[PATCH][DAGCombiner] Teach how to combine a pair of shuffles into a single shuffle if the resulting mask is legal.
Andrea Di Biagio
andrea.dibiagio at gmail.com
Tue Jul 8 07:51:53 PDT 2014
Hi,
this patch teaches the DAGCombiner how to fold shuffles according to rule:
(shuffle (shuffle (X, Undef, M0)), Undef, M1) -> (shuffle X, Undef, M2)
We do this only if M2 is legal according to method 'isShuffleMaskLegal'.
DAGCombiner already implemented some rules to simplify a sequence of
shuffles into the vector operand according to rule
(shuffle (shuffle (X, Undef, M0), Undef, M1) -> X
This patch extends that logic to also cover the case where the two
shuffles in input don't perform an identity operation. To avoid
introducing illegal shuffles that are then expanded during
Legalization into a sub-optimal sequence of instructions, the new rule
would only trigger if if the mask of the resulting shufflevector is
legal (this is done calling method 'TLI.isShuffleMaskLegal').
In my experiments, this patch improves how shuffles are selected on
x86 (see for example my new tests swizzle-2.ll and swizzle-avx2.ll).
However, in most cases - related to 128-bit vector swizzle oeprations
- the backend already knows how to efficiently combine PSHUFD dag
nodes thanks to the new target combine rules added at revisions 211889
and 211890.
In general, this patch has clearly the advantage of being target
independent, since it works on target independent shuffle dag nodes.
That means, potentially all targets (not only x86) can take advantage
of this new rule.
Second, most shuffle pairs can be safely combined before we run the
legalizer on vector operations. This allows us to combine/simplify
some dag nodes earlier in the process and not only immediately before
instruction selection. That said, this patch is _not_ meant to be a
replacement for the target specific combine rules already available on
x86; the backend might still potentially introduce new shuffles during
legalization stage. Another reason why on x86 we still need target
specific combine rules is because this new rule is very simple and
avoids to aggressively optimize shuffles; on the other hand, rules
added at 211890 tends to be more aggressive since they know how to see
through bitcast operations when folding shuffles.
I only tested this patch on x86. I had a very quick look at the
codegen for test swizzle-2.ll when passing an ARM triple (I tested
only armv7-linux).
I saw that the code generated is overall better with this patch (i.e.
smaller - with less shuffles).
Please let me know if ok to submit.
Thanks,
Andrea
-------------- next part --------------
Index: lib/CodeGen/SelectionDAG/DAGCombiner.cpp
===================================================================
--- lib/CodeGen/SelectionDAG/DAGCombiner.cpp (revision 212535)
+++ lib/CodeGen/SelectionDAG/DAGCombiner.cpp (working copy)
@@ -10670,6 +10670,8 @@
assert(OtherSV->getOperand(0).getValueType() == VT &&
"Shuffle types don't match");
+ SmallVector<int, 4> Mask;
+ // Compute the combined shuffle mask.
for (unsigned i = 0; i != NumElts; ++i) {
int Idx = SVN->getMaskElt(i);
assert(Idx < (int)NumElts && "Index references undef operand");
@@ -10677,13 +10679,29 @@
// shuffle. Adopt the incoming index.
if (Idx >= 0)
Idx = OtherSV->getMaskElt(Idx);
+ Mask.push_back(Idx);
+ }
+ bool IsIdentityMask = true;
+ for (unsigned i = 0; i != NumElts && IsIdentityMask; ++i) {
+ // Skip Undefs.
+ if (Mask[i] < 0)
+ continue;
+
// The combined shuffle must map each index to itself.
- if (Idx >= 0 && (unsigned)Idx != i)
- return SDValue();
+ IsIdentityMask = (unsigned)Mask[i] == i;
}
- return OtherSV->getOperand(0);
+ if (IsIdentityMask)
+ // optimize shuffle(shuffle(x, undef), undef) -> x.
+ return OtherSV->getOperand(0);
+
+ // It may still be beneficial to combine the two shuffles if the
+ // resulting shuffle is legal.
+ // shuffle(shuffle(x, undef, M1), undef, M2) -> shuffle(x, undef, M3).
