[X86][AVX] Fold extract_subvector(broadcast(x),c) -> extract_subvecto…

…r(broadcast(x),0) iff c != 0 If we're extracting an upper subvector from a broadcast we're better off extracting the lowest subvector instead as it avoids an actual extract instruction and might help SimplifyDemandedVectorElts further simplify the code.
llvm · May 24, 2020 · e508d64 · e508d64
1 parent 57bb478
commit e508d64
Show file tree

Hide file tree

Showing 9 changed files with 207 additions and 273 deletions.
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -47290,6 +47290,13 @@ static SDValue combineExtractSubvector(SDNode *N, SelectionDAG &DAG,
     }
   }
 
+  // If we're extracting an upper subvector from a broadcast we should just
+  // extract the lowest subvector instead which should allow
+  // SimplifyDemandedVectorElts do more simplifications.
+  if (IdxVal != 0 && (InVec.getOpcode() == X86ISD::VBROADCAST ||
+                      InVec.getOpcode() == X86ISD::VBROADCAST_LOAD))
+    return extractSubVector(InVec, 0, DAG, SDLoc(N), VT.getSizeInBits());
+
   // If we're extracting the lowest subvector and we're the only user,
   // we may be able to perform this with a smaller vector width.
   if (IdxVal == 0 && InVec.hasOneUse()) {

diff --git a/llvm/test/CodeGen/X86/pr45443.ll b/llvm/test/CodeGen/X86/pr45443.ll
@@ -8,11 +8,9 @@ define <16 x float> @PR45443() {
 ; X86-NEXT:    vpbroadcastd {{.*#+}} zmm1 = [2181038080,2181038080,2181038080,2181038080,2181038080,2181038080,2181038080,2181038080,2181038080,2181038080,2181038080,2181038080,2181038080,2181038080,2181038080,2181038080]
 ; X86-NEXT:    vfmadd231ps {{.*#+}} zmm0 = (zmm0 * mem) + zmm0
 ; X86-NEXT:    vpcmpltud {{\.LCPI.*}}{1to16}, %zmm1, %k1
-; X86-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
-; X86-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [16777215,16777215,16777215,16777215,16777215,16777215,16777215,16777215]
-; X86-NEXT:    vpand %ymm3, %ymm2, %ymm2
-; X86-NEXT:    vpand %ymm3, %ymm1, %ymm1
-; X86-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm1
+; X86-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [16777215,16777215,16777215,16777215,16777215,16777215,16777215,16777215]
+; X86-NEXT:    vpand %ymm2, %ymm1, %ymm1
+; X86-NEXT:    vinserti64x4 $1, %ymm1, %zmm1, %zmm1
 ; X86-NEXT:    vptestmd %zmm1, %zmm1, %k1 {%k1}
 ; X86-NEXT:    vbroadcastss {{\.LCPI.*}}, %zmm0 {%k1}
 ; X86-NEXT:    retl
@@ -22,11 +20,9 @@ define <16 x float> @PR45443() {
 ; X64-NEXT:    vpbroadcastd {{.*#+}} zmm1 = [2181038080,2181038080,2181038080,2181038080,2181038080,2181038080,2181038080,2181038080,2181038080,2181038080,2181038080,2181038080,2181038080,2181038080,2181038080,2181038080]
 ; X64-NEXT:    vfmadd231ps {{.*#+}} zmm0 = (zmm0 * mem) + zmm0
 ; X64-NEXT:    vpcmpltud {{.*}}(%rip){1to16}, %zmm1, %k1
-; X64-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
-; X64-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [16777215,16777215,16777215,16777215,16777215,16777215,16777215,16777215]
-; X64-NEXT:    vpand %ymm3, %ymm2, %ymm2
-; X64-NEXT:    vpand %ymm3, %ymm1, %ymm1
-; X64-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm1
+; X64-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [16777215,16777215,16777215,16777215,16777215,16777215,16777215,16777215]
+; X64-NEXT:    vpand %ymm2, %ymm1, %ymm1
+; X64-NEXT:    vinserti64x4 $1, %ymm1, %zmm1, %zmm1
 ; X64-NEXT:    vptestmd %zmm1, %zmm1, %k1 {%k1}
 ; X64-NEXT:    vbroadcastss {{.*}}(%rip), %zmm0 {%k1}
 ; X64-NEXT:    retq

