tcg: Add gvec expanders for variable shift

The gvec expanders perform a modulo on the shift count.  If the target
requires alternate behaviour, then it cannot use the generic gvec
expanders anyway, and will have to have its own custom code.

Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
This commit is contained in:
Richard Henderson 2019-04-13 20:42:37 -10:00
parent 37ee55a081
commit 5ee5c14cac
6 changed files with 384 additions and 0 deletions

View File

@ -725,6 +725,150 @@ void HELPER(gvec_sar64i)(void *d, void *a, uint32_t desc)
clear_high(d, oprsz, desc); clear_high(d, oprsz, desc);
} }
void HELPER(gvec_shl8v)(void *d, void *a, void *b, uint32_t desc)
{
intptr_t oprsz = simd_oprsz(desc);
intptr_t i;
for (i = 0; i < oprsz; i += sizeof(uint8_t)) {
uint8_t sh = *(uint8_t *)(b + i) & 7;
*(uint8_t *)(d + i) = *(uint8_t *)(a + i) << sh;
}
clear_high(d, oprsz, desc);
}
void HELPER(gvec_shl16v)(void *d, void *a, void *b, uint32_t desc)
{
intptr_t oprsz = simd_oprsz(desc);
intptr_t i;
for (i = 0; i < oprsz; i += sizeof(uint16_t)) {
uint8_t sh = *(uint16_t *)(b + i) & 15;
*(uint16_t *)(d + i) = *(uint16_t *)(a + i) << sh;
}
clear_high(d, oprsz, desc);
}
void HELPER(gvec_shl32v)(void *d, void *a, void *b, uint32_t desc)
{
intptr_t oprsz = simd_oprsz(desc);
intptr_t i;
for (i = 0; i < oprsz; i += sizeof(uint32_t)) {
uint8_t sh = *(uint32_t *)(b + i) & 31;
*(uint32_t *)(d + i) = *(uint32_t *)(a + i) << sh;
}
clear_high(d, oprsz, desc);
}
void HELPER(gvec_shl64v)(void *d, void *a, void *b, uint32_t desc)
{
intptr_t oprsz = simd_oprsz(desc);
intptr_t i;
for (i = 0; i < oprsz; i += sizeof(uint64_t)) {
uint8_t sh = *(uint64_t *)(b + i) & 63;
*(uint64_t *)(d + i) = *(uint64_t *)(a + i) << sh;
}
clear_high(d, oprsz, desc);
}
void HELPER(gvec_shr8v)(void *d, void *a, void *b, uint32_t desc)
{
intptr_t oprsz = simd_oprsz(desc);
intptr_t i;
for (i = 0; i < oprsz; i += sizeof(uint8_t)) {
uint8_t sh = *(uint8_t *)(b + i) & 7;
*(uint8_t *)(d + i) = *(uint8_t *)(a + i) >> sh;
}
clear_high(d, oprsz, desc);
}
void HELPER(gvec_shr16v)(void *d, void *a, void *b, uint32_t desc)
{
intptr_t oprsz = simd_oprsz(desc);
intptr_t i;
for (i = 0; i < oprsz; i += sizeof(uint16_t)) {
uint8_t sh = *(uint16_t *)(b + i) & 15;
*(uint16_t *)(d + i) = *(uint16_t *)(a + i) >> sh;
}
clear_high(d, oprsz, desc);
}
void HELPER(gvec_shr32v)(void *d, void *a, void *b, uint32_t desc)
{
intptr_t oprsz = simd_oprsz(desc);
intptr_t i;
for (i = 0; i < oprsz; i += sizeof(uint32_t)) {
uint8_t sh = *(uint32_t *)(b + i) & 31;
*(uint32_t *)(d + i) = *(uint32_t *)(a + i) >> sh;
}
clear_high(d, oprsz, desc);
}
void HELPER(gvec_shr64v)(void *d, void *a, void *b, uint32_t desc)
{
intptr_t oprsz = simd_oprsz(desc);
intptr_t i;
for (i = 0; i < oprsz; i += sizeof(uint64_t)) {
uint8_t sh = *(uint64_t *)(b + i) & 63;
*(uint64_t *)(d + i) = *(uint64_t *)(a + i) >> sh;
}
clear_high(d, oprsz, desc);
}
void HELPER(gvec_sar8v)(void *d, void *a, void *b, uint32_t desc)
{
intptr_t oprsz = simd_oprsz(desc);
intptr_t i;
for (i = 0; i < oprsz; i += sizeof(vec8)) {
uint8_t sh = *(uint8_t *)(b + i) & 7;
*(int8_t *)(d + i) = *(int8_t *)(a + i) >> sh;
}
clear_high(d, oprsz, desc);
}
void HELPER(gvec_sar16v)(void *d, void *a, void *b, uint32_t desc)
{
intptr_t oprsz = simd_oprsz(desc);
intptr_t i;
for (i = 0; i < oprsz; i += sizeof(int16_t)) {
uint8_t sh = *(uint16_t *)(b + i) & 15;
*(int16_t *)(d + i) = *(int16_t *)(a + i) >> sh;
}
clear_high(d, oprsz, desc);
}
void HELPER(gvec_sar32v)(void *d, void *a, void *b, uint32_t desc)
{
intptr_t oprsz = simd_oprsz(desc);
intptr_t i;
for (i = 0; i < oprsz; i += sizeof(vec32)) {
uint8_t sh = *(uint32_t *)(b + i) & 31;
*(int32_t *)(d + i) = *(int32_t *)(a + i) >> sh;
}
clear_high(d, oprsz, desc);
}
void HELPER(gvec_sar64v)(void *d, void *a, void *b, uint32_t desc)
{
intptr_t oprsz = simd_oprsz(desc);
intptr_t i;
for (i = 0; i < oprsz; i += sizeof(vec64)) {
uint8_t sh = *(uint64_t *)(b + i) & 63;
*(int64_t *)(d + i) = *(int64_t *)(a + i) >> sh;
}
clear_high(d, oprsz, desc);
}
/* If vectors are enabled, the compiler fills in -1 for true. /* If vectors are enabled, the compiler fills in -1 for true.
Otherwise, we must take care of this by hand. */ Otherwise, we must take care of this by hand. */
#ifdef CONFIG_VECTOR16 #ifdef CONFIG_VECTOR16

