target/arm: Pass separate addend to {U, S}DOT helpers

For SVE, we potentially have a 4th argument coming from the
movprfx instruction.  Currently we do not optimize movprfx,
so the problem is not visible.

Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210525010358.152808-50-richard.henderson@linaro.org
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
master
Richard Henderson 2021-05-24 18:03:15 -07:00 committed by Peter Maydell
parent 77e786bb95
commit bc2bd6974e
6 changed files with 109 additions and 76 deletions

View File

@ -608,15 +608,19 @@ DEF_HELPER_FLAGS_5(sve2_sqrdmlah_d, TCG_CALL_NO_RWG,
DEF_HELPER_FLAGS_5(sve2_sqrdmlsh_d, TCG_CALL_NO_RWG,
void, ptr, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_sdot_b, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_udot_b, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_sdot_h, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_udot_h, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_5(gvec_sdot_b, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_5(gvec_udot_b, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_5(gvec_sdot_h, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_5(gvec_udot_h, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_sdot_idx_b, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_udot_idx_b, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_sdot_idx_h, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_udot_idx_h, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_5(gvec_sdot_idx_b, TCG_CALL_NO_RWG,
void, ptr, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_5(gvec_udot_idx_b, TCG_CALL_NO_RWG,
void, ptr, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_5(gvec_sdot_idx_h, TCG_CALL_NO_RWG,
void, ptr, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_5(gvec_udot_idx_h, TCG_CALL_NO_RWG,
void, ptr, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_5(gvec_fcaddh, TCG_CALL_NO_RWG,
void, ptr, ptr, ptr, ptr, i32)

View File

@ -756,12 +756,13 @@ UMIN_zzi 00100101 .. 101 011 110 ........ ..... @rdn_i8u
MUL_zzi 00100101 .. 110 000 110 ........ ..... @rdn_i8s
# SVE integer dot product (unpredicated)
DOT_zzz 01000100 1 sz:1 0 rm:5 00000 u:1 rn:5 rd:5 ra=%reg_movprfx
DOT_zzzz 01000100 1 sz:1 0 rm:5 00000 u:1 rn:5 rd:5 \
ra=%reg_movprfx
# SVE integer dot product (indexed)
DOT_zzx 01000100 101 index:2 rm:3 00000 u:1 rn:5 rd:5 \
DOT_zzxw 01000100 101 index:2 rm:3 00000 u:1 rn:5 rd:5 \
sz=0 ra=%reg_movprfx
DOT_zzx 01000100 111 index:1 rm:4 00000 u:1 rn:5 rd:5 \
DOT_zzxw 01000100 111 index:1 rm:4 00000 u:1 rn:5 rd:5 \
sz=1 ra=%reg_movprfx
# SVE floating-point complex add (predicated)

View File

@ -683,6 +683,17 @@ static void gen_gvec_op3_qc(DisasContext *s, bool is_q, int rd, int rn,
tcg_temp_free_ptr(qc_ptr);
}
/* Expand a 4-operand operation using an out-of-line helper. */
static void gen_gvec_op4_ool(DisasContext *s, bool is_q, int rd, int rn,
int rm, int ra, int data, gen_helper_gvec_4 *fn)
{
tcg_gen_gvec_4_ool(vec_full_reg_offset(s, rd),
vec_full_reg_offset(s, rn),
vec_full_reg_offset(s, rm),
vec_full_reg_offset(s, ra),
is_q ? 16 : 8, vec_full_reg_size(s), data, fn);
}
/* Set ZF and NF based on a 64 bit result. This is alas fiddlier
* than the 32 bit equivalent.
*/
@ -12183,7 +12194,7 @@ static void disas_simd_three_reg_same_extra(DisasContext *s, uint32_t insn)
return;
case 0x2: /* SDOT / UDOT */
gen_gvec_op3_ool(s, is_q, rd, rn, rm, 0,
gen_gvec_op4_ool(s, is_q, rd, rn, rm, rd, 0,
u ? gen_helper_gvec_udot_b : gen_helper_gvec_sdot_b);
return;
@ -13442,7 +13453,7 @@ static void disas_simd_indexed(DisasContext *s, uint32_t insn)
switch (16 * u + opcode) {
case 0x0e: /* SDOT */
case 0x1e: /* UDOT */
gen_gvec_op3_ool(s, is_q, rd, rn, rm, index,
gen_gvec_op4_ool(s, is_q, rd, rn, rm, rd, index,
u ? gen_helper_gvec_udot_idx_b
: gen_helper_gvec_sdot_idx_b);
return;

