[x86] add more tests for lowerShuffleWithUndefHalf; NFC

Some other transform is creating the opposite form and causing 
an infinite loop if we try to split some of these.

llvm-svn: 352327

1 files changed, 90 insertions, 0 deletions

diff --git a/llvm/test/CodeGen/X86/vector-shuffle-256-v8.ll b/llvm/test/CodeGen/X86/vector-shuffle-256-v8.ll
index b2187397bb1..1404a1d7e2b 100644
--- a/llvm/test/CodeGen/X86/vector-shuffle-256-v8.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-256-v8.ll
@@ -2929,6 +2929,96 @@ define <8 x float> @unpckh_v8f32(<8 x float> %x, <8 x float> %y) {
   ret <8 x float> %unpckh
 }
 
+; FIXME: AVX1 lowering is better than AVX2 (and AVX512?)
+; Alternate form of the above - make sure we don't have conflicting transforms.
+
+define <8 x i32> @blend_perm_v8i32(<8 x i32> %x, <8 x i32> %y) {
+; AVX1-LABEL: blend_perm_v8i32:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm1
+; AVX1-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: blend_perm_v8i32:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
+; AVX2-NEXT:    vmovaps {{.*#+}} ymm1 = <2,6,3,7,u,u,u,u>
+; AVX2-NEXT:    vpermps %ymm0, %ymm1, %ymm0
+; AVX2-NEXT:    retq
+;
+; AVX512VL-LABEL: blend_perm_v8i32:
+; AVX512VL:       # %bb.0:
+; AVX512VL-NEXT:    vmovdqa {{.*#+}} ymm2 = <2,14,3,15,u,u,u,u>
+; AVX512VL-NEXT:    vpermt2d %ymm1, %ymm2, %ymm0
+; AVX512VL-NEXT:    retq
+  %unpckh = shufflevector <8 x i32> %x, <8 x i32> %y, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 14, i32 15>
+  %r = shufflevector <8 x i32> %unpckh, <8 x i32> undef, <8 x i32> <i32 2, i32 6, i32 3, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>
+  ret <8 x i32> %r
+}
+
+; FIXME: Same as above but with floats. AVX1 lowering is better than AVX2 (and AVX512?)
+
+define <8 x float> @blend_perm_v8f32(<8 x float> %x, <8 x float> %y) {
+; AVX1-LABEL: blend_perm_v8f32:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm1
+; AVX1-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: blend_perm_v8f32:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
+; AVX2-NEXT:    vmovaps {{.*#+}} ymm1 = <2,6,3,7,u,u,u,u>
+; AVX2-NEXT:    vpermps %ymm0, %ymm1, %ymm0
+; AVX2-NEXT:    retq
+;
+; AVX512VL-LABEL: blend_perm_v8f32:
+; AVX512VL:       # %bb.0:
+; AVX512VL-NEXT:    vmovaps {{.*#+}} ymm2 = <2,14,3,15,u,u,u,u>
+; AVX512VL-NEXT:    vpermt2ps %ymm1, %ymm2, %ymm0
+; AVX512VL-NEXT:    retq
+  %unpckh = shufflevector <8 x float> %x, <8 x float> %y, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 14, i32 15>
+  %r = shufflevector <8 x float> %unpckh, <8 x float> undef, <8 x i32> <i32 2, i32 6, i32 3, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>
+  ret <8 x float> %r
+}
+
+; FIXME: AVX1 lowering is better than AVX2/AVX512.
+; Another variation of the above - make sure we don't have conflicting transforms.
+
+define <8 x i32> @unpckh_v8i32_unary(<8 x i32> %x) {
+; AVX1-LABEL: unpckh_v8i32_unary:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; AVX1-NEXT:    retq
+;
+; AVX2OR512VL-LABEL: unpckh_v8i32_unary:
+; AVX2OR512VL:       # %bb.0:
+; AVX2OR512VL-NEXT:    vmovaps {{.*#+}} ymm1 = <2,6,3,7,u,u,u,u>
+; AVX2OR512VL-NEXT:    vpermps %ymm0, %ymm1, %ymm0
+; AVX2OR512VL-NEXT:    retq
+  %r = shufflevector <8 x i32> %x, <8 x i32> undef, <8 x i32> <i32 2, i32 6, i32 3, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>
+  ret <8 x i32> %r
+}
+
+; FIXME: Same as above but with floats. AVX1 lowering is better than AVX2/AVX512.
+
+define <8 x float> @unpckh_v8f32_unary(<8 x float> %x) {
+; AVX1-LABEL: unpckh_v8f32_unary:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; AVX1-NEXT:    retq
+;
+; AVX2OR512VL-LABEL: unpckh_v8f32_unary:
+; AVX2OR512VL:       # %bb.0:
+; AVX2OR512VL-NEXT:    vmovaps {{.*#+}} ymm1 = <2,6,3,7,u,u,u,u>
+; AVX2OR512VL-NEXT:    vpermps %ymm0, %ymm1, %ymm0
+; AVX2OR512VL-NEXT:    retq
+  %r = shufflevector <8 x float> %x, <8 x float> undef, <8 x i32> <i32 2, i32 6, i32 3, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>
+  ret <8 x float> %r
+}
+
 ; FIXME: Why are integer and FP (below) lowering different for AVX1?
 
 define <8 x i32> @lowhalf_v8i32(<8 x i32> %x, <8 x i32> %y) {