View File

@ -254,6 +254,21 @@ DEF_HELPER_FLAGS_3(gvec_sar16i, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
DEF_HELPER_FLAGS_3(gvec_sar32i, TCG_CALL_NO_RWG, void, ptr, ptr, i32) DEF_HELPER_FLAGS_3(gvec_sar32i, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
DEF_HELPER_FLAGS_3(gvec_sar64i, TCG_CALL_NO_RWG, void, ptr, ptr, i32) DEF_HELPER_FLAGS_3(gvec_sar64i, TCG_CALL_NO_RWG, void, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_shl8v, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_shl16v, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_shl32v, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_shl64v, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_shr8v, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_shr16v, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_shr32v, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_shr64v, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_sar8v, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_sar16v, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_sar32v, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_sar64v, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_eq8, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) DEF_HELPER_FLAGS_4(gvec_eq8, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_eq16, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) DEF_HELPER_FLAGS_4(gvec_eq16, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)
DEF_HELPER_FLAGS_4(gvec_eq32, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) DEF_HELPER_FLAGS_4(gvec_eq32, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32)

View File

@ -2555,6 +2555,201 @@ void tcg_gen_gvec_sari(unsigned vece, uint32_t dofs, uint32_t aofs,
} }
} }
/*
* Expand D = A << (B % element bits)
*
* Unlike scalar shifts, where it is easy for the target front end
* to include the modulo as part of the expansion. If the target
* naturally includes the modulo as part of the operation, great!
* If the target has some other behaviour from out-of-range shifts,
* then it could not use this function anyway, and would need to
* do it's own expansion with custom functions.
*/
static void tcg_gen_shlv_mod_vec(unsigned vece, TCGv_vec d,
TCGv_vec a, TCGv_vec b)
{
TCGv_vec t = tcg_temp_new_vec_matching(d);
tcg_gen_dupi_vec(vece, t, (8 << vece) - 1);
tcg_gen_and_vec(vece, t, t, b);
tcg_gen_shlv_vec(vece, d, a, t);
tcg_temp_free_vec(t);
}
static void tcg_gen_shl_mod_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
{
TCGv_i32 t = tcg_temp_new_i32();
tcg_gen_andi_i32(t, b, 31);
tcg_gen_shl_i32(d, a, t);
tcg_temp_free_i32(t);
}
static void tcg_gen_shl_mod_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
{
TCGv_i64 t = tcg_temp_new_i64();
tcg_gen_andi_i64(t, b, 63);
tcg_gen_shl_i64(d, a, t);
tcg_temp_free_i64(t);
}
void tcg_gen_gvec_shlv(unsigned vece, uint32_t dofs, uint32_t aofs,
uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
{
static const TCGOpcode vecop_list[] = { INDEX_op_shlv_vec, 0 };
static const GVecGen3 g[4] = {
{ .fniv = tcg_gen_shlv_mod_vec,
.fno = gen_helper_gvec_shl8v,
.opt_opc = vecop_list,
.vece = MO_8 },
{ .fniv = tcg_gen_shlv_mod_vec,
.fno = gen_helper_gvec_shl16v,
.opt_opc = vecop_list,
.vece = MO_16 },
{ .fni4 = tcg_gen_shl_mod_i32,
.fniv = tcg_gen_shlv_mod_vec,
.fno = gen_helper_gvec_shl32v,
.opt_opc = vecop_list,
.vece = MO_32 },
{ .fni8 = tcg_gen_shl_mod_i64,
.fniv = tcg_gen_shlv_mod_vec,
.fno = gen_helper_gvec_shl64v,
.opt_opc = vecop_list,
.prefer_i64 = TCG_TARGET_REG_BITS == 64,