View File

@ -230,7 +230,7 @@ static bool trans_VCADD(DisasContext *s, arg_VCADD *a)
static bool trans_VDOT(DisasContext *s, arg_VDOT *a)
{
int opr_sz;
gen_helper_gvec_3 *fn_gvec;
gen_helper_gvec_4 *fn_gvec;
if (!dc_isar_feature(aa32_dp, s)) {
return false;
@ -252,9 +252,10 @@ static bool trans_VDOT(DisasContext *s, arg_VDOT *a)
opr_sz = (1 + a->q) * 8;
fn_gvec = a->u ? gen_helper_gvec_udot_b : gen_helper_gvec_sdot_b;
tcg_gen_gvec_3_ool(vfp_reg_offset(1, a->vd),
tcg_gen_gvec_4_ool(vfp_reg_offset(1, a->vd),
vfp_reg_offset(1, a->vn),
vfp_reg_offset(1, a->vm),
vfp_reg_offset(1, a->vd),
opr_sz, opr_sz, 0, fn_gvec);
return true;
}
@ -332,7 +333,7 @@ static bool trans_VCMLA_scalar(DisasContext *s, arg_VCMLA_scalar *a)
static bool trans_VDOT_scalar(DisasContext *s, arg_VDOT_scalar *a)
{
gen_helper_gvec_3 *fn_gvec;
gen_helper_gvec_4 *fn_gvec;
int opr_sz;
TCGv_ptr fpst;
@ -357,9 +358,10 @@ static bool trans_VDOT_scalar(DisasContext *s, arg_VDOT_scalar *a)
fn_gvec = a->u ? gen_helper_gvec_udot_idx_b : gen_helper_gvec_sdot_idx_b;
opr_sz = (1 + a->q) * 8;
fpst = fpstatus_ptr(FPST_STD);
tcg_gen_gvec_3_ool(vfp_reg_offset(1, a->vd),
tcg_gen_gvec_4_ool(vfp_reg_offset(1, a->vd),
vfp_reg_offset(1, a->vn),
vfp_reg_offset(1, a->rm),
vfp_reg_offset(1, a->vd),
opr_sz, opr_sz, a->index, fn_gvec);
tcg_temp_free_ptr(fpst);
return true;

View File

@ -3800,28 +3800,29 @@ DO_ZZI(UMIN, umin)
#undef DO_ZZI
static bool trans_DOT_zzz(DisasContext *s, arg_DOT_zzz *a)
static bool trans_DOT_zzzz(DisasContext *s, arg_DOT_zzzz *a)
{
static gen_helper_gvec_3 * const fns[2][2] = {
static gen_helper_gvec_4 * const fns[2][2] = {
{ gen_helper_gvec_sdot_b, gen_helper_gvec_sdot_h },
{ gen_helper_gvec_udot_b, gen_helper_gvec_udot_h }
};
if (sve_access_check(s)) {
gen_gvec_ool_zzz(s, fns[a->u][a->sz], a->rd, a->rn, a->rm, 0);
gen_gvec_ool_zzzz(s, fns[a->u][a->sz], a->rd, a->rn, a->rm, a->ra, 0);
}
return true;
}
static bool trans_DOT_zzx(DisasContext *s, arg_DOT_zzx *a)
static bool trans_DOT_zzxw(DisasContext *s, arg_DOT_zzxw *a)
{
static gen_helper_gvec_3 * const fns[2][2] = {
static gen_helper_gvec_4 * const fns[2][2] = {
{ gen_helper_gvec_sdot_idx_b, gen_helper_gvec_sdot_idx_h },
{ gen_helper_gvec_udot_idx_b, gen_helper_gvec_udot_idx_h }
};
if (sve_access_check(s)) {
gen_gvec_ool_zzz(s, fns[a->u][a->sz], a->rd, a->rn, a->rm, a->index);
gen_gvec_ool_zzzz(s, fns[a->u][a->sz], a->rd, a->rn, a->rm,
a->ra, a->index);
}
return true;
}

