[llvm] r290865 - [AVX-512] Add support for pushing bitcasts through INSERT_SUBVEC in order to select a masked operation.

[Craig Topper via llvm-commits]

Author: ctopper
Date: Mon Jan  2 23:46:02 2017
New Revision: 290865

URL: http://llvm.org/viewvc/llvm-project?rev=290865&view=rev
Log:
[AVX-512] Add support for pushing bitcasts through INSERT_SUBVEC in order to select a masked operation.

Modified:
    llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
    llvm/trunk/test/CodeGen/X86/avx512vl-intrinsics-upgrade.ll

Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=290865&r1=290864&r2=290865&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Mon Jan  2 23:46:02 2017
@@ -28685,6 +28685,29 @@ static bool combineBitcastForMaskedOp(SD
     return BitcastAndCombineShuffle(Opcode, Op.getOperand(0), Op.getOperand(1),
                                     Op.getOperand(2));
   }
+  case ISD::INSERT_SUBVECTOR: {
+    unsigned EltSize = EltVT.getSizeInBits();
+    if (EltSize != 32 && EltSize != 64)
+      return false;
+    MVT OpEltVT = Op.getSimpleValueType().getVectorElementType();
+    // Only change element size, not type.
+    if (VT.isInteger() != OpEltVT.isInteger())
+      return false;
+    uint64_t Imm = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
+    Imm = (Imm * OpEltVT.getSizeInBits()) / EltSize;
+    SDValue Op0 = DAG.getBitcast(VT, Op.getOperand(0));
+    DCI.AddToWorklist(Op0.getNode());
+    // Op1 needs to be bitcasted to a smaller vector with the same element type.
+    SDValue Op1 = Op.getOperand(1);
+    MVT Op1VT = MVT::getVectorVT(EltVT,
+                            Op1.getSimpleValueType().getSizeInBits() / EltSize);
+    Op1 = DAG.getBitcast(Op1VT, Op1);
+    DCI.AddToWorklist(Op1.getNode());
+    DCI.CombineTo(OrigOp.getNode(),
+                  DAG.getNode(Opcode, DL, VT, Op0, Op1,
+                              DAG.getConstant(Imm, DL, MVT::i8)));
+    return true;
+  }
   }
 
   return false;

Modified: llvm/trunk/test/CodeGen/X86/avx512vl-intrinsics-upgrade.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/avx512vl-intrinsics-upgrade.ll?rev=290865&r1=290864&r2=290865&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/avx512vl-intrinsics-upgrade.ll (original)
+++ llvm/trunk/test/CodeGen/X86/avx512vl-intrinsics-upgrade.ll Mon Jan  2 23:46:02 2017
@@ -4798,12 +4798,12 @@ declare <8 x float> @llvm.x86.avx512.mas
 define <8 x float>@test_int_x86_avx512_mask_insertf32x4_256(<8 x float> %x0, <4 x float> %x1, <8 x float> %x3, i8 %x4) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_insertf32x4_256:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vinsertf32x4 $1, %xmm1, %ymm0, %ymm0 ## encoding: [0x62,0xf3,0x7d,0x28,0x18,0xc1,0x01]
+; CHECK-NEXT:    vinsertf32x4 $1, %xmm1, %ymm0, %ymm3 ## encoding: [0x62,0xf3,0x7d,0x28,0x18,0xd9,0x01]
 ; CHECK-NEXT:    kmovw %edi, %k1 ## encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vblendmps %ymm0, %ymm2, %ymm1 {%k1} ## encoding: [0x62,0xf2,0x6d,0x29,0x65,0xc8]
-; CHECK-NEXT:    vmovaps %ymm0, %ymm2 {%k1} {z} ## encoding: [0x62,0xf1,0x7c,0xa9,0x28,0xd0]
-; CHECK-NEXT:    vaddps %ymm0, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf4,0x58,0xc0]
-; CHECK-NEXT:    vaddps %ymm0, %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xec,0x58,0xc0]
+; CHECK-NEXT:    vinsertf32x4 $1, %xmm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf3,0x7d,0x29,0x18,0xd1,0x01]
+; CHECK-NEXT:    vinsertf32x4 $1, %xmm1, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf3,0x7d,0xa9,0x18,0xc1,0x01]
+; CHECK-NEXT:    vaddps %ymm3, %ymm2, %ymm1 ## EVEX TO VEX Compression encoding: [0xc5,0xec,0x58,0xcb]
+; CHECK-NEXT:    vaddps %ymm1, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfc,0x58,0xc1]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
   %res = call <8 x float> @llvm.x86.avx512.mask.insertf32x4.256(<8 x float> %x0, <4 x float> %x1, i32 1, <8 x float> %x3, i8 %x4)
   %res1 = call <8 x float> @llvm.x86.avx512.mask.insertf32x4.256(<8 x float> %x0, <4 x float> %x1, i32 1, <8 x float> %x3, i8 -1)
@@ -4818,12 +4818,12 @@ declare <8 x i32> @llvm.x86.avx512.mask.
 define <8 x i32>@test_int_x86_avx512_mask_inserti32x4_256(<8 x i32> %x0, <4 x i32> %x1, <8 x i32> %x3, i8 %x4) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_inserti32x4_256:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vinserti32x4 $1, %xmm1, %ymm0, %ymm0 ## encoding: [0x62,0xf3,0x7d,0x28,0x38,0xc1,0x01]
+; CHECK-NEXT:    vinserti32x4 $1, %xmm1, %ymm0, %ymm3 ## encoding: [0x62,0xf3,0x7d,0x28,0x38,0xd9,0x01]
 ; CHECK-NEXT:    kmovw %edi, %k1 ## encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vpblendmd %ymm0, %ymm2, %ymm1 {%k1} ## encoding: [0x62,0xf2,0x6d,0x29,0x64,0xc8]
-; CHECK-NEXT:    vmovdqa32 %ymm0, %ymm2 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xa9,0x6f,0xd0]
-; CHECK-NEXT:    vpaddd %ymm0, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf5,0xfe,0xc0]
-; CHECK-NEXT:    vpaddd %ymm0, %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0xfe,0xc0]
+; CHECK-NEXT:    vinserti32x4 $1, %xmm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf3,0x7d,0x29,0x38,0xd1,0x01]
+; CHECK-NEXT:    vinserti32x4 $1, %xmm1, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf3,0x7d,0xa9,0x38,0xc1,0x01]
+; CHECK-NEXT:    vpaddd %ymm3, %ymm2, %ymm1 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0xfe,0xcb]
+; CHECK-NEXT:    vpaddd %ymm1, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xfe,0xc1]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
 
   %res = call <8 x i32> @llvm.x86.avx512.mask.inserti32x4.256(<8 x i32> %x0, <4 x i32> %x1, i32 1, <8 x i32> %x3, i8 %x4)