_gvec_uaba(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz);
+
/*
* Forward to the isar_feature_* tests given a DisasContext pointer.
*/
diff --git a/target/arm/neon_helper.c b/target/arm/neon_helper.c
index 448be93fa1..2ef75e04c8 100644
--- a/target/arm/neon_helper.c
+++ b/target/arm/neon_helper.c
@@ -576,16 +576,6 @@ NEON_POP(pmax_s16, neon_s16, 2)
NEON_POP(pmax_u16, neon_u16, 2)
#undef NEON_FN
-#define NEON_FN(dest, src1, src2) \
- dest = (src1 > src2) ? (src1 - src2) : (src2 - src1)
-NEON_VOP(abd_s8, neon_s8, 4)
-NEON_VOP(abd_u8, neon_u8, 4)
-NEON_VOP(abd_s16, neon_s16, 2)
-NEON_VOP(abd_u16, neon_u16, 2)
-NEON_VOP(abd_s32, neon_s32, 1)
-NEON_VOP(abd_u32, neon_u32, 1)
-#undef NEON_FN
-
#define NEON_FN(dest, src1, src2) do { \
int8_t tmp; \
tmp = (int8_t)src2; \
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index 38f72bf550..da140d8b91 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -11214,6 +11214,13 @@ static void disas_simd_3same_int(DisasContext *s, uint32_t insn)
gen_gvec_fn3(s, is_q, rd, rn, rm, gen_gvec_sabd, size);
}
return;
+ case 0xf: /* SABA, UABA */
+ if (u) {
+ gen_gvec_fn3(s, is_q, rd, rn, rm, gen_gvec_uaba, size);
+ } else {
+ gen_gvec_fn3(s, is_q, rd, rn, rm, gen_gvec_saba, size);
+ }
+ return;
case 0x10: /* ADD, SUB */
if (u) {
gen_gvec_fn3(s, is_q, rd, rn, rm, tcg_gen_gvec_sub, size);
@@ -11346,16 +11353,6 @@ static void disas_simd_3same_int(DisasContext *s, uint32_t insn)
genenvfn = fns[size][u];
break;
}
- case 0xf: /* SABA, UABA */
- {
- static NeonGenTwoOpFn * const fns[3][2] = {
- { gen_helper_neon_abd_s8, gen_helper_neon_abd_u8 },
- { gen_helper_neon_abd_s16, gen_helper_neon_abd_u16 },
- { gen_helper_neon_abd_s32, gen_helper_neon_abd_u32 },
- };
- genfn = fns[size][u];
- break;
- }
case 0x16: /* SQDMULH, SQRDMULH */
{
static NeonGenTwoOpEnvFn * const fns[2][2] = {
diff --git a/target/arm/translate.c b/target/arm/translate.c
index e0c4de2898..4af52ab7e8 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -5494,6 +5494,124 @@ void gen_gvec_uabd(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
tcg_gen_gvec_3(rd_ofs, rn_ofs, rm_ofs, opr_sz, max_sz, &ops[vece]);
}
+static void gen_saba_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
+{
+ TCGv_i32 t = tcg_temp_new_i32();
+ gen_sabd_i32(t, a, b);
+ tcg_gen_add_i32(d, d, t);
+ tcg_temp_free_i32(t);
+}
+
+static void gen_saba_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
+{
+ TCGv_i64 t = tcg_temp_new_i64();
+ gen_sabd_i64(t, a, b);
+ tcg_gen_add_i64(d, d, t);
+ tcg_temp_free_i64(t);
+}
+
+static void gen_saba_vec(unsigned vece, TCGv_vec d, TCGv_vec a, TCGv_vec b)
+{
+ TCGv_vec t = tcg_temp_new_vec_matching(d);
+ gen_sabd_vec(vece, t, a, b);
+ tcg_gen_add_vec(vece, d, d, t);
+ tcg_temp_free_vec(t);
+}
+
+void gen_gvec_saba(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz)
+{
+ static const TCGOpcode vecop_list[] = {
+ INDEX_op_sub_vec, INDEX_op_add_vec,
+ INDEX_op_smin_vec, INDEX_op_smax_vec, 0
+ };
+ static const GVecGen3 ops[4] = {
+ { .fniv = gen_saba_vec,
+ .fno = gen_helper_gvec_saba_b,
+ .opt_opc = vecop_list,
+ .load_dest = true,
+ .vece = MO_8 },
+ { .fniv = gen_saba_vec,
+ .fno = gen_helper_gvec_saba_h,
+ .opt_opc = vecop_list,
+ .load_dest = true,
+ .vece = MO_16 },
+ { .fni4 = gen_saba_i32,
+ .fniv = gen_saba_vec,
+ .fno = gen_helper_gvec_saba_s,
+ .opt_opc = vecop_list,
+ .load_dest = true,
+ .vece = MO_32 },
+ { .fni8 = gen_saba_i64,
+ .fniv = gen_saba_vec,
+ .fno = gen_helper_gvec_saba_d,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ .opt_opc = vecop_list,
+ .load_dest = true,
+ .vece = MO_64 },
