[DAGCombiner] Improved target independent vector shuffle combine rule.
This patch improves the existing algorithm in DAGCombiner that
attempts to fold shuffles according to rule:
shuffle(shuffle(x, y, M1), undef, M2) -> shuffle(y, undef, M3)
Before this change, there were cases where the DAGCombiner conservatively
avoided folding shuffles even if the resulting mask would have been legal.
That is because the algorithm wrongly assumed that commuting
an illegal shuffle mask would always produce an illegal mask.
With this change, we now correctly compute the commuted shuffle mask before
calling method 'isShuffleMaskLegal' on it.
On X86, this improves for example the codegen for the following function:
define <4 x i32> @test(<4 x i32> %A, <4 x i32> %B) {
%1 = shufflevector <4 x i32> %B, <4 x i32> %A, <4 x i32> <i32 1, i32 2, i32 6, i32 7>
%2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 2, i32 3>
ret <4 x i32> %2
}
Before this change the X86 backend (-mcpu=corei7) generated
the following assembly code for function @test:
shufps $-23, %xmm0, %xmm1 # xmm1 = xmm1[1,2],xmm0[2,3]
movhlps %xmm1, %xmm1 # xmm1 = xmm1[1,1]
movaps %xmm1, %xmm0
Now we produce:
movhlps %xmm0, %xmm0 # xmm0 = xmm0[1,1]
Added extra test cases in combine-vec-shuffle-2.ll to verify that we correctly
fold according to the above-mentioned rule.
llvm-svn: 215555
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index e8cbd8a..603ccb0 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -10838,16 +10838,30 @@
// It may still be beneficial to combine the two shuffles if the
// resulting shuffle is legal.
- if (TLI.isTypeLegal(VT) && TLI.isShuffleMaskLegal(Mask, VT)) {
- if (!CommuteOperands)
- // shuffle(shuffle(x, undef, M1), undef, M2) -> shuffle(x, undef, M3).
- // shuffle(shuffle(x, y, M1), undef, M2) -> shuffle(x, undef, M3)
- return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(0), N1,
- &Mask[0]);
-
- // shuffle(shuffle(x, y, M1), undef, M2) -> shuffle(undef, y, M3)
- return DAG.getVectorShuffle(VT, SDLoc(N), N1, N0->getOperand(1),
- &Mask[0]);
+ if (TLI.isTypeLegal(VT)) {
+ if (!CommuteOperands) {
+ if (TLI.isShuffleMaskLegal(Mask, VT))
+ // shuffle(shuffle(x, undef, M1), undef, M2) -> shuffle(x, undef, M3).
+ // shuffle(shuffle(x, y, M1), undef, M2) -> shuffle(x, undef, M3)
+ return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(0), N1,
+ &Mask[0]);
+ } else {
+ // Compute the commuted shuffle mask.
+ for (unsigned i = 0; i != NumElts; ++i) {
+ int idx = Mask[i];
+ if (idx < 0)
+ continue;
+ else if (idx < (int)NumElts)
+ Mask[i] = idx + NumElts;
+ else
+ Mask[i] = idx - NumElts;
+ }
+
+ if (TLI.isShuffleMaskLegal(Mask, VT))
+ // shuffle(shuffle(x, y, M1), undef, M2) -> shuffle(y, undef, M3)
+ return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(1), N1,
+ &Mask[0]);
+ }
}
}
diff --git a/llvm/test/CodeGen/X86/combine-vec-shuffle-2.ll b/llvm/test/CodeGen/X86/combine-vec-shuffle-2.ll
index 877d382..872384c 100644
--- a/llvm/test/CodeGen/X86/combine-vec-shuffle-2.ll