+ if (TLI.isShuffleMaskLegal(Mask, VT))
+ return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(0), N1,
+ &Mask[0]);
}
return SDValue();
Index: test/CodeGen/X86/swizzle-2.ll
===================================================================
--- test/CodeGen/X86/swizzle-2.ll (revision 0)
+++ test/CodeGen/X86/swizzle-2.ll (working copy)
@@ -0,0 +1,515 @@
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mcpu=corei7 | FileCheck %s
+
+; Test that we correctly fold a shuffle that performs a swizzle of another
+; shuffle node according to the rule
+; shuffle (shuffle (x, undef, M0), undef, M1) -> shuffle(x, undef, M2)
+;
+; We only do this if the resulting mask is legal to avoid introducing an
+; illegal shuffle that is expanded into a sub-optimal sequence of instructions
+; during lowering stage.
+
+
+define <4 x i32> @swizzle_1(<4 x i32> %v) {
+ %1 = shufflevector <4 x i32> %v, <4 x i32> undef, <4 x i32> <i32 3, i32 2, i32 0, i32 1>
+ %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 3, i32 2, i32 0, i32 1>
+ ret <4 x i32> %2
+}
+; CHECK-LABEL: swizzle_1
+; Mask: [1,0,3,2]
+; CHECK: pshufd $-79
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x i32> @swizzle_2(<4 x i32> %v) {
+ %1 = shufflevector <4 x i32> %v, <4 x i32> undef, <4 x i32> <i32 3, i32 1, i32 0, i32 2>
+ %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 3, i32 1, i32 0, i32 2>
+ ret <4 x i32> %2
+}
+; CHECK-LABEL: swizzle_2
+; Mask: [2,1,3,0]
+; CHECK: pshufd $54
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x i32> @swizzle_3(<4 x i32> %v) {
+ %1 = shufflevector <4 x i32> %v, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 1, i32 0>
+ %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 1, i32 0>
+ ret <4 x i32> %2
+}
+; CHECK-LABEL: swizzle_3
+; Mask: [1,0,3,2]
+; CHECK: pshufd $-79
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x i32> @swizzle_4(<4 x i32> %v) {
+ %1 = shufflevector <4 x i32> %v, <4 x i32> undef, <4 x i32> <i32 2, i32 1, i32 3, i32 0>
+ %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 2, i32 1, i32 3, i32 0>
+ ret <4 x i32> %2
+}
+; CHECK-LABEL: swizzle_4
+; Mask: [3,1,0,2]
+; CHECK: pshufd $-121
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x i32> @swizzle_5(<4 x i32> %v) {
+ %1 = shufflevector <4 x i32> %v, <4 x i32> undef, <4 x i32> <i32 1, i32 2, i32 3, i32 0>
+ %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 1, i32 2, i32 3, i32 0>
+ ret <4 x i32> %2
+}
+; CHECK-LABEL: swizzle_5
+; Mask: [2,3,0,1]
+; CHECK: pshufd $78
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x i32> @swizzle_6(<4 x i32> %v) {
+ %1 = shufflevector <4 x i32> %v, <4 x i32> undef, <4 x i32> <i32 1, i32 2, i32 0, i32 3>
+ %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 1, i32 2, i32 0, i32 3>
+ ret <4 x i32> %2
+}
+; CHECK-LABEL: swizzle_6
+; Mask: [2,0,1,3]
+; CHECK: pshufd $-46
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x i32> @swizzle_7(<4 x i32> %v) {
+ %1 = shufflevector <4 x i32> %v, <4 x i32> undef, <4 x i32> <i32 0, i32 3, i32 1, i32 2>
+ %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 0, i32 3, i32 1, i32 2>
+ ret <4 x i32> %2
+}
+; CHECK-LABEL: swizzle_7
+; Mask: [0,2,3,1]
+; CHECK: pshufd $120
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x i32> @swizzle_8(<4 x i32> %v) {
+ %1 = shufflevector <4 x i32> %v, <4 x i32> undef, <4 x i32> <i32 3, i32 0, i32 2, i32 1>
+ %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 3, i32 0, i32 2, i32 1>
+ ret <4 x i32> %2
+}
+; CHECK-LABEL: swizzle_8
+; Mask: [1,3,2,0]
+; CHECK: pshufd $45
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x i32> @swizzle_9(<4 x i32> %v) {
+ %1 = shufflevector <4 x i32> %v, <4 x i32> undef, <4 x i32> <i32 3, i32 0, i32 1, i32 2>
+ %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 3, i32 0, i32 1, i32 2>
+ ret <4 x i32> %2
