On Wed, Oct 1, 2014 at 9:03 PM, Uros Bizjak <[email protected]> wrote:
>> And now the expand_vec_perm_palignr improvement, tested
>> with GCC_TEST_RUN_EXPENSIVE=1 make check-gcc \
>> RUNTESTFLAGS='--target_board=unix/-mavx2 dg-torture.exp=vshuf*.c'
>> E.g.
>> typedef unsigned long long V __attribute__ ((vector_size (32)));
>> extern void abort (void);
>> V a, b, c, d;
>> void test_14 (void)
>> {
>> V mask = { 6, 1, 3, 4 };
>> int i;
>> c = __builtin_shuffle (a, mask);
>> d = __builtin_shuffle (a, b, mask);
>> }
>> (distilled from test 15 in vshuf-v4di.c) results in:
>> - vmovdqa a(%rip), %ymm0
>> - vpermq $54, %ymm0, %ymm1
>> - vpshufb .LC1(%rip), %ymm0, %ymm0
>> - vmovdqa %ymm1, c(%rip)
>> - vmovdqa b(%rip), %ymm1
>> - vpshufb .LC0(%rip), %ymm1, %ymm1
>> - vpermq $78, %ymm1, %ymm1
>> - vpor %ymm1, %ymm0, %ymm0
>> + vmovdqa a(%rip), %ymm1
>> + vpermq $54, %ymm1, %ymm0
>> + vmovdqa %ymm0, c(%rip)
>> + vmovdqa b(%rip), %ymm0
>> + vpalignr $8, %ymm1, %ymm0, %ymm0
>> + vpermq $99, %ymm0, %ymm0
>> vmovdqa %ymm0, d(%rip)
>> vzeroupper
>> ret
>> change (and two fewer .rodata constants).
>
> On a related note, I would like to point out that
> gcc.target/i386/pr61403.c also fails to generate blend insn with
> -mavx2. The new insn sequence includes lots of new vpshufb insns with
> memory access.
Following patch fixes the failure:
--cut here--
Index: i386.c
===================================================================
--- i386.c (revision 215802)
+++ i386.c (working copy)
@@ -43407,8 +43407,10 @@ expand_vec_perm_pblendv (struct expand_vec_perm_d
AVX and AVX2 as they require more than 2 instructions. */
if (d->one_operand_p)
return false;
- if (TARGET_SSE4_1 && GET_MODE_SIZE (vmode) == 16)
+ if (TARGET_AVX2 && GET_MODE_SIZE (vmode) == 32)
;
+ else if (TARGET_SSE4_1 && GET_MODE_SIZE (vmode) == 16)
+ ;
else
return false;
--cut here--
The comment above expand_vec_perm_pblendv claims that:
/* Use the same checks as in expand_vec_perm_blend, but skipping
AVX and AVX2 as they require more than 2 instructions. */
However, I see a significant reduction in vpshufb and vpor
instructions (33->16 and 22->11), and 6 new vblendps insns.
BTW: I have no access to avx2 target, so I can't test the patch with a
runtime tests. OTOH, it doesn't ICE for "GCC_TEST_RUN_EXPENSIVE=1 make
check-gcc RUNTESTFLAGS='--target_board=unix/-mavx2
dg-torture.exp=vshuf*.c'".
Jakub, what do you think?
Uros.
