[X86][SSE] Enable initial support for domain crossing at high shuffle combine depths.

As discussed on D27692, this permits another domain to be used to combine a shuffle at high depths.

We currently set the required depth at 4 or more combined shuffles, this is probably too high for most targets but is a good starting point and already helps avoid a number of costly variable shuffles.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295608 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 88651a2d22790679186748192e11c220dbbd51c0..4d660acbc20a407251add58b495e7a4260ae97eb 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -26971,9 +26971,9 @@ static bool combineX86ShuffleChain(ArrayRef<SDValue> Inputs, SDValue Root,
   unsigned Shuffle, PermuteImm;
 
   // Which shuffle domains are permitted?
-  // TODO - Allow either domain after a threshold depth.
-  bool AllowFloatDomain = FloatDomain;
-  bool AllowIntDomain = !FloatDomain;
+  // Permit domain crossing at higher combine depths.
+  bool AllowFloatDomain = FloatDomain || (Depth > 3);
+  bool AllowIntDomain = !FloatDomain || (Depth > 3);
 
   if (UnaryShuffle) {
     // If we are shuffling a X86ISD::VZEXT_LOAD then we can use the load

diff --git a/test/CodeGen/X86/shuffle-vs-trunc-512.ll b/test/CodeGen/X86/shuffle-vs-trunc-512.ll
index 8f9b8c156d3c1ec677dae77dd1b50cf983dae228..aaf1ffe8adc03beacaa5c9e46c5e64d8eb77e61b 100644 (file)
--- a/test/CodeGen/X86/shuffle-vs-trunc-512.ll
+++ b/test/CodeGen/X86/shuffle-vs-trunc-512.ll
@@ -100,11 +100,11 @@ define void @trunc_v32i16_to_v32i8(<64 x i8>* %L, <32 x i8>* %S) nounwind {
 define void @shuffle_v32i16_to_v16i16(<32 x i16>* %L, <16 x i16>* %S) nounwind {
 ; AVX512F-LABEL: shuffle_v32i16_to_v16i16:
 ; AVX512F:       # BB#0:
-; AVX512F-NEXT:    vmovdqa (%rdi), %ymm0
-; AVX512F-NEXT:    vmovdqa 32(%rdi), %ymm1
-; AVX512F-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[0,1,4,5,4,5,6,7,0,1,4,5,8,9,12,13,16,17,20,21,20,21,22,23,16,17,20,21,24,25,28,29]
-; AVX512F-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15,16,17,20,21,24,25,28,29,24,25,28,29,28,29,30,31]
-; AVX512F-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
+; AVX512F-NEXT:    vpshuflw {{.*#+}} ymm0 = mem[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512F-NEXT:    vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]
+; AVX512F-NEXT:    vpshuflw {{.*#+}} ymm1 = mem[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512F-NEXT:    vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]
+; AVX512F-NEXT:    vshufps {{.*#+}} ymm0 = ymm1[0,2],ymm0[0,2],ymm1[4,6],ymm0[4,6]
 ; AVX512F-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX512F-NEXT:    vmovdqa %ymm0, (%rsi)
 ; AVX512F-NEXT:    retq
@@ -113,11 +113,9 @@ define void @shuffle_v32i16_to_v16i16(<32 x i16>* %L, <16 x i16>* %S) nounwind {
 ; AVX512VL:       # BB#0:
 ; AVX512VL-NEXT:    vpshuflw {{.*#+}} ymm0 = mem[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
 ; AVX512VL-NEXT:    vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]
-; AVX512VL-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[0,1,0,2,4,5,4,6]
 ; AVX512VL-NEXT:    vpshuflw {{.*#+}} ymm1 = mem[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
 ; AVX512VL-NEXT:    vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]
-; AVX512VL-NEXT:    vpshufd {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7]
-; AVX512VL-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5],ymm0[6,7]
+; AVX512VL-NEXT:    vshufps {{.*#+}} ymm0 = ymm1[0,2],ymm0[0,2],ymm1[4,6],ymm0[4,6]
 ; AVX512VL-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX512VL-NEXT:    vmovdqa %ymm0, (%rsi)
 ; AVX512VL-NEXT:    retq
@@ -126,9 +124,11 @@ define void @shuffle_v32i16_to_v16i16(<32 x i16>* %L, <16 x i16>* %S) nounwind {
 ; AVX512BW:       # BB#0:
 ; AVX512BW-NEXT:    vmovdqu16 (%rdi), %zmm0
 ; AVX512BW-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[0,1,4,5,4,5,6,7,0,1,4,5,8,9,12,13,16,17,20,21,20,21,22,23,16,17,20,21,24,25,28,29]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15,16,17,20,21,24,25,28,29,24,25,28,29,28,29,30,31]
-; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
+; AVX512BW-NEXT:    vpshuflw {{.*#+}} ymm1 = ymm1[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512BW-NEXT:    vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]
+; AVX512BW-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm0[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512BW-NEXT:    vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]
+; AVX512BW-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[0,2],ymm1[0,2],ymm0[4,6],ymm1[4,6]
 ; AVX512BW-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX512BW-NEXT:    vmovdqa %ymm0, (%rsi)
 ; AVX512BW-NEXT:    retq