[llvm] df5ea2b - [X86] Add shuffles showing failure to use PERMUTE(BLEND(X,Y))

[Simon Pilgrim via llvm-commits]

Author: Simon Pilgrim
Date: 2022-05-15T14:31:32+01:00
New Revision: df5ea2b12d482986269b74f7c9d2ccb8c5c90676

URL: https://github.com/llvm/llvm-project/commit/df5ea2b12d482986269b74f7c9d2ccb8c5c90676
DIFF: https://github.com/llvm/llvm-project/commit/df5ea2b12d482986269b74f7c9d2ccb8c5c90676.diff

LOG: [X86] Add shuffles showing failure to use PERMUTE(BLEND(X,Y))

One AVX2+ targets we have a immediate VPERMQ/PD cross-lane permute thats better than relying on a pair of VPERM2F128 cross-lanes to feed a blend.

Reported on discourse

Added: 
    

Modified: 
    llvm/test/CodeGen/X86/vector-shuffle-256-v4.ll

Removed: 
    


################################################################################
diff  --git a/llvm/test/CodeGen/X86/vector-shuffle-256-v4.ll b/llvm/test/CodeGen/X86/vector-shuffle-256-v4.ll
index 6bebb91b7afe..5bbc6ae15908 100644
--- a/llvm/test/CodeGen/X86/vector-shuffle-256-v4.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-256-v4.ll
@@ -576,6 +576,23 @@ define <4 x double> @shuffle_v4f64_0415(<4 x double> %a, <4 x double> %b) {
   ret <4 x double> %shuffle
 }
 
+define <4 x double> @shuffle_v4f64_2741(<4 x double> %a, <4 x double> %b) {
+; AVX1OR2-LABEL: shuffle_v4f64_2741:
+; AVX1OR2:       # %bb.0:
+; AVX1OR2-NEXT:    vperm2f128 {{.*#+}} ymm2 = ymm1[2,3],ymm0[0,1]
+; AVX1OR2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[0,1]
+; AVX1OR2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5],ymm2[6,7]
+; AVX1OR2-NEXT:    retq
+;
+; AVX512VL-LABEL: shuffle_v4f64_2741:
+; AVX512VL:       # %bb.0:
+; AVX512VL-NEXT:    vmovapd {{.*#+}} ymm2 = [2,7,4,1]
+; AVX512VL-NEXT:    vpermt2pd %ymm1, %ymm2, %ymm0
+; AVX512VL-NEXT:    retq
+  %shuffle = shufflevector <4 x double> %a, <4 x double> %b, <4 x i32> <i32 2, i32 7, i32 4, i32 1>
+  ret <4 x double> %shuffle
+}
+
 define <4 x double> @shuffle_v4f64_u062(<4 x double> %a, <4 x double> %b) {
 ; ALL-LABEL: shuffle_v4f64_u062:
 ; ALL:       # %bb.0:
@@ -1361,6 +1378,23 @@ define <4 x i64> @shuffle_v4i64_0415(<4 x i64> %a, <4 x i64> %b) {
   ret <4 x i64> %shuffle
 }
 
+define <4 x i64> @shuffle_v4i64_2741(<4 x i64> %a, <4 x i64> %b) {
+; AVX1OR2-LABEL: shuffle_v4i64_2741:
+; AVX1OR2:       # %bb.0:
+; AVX1OR2-NEXT:    vperm2f128 {{.*#+}} ymm2 = ymm1[2,3],ymm0[0,1]
+; AVX1OR2-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[0,1]
+; AVX1OR2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5],ymm2[6,7]
+; AVX1OR2-NEXT:    retq
+;
+; AVX512VL-LABEL: shuffle_v4i64_2741:
+; AVX512VL:       # %bb.0:
+; AVX512VL-NEXT:    vmovdqa {{.*#+}} ymm2 = [2,7,4,1]
+; AVX512VL-NEXT:    vpermt2q %ymm1, %ymm2, %ymm0
+; AVX512VL-NEXT:    retq
+  %shuffle = shufflevector <4 x i64> %a, <4 x i64> %b, <4 x i32> <i32 2, i32 7, i32 4, i32 1>
+  ret <4 x i64> %shuffle
+}
+
 define <4 x i64> @shuffle_v4i64_z4z6(<4 x i64> %a) {
 ; AVX1-LABEL: shuffle_v4i64_z4z6:
 ; AVX1:       # %bb.0: