[X86] Use PSADBW for v8i8 addition reductions.

[craig.topper]

Summary

Improves the 8 byte case from PR42674.

[RKSimon]

llvm/lib/Target/X86/X86ISelLowering.cpp

35451 ↗

(On Diff #214576)

we can easily support 2i8/4i8 as well by replacing this with an insertion into a zero v16i8 vector

[craig.topper]

That doesn’t seem profitable for v2i8. We’d be better off extracting both elements and doing a scalar add. For v4i8, I’m not sure. Psadbw is 5 cycles on some CPUs if I remember right, the normal expansion is probably faster on those CPUs.

[RKSimon]

OK, let's just go for v8i8

llvm/trunk/lib/Target/X86/X86ISelLowering.cpp

static SDValue combineReductionToHorizontal(SDNode *ExtElt, SelectionDAG &DAG,
   assert(isNullConstant(Index) &&
          "Reduction doesn't end in an extract from index 0");
 
   EVT VT = ExtElt->getValueType(0);
   EVT VecVT = Rdx.getValueType();
   if (VecVT.getScalarType() != VT)
     return SDValue();
 
+  SDLoc DL(ExtElt);
+
+  if (VecVT == MVT::v8i8) {
+    // Pad with undef.
+    Rdx = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, Rdx,
+                      DAG.getUNDEF(VecVT));
+    Rdx = DAG.getNode(X86ISD::PSADBW, DL, MVT::v2i64, Rdx,
+                      DAG.getConstant(0, DL, MVT::v16i8));
+    Rdx = DAG.getBitcast(MVT::v16i8, Rdx);
+    return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, Rdx, Index);
+  }
+
   // Must be a >=128-bit vector with pow2 elements.
   if ((VecVT.getSizeInBits() % 128) != 0 ||
       !isPowerOf2_32(VecVT.getVectorNumElements()))
     return SDValue();
 
-  SDLoc DL(ExtElt);
-
   // vXi8 reduction - sum lo/hi halves then use PSADBW.
   if (VT == MVT::i8) {
     while (Rdx.getValueSizeInBits() > 128) {
       unsigned HalfSize = VecVT.getSizeInBits() / 2;
       unsigned HalfElts = VecVT.getVectorNumElements() / 2;
       SDValue Lo = extractSubVector(Rdx, 0, DAG, DL, HalfSize);
       SDValue Hi = extractSubVector(Rdx, HalfElts, DAG, DL, HalfSize);
       Rdx = DAG.getNode(ISD::ADD, DL, Lo.getValueType(), Lo, Hi);

llvm/trunk/test/CodeGen/X86/vector-reduce-add.ll

; AVX512VL-NEXT:    retq
   %a0 = load <4 x i8>, <4 x i8>* %p
   %1 = call i8 @llvm.experimental.vector.reduce.add.v4i8(<4 x i8> %a0)
   ret i8 %1
 }
 
 define i8 @test_v8i8(<8 x i8> %a0) {
 ; SSE2-LABEL: test_v8i8:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    movdqa %xmm1, %xmm0
-; SSE2-NEXT:    psrld $16, %xmm0
-; SSE2-NEXT:    paddb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $8, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
+; SSE2-NEXT:    pxor %xmm1, %xmm1
+; SSE2-NEXT:    psadbw %xmm0, %xmm1
 ; SSE2-NEXT:    movd %xmm1, %eax
 ; SSE2-NEXT:    # kill: def $al killed $al killed $eax
 ; SSE2-NEXT:    retq
 ;
 ; SSE41-LABEL: test_v8i8:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; SSE41-NEXT:    paddb %xmm0, %xmm1
-; SSE41-NEXT:    movdqa %xmm1, %xmm0
-; SSE41-NEXT:    psrld $16, %xmm0
-; SSE41-NEXT:    paddb %xmm1, %xmm0
-; SSE41-NEXT:    movdqa %xmm0, %xmm1
-; SSE41-NEXT:    psrlw $8, %xmm1
-; SSE41-NEXT:    paddb %xmm0, %xmm1
+; SSE41-NEXT:    pxor %xmm1, %xmm1
+; SSE41-NEXT:    psadbw %xmm0, %xmm1
 ; SSE41-NEXT:    pextrb $0, %xmm1, %eax
 ; SSE41-NEXT:    # kill: def $al killed $al killed $eax
 ; SSE41-NEXT:    retq
 ;
 ; AVX-LABEL: test_v8i8:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; AVX-NEXT:    vpaddb %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    vpsrld $16, %xmm0, %xmm1
-; AVX-NEXT:    vpaddb %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    vpsrlw $8, %xmm0, %xmm1
-; AVX-NEXT:    vpaddb %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; AVX-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
 ; AVX-NEXT:    vpextrb $0, %xmm0, %eax
 ; AVX-NEXT:    # kill: def $al killed $al killed $eax
 ; AVX-NEXT:    retq
 ;
 ; AVX512-LABEL: test_v8i8:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; AVX512-NEXT:    vpaddb %xmm1, %xmm0, %xmm0
-; AVX512-NEXT:    vpsrld $16, %xmm0, %xmm1
-; AVX512-NEXT:    vpaddb %xmm1, %xmm0, %xmm0
-; AVX512-NEXT:    vpsrlw $8, %xmm0, %xmm1
-; AVX512-NEXT:    vpaddb %xmm1, %xmm0, %xmm0
+; AVX512-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; AVX512-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
 ; AVX512-NEXT:    vpextrb $0, %xmm0, %eax
 ; AVX512-NEXT:    # kill: def $al killed $al killed $eax
 ; AVX512-NEXT:    retq
   %1 = call i8 @llvm.experimental.vector.reduce.add.v8i8(<8 x i8> %a0)
   ret i8 %1
 }
 