.vece = MO_64 },
};
tcg_debug_assert(vece <= MO_64);
tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
}
/*
* Similarly for logical right shifts.
*/
static void tcg_gen_shrv_mod_vec(unsigned vece, TCGv_vec d,
TCGv_vec a, TCGv_vec b)
{
TCGv_vec t = tcg_temp_new_vec_matching(d);
tcg_gen_dupi_vec(vece, t, (8 << vece) - 1);
tcg_gen_and_vec(vece, t, t, b);
tcg_gen_shrv_vec(vece, d, a, t);
tcg_temp_free_vec(t);
}
static void tcg_gen_shr_mod_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
{
TCGv_i32 t = tcg_temp_new_i32();
tcg_gen_andi_i32(t, b, 31);
tcg_gen_shr_i32(d, a, t);
tcg_temp_free_i32(t);
}
static void tcg_gen_shr_mod_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
{
TCGv_i64 t = tcg_temp_new_i64();
tcg_gen_andi_i64(t, b, 63);
tcg_gen_shr_i64(d, a, t);
tcg_temp_free_i64(t);
}
void tcg_gen_gvec_shrv(unsigned vece, uint32_t dofs, uint32_t aofs,
uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
{
static const TCGOpcode vecop_list[] = { INDEX_op_shrv_vec, 0 };
static const GVecGen3 g[4] = {
{ .fniv = tcg_gen_shrv_mod_vec,
.fno = gen_helper_gvec_shr8v,
.opt_opc = vecop_list,
.vece = MO_8 },
{ .fniv = tcg_gen_shrv_mod_vec,
.fno = gen_helper_gvec_shr16v,
.opt_opc = vecop_list,
.vece = MO_16 },
{ .fni4 = tcg_gen_shr_mod_i32,
.fniv = tcg_gen_shrv_mod_vec,
.fno = gen_helper_gvec_shr32v,
.opt_opc = vecop_list,
.vece = MO_32 },
{ .fni8 = tcg_gen_shr_mod_i64,
.fniv = tcg_gen_shrv_mod_vec,
.fno = gen_helper_gvec_shr64v,
.opt_opc = vecop_list,
.prefer_i64 = TCG_TARGET_REG_BITS == 64,
.vece = MO_64 },
};
tcg_debug_assert(vece <= MO_64);
tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
}
/*
* Similarly for arithmetic right shifts.
*/
static void tcg_gen_sarv_mod_vec(unsigned vece, TCGv_vec d,
TCGv_vec a, TCGv_vec b)
{
TCGv_vec t = tcg_temp_new_vec_matching(d);
tcg_gen_dupi_vec(vece, t, (8 << vece) - 1);
tcg_gen_and_vec(vece, t, t, b);
tcg_gen_sarv_vec(vece, d, a, t);
tcg_temp_free_vec(t);
}
static void tcg_gen_sar_mod_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
{
TCGv_i32 t = tcg_temp_new_i32();
tcg_gen_andi_i32(t, b, 31);
tcg_gen_sar_i32(d, a, t);
tcg_temp_free_i32(t);
}
static void tcg_gen_sar_mod_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
{
TCGv_i64 t = tcg_temp_new_i64();
tcg_gen_andi_i64(t, b, 63);
tcg_gen_sar_i64(d, a, t);
tcg_temp_free_i64(t);
}
void tcg_gen_gvec_sarv(unsigned vece, uint32_t dofs, uint32_t aofs,
uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
{
static const TCGOpcode vecop_list[] = { INDEX_op_sarv_vec, 0 };
static const GVecGen3 g[4] = {
{ .fniv = tcg_gen_sarv_mod_vec,
.fno = gen_helper_gvec_sar8v,
.opt_opc = vecop_list,
.vece = MO_8 },
{ .fniv = tcg_gen_sarv_mod_vec,
.fno = gen_helper_gvec_sar16v,
.opt_opc = vecop_list,
.vece = MO_16 },
{ .fni4 = tcg_gen_sar_mod_i32,
.fniv = tcg_gen_sarv_mod_vec,
.fno = gen_helper_gvec_sar32v,
.opt_opc = vecop_list,
.vece = MO_32 },
{ .fni8 = tcg_gen_sar_mod_i64,
.fniv = tcg_gen_sarv_mod_vec,
.fno = gen_helper_gvec_sar64v,
.opt_opc = vecop_list,
.prefer_i64 = TCG_TARGET_REG_BITS == 64,
.vece = MO_64 },
};
tcg_debug_assert(vece <= MO_64);
tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
}
/* Expand OPSZ bytes worth of three-operand operations using i32 elements. */ /* Expand OPSZ bytes worth of three-operand operations using i32 elements. */
static void expand_cmp_i32(uint32_t dofs, uint32_t aofs, uint32_t bofs, static void expand_cmp_i32(uint32_t dofs, uint32_t aofs, uint32_t bofs,
uint32_t oprsz, TCGCond cond) uint32_t oprsz, TCGCond cond)

