[x86] narrow extract subvector of vector select

This is a potentially large perf win for AVX1 targets because of the way we
auto-vectorize to 256-bit but then expect the backend to legalize/optimize
for the half-implemented AVX1 ISA.

On the motivating example from PR37428 (even though this patch doesn't solve
the vector shift issue):
https://bugs.llvm.org/show_bug.cgi?id=37428
...there's a 16% speedup when compiling with "-mavx" (perf tested on Haswell)
because we eliminate the remaining 256-bit vblendv ops.

I added comments on a couple of tests that require further work. If we have
256-bit logic ops separating the vselect and extract, we should probably narrow
everything to 128-bit, but that requires a larger pattern match.

Differential Revision: https://reviews.llvm.org/D62969

llvm-svn: 362797

1 files changed, 26 insertions, 36 deletions

diff --git a/llvm/test/CodeGen/X86/vector-reduce-smin-widen.ll b/llvm/test/CodeGen/X86/vector-reduce-smin-widen.ll
index 0a6ada8dd3d..aacf799bbed 100644
--- a/llvm/test/CodeGen/X86/vector-reduce-smin-widen.ll
+++ b/llvm/test/CodeGen/X86/vector-reduce-smin-widen.ll
@@ -148,13 +148,11 @@ define i64 @test_v4i64(<4 x i64> %a0) {
 ; AVX1-LABEL: test_v4i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm3
-; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
-; AVX1-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
+; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm2
+; AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    vpermilps {{.*#+}} xmm1 = xmm0[2,3,0,1]
 ; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm2
-; AVX1-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
+; AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    vmovq %xmm0, %rax
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    retq
@@ -321,18 +319,15 @@ define i64 @test_v8i64(<8 x i64> %a0) {
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm2
 ; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm3
-; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm3
-; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
-; AVX1-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm4
+; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm5
+; AVX1-NEXT:    vblendvpd %xmm5, %xmm0, %xmm1, %xmm0
+; AVX1-NEXT:    vblendvpd %xmm4, %xmm2, %xmm3, %xmm1
 ; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm2
-; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm3
-; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm2
-; AVX1-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
+; AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    vpermilps {{.*#+}} xmm1 = xmm0[2,3,0,1]
 ; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm2
-; AVX1-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
+; AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    vmovq %xmm0, %rax
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    retq
@@ -610,32 +605,27 @@ define i64 @test_v16i64(<16 x i64> %a0) {
 ;
 ; AVX1-LABEL: test_v16i64:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm4
-; AVX1-NEXT:    vextractf128 $1, %ymm3, %xmm5
-; AVX1-NEXT:    vpcmpgtq %xmm4, %xmm5, %xmm4
-; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm3, %xmm5
-; AVX1-NEXT:    vinsertf128 $1, %xmm4, %ymm5, %ymm4
-; AVX1-NEXT:    vblendvpd %ymm4, %ymm1, %ymm3, %ymm1
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
+; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm3, %xmm8
+; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm2, %xmm9
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm11
+; AVX1-NEXT:    vextractf128 $1, %ymm3, %xmm7
+; AVX1-NEXT:    vpcmpgtq %xmm11, %xmm7, %xmm10
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm5
 ; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm4
-; AVX1-NEXT:    vpcmpgtq %xmm3, %xmm4, %xmm3
-; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm2, %xmm4
-; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm4, %ymm3
-; AVX1-NEXT:    vblendvpd %ymm3, %ymm0, %ymm2, %ymm0
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm3
-; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm3
-; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
-; AVX1-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpgtq %xmm5, %xmm4, %xmm6
+; AVX1-NEXT:    vblendvpd %xmm6, %xmm5, %xmm4, %xmm4
+; AVX1-NEXT:    vblendvpd %xmm10, %xmm11, %xmm7, %xmm5
+; AVX1-NEXT:    vpcmpgtq %xmm4, %xmm5, %xmm6
+; AVX1-NEXT:    vblendvpd %xmm9, %xmm0, %xmm2, %xmm0
+; AVX1-NEXT:    vblendvpd %xmm8, %xmm1, %xmm3, %xmm1
+; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm2
+; AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
+; AVX1-NEXT:    vblendvpd %xmm6, %xmm4, %xmm5, %xmm1
 ; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm2
-; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm3
-; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm2
-; AVX1-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
+; AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    vpermilps {{.*#+}} xmm1 = xmm0[2,3,0,1]
 ; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm2
-; AVX1-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
+; AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    vmovq %xmm0, %rax
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    retq