[llvm] e7e204a - [X86][AVX] Attempt to lower v16i32/v16f32 shuffles with lowerShuffleAsRepeatedMaskAndLanePermute

[Simon Pilgrim via llvm-commits]

Author: Simon Pilgrim
Date: 2020-06-23T18:33:50+01:00
New Revision: e7e204a373683a09978be2b1df8783ba26fe43a4

URL: https://github.com/llvm/llvm-project/commit/e7e204a373683a09978be2b1df8783ba26fe43a4
DIFF: https://github.com/llvm/llvm-project/commit/e7e204a373683a09978be2b1df8783ba26fe43a4.diff

LOG: [X86][AVX] Attempt to lower v16i32/v16f32 shuffles with lowerShuffleAsRepeatedMaskAndLanePermute

Avoids prematurely creating permps/permd variable shuffles.

Fixes PR46249

Added: 
    

Modified: 
    llvm/lib/Target/X86/X86ISelLowering.cpp
    llvm/test/CodeGen/X86/vector-shuffle-512-v16.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index a6bb4155e130..c3e89854bc48 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -17151,6 +17151,12 @@ static SDValue lowerV16F32Shuffle(const SDLoc &DL, ArrayRef<int> Mask,
     return lowerShuffleWithSHUFPS(DL, MVT::v16f32, RepeatedMask, V1, V2, DAG);
   }
 
+  // Try to create an in-lane repeating shuffle mask and then shuffle the
+  // results into the target lanes.
+  if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute(
+          DL, MVT::v16f32, V1, V2, Mask, Subtarget, DAG))
+    return V;
+
   // If we have a single input shuffle with 
diff erent shuffle patterns in the
   // 128-bit lanes and don't lane cross, use variable mask VPERMILPS.
   if (V2.isUndef() &&
@@ -17288,6 +17294,13 @@ static SDValue lowerV16I32Shuffle(const SDLoc &DL, ArrayRef<int> Mask,
                                             CastV1, CastV2, DAG);
     return DAG.getBitcast(MVT::v16i32, ShufPS);
   }
+
+  // Try to create an in-lane repeating shuffle mask and then shuffle the
+  // results into the target lanes.
+  if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute(
+          DL, MVT::v16i32, V1, V2, Mask, Subtarget, DAG))
+    return V;
+
   // If we have AVX512F support, we can use VEXPAND.
   if (SDValue V = lowerShuffleToEXPAND(DL, MVT::v16i32, Zeroable, Mask, V1, V2,
                                        DAG, Subtarget))
@@ -17296,6 +17309,7 @@ static SDValue lowerV16I32Shuffle(const SDLoc &DL, ArrayRef<int> Mask,
   if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v16i32, V1, V2, Mask,
                                           Zeroable, Subtarget, DAG))
     return Blend;
+
   return lowerShuffleWithPERMV(DL, MVT::v16i32, Mask, V1, V2, DAG);
 }
 

diff  --git a/llvm/test/CodeGen/X86/vector-shuffle-512-v16.ll b/llvm/test/CodeGen/X86/vector-shuffle-512-v16.ll
index be79b96a9b64..bcd3a7b85150 100644
--- a/llvm/test/CodeGen/X86/vector-shuffle-512-v16.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-512-v16.ll
@@ -264,8 +264,8 @@ define <16 x float> @shuffle_v16f32_0f_1f_0e_16_0d_1d_04_1e_0b_1b_0a_1a_09_19_08
 define <16 x i32> @shuffle_v16i32_0b_0a_09_08_0f_0e_0d_0c_03_02_01_00_07_06_05_04(<16 x i32> %a) {
 ; ALL-LABEL: shuffle_v16i32_0b_0a_09_08_0f_0e_0d_0c_03_02_01_00_07_06_05_04:
 ; ALL:       # %bb.0:
-; ALL-NEXT:    vmovaps {{.*#+}} zmm1 = [11,10,9,8,15,14,13,12,3,2,1,0,7,6,5,4]
-; ALL-NEXT:    vpermps %zmm0, %zmm1, %zmm0
+; ALL-NEXT:    vpshufd {{.*#+}} zmm0 = zmm0[3,2,1,0,7,6,5,4,11,10,9,8,15,14,13,12]
+; ALL-NEXT:    vshufi64x2 {{.*#+}} zmm0 = zmm0[4,5,6,7,0,1,2,3]
 ; ALL-NEXT:    retq
   %1 = shufflevector <16 x i32> %a, <16 x i32> undef, <16 x i32> <i32 11, i32 10, i32 9, i32 8, i32 15, i32 14, i32 13, i32 12, i32 3, i32 2, i32 1, i32 0, i32 7, i32 6, i32 5, i32 4>
   ret <16 x i32> %1
@@ -274,8 +274,8 @@ define <16 x i32> @shuffle_v16i32_0b_0a_09_08_0f_0e_0d_0c_03_02_01_00_07_06_05_0
 define <16 x float> @shuffle_v16f32_0b_0a_09_08_0f_0e_0d_0c_03_02_01_00_07_06_05_04(<16 x float> %a) {
 ; ALL-LABEL: shuffle_v16f32_0b_0a_09_08_0f_0e_0d_0c_03_02_01_00_07_06_05_04:
 ; ALL:       # %bb.0:
-; ALL-NEXT:    vmovaps {{.*#+}} zmm1 = [11,10,9,8,15,14,13,12,3,2,1,0,7,6,5,4]
-; ALL-NEXT:    vpermps %zmm0, %zmm1, %zmm0
+; ALL-NEXT:    vpermilps {{.*#+}} zmm0 = zmm0[3,2,1,0,7,6,5,4,11,10,9,8,15,14,13,12]
+; ALL-NEXT:    vshuff64x2 {{.*#+}} zmm0 = zmm0[4,5,6,7,0,1,2,3]
 ; ALL-NEXT:    retq
   %1 = shufflevector <16 x float> %a, <16 x float> undef, <16 x i32> <i32 11, i32 10, i32 9, i32 8, i32 15, i32 14, i32 13, i32 12, i32 3, i32 2, i32 1, i32 0, i32 7, i32 6, i32 5, i32 4>
   ret <16 x float> %1