+}
+; CHECK-LABEL: swizzle_9
+; Mask: [2,3,0,1]
+; CHECK: pshufd $78
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x i32> @swizzle_10(<4 x i32> %v) {
+ %1 = shufflevector <4 x i32> %v, <4 x i32> undef, <4 x i32> <i32 2, i32 0, i32 1, i32 3>
+ %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 2, i32 0, i32 1, i32 3>
+ ret <4 x i32> %2
+}
+; CHECK-LABEL: swizzle_10
+; Mask: [1,2,0,3]
+; CHECK: pshufd $-55
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x i32> @swizzle_11(<4 x i32> %v) {
+ %1 = shufflevector <4 x i32> %v, <4 x i32> undef, <4 x i32> <i32 2, i32 0, i32 3, i32 1>
+ %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 2, i32 0, i32 3, i32 1>
+ ret <4 x i32> %2
+}
+; CHECK-LABEL: swizzle_11
+; Mask: [3,2,1,0]
+; CHECK: pshufd $27
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x i32> @swizzle_12(<4 x i32> %v) {
+ %1 = shufflevector <4 x i32> %v, <4 x i32> undef, <4 x i32> <i32 0, i32 2, i32 3, i32 1>
+ %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 0, i32 2, i32 3, i32 1>
+ ret <4 x i32> %2
+}
+; CHECK-LABEL: swizzle_12
+; Mask: [0,3,1,2]
+; CHECK: pshufd $-100
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x i32> @swizzle_13(<4 x i32> %v) {
+ %1 = shufflevector <4 x i32> %v, <4 x i32> undef, <4 x i32> <i32 1, i32 3, i32 0, i32 2>
+ %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 1, i32 3, i32 0, i32 2>
+ ret <4 x i32> %2
+}
+; CHECK-LABEL: swizzle_13
+; Mask: [3,2,1,0]
+; CHECK: pshufd $27
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x i32> @swizzle_14(<4 x i32> %v) {
+ %1 = shufflevector <4 x i32> %v, <4 x i32> undef, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
+ %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
+ ret <4 x i32> %2
+}
+; CHECK-LABEL: swizzle_14
+; Mask: [3,0,2,1]
+; CHECK: pshufd $99
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x float> @swizzle_15(<4 x float> %v) {
+ %1 = shufflevector <4 x float> %v, <4 x float> undef, <4 x i32> <i32 3, i32 2, i32 0, i32 1>
+ %2 = shufflevector <4 x float> %1, <4 x float> undef, <4 x i32> <i32 3, i32 2, i32 0, i32 1>
+ ret <4 x float> %2
+}
+; CHECK-LABEL: swizzle_15
+; Mask: [1,0,3,2]
+; CHECK: pshufd $-79
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x float> @swizzle_16(<4 x float> %v) {
+ %1 = shufflevector <4 x float> %v, <4 x float> undef, <4 x i32> <i32 3, i32 1, i32 0, i32 2>
+ %2 = shufflevector <4 x float> %1, <4 x float> undef, <4 x i32> <i32 3, i32 1, i32 0, i32 2>
+ ret <4 x float> %2
+}
+; CHECK-LABEL: swizzle_16
+; Mask: [2,1,3,0]
+; CHECK: pshufd $54
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x float> @swizzle_17(<4 x float> %v) {
+ %1 = shufflevector <4 x float> %v, <4 x float> undef, <4 x i32> <i32 2, i32 3, i32 1, i32 0>
+ %2 = shufflevector <4 x float> %1, <4 x float> undef, <4 x i32> <i32 2, i32 3, i32 1, i32 0>
+ ret <4 x float> %2
+}
+; CHECK-LABEL: swizzle_17
+; Mask: [1,0,3,2]
+; CHECK: pshufd $-79
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x float> @swizzle_18(<4 x float> %v) {
+ %1 = shufflevector <4 x float> %v, <4 x float> undef, <4 x i32> <i32 2, i32 1, i32 3, i32 0>
+ %2 = shufflevector <4 x float> %1, <4 x float> undef, <4 x i32> <i32 2, i32 1, i32 3, i32 0>
+ ret <4 x float> %2
+}
+; CHECK-LABEL: swizzle_18
+; Mask: [3,1,0,2]
+; CHECK: pshufd $-121
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x float> @swizzle_19(<4 x float> %v) {
+ %1 = shufflevector <4 x float> %v, <4 x float> undef, <4 x i32> <i32 1, i32 2, i32 3, i32 0>
+ %2 = shufflevector <4 x float> %1, <4 x float> undef, <4 x i32> <i32 1, i32 2, i32 3, i32 0>
+ ret <4 x float> %2
+}
+; CHECK-LABEL: swizzle_19
+; Mask: [2,3,0,1]
+; CHECK: pshufd $78
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x float> @swizzle_20(<4 x float> %v) {
+ %1 = shufflevector <4 x float> %v, <4 x float> undef, <4 x i32> <i32 1, i32 2, i32 0, i32 3>