+ };
+ tcg_gen_gvec_3(rd_ofs, rn_ofs, rm_ofs, opr_sz, max_sz, &ops[vece]);
+}
+
+static void gen_uaba_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
+{
+ TCGv_i32 t = tcg_temp_new_i32();
+ gen_uabd_i32(t, a, b);
+ tcg_gen_add_i32(d, d, t);
+ tcg_temp_free_i32(t);
+}
+
+static void gen_uaba_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
+{
+ TCGv_i64 t = tcg_temp_new_i64();
+ gen_uabd_i64(t, a, b);
+ tcg_gen_add_i64(d, d, t);
+ tcg_temp_free_i64(t);
+}
+
+static void gen_uaba_vec(unsigned vece, TCGv_vec d, TCGv_vec a, TCGv_vec b)
+{
+ TCGv_vec t = tcg_temp_new_vec_matching(d);
+ gen_uabd_vec(vece, t, a, b);
+ tcg_gen_add_vec(vece, d, d, t);
+ tcg_temp_free_vec(t);
+}
+
+void gen_gvec_uaba(unsigned vece, uint32_t rd_ofs, uint32_t rn_ofs,
+ uint32_t rm_ofs, uint32_t opr_sz, uint32_t max_sz)
+{
+ static const TCGOpcode vecop_list[] = {
+ INDEX_op_sub_vec, INDEX_op_add_vec,
+ INDEX_op_umin_vec, INDEX_op_umax_vec, 0
+ };
+ static const GVecGen3 ops[4] = {
+ { .fniv = gen_uaba_vec,
+ .fno = gen_helper_gvec_uaba_b,
+ .opt_opc = vecop_list,
+ .load_dest = true,
+ .vece = MO_8 },
+ { .fniv = gen_uaba_vec,
+ .fno = gen_helper_gvec_uaba_h,
+ .opt_opc = vecop_list,
+ .load_dest = true,
+ .vece = MO_16 },
+ { .fni4 = gen_uaba_i32,
+ .fniv = gen_uaba_vec,
+ .fno = gen_helper_gvec_uaba_s,
+ .opt_opc = vecop_list,
+ .load_dest = true,
+ .vece = MO_32 },
+ { .fni8 = gen_uaba_i64,
+ .fniv = gen_uaba_vec,
+ .fno = gen_helper_gvec_uaba_d,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ .opt_opc = vecop_list,
+ .load_dest = true,
+ .vece = MO_64 },
+ };
+ tcg_gen_gvec_3(rd_ofs, rn_ofs, rm_ofs, opr_sz, max_sz, &ops[vece]);
+}
+
/* Translate a NEON data processing instruction. Return nonzero if the
instruction is invalid.
We process data in a mixture of 32-bit and 64-bit chunks.
@@ -5772,6 +5890,16 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
vec_size, vec_size);
}
return 0;
+
+ case NEON_3R_VABA:
+ if (u) {
+ gen_gvec_uaba(size, rd_ofs, rn_ofs, rm_ofs,
+ vec_size, vec_size);
+ } else {
+ gen_gvec_saba(size, rd_ofs, rn_ofs, rm_ofs,
+ vec_size, vec_size);
+ }
+ return 0;
}
if (size == 3) {
@@ -5902,12 +6030,6 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
case NEON_3R_VQRSHL:
GEN_NEON_INTEGER_OP_ENV(qrshl);
break;
- case NEON_3R_VABA:
- GEN_NEON_INTEGER_OP(abd);
- tcg_temp_free_i32(tmp2);
- tmp2 = neon_load_reg(rd, pass);
- gen_neon_add(size, tmp, tmp2);
- break;
case NEON_3R_VPMAX:
GEN_NEON_INTEGER_OP(pmax);
break;
diff --git a/target/arm/vec_helper.c b/target/arm/vec_helper.c
index a4972d02fc..1be41a8baf 100644
--- a/target/arm/vec_helper.c
+++ b/target/arm/vec_helper.c
@@ -1431,3 +1431,27 @@ DO_ABD(gvec_uabd_s, uint32_t)
DO_ABD(gvec_uabd_d, uint64_t)
#undef DO_ABD
+
+#define DO_ABA(NAME, TYPE) \
+void HELPER(NAME)(void *vd, void *vn, void *vm, uint32_t desc) \
+{ \
+ intptr_t i, opr_sz = simd_oprsz(desc); \
+ TYPE *d = vd, *n = vn, *m = vm; \
+ \
+ for (i = 0; i < opr_sz / sizeof(TYPE); ++i) { \
+ d[i] += n[i] < m[i] ? m[i] - n[i] : n[i] - m[i]; \
+ } \
+ clear_tail(d, opr_sz, simd_maxsz(desc)); \
+}
+
+DO_ABA(gvec_saba_b, int8_t)
+DO_ABA(gvec_saba_h, int16_t)
+DO_ABA(gvec_saba_s, int32_t)
+DO_ABA(gvec_saba_d, int64_t)
+
+DO_ABA(gvec_uaba_b, uint8_t)
+DO_ABA(gvec_uaba_h, uint16_t)
+DO_ABA(gvec_uaba_s, uint32_t)
+DO_ABA(gvec_uaba_d, uint64_t)
+
+#undef DO_ABD
--
2.20.1
^ permalink raw reply related [flat|nested] 18+ messages in thread* Re: [PATCH v2 00/15] target/arm: partial vector cleanup
2020-05-02 22:44 [PATCH v2 00/15] target/arm: partial vector cleanup Richard Henderson
` (14 preceding siblings ...)