+++ b/llvm/test/CodeGen/X86/combine-vec-shuffle-2.ll
@@ -204,8 +204,8 @@
ret <4 x i32> %2
}
; CHECK-LABEL: test18
-; CHECK: blendps $11
-; CHECK-NEXT: pshufd $-59
+; CHECK-NOT: blendps
+; CHECK: pshufd {{.*}} # xmm0 = xmm1[1,1,0,3]
; CHECK-NEXT: ret
define <4 x i32> @test19(<4 x i32> %A, <4 x i32> %B) {
@@ -240,6 +240,8 @@
; CHECK-NEXT: pshufd $-60
; CHECK-NEXT: ret
+; Test that we correctly combine shuffles according to rule
+; shuffle(shuffle(x, y), undef) -> shuffle(y, undef)
define <4 x i32> @test22(<4 x i32> %A, <4 x i32> %B) {
%1 = shufflevector <4 x i32> %A, <4 x i32> %B, <4 x i32> <i32 4, i32 5, i32 2, i32 7>
@@ -247,7 +249,69 @@
ret <4 x i32> %2
}
; CHECK-LABEL: test22
-; CHECK: blendps $11
-; CHECK-NEXT: pshufd $-43
+; CHECK-NOT: blendps
+; CHECK: pshufd {{.*}} # xmm0 = xmm1[1,1,1,3]
+; CHECK-NEXT: ret
+
+define <4 x i32> @test23(<4 x i32> %A, <4 x i32> %B) {
+ %1 = shufflevector <4 x i32> %A, <4 x i32> %B, <4 x i32> <i32 4, i32 5, i32 2, i32 7>
+ %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 3>
+ ret <4 x i32> %2
+}
+; CHECK-LABEL: test23
+; CHECK-NOT: blendps
+; CHECK: pshufd {{.*}} # xmm0 = xmm1[0,1,0,3]
+; CHECK-NEXT: ret
+
+define <4 x i32> @test24(<4 x i32> %A, <4 x i32> %B) {
+ %1 = shufflevector <4 x i32> %A, <4 x i32> %B, <4 x i32> <i32 4, i32 1, i32 6, i32 7>
+ %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 0, i32 3, i32 2, i32 4>
+ ret <4 x i32> %2
+}
+; CHECK-LABEL: test24
+; CHECK-NOT: blendps
+; CHECK: pshufd {{.*}} # xmm0 = xmm1[0,3,2,0]
+; CHECK-NEXT: ret
+
+define <4 x i32> @test25(<4 x i32> %A, <4 x i32> %B) {
+ %1 = shufflevector <4 x i32> %B, <4 x i32> %A, <4 x i32> <i32 1, i32 5, i32 2, i32 4>
+ %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 3, i32 1, i32 3, i32 1>
+ ret <4 x i32> %2
+}
+; CHECK-LABEL: test25
+; CHECK-NOT: shufps
+; CHECK: pshufd {{.*}} # xmm0 = xmm0[0,1,0,1]
+; CHECK-NEXT: ret
+
+define <4 x i32> @test26(<4 x i32> %A, <4 x i32> %B) {
+ %1 = shufflevector <4 x i32> %B, <4 x i32> %A, <4 x i32> <i32 1, i32 2, i32 6, i32 7>
+ %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 2, i32 3>
+ ret <4 x i32> %2
+}
+; CHECK-LABEL: test26
+; CHECK-NOT: shufps
+; CHECK: movhlps {{.*}} # xmm0 = xmm0[1,1]
+; CHECK-NEXT: ret
+
+define <4 x i32> @test27(<4 x i32> %A, <4 x i32> %B) {
+ %1 = shufflevector <4 x i32> %B, <4 x i32> %A, <4 x i32> <i32 2, i32 1, i32 5, i32 4>
+ %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 3, i32 2, i32 3, i32 2>
+ ret <4 x i32> %2
+}
+; CHECK-LABEL: test27
+; CHECK-NOT: shufps
+; CHECK-NOT: movhlps
+; CHECK: pshufd {{.*}} # xmm0 = xmm0[0,1,0,1]
+; CHECK-NEXT: ret
+
+define <4 x i32> @test28(<4 x i32> %A, <4 x i32> %B) {
+ %1 = shufflevector <4 x i32> %B, <4 x i32> %A, <4 x i32> <i32 1, i32 2, i32 4, i32 5>
+ %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 3, i32 2>
+ ret <4 x i32> %2
+}
+; CHECK-LABEL: test28
+; CHECK-NOT: shufps
+; CHECK-NOT: movhlps
+; CHECK: pshufd {{.*}} # xmm0 = xmm0[0,1,1,0]
; CHECK-NEXT: ret
