[X86][AVX] lowerShuffleAsLanePermuteAndPermute - incomplete lane shuffle mask

[RKSimon]

Bugzilla Link	40730
Resolution	FIXED
Resolved on	Feb 18, 2019 03:22
Version	trunk
OS	Windows NT
Blocks	#39678
CC	@topperc,@zmodem,@RKSimon,@rotateright
Fixed by commit(s)	r354034,r354117

Extended Description

https://gcc.godbolt.org/z/tLJJE0

define <8 x i32> @​shuffle_v8i32_0dcd3f14(<8 x i32> %a, <8 x i32> %b) {
%shuffle = shufflevector <8 x i32> %a, <8 x i32> %b, <8 x i32> <i32 0, i32 13, i32 12, i32 13, i32 3, i32 15, i32 1, i32 4>
ret <8 x i32> %shuffle
}

define <8 x i32> @​shuffle_v8i32_0dcd3f14_constant(<8 x i32> %a0) {
%res = shufflevector <8 x i32> %a0, <8 x i32> <i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16>, <8 x i32> <i32 0, i32 13, i32 12, i32 13, i32 3, i32 15, i32 1, i32 4>
ret <8 x i32> %res
}

When the shuffle gets lowered, the constant argument gets incorrectly folded. This appears to be due to the shuffle mask that (correctly) lowers to vperm2f128, containing undef elements in the wrong place that allows undef propagation that leads to incorrect constant folding.

shuffle_v8i32_0dcd3f14:
vextractf128 $1, %ymm0, %xmm2
vblendps $1, %xmm2, %xmm0, %xmm2 # xmm2 = xmm2[0],xmm0[1,2,3]
vpermilps $23, %xmm2, %xmm2 # xmm2 = xmm2[3,1,1,0]
vinsertf128 $1, %xmm2, %ymm0, %ymm0
vperm2f128 $17, %ymm0, %ymm1, %ymm1 # ymm1 = ymm1[2,3,2,3]
vpermilpd $4, %ymm1, %ymm1 # ymm1 = ymm1[0,0,3,2]
vblendps $209, %ymm0, %ymm1, %ymm0 # ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4],ymm1[5],ymm0[6,7]
retq

.LCPI0_0:
.quad 60129542157 # 0x‭0000000E0000000D‬
.quad 60129542157 # 0x‭0000000E0000000D‬
.zero 8 <-- INCORRECT - should be 0x‭0000000F00000000
.quad 60129542157 # 0x‭0000000E0000000D‬
shuffle_v8i32_0dcd3f14_constant:
vextractf128 $1, %ymm0, %xmm1
vblendps $1, %xmm1, %xmm0, %xmm1 # xmm1 = xmm1[0],xmm0[1,2,3]
vpermilps $23, %xmm1, %xmm1 # xmm1 = xmm1[3,1,1,0]
vinsertf128 $1, %xmm1, %ymm0, %ymm0
vblendps $46, .LCPI0_0(%rip), %ymm0, %ymm0 # ymm0 = ymm0[0],mem[1,2,3],ymm0[4],mem[5],ymm0[6,7]
retq

Reduced from an internal fuzz test.

[RKSimon]

assigned to @RKSimon

[RKSimon]

The issue looks like it first appeared in rL344446, so is a regression in the 8.0 branch

[RKSimon]

https://gcc.godbolt.org/z/9BAQca contains the generic shuffle as well as the constant folding bugged version

[RKSimon]

Test case added at rL354034

[RKSimon]

That should've been:

.LCPI0_0:
.quad 60129542157 # 0x0000000E0000000D
.quad 60129542157 # 0x0000000E0000000D
.zero 8 <-- INCORRECT - should be 0x0000001000000000
.quad 60129542157 # 0x0000000E0000000D
shuffle_v8i32_0dcd3f14_constant:
vextractf128 $1, %ymm0, %xmm1
vblendps $1, %xmm1, %xmm0, %xmm1 # xmm1 = xmm1[0],xmm0[1,2,3]
vpermilps $23, %xmm1, %xmm1 # xmm1 = xmm1[3,1,1,0]
vinsertf128 $1, %xmm1, %ymm0, %ymm0
vblendps $46, .LCPI0_0(%rip), %ymm0, %ymm0 # ymm0 = ymm0[0],mem[1,2,3],ymm0[4],mem[5],ymm0[6,7]
retq

[RKSimon]

https://reviews.llvm.org/D58237

[RKSimon]

Fixed in trunk at rL354117

@​Hans - please give it a while and then cherry pick r354034 + r354117

[zmodem]

Fixed in trunk at rL354117

@​Hans - please give it a while and then cherry pick r354034 + r354117

Thanks! Merged them together in r354260. Please let me know if there are any follow-ups.