View File

@ -318,6 +318,17 @@ void tcg_gen_gvec_shri(unsigned vece, uint32_t dofs, uint32_t aofs,
void tcg_gen_gvec_sari(unsigned vece, uint32_t dofs, uint32_t aofs, void tcg_gen_gvec_sari(unsigned vece, uint32_t dofs, uint32_t aofs,
int64_t shift, uint32_t oprsz, uint32_t maxsz); int64_t shift, uint32_t oprsz, uint32_t maxsz);
/*
* Perform vector shift by vector element, modulo the element size.
* E.g. D[i] = A[i] << (B[i] % (8 << vece)).
*/
void tcg_gen_gvec_shlv(unsigned vece, uint32_t dofs, uint32_t aofs,
uint32_t bofs, uint32_t oprsz, uint32_t maxsz);
void tcg_gen_gvec_shrv(unsigned vece, uint32_t dofs, uint32_t aofs,
uint32_t bofs, uint32_t oprsz, uint32_t maxsz);
void tcg_gen_gvec_sarv(unsigned vece, uint32_t dofs, uint32_t aofs,
uint32_t bofs, uint32_t oprsz, uint32_t maxsz);
void tcg_gen_gvec_cmp(TCGCond cond, unsigned vece, uint32_t dofs, void tcg_gen_gvec_cmp(TCGCond cond, unsigned vece, uint32_t dofs,
uint32_t aofs, uint32_t bofs, uint32_t aofs, uint32_t bofs,
uint32_t oprsz, uint32_t maxsz); uint32_t oprsz, uint32_t maxsz);

View File

@ -583,3 +583,18 @@ void tcg_gen_umax_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b)
{ {
do_op3(vece, r, a, b, INDEX_op_umax_vec); do_op3(vece, r, a, b, INDEX_op_umax_vec);
} }
void tcg_gen_shlv_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b)
{
do_op3(vece, r, a, b, INDEX_op_shlv_vec);
}
void tcg_gen_shrv_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b)
{
do_op3(vece, r, a, b, INDEX_op_shrv_vec);
}
void tcg_gen_sarv_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b)
{
do_op3(vece, r, a, b, INDEX_op_sarv_vec);
}

View File

@ -986,6 +986,10 @@ void tcg_gen_shli_vec(unsigned vece, TCGv_vec r, TCGv_vec a, int64_t i);
void tcg_gen_shri_vec(unsigned vece, TCGv_vec r, TCGv_vec a, int64_t i); void tcg_gen_shri_vec(unsigned vece, TCGv_vec r, TCGv_vec a, int64_t i);
void tcg_gen_sari_vec(unsigned vece, TCGv_vec r, TCGv_vec a, int64_t i); void tcg_gen_sari_vec(unsigned vece, TCGv_vec r, TCGv_vec a, int64_t i);
void tcg_gen_shlv_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec s);
void tcg_gen_shrv_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec s);
void tcg_gen_sarv_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec s);
void tcg_gen_cmp_vec(TCGCond cond, unsigned vece, TCGv_vec r, void tcg_gen_cmp_vec(TCGCond cond, unsigned vece, TCGv_vec r,
TCGv_vec a, TCGv_vec b); TCGv_vec a, TCGv_vec b);