2020-05-02 22:45 ` [PATCH v2 15/15] target/arm: Vectorize SABA/UABA Richard Henderson
@ 2020-05-05 10:38 ` Peter Maydell
2020-05-05 14:00 ` Richard Henderson
15 siblings, 1 reply; 18+ messages in thread
From: Peter Maydell @ 2020-05-05 10:38 UTC (permalink / raw)
To: Richard Henderson; +Cc: QEMU Developers
On Sat, 2 May 2020 at 23:45, Richard Henderson
<richard.henderson@linaro.org> wrote:
>
> This is not complete, but shows the direction I'd like to go.
>
> It may well help what Peter is doing with the neon decodetree
> conversion. It may be helpful to apply before the conversion
> in order to reduce the number of special cases. As may
> continuing with the cleanup; I'll probably work on that more
> next week.
>
> Version 2 extracts more bits from my sve2 branch. There's
> still more to pull back, especially for crypto_helper.c, where
> there are also tail clearing bugs to fix.
This doesn't apply on master any more as a result of the
first lump of decodetree stuff going in. Also patch 2
doesn't seem to compile:
/home/petmay01/linaro/qemu-from-laptop/qemu/target/arm/translate.c: In
function ‘gen_gvec_srshr’:
/home/petmay01/linaro/qemu-from-laptop/qemu/target/arm/translate.c:4107:9:
error: implicit declaration of function ‘tcg_gen_gvec_dup_imm’; did
you mean ‘tcg_gen_gvec_dup_i64’?
[-Werror=implicit-function-declaration]
tcg_gen_gvec_dup_imm(vece, rd_ofs, opr_sz, max_sz, 0);
^~~~~~~~~~~~~~~~~~~~
tcg_gen_gvec_dup_i64
/home/petmay01/linaro/qemu-from-laptop/qemu/target/arm/translate.c:4107:9:
error: nested extern declaration of ‘tcg_gen_gvec_dup_imm’
[-Werror=nested-externs]
Any chance you could do a rebase and resend?
thanks
-- PMM
^ permalink raw reply [flat|nested] 18+ messages in thread* Re: [PATCH v2 00/15] target/arm: partial vector cleanup
2020-05-05 10:38 ` [PATCH v2 00/15] target/arm: partial vector cleanup Peter Maydell
@ 2020-05-05 14:00 ` Richard Henderson
0 siblings, 0 replies; 18+ messages in thread
From: Richard Henderson @ 2020-05-05 14:00 UTC (permalink / raw)
To: Peter Maydell; +Cc: QEMU Developers
On 5/5/20 3:38 AM, Peter Maydell wrote:
> This doesn't apply on master any more as a result of the
> first lump of decodetree stuff going in. Also patch 2
> doesn't seem to compile:
>
> /home/petmay01/linaro/qemu-from-laptop/qemu/target/arm/translate.c: In
> function ‘gen_gvec_srshr’:
> /home/petmay01/linaro/qemu-from-laptop/qemu/target/arm/translate.c:4107:9:
> error: implicit declaration of function ‘tcg_gen_gvec_dup_imm’; did
> you mean ‘tcg_gen_gvec_dup_i64’?
> [-Werror=implicit-function-declaration]
> tcg_gen_gvec_dup_imm(vece, rd_ofs, opr_sz, max_sz, 0);
> ^~~~~~~~~~~~~~~~~~~~
> tcg_gen_gvec_dup_i64
> /home/petmay01/linaro/qemu-from-laptop/qemu/target/arm/translate.c:4107:9:
> error: nested extern declaration of ‘tcg_gen_gvec_dup_imm’
> [-Werror=nested-externs]
>
> Any chance you could do a rebase and resend?
Ah yes, it seems built on top of tcg-next. Time for a pull there too, lest
this have to get changed yet again.
r~
^ permalink raw reply [flat|nested] 18+ messages in thread