[llvm] r312462 - [X86] Add a combine to recognize when we have two insert subvectors that together write the whole vector, but the starting vector isn't undef.

[Craig Topper via llvm-commits]

Author: ctopper
Date: Sun Sep  3 18:13:36 2017
New Revision: 312462

URL: http://llvm.org/viewvc/llvm-project?rev=312462&view=rev
Log:
[X86] Add a combine to recognize when we have two insert subvectors that together write the whole vector, but the starting vector isn't undef.

In this case we should replace the starting vector with undef.

Modified:
    llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
    llvm/trunk/test/CodeGen/X86/clear_upper_vector_element_bits.ll
    llvm/trunk/test/CodeGen/X86/insertelement-zero.ll

Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=312462&r1=312461&r2=312462&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Sun Sep  3 18:13:36 2017
@@ -35750,6 +35750,18 @@ static SDValue combineInsertSubvector(SD
         return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT,
                            getZeroVector(OpVT, Subtarget, DAG, dl), SubVec2,
                            Vec.getOperand(2));
+
+      // If we are inserting into both halves of the vector, the starting
+      // vector should be undef. If it isn't, make it so. Only do this if the
+      // the early insert has no other uses.
+      // TODO: Should this be a generic DAG combine?
+      if (!Vec.getOperand(0).isUndef() && Vec.hasOneUse()) {
+        Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, DAG.getUNDEF(OpVT),
+                          SubVec2, Vec.getOperand(2));
+        DCI.AddToWorklist(Vec.getNode());
+        return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Vec, SubVec, Idx);
+
+      }
     }
   }
 

Modified: llvm/trunk/test/CodeGen/X86/clear_upper_vector_element_bits.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/clear_upper_vector_element_bits.ll?rev=312462&r1=312461&r2=312462&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/clear_upper_vector_element_bits.ll (original)
+++ llvm/trunk/test/CodeGen/X86/clear_upper_vector_element_bits.ll Sun Sep  3 18:13:36 2017
@@ -758,7 +758,6 @@ define <16 x i16> @_clearupper16xi16b(<1
 ; AVX:       # BB#0:
 ; AVX-NEXT:    vmovaps {{.*#+}} xmm1 = [255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0]
 ; AVX-NEXT:    vandps %xmm1, %xmm0, %xmm2
-; AVX-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm0[4,5,6,7]
 ; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm0
 ; AVX-NEXT:    vandps %xmm1, %xmm0, %xmm0
 ; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm2, %ymm0

Modified: llvm/trunk/test/CodeGen/X86/insertelement-zero.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/insertelement-zero.ll?rev=312462&r1=312461&r2=312462&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/insertelement-zero.ll (original)
+++ llvm/trunk/test/CodeGen/X86/insertelement-zero.ll Sun Sep  3 18:13:36 2017
@@ -473,7 +473,6 @@ define <32 x i8> @insert_v32i8_z12345678
 ; AVX1-NEXT:    xorl %eax, %eax
 ; AVX1-NEXT:    vpinsrb $0, %eax, %xmm0, %xmm1
 ; AVX1-NEXT:    vpinsrb $15, %eax, %xmm1, %xmm1
-; AVX1-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm0[4,5,6,7]
 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
 ; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
 ; AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5,6],xmm2[7]
@@ -485,7 +484,6 @@ define <32 x i8> @insert_v32i8_z12345678
 ; AVX2-NEXT:    xorl %eax, %eax
 ; AVX2-NEXT:    vpinsrb $0, %eax, %xmm0, %xmm1
 ; AVX2-NEXT:    vpinsrb $15, %eax, %xmm1, %xmm1
-; AVX2-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm0[4,5,6,7]
 ; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm0
 ; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
 ; AVX2-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5,6],xmm2[7]