+ %2 = shufflevector <4 x float> %1, <4 x float> undef, <4 x i32> <i32 1, i32 2, i32 0, i32 3>
+ ret <4 x float> %2
+}
+; CHECK-LABEL: swizzle_20
+; Mask: [2,0,1,3]
+; CHECK: pshufd $-46
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x float> @swizzle_21(<4 x float> %v) {
+ %1 = shufflevector <4 x float> %v, <4 x float> undef, <4 x i32> <i32 0, i32 3, i32 1, i32 2>
+ %2 = shufflevector <4 x float> %1, <4 x float> undef, <4 x i32> <i32 0, i32 3, i32 1, i32 2>
+ ret <4 x float> %2
+}
+; CHECK-LABEL: swizzle_21
+; Mask: [0,2,3,1]
+; CHECK: pshufd $120
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x float> @swizzle_22(<4 x float> %v) {
+ %1 = shufflevector <4 x float> %v, <4 x float> undef, <4 x i32> <i32 3, i32 0, i32 2, i32 1>
+ %2 = shufflevector <4 x float> %1, <4 x float> undef, <4 x i32> <i32 3, i32 0, i32 2, i32 1>
+ ret <4 x float> %2
+}
+; CHECK-LABEL: swizzle_22
+; Mask: [1,3,2,0]
+; CHECK: pshufd $45
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x float> @swizzle_23(<4 x float> %v) {
+ %1 = shufflevector <4 x float> %v, <4 x float> undef, <4 x i32> <i32 3, i32 0, i32 1, i32 2>
+ %2 = shufflevector <4 x float> %1, <4 x float> undef, <4 x i32> <i32 3, i32 0, i32 1, i32 2>
+ ret <4 x float> %2
+}
+; CHECK-LABEL: swizzle_23
+; Mask: [2,3,0,1]
+; CHECK: pshufd $78
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x float> @swizzle_24(<4 x float> %v) {
+ %1 = shufflevector <4 x float> %v, <4 x float> undef, <4 x i32> <i32 2, i32 0, i32 1, i32 3>
+ %2 = shufflevector <4 x float> %1, <4 x float> undef, <4 x i32> <i32 2, i32 0, i32 1, i32 3>
+ ret <4 x float> %2
+}
+; CHECK-LABEL: swizzle_24
+; Mask: [1,2,0,3]
+; CHECK: pshufd $-55
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x float> @swizzle_25(<4 x float> %v) {
+ %1 = shufflevector <4 x float> %v, <4 x float> undef, <4 x i32> <i32 2, i32 0, i32 3, i32 1>
+ %2 = shufflevector <4 x float> %1, <4 x float> undef, <4 x i32> <i32 2, i32 0, i32 3, i32 1>
+ ret <4 x float> %2
+}
+; CHECK-LABEL: swizzle_25
+; Mask: [3,2,1,0]
+; CHECK: pshufd $27
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x float> @swizzle_26(<4 x float> %v) {
+ %1 = shufflevector <4 x float> %v, <4 x float> undef, <4 x i32> <i32 0, i32 2, i32 3, i32 1>
+ %2 = shufflevector <4 x float> %1, <4 x float> undef, <4 x i32> <i32 0, i32 2, i32 3, i32 1>
+ ret <4 x float> %2
+}
+; CHECK-LABEL: swizzle_26
+; Mask: [0,3,1,2]
+; CHECK: pshufd $-100
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x float> @swizzle_27(<4 x float> %v) {
+ %1 = shufflevector <4 x float> %v, <4 x float> undef, <4 x i32> <i32 1, i32 3, i32 0, i32 2>
+ %2 = shufflevector <4 x float> %1, <4 x float> undef, <4 x i32> <i32 1, i32 3, i32 0, i32 2>
+ ret <4 x float> %2
+}
+; CHECK-LABEL: swizzle_27
+; Mask: [3,2,1,0]
+; CHECK: pshufd $27
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x float> @swizzle_28(<4 x float> %v) {
+ %1 = shufflevector <4 x float> %v, <4 x float> undef, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
+ %2 = shufflevector <4 x float> %1, <4 x float> undef, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
+ ret <4 x float> %2
+}
+; CHECK-LABEL: swizzle_28
+; Mask: [3,0,2,1]
+; CHECK: pshufd $99
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+
+define <4 x float> @swizzle_29(<4 x float> %v) {
+ %1 = shufflevector <4 x float> %v, <4 x float> undef, <4 x i32> <i32 3, i32 1, i32 2, i32 0>
+ %2 = shufflevector <4 x float> %1, <4 x float> undef, <4 x i32> <i32 1, i32 0, i32 2, i32 3>
+ ret <4 x float> %2
+}
+; CHECK-LABEL: swizzle_29
+; Mask: [1,3,2,0]
+; CHECK: pshufd $45
+; CHECK-NOT: pshufd
+; CHECK-NEXT: ret
+
+; Make sure that we combine the shuffles from each function below into a single
+; legal shuffle (either pshuflw or pshufb depending on the masks).