View File

@ -375,71 +375,76 @@ void HELPER(sve2_sqrdmlsh_d)(void *vd, void *vn, void *vm,
* All elements are treated equally, no matter where they are.
*/
void HELPER(gvec_sdot_b)(void *vd, void *vn, void *vm, uint32_t desc)
void HELPER(gvec_sdot_b)(void *vd, void *vn, void *vm, void *va, uint32_t desc)
{
intptr_t i, opr_sz = simd_oprsz(desc);
int32_t *d = vd;
int32_t *d = vd, *a = va;
int8_t *n = vn, *m = vm;
for (i = 0; i < opr_sz / 4; ++i) {
d[i] += n[i * 4 + 0] * m[i * 4 + 0]
+ n[i * 4 + 1] * m[i * 4 + 1]
+ n[i * 4 + 2] * m[i * 4 + 2]
+ n[i * 4 + 3] * m[i * 4 + 3];
d[i] = (a[i] +
n[i * 4 + 0] * m[i * 4 + 0] +
n[i * 4 + 1] * m[i * 4 + 1] +
n[i * 4 + 2] * m[i * 4 + 2] +
n[i * 4 + 3] * m[i * 4 + 3]);
}
clear_tail(d, opr_sz, simd_maxsz(desc));
}
void HELPER(gvec_udot_b)(void *vd, void *vn, void *vm, uint32_t desc)
void HELPER(gvec_udot_b)(void *vd, void *vn, void *vm, void *va, uint32_t desc)
{
intptr_t i, opr_sz = simd_oprsz(desc);
uint32_t *d = vd;
uint32_t *d = vd, *a = va;
uint8_t *n = vn, *m = vm;
for (i = 0; i < opr_sz / 4; ++i) {
d[i] += n[i * 4 + 0] * m[i * 4 + 0]
+ n[i * 4 + 1] * m[i * 4 + 1]
+ n[i * 4 + 2] * m[i * 4 + 2]
+ n[i * 4 + 3] * m[i * 4 + 3];
d[i] = (a[i] +
n[i * 4 + 0] * m[i * 4 + 0] +
n[i * 4 + 1] * m[i * 4 + 1] +
n[i * 4 + 2] * m[i * 4 + 2] +
n[i * 4 + 3] * m[i * 4 + 3]);
}
clear_tail(d, opr_sz, simd_maxsz(desc));
}
void HELPER(gvec_sdot_h)(void *vd, void *vn, void *vm, uint32_t desc)
void HELPER(gvec_sdot_h)(void *vd, void *vn, void *vm, void *va, uint32_t desc)
{
intptr_t i, opr_sz = simd_oprsz(desc);
int64_t *d = vd;
int64_t *d = vd, *a = va;
int16_t *n = vn, *m = vm;
for (i = 0; i < opr_sz / 8; ++i) {
d[i] += (int64_t)n[i * 4 + 0] * m[i * 4 + 0]
+ (int64_t)n[i * 4 + 1] * m[i * 4 + 1]
+ (int64_t)n[i * 4 + 2] * m[i * 4 + 2]
+ (int64_t)n[i * 4 + 3] * m[i * 4 + 3];
d[i] = (a[i] +
(int64_t)n[i * 4 + 0] * m[i * 4 + 0] +
(int64_t)n[i * 4 + 1] * m[i * 4 + 1] +
(int64_t)n[i * 4 + 2] * m[i * 4 + 2] +
(int64_t)n[i * 4 + 3] * m[i * 4 + 3]);
}
clear_tail(d, opr_sz, simd_maxsz(desc));
}
void HELPER(gvec_udot_h)(void *vd, void *vn, void *vm, uint32_t desc)
void HELPER(gvec_udot_h)(void *vd, void *vn, void *vm, void *va, uint32_t desc)
{
intptr_t i, opr_sz = simd_oprsz(desc);
uint64_t *d = vd;
uint64_t *d = vd, *a = va;
uint16_t *n = vn, *m = vm;
for (i = 0; i < opr_sz / 8; ++i) {
d[i] += (uint64_t)n[i * 4 + 0] * m[i * 4 + 0]