 define i8 @test_v8i8_load(<8 x i8>* %p) {
 ; SSE2-LABEL: test_v8i8_load:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movq {{.*#+}} xmm0 = mem[0],zero
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    movdqa %xmm1, %xmm0
-; SSE2-NEXT:    psrld $16, %xmm0
-; SSE2-NEXT:    paddb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $8, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
+; SSE2-NEXT:    pxor %xmm1, %xmm1
+; SSE2-NEXT:    psadbw %xmm0, %xmm1
 ; SSE2-NEXT:    movd %xmm1, %eax
 ; SSE2-NEXT:    # kill: def $al killed $al killed $eax
 ; SSE2-NEXT:    retq
 ;
 ; SSE41-LABEL: test_v8i8_load:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movq {{.*#+}} xmm0 = mem[0],zero
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; SSE41-NEXT:    paddb %xmm0, %xmm1
-; SSE41-NEXT:    movdqa %xmm1, %xmm0
-; SSE41-NEXT:    psrld $16, %xmm0
-; SSE41-NEXT:    paddb %xmm1, %xmm0
-; SSE41-NEXT:    movdqa %xmm0, %xmm1
-; SSE41-NEXT:    psrlw $8, %xmm1
-; SSE41-NEXT:    paddb %xmm0, %xmm1
+; SSE41-NEXT:    pxor %xmm1, %xmm1
+; SSE41-NEXT:    psadbw %xmm0, %xmm1
 ; SSE41-NEXT:    pextrb $0, %xmm1, %eax
 ; SSE41-NEXT:    # kill: def $al killed $al killed $eax
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: test_v8i8_load:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
-; AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; AVX1-NEXT:    vpaddb %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
-; AVX1-NEXT:    vpaddb %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm1
-; AVX1-NEXT:    vpaddb %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpextrb $0, %xmm0, %eax
-; AVX1-NEXT:    # kill: def $al killed $al killed $eax
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: test_v8i8_load:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    movq (%rdi), %rax
-; AVX2-NEXT:    vmovq %rax, %xmm0
-; AVX2-NEXT:    shrq $32, %rax
-; AVX2-NEXT:    vmovd %eax, %xmm1
-; AVX2-NEXT:    vpbroadcastd %xmm1, %xmm1
-; AVX2-NEXT:    vpaddb %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpsrld $16, %xmm0, %xmm1
-; AVX2-NEXT:    vpaddb %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm1
-; AVX2-NEXT:    vpaddb %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpextrb $0, %xmm0, %eax
-; AVX2-NEXT:    # kill: def $al killed $al killed $eax
-; AVX2-NEXT:    retq
-;
-; AVX512BW-LABEL: test_v8i8_load:
-; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    movq (%rdi), %rax
-; AVX512BW-NEXT:    vmovq %rax, %xmm0
-; AVX512BW-NEXT:    shrq $32, %rax
-; AVX512BW-NEXT:    vmovd %eax, %xmm1
-; AVX512BW-NEXT:    vpbroadcastd %xmm1, %xmm1
-; AVX512BW-NEXT:    vpaddb %xmm1, %xmm0, %xmm0
-; AVX512BW-NEXT:    vpsrld $16, %xmm0, %xmm1
-; AVX512BW-NEXT:    vpaddb %xmm1, %xmm0, %xmm0
-; AVX512BW-NEXT:    vpsrlw $8, %xmm0, %xmm1
-; AVX512BW-NEXT:    vpaddb %xmm1, %xmm0, %xmm0
-; AVX512BW-NEXT:    vpextrb $0, %xmm0, %eax
-; AVX512BW-NEXT:    # kill: def $al killed $al killed $eax
-; AVX512BW-NEXT:    retq
+; AVX-LABEL: test_v8i8_load:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
+; AVX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; AVX-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    vpextrb $0, %xmm0, %eax
+; AVX-NEXT:    # kill: def $al killed $al killed $eax
+; AVX-NEXT:    retq
 ;
-; AVX512VL-LABEL: test_v8i8_load:
-; AVX512VL:       # %bb.0:
-; AVX512VL-NEXT:    movq (%rdi), %rax
-; AVX512VL-NEXT:    vmovq %rax, %xmm0
-; AVX512VL-NEXT:    shrq $32, %rax
-; AVX512VL-NEXT:    vpbroadcastd %eax, %xmm1
-; AVX512VL-NEXT:    vpaddb %xmm1, %xmm0, %xmm0
-; AVX512VL-NEXT:    vpsrld $16, %xmm0, %xmm1
-; AVX512VL-NEXT:    vpaddb %xmm1, %xmm0, %xmm0
-; AVX512VL-NEXT:    vpsrlw $8, %xmm0, %xmm1
-; AVX512VL-NEXT:    vpaddb %xmm1, %xmm0, %xmm0
-; AVX512VL-NEXT:    vpextrb $0, %xmm0, %eax
-; AVX512VL-NEXT:    # kill: def $al killed $al killed $eax
-; AVX512VL-NEXT:    retq
+; AVX512-LABEL: test_v8i8_load:
+; AVX512:       # %bb.0:
+; AVX512-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
+; AVX512-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; AVX512-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
+; AVX512-NEXT:    vpextrb $0, %xmm0, %eax
+; AVX512-NEXT:    # kill: def $al killed $al killed $eax
+; AVX512-NEXT:    retq
   %a0 = load <8 x i8>, <8 x i8>* %p
   %1 = call i8 @llvm.experimental.vector.reduce.add.v8i8(<8 x i8> %a0)
   ret i8 %1
 }
 
 define i8 @test_v16i8(<16 x i8> %a0) {
 ; SSE2-LABEL: test_v16i8:
 ; SSE2:       # %bb.0: