> -----Original Message-----
> From: Jakub Jelinek <[email protected]>
> Sent: Tuesday, May 19, 2026 3:43 PM
> To: Uros Bizjak <[email protected]>; Liu, Hongtao <[email protected]>
> Cc: [email protected]
> Subject: [PATCH] i386: Use vpermilps for some non-const permutations
> [PR125357]
>
> Hi!
>
> We don't use vpermilps insn for V4S[IF]mode variable permutations on
> TARGET_AVX without TARGET_AVX512*. For TARGET_AVX512* there are
> plenty of permutation instructions already. For TARGET_AVX2, the function
> has special cases for one_operand_shuffle for V8SImode/V8SFmode and
> emits reasonable code, but for V4SImode/V4SFmode with TARGET_AVX2 it
> handles those using V8SImode/V8SFmode as two operand shuffle, which
> requires
> 2 preparation instructions, vpermd and one finalization instruction.
> And for !TARGET_AVX2 && TARGET_AVX we just emit terrible code for these.
>
> So, the following patch uses vpermilps for V4S[IF]mode one_operand_shuffle.
>
> Trying to handle V8S[IF]mode is not worth it, for TARGET_AVX2 we already
> emit good code (see above) and for !TARGET_AVX2 && TARGET_AVX
> V8SImode mask is not valid vector mode, so we emit terrible code no matter
> what.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
Ok.
>
> 2026-05-19 Jakub Jelinek <[email protected]>
>
> PR target/125357
> * config/i386/i386-expand.cc (ix86_expand_vec_perm): For
> one_operand_shuffle if TARGET_AVX and not TARGET_AVX512F use
> vpermilps for V4SImode/V4SFmode. Formatting fix.
>
> * gcc.target/i386/avx-pr125357.c: New test.
> * gcc.target/i386/avx2-pr125357.c: New test.
>
> --- gcc/config/i386/i386-expand.cc.jj 2026-05-18 09:47:52.032945399
> +0200
> +++ gcc/config/i386/i386-expand.cc 2026-05-18 11:29:06.397822007
> +0200
> @@ -5578,7 +5578,7 @@ ix86_expand_vec_perm (rtx operands[])
> switch (mode)
> {
> case E_V16SImode:
> - gen =gen_avx512f_permvarv16si;
> + gen = gen_avx512f_permvarv16si;
> break;
> case E_V16SFmode:
> gen = gen_avx512f_permvarv16sf;
> @@ -5702,6 +5702,8 @@ ix86_expand_vec_perm (rtx operands[])
> return;
>
> case E_V4SImode:
> + if (one_operand_shuffle)
> + break; /* Handled below for TARGET_AVX. */
> /* By combining the two 128-bit input vectors into one 256-bit
> input vector, we can use VPERMD and VPERMPS for the full
> two-operand shuffle. */
> @@ -5714,6 +5716,8 @@ ix86_expand_vec_perm (rtx operands[])
> return;
>
> case E_V4SFmode:
> + if (one_operand_shuffle)
> + break; /* Handled below for TARGET_AVX. */
> t1 = gen_reg_rtx (V8SFmode);
> t2 = gen_reg_rtx (V8SImode);
> mask = gen_lowpart (V4SImode, mask); @@ -5820,6 +5824,22 @@
> ix86_expand_vec_perm (rtx operands[])
> }
> }
>
> + if (TARGET_AVX
> + && one_operand_shuffle
> + && (mode == V4SImode || mode == V4SFmode))
> + {
> + if (mode == V4SImode)
> + {
> + op0 = gen_lowpart (V4SFmode, op0);
> + t1 = gen_reg_rtx (V4SFmode);
> + emit_insn (gen_avx_vpermilvarv4sf3 (t1, op0, mask));
> + emit_move_insn (target, gen_lowpart (mode, t1));
> + }
> + else
> + emit_insn (gen_avx_vpermilvarv4sf3 (target, op0, mask));
> + return;
> + }
> +
> if (TARGET_XOP)
> {
> /* The XOP VPPERM insn supports three inputs. By ignoring the
> --- gcc/testsuite/gcc.target/i386/avx-pr125357.c.jj 2026-05-18
> 11:38:49.440915202 +0200
> +++ gcc/testsuite/gcc.target/i386/avx-pr125357.c 2026-05-18
> 11:38:27.018296200 +0200
> @@ -0,0 +1,19 @@
> +/* PR target/125357 */
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -mavx -mno-avx2 -mno-xop" } */
> +/* { dg-final { scan-assembler-times "\tvpermilps\t" 2 } } */
> +
> +typedef int v4si __attribute__((vector_size (16))); typedef float v4sf
> +__attribute__((vector_size (16)));
> +
> +v4si
> +foo (v4si x, v4si y)
> +{
> + return __builtin_shuffle (x, y);
> +}
> +
> +v4sf
> +bar (v4sf x, v4si y)
> +{
> + return __builtin_shuffle (x, y);
> +}
> --- gcc/testsuite/gcc.target/i386/avx2-pr125357.c.jj 2026-05-18
> 11:39:01.767705755 +0200
> +++ gcc/testsuite/gcc.target/i386/avx2-pr125357.c 2026-05-18
> 11:39:15.817467030 +0200
> @@ -0,0 +1,19 @@
> +/* PR target/125357 */
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -mavx2 -mno-avx512f -mno-xop" } */
> +/* { dg-final { scan-assembler-times "\tvpermilps\t" 2 } } */
> +
> +typedef int v4si __attribute__((vector_size (16))); typedef float v4sf
> +__attribute__((vector_size (16)));
> +
> +v4si
> +foo (v4si x, v4si y)
> +{
> + return __builtin_shuffle (x, y);
> +}
> +
> +v4sf
> +bar (v4sf x, v4si y)
> +{
> + return __builtin_shuffle (x, y);
> +}
>
> Jakub
