[SelectionDAG] Simplify extract_subvector( insert_subvector ( Vec, In, Idx ), Idx ) -> In

[RKSimon]

Summary

If we are extracting a subvector that has just been inserted then we should just use the original inserted subvector.

This has come up in certain several x86 shuffle lowering cases where we are crossing 128-bit lanes.

[andreadb]

Hi Simon,

LGTM. Thanks!

lib/CodeGen/SelectionDAG/SelectionDAG.cpp

3758

extra spaces?

lib/CodeGen/SelectionDAG/SelectionDAG.cpp

if (VT.isSimple() && N1.getValueType().isSimple()) {
         assert((VT.getVectorNumElements() + N2C->getZExtValue()
                 <= N1.getValueType().getVectorNumElements())
                && "Extract subvector overflow!");
       }
 
       // Trivial extraction.
       if (VT.getSimpleVT() == N1.getSimpleValueType())
         return N1;
+
+      // EXTRACT_SUBVECTOR of INSERT_SUBVECTOR is often created
+      // during shuffle legalization.
+      if (N1.getOpcode() == ISD::INSERT_SUBVECTOR && N2 == N1.getOperand(2) &&
+          VT == N1.getOperand(1).getValueType())
+          return N1.getOperand(1);
     }
     break;
   }
 
   // Perform trivial constant folding.
   if (SDValue SV =
           FoldConstantArithmetic(Opcode, DL, VT, N1.getNode(), N2.getNode()))
     return SV;

test/CodeGen/X86/vector-shuffle-256-v16.ll

; AVX2-NEXT:    retq
   %shuffle = shufflevector <16 x i16> %a, <16 x i16> %b, <16 x i32> <i32 27, i32 26, i32 1, i32 29, i32 26, i32 23, i32 11, i32 16, i32 1, i32 9, i32 16, i32 28, i32 13, i32 4, i32 0, i32 24>
   ret <16 x i16> %shuffle
 }
 
 define <16 x i16> @insert_dup_mem_v16i16_i32(i32* %ptr) {
 ; AVX1-LABEL: insert_dup_mem_v16i16_i32:
 ; AVX1:       # BB#0:
 ; AVX1-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; AVX1-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,0,0,0,4,5,6,7]
-; AVX1-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]
+; AVX1-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1]
 ; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: insert_dup_mem_v16i16_i32:
 ; AVX2:       # BB#0:
 ; AVX2-NEXT:    vpbroadcastw (%rdi), %ymm0
 ; AVX2-NEXT:    retq
   %tmp = load i32, i32* %ptr, align 4
   %tmp1 = insertelement <4 x i32> zeroinitializer, i32 %tmp, i32 0
   %tmp2 = bitcast <4 x i32> %tmp1 to <8 x i16>
   %tmp3 = shufflevector <8 x i16> %tmp2, <8 x i16> undef, <16 x i32> zeroinitializer
   ret <16 x i16> %tmp3
 }
 
 define <16 x i16> @insert_dup_mem_v16i16_sext_i16(i16* %ptr) {
 ; AVX1-LABEL: insert_dup_mem_v16i16_sext_i16:
 ; AVX1:       # BB#0:
 ; AVX1-NEXT:    movswl (%rdi), %eax
 ; AVX1-NEXT:    vmovd %eax, %xmm0
-; AVX1-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,0,0,0,4,5,6,7]
-; AVX1-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]
+; AVX1-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1]
 ; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: insert_dup_mem_v16i16_sext_i16:
 ; AVX2:       # BB#0:
 ; AVX2-NEXT:    movswl (%rdi), %eax
 ; AVX2-NEXT:    vmovd %eax, %xmm0
 ; AVX2-NEXT:    vpbroadcastw %xmm0, %ymm0
 ; AVX2-NEXT:    retq
   %tmp = load i16, i16* %ptr, align 2
   %tmp1 = sext i16 %tmp to i32
   %tmp2 = insertelement <4 x i32> zeroinitializer, i32 %tmp1, i32 0
   %tmp3 = bitcast <4 x i32> %tmp2 to <8 x i16>
   %tmp4 = shufflevector <8 x i16> %tmp3, <8 x i16> undef, <16 x i32> zeroinitializer
   ret <16 x i16> %tmp4
 }
 
 define <16 x i16> @insert_dup_elt1_mem_v16i16_i32(i32* %ptr) #0 {
 ; AVX1-LABEL: insert_dup_elt1_mem_v16i16_i32:
 ; AVX1:       # BB#0:
 ; AVX1-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; AVX1-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[1,1,1,1,4,5,6,7]
-; AVX1-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]
+; AVX1-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[2,3,2,3,2,3,2,3,2,3,2,3,2,3,2,3]
 ; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: insert_dup_elt1_mem_v16i16_i32:
 ; AVX2:       # BB#0:
 ; AVX2-NEXT:    vpbroadcastw 2(%rdi), %ymm0
 ; AVX2-NEXT:    retq
   %tmp = load i32, i32* %ptr, align 4
   %tmp1 = insertelement <4 x i32> zeroinitializer, i32 %tmp, i32 0
   %tmp2 = bitcast <4 x i32> %tmp1 to <8 x i16>
   %tmp3 = shufflevector <8 x i16> %tmp2, <8 x i16> undef, <16 x i32> <i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1>
   ret <16 x i16> %tmp3
 }
 
 define <16 x i16> @insert_dup_elt3_mem_v16i16_i32(i32* %ptr) #0 {
 ; AVX1-LABEL: insert_dup_elt3_mem_v16i16_i32:
 ; AVX1:       # BB#0:
 ; AVX1-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; AVX1-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,0,1,1]
-; AVX1-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[3,3,3,3,4,5,6,7]
-; AVX1-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]
+; AVX1-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[2,3,2,3,2,3,2,3,2,3,2,3,2,3,2,3]
 ; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: insert_dup_elt3_mem_v16i16_i32:
 ; AVX2:       # BB#0:
 ; AVX2-NEXT:    vpbroadcastw 2(%rdi), %ymm0
 ; AVX2-NEXT:    retq
   %tmp = load i32, i32* %ptr, align 4
   %tmp1 = insertelement <4 x i32> zeroinitializer, i32 %tmp, i32 1
   %tmp2 = bitcast <4 x i32> %tmp1 to <8 x i16>
   %tmp3 = shufflevector <8 x i16> %tmp2, <8 x i16> undef, <16 x i32> <i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3>
   ret <16 x i16> %tmp3
 }