+ (uint64_t)n[i * 4 + 1] * m[i * 4 + 1]
+ (uint64_t)n[i * 4 + 2] * m[i * 4 + 2]
+ (uint64_t)n[i * 4 + 3] * m[i * 4 + 3];
d[i] = (a[i] +
(uint64_t)n[i * 4 + 0] * m[i * 4 + 0] +
(uint64_t)n[i * 4 + 1] * m[i * 4 + 1] +
(uint64_t)n[i * 4 + 2] * m[i * 4 + 2] +
(uint64_t)n[i * 4 + 3] * m[i * 4 + 3]);
}
clear_tail(d, opr_sz, simd_maxsz(desc));
}
void HELPER(gvec_sdot_idx_b)(void *vd, void *vn, void *vm, uint32_t desc)
void HELPER(gvec_sdot_idx_b)(void *vd, void *vn, void *vm,
void *va, uint32_t desc)
{
intptr_t i, segend, opr_sz = simd_oprsz(desc), opr_sz_4 = opr_sz / 4;
intptr_t index = simd_data(desc);
int32_t *d = vd;
int32_t *d = vd, *a = va;
int8_t *n = vn;
int8_t *m_indexed = (int8_t *)vm + H4(index) * 4;
@ -455,10 +460,11 @@ void HELPER(gvec_sdot_idx_b)(void *vd, void *vn, void *vm, uint32_t desc)
int8_t m3 = m_indexed[i * 4 + 3];
do {
d[i] += n[i * 4 + 0] * m0
+ n[i * 4 + 1] * m1
+ n[i * 4 + 2] * m2
+ n[i * 4 + 3] * m3;
d[i] = (a[i] +
n[i * 4 + 0] * m0 +
n[i * 4 + 1] * m1 +
n[i * 4 + 2] * m2 +
n[i * 4 + 3] * m3);
} while (++i < segend);
segend = i + 4;
} while (i < opr_sz_4);
@ -466,11 +472,12 @@ void HELPER(gvec_sdot_idx_b)(void *vd, void *vn, void *vm, uint32_t desc)
clear_tail(d, opr_sz, simd_maxsz(desc));
}
void HELPER(gvec_udot_idx_b)(void *vd, void *vn, void *vm, uint32_t desc)
void HELPER(gvec_udot_idx_b)(void *vd, void *vn, void *vm,
void *va, uint32_t desc)
{
intptr_t i, segend, opr_sz = simd_oprsz(desc), opr_sz_4 = opr_sz / 4;
intptr_t index = simd_data(desc);
uint32_t *d = vd;
uint32_t *d = vd, *a = va;
uint8_t *n = vn;
uint8_t *m_indexed = (uint8_t *)vm + H4(index) * 4;
@ -486,10 +493,11 @@ void HELPER(gvec_udot_idx_b)(void *vd, void *vn, void *vm, uint32_t desc)
uint8_t m3 = m_indexed[i * 4 + 3];
do {
d[i] += n[i * 4 + 0] * m0
+ n[i * 4 + 1] * m1
+ n[i * 4 + 2] * m2
+ n[i * 4 + 3] * m3;
d[i] = (a[i] +
n[i * 4 + 0] * m0 +
n[i * 4 + 1] * m1 +
n[i * 4 + 2] * m2 +
n[i * 4 + 3] * m3);
} while (++i < segend);
segend = i + 4;
} while (i < opr_sz_4);
@ -497,11 +505,12 @@ void HELPER(gvec_udot_idx_b)(void *vd, void *vn, void *vm, uint32_t desc)
clear_tail(d, opr_sz, simd_maxsz(desc));
}
void HELPER(gvec_sdot_idx_h)(void *vd, void *vn, void *vm, uint32_t desc)
void HELPER(gvec_sdot_idx_h)(void *vd, void *vn, void *vm,
void *va, uint32_t desc)
{
intptr_t i, opr_sz = simd_oprsz(desc), opr_sz_8 = opr_sz / 8;
intptr_t index = simd_data(desc);
int64_t *d = vd;
int64_t *d = vd, *a = va;
int16_t *n = vn;
int16_t *m_indexed = (int16_t *)vm + index * 4;
@ -509,30 +518,33 @@ void HELPER(gvec_sdot_idx_h)(void *vd, void *vn, void *vm, uint32_t desc)
* Process the entire segment all at once, writing back the results
* only after we've consumed all of the inputs.
*/
for (i = 0; i < opr_sz_8 ; i += 2) {
uint64_t d0, d1;
for (i = 0; i < opr_sz_8; i += 2) {
int64_t d0, d1;
d0 = n[i * 4 + 0] * (int64_t)m_indexed[i * 4 + 0];
d0 = a[i + 0];
d0 += n[i * 4 + 0] * (int64_t)m_indexed[i * 4 + 0];
d0 += n[i * 4 + 1] * (int64_t)m_indexed[i * 4 + 1];
d0 += n[i * 4 + 2] * (int64_t)m_indexed[i * 4 + 2];
d0 += n[i * 4 + 3] * (int64_t)m_indexed[i * 4 + 3];
d1 = n[i * 4 + 4] * (int64_t)m_indexed[i * 4 + 0];
d1 = a[i + 1];
d1 += n[i * 4 + 4] * (int64_t)m_indexed[i * 4 + 0];
d1 += n[i * 4 + 5] * (int64_t)m_indexed[i * 4 + 1];
d1 += n[i * 4 + 6] * (int64_t)m_indexed[i * 4 + 2];
d1 += n[i * 4 + 7] * (int64_t)m_indexed[i * 4 + 3];
d[i + 0] += d0;
d[i + 1] += d1;
d[i + 0] = d0;
d[i + 1] = d1;
}
clear_tail(d, opr_sz, simd_maxsz(desc));
}
void HELPER(gvec_udot_idx_h)(void *vd, void *vn, void *vm, uint32_t desc)
void HELPER(gvec_udot_idx_h)(void *vd, void *vn, void *vm,
void *va, uint32_t desc)
{
intptr_t i, opr_sz = simd_oprsz(desc), opr_sz_8 = opr_sz / 8;
intptr_t index = simd_data(desc);
uint64_t *d = vd;
uint64_t *d = vd, *a = va;
uint16_t *n = vn;
uint16_t *m_indexed = (uint16_t *)vm + index * 4;
@ -540,22 +552,24 @@ void HELPER(gvec_udot_idx_h)(void *vd, void *vn, void *vm, uint32_t desc)
* Process the entire segment all at once, writing back the results
* only after we've consumed all of the inputs.
*/
for (i = 0; i < opr_sz_8 ; i += 2) {
for (i = 0; i < opr_sz_8; i += 2) {
uint64_t d0, d1;
d0 = n[i * 4 + 0] * (uint64_t)m_indexed[i * 4 + 0];
d0 = a[i + 0];
d0 += n[i * 4 + 0] * (uint64_t)m_indexed[i * 4 + 0];
d0 += n[i * 4 + 1] * (uint64_t)m_indexed[i * 4 + 1];
d0 += n[i * 4 + 2] * (uint64_t)m_indexed[i * 4 + 2];
d0 += n[i * 4 + 3] * (uint64_t)m_indexed[i * 4 + 3];
d1 = n[i * 4 + 4] * (uint64_t)m_indexed[i * 4 + 0];
d1 = a[i + 1];
d1 += n[i * 4 + 4] * (uint64_t)m_indexed[i * 4 + 0];
d1 += n[i * 4 + 5] * (uint64_t)m_indexed[i * 4 + 1];
d1 += n[i * 4 + 6] * (uint64_t)m_indexed[i * 4 + 2];
d1 += n[i * 4 + 7] * (uint64_t)m_indexed[i * 4 + 3];
d[i + 0] += d0;
d[i + 1] += d1;
d[i + 0] = d0;
d[i + 1] = d1;
}
clear_tail(d, opr_sz, simd_maxsz(desc));
}