diff --git a/llvm/test/CodeGen/X86/vector-fshl-rot-256.ll b/llvm/test/CodeGen/X86/vector-fshl-rot-256.ll
@@ -587,10 +587,9 @@ define <4 x i64> @splatvar_funnnel_v4i64(<4 x i64> %x, <4 x i64> %amt) nounwind
 ;
 ; XOPAVX2-LABEL: splatvar_funnnel_v4i64:
 ; XOPAVX2:       # %bb.0:
-; XOPAVX2-NEXT:    vpbroadcastq %xmm1, %ymm1
 ; XOPAVX2-NEXT:    vextracti128 $1, %ymm0, %xmm2
-; XOPAVX2-NEXT:    vextracti128 $1, %ymm1, %xmm3
-; XOPAVX2-NEXT:    vprotq %xmm3, %xmm2, %xmm2
+; XOPAVX2-NEXT:    vpbroadcastq %xmm1, %xmm1
+; XOPAVX2-NEXT:    vprotq %xmm1, %xmm2, %xmm2
 ; XOPAVX2-NEXT:    vprotq %xmm1, %xmm0, %xmm0
 ; XOPAVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm0
 ; XOPAVX2-NEXT:    retq
@@ -671,10 +670,9 @@ define <8 x i32> @splatvar_funnnel_v8i32(<8 x i32> %x, <8 x i32> %amt) nounwind
 ;
 ; XOPAVX2-LABEL: splatvar_funnnel_v8i32:
 ; XOPAVX2:       # %bb.0:
-; XOPAVX2-NEXT:    vpbroadcastd %xmm1, %ymm1
 ; XOPAVX2-NEXT:    vextracti128 $1, %ymm0, %xmm2
-; XOPAVX2-NEXT:    vextracti128 $1, %ymm1, %xmm3
-; XOPAVX2-NEXT:    vprotd %xmm3, %xmm2, %xmm2
+; XOPAVX2-NEXT:    vpbroadcastd %xmm1, %xmm1
+; XOPAVX2-NEXT:    vprotd %xmm1, %xmm2, %xmm2
 ; XOPAVX2-NEXT:    vprotd %xmm1, %xmm0, %xmm0
 ; XOPAVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm0
 ; XOPAVX2-NEXT:    retq
@@ -741,10 +739,9 @@ define <16 x i16> @splatvar_funnnel_v16i16(<16 x i16> %x, <16 x i16> %amt) nounw
 ;
 ; XOPAVX2-LABEL: splatvar_funnnel_v16i16:
 ; XOPAVX2:       # %bb.0:
-; XOPAVX2-NEXT:    vpbroadcastw %xmm1, %ymm1
 ; XOPAVX2-NEXT:    vextracti128 $1, %ymm0, %xmm2
-; XOPAVX2-NEXT:    vextracti128 $1, %ymm1, %xmm3
-; XOPAVX2-NEXT:    vprotw %xmm3, %xmm2, %xmm2
+; XOPAVX2-NEXT:    vpbroadcastw %xmm1, %xmm1
+; XOPAVX2-NEXT:    vprotw %xmm1, %xmm2, %xmm2
 ; XOPAVX2-NEXT:    vprotw %xmm1, %xmm0, %xmm0
 ; XOPAVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm0
 ; XOPAVX2-NEXT:    retq
@@ -891,10 +888,9 @@ define <32 x i8> @splatvar_funnnel_v32i8(<32 x i8> %x, <32 x i8> %amt) nounwind
 ;
 ; XOPAVX2-LABEL: splatvar_funnnel_v32i8:
 ; XOPAVX2:       # %bb.0:
-; XOPAVX2-NEXT:    vpbroadcastb %xmm1, %ymm1
 ; XOPAVX2-NEXT:    vextracti128 $1, %ymm0, %xmm2
-; XOPAVX2-NEXT:    vextracti128 $1, %ymm1, %xmm3
-; XOPAVX2-NEXT:    vprotb %xmm3, %xmm2, %xmm2
+; XOPAVX2-NEXT:    vpbroadcastb %xmm1, %xmm1
+; XOPAVX2-NEXT:    vprotb %xmm1, %xmm2, %xmm2
 ; XOPAVX2-NEXT:    vprotb %xmm1, %xmm0, %xmm0
 ; XOPAVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm0
 ; XOPAVX2-NEXT:    retq