+
+define <8 x i16> @swizzle_30(<8 x i16> %v) {
+ %1 = shufflevector <8 x i16> %v, <8 x i16> undef, <8 x i32> <i32 3, i32 1, i32 2, i32 0, i32 7, i32 5, i32 6, i32 4>
+ %2 = shufflevector <8 x i16> %1, <8 x i16> undef, <8 x i32> <i32 1, i32 0, i32 2, i32 3, i32 7, i32 5, i32 6, i32 4>
+ ret <8 x i16> %2
+}
+; CHECK-LABEL: swizzle_30
+; Mask: [1,3,2,0,5,7,6,4]
+; CHECK: pshuflw $45
+; CHECK-NOT: pshufb
+; CHECK-NEXT: ret
+
+
+define <8 x i16> @swizzle_31(<8 x i16> %v) {
+ %1 = shufflevector <8 x i16> %v, <8 x i16> undef, <8 x i32> <i32 3, i32 0, i32 2, i32 1, i32 7, i32 5, i32 6, i32 4>
+ %2 = shufflevector <8 x i16> %1, <8 x i16> undef, <8 x i32> <i32 3, i32 0, i32 2, i32 1, i32 7, i32 5, i32 6, i32 4>
+ ret <8 x i16> %2
+}
+; CHECK-LABEL: swizzle_31
+; Mask: [1,3,2,0,4,5,6,7]
+; CHECK: pshuflw $45
+; CHECK-NOT: pshufb
+; CHECK: ret
+
+
+define <8 x i16> @swizzle_32(<8 x i16> %v) {
+ %1 = shufflevector <8 x i16> %v, <8 x i16> undef, <8 x i32> <i32 1, i32 2, i32 3, i32 0, i32 7, i32 5, i32 6, i32 4>
+ %2 = shufflevector <8 x i16> %1, <8 x i16> undef, <8 x i32> <i32 1, i32 2, i32 3, i32 0, i32 7, i32 5, i32 6, i32 4>
+ ret <8 x i16> %2
+}
+; CHECK-LABEL: swizzle_32
+; Mask: [2,3,0,1,4,5,6,7] --> equivalent to pshufd mask [1,0,2,3]
+; CHECK: pshufd $-31
+; CHECK-NOT: pshufb
+; CHECK: ret
+
+define <8 x i16> @swizzle_33(<8 x i16> %v) {
+ %1 = shufflevector <8 x i16> %v, <8 x i16> undef, <8 x i32> <i32 4, i32 6, i32 5, i32 7, i32 2, i32 3, i32 1, i32 0>
+ %2 = shufflevector <8 x i16> %1, <8 x i16> undef, <8 x i32> <i32 4, i32 6, i32 5, i32 7, i32 2, i32 3, i32 1, i32 0>
+ ret <8 x i16> %2
+}
+; CHECK-LABEL: swizzle_33
+; CHECK: pshufb
+; CHECK-NOT: pshufb
+; CHECK-NOT: shufpd
+; CHECK: ret
+
+
+define <8 x i16> @swizzle_34(<8 x i16> %v) {
+ %1 = shufflevector <8 x i16> %v, <8 x i16> undef, <8 x i32> <i32 4, i32 7, i32 6, i32 5, i32 1, i32 2, i32 0, i32 3>
+ %2 = shufflevector <8 x i16> %1, <8 x i16> undef, <8 x i32> <i32 4, i32 7, i32 6, i32 5, i32 1, i32 2, i32 0, i32 3>
+ ret <8 x i16> %2
+}
+; CHECK-LABEL: swizzle_34
+; CHECK: pshufb
+; CHECK-NOT: pshufb
+; CHECK-NOT: shufpd
+; CHECK: ret
+
+
+define <8 x i16> @swizzle_35(<8 x i16> %v) {
+ %1 = shufflevector <8 x i16> %v, <8 x i16> undef, <8 x i32> <i32 7, i32 4, i32 6, i32 5, i32 1, i32 3, i32 0, i32 2>
+ %2 = shufflevector <8 x i16> %1, <8 x i16> undef, <8 x i32> <i32 7, i32 4, i32 6, i32 5, i32 1, i32 3, i32 0, i32 2>
+ ret <8 x i16> %2
+}
+; CHECK-LABEL: swizzle_35
+; CHECK: pshufb
+; CHECK-NOT: pshufb
+; CHECK: ret
+
+
+define <8 x i16> @swizzle_36(<8 x i16> %v) {
+ %1 = shufflevector <8 x i16> %v, <8 x i16> undef, <8 x i32> <i32 4, i32 6, i32 7, i32 5, i32 0, i32 1, i32 3, i32 2>
+ %2 = shufflevector <8 x i16> %1, <8 x i16> undef, <8 x i32> <i32 4, i32 6, i32 7, i32 5, i32 0, i32 1, i32 3, i32 2>
+ ret <8 x i16> %2
+}
+; CHECK-LABEL: swizzle_36
+; CHECK: pshufb
+; CHECK-NOT: pshufb
+; CHECK-NOT: shufpd
+; CHECK: ret
+
+
+define <8 x i16> @swizzle_37(<8 x i16> %v) {
+ %1 = shufflevector <8 x i16> %v, <8 x i16> undef, <8 x i32> <i32 1, i32 0, i32 3, i32 2, i32 7, i32 5, i32 6, i32 4>
+ %2 = shufflevector <8 x i16> %1, <8 x i16> undef, <8 x i32> <i32 1, i32 0, i32 3, i32 2, i32 7, i32 4, i32 6, i32 5>
+ ret <8 x i16> %2
+}
+; CHECK-LABEL: swizzle_37
+; Mask: [0,1,2,3,4,7,6,5]
+; CHECK: pshufhw $108
+; CHECK-NOT: pshufb
+; CHECK: ret
+
+
+define <8 x i16> @swizzle_38(<8 x i16> %v) {
+ %1 = shufflevector <8 x i16> %v, <8 x i16> undef, <8 x i32> <i32 5, i32 6, i32 4, i32 7, i32 0, i32 2, i32 1, i32 3>
+ %2 = shufflevector <8 x i16> %1, <8 x i16> undef, <8 x i32> <i32 5, i32 6, i32 4, i32 7, i32 0, i32 2, i32 1, i32 3>
+ ret <8 x i16> %2
+}
+; CHECK-LABEL: swizzle_38
+; CHECK: pshufb
+; CHECK-NOT: pshufb
+; CHECK-NOT: shufpd
+; CHECK: ret
+
+
+define <8 x i16> @swizzle_39(<8 x i16> %v) {
+ %1 = shufflevector <8 x i16> %v, <8 x i16> undef, <8 x i32> <i32 5, i32 4, i32 6, i32 7, i32 3, i32 2, i32 1, i32 0>
+ %2 = shufflevector <8 x i16> %1, <8 x i16> undef, <8 x i32> <i32 5, i32 4, i32 6, i32 7, i32 3, i32 2, i32 1, i32 0>
+ ret <8 x i16> %2
+}
+; CHECK-LABEL: swizzle_39
+; CHECK: pshufb
+; CHECK-NOT: pshufb
+; CHECK-NOT: shufpd
+; CHECK: ret
+
+
+define <8 x i16> @swizzle_40(<8 x i16> %v) {
+ %1 = shufflevector <8 x i16> %v, <8 x i16> undef, <8 x i32> <i32 6, i32 4, i32 7, i32 5, i32 1, i32 0, i32 3, i32 2>
+ %2 = shufflevector <8 x i16> %1, <8 x i16> undef, <8 x i32> <i32 6, i32 4, i32 7, i32 5, i32 1, i32 0, i32 3, i32 2>
+ ret <8 x i16> %2
+}
+; CHECK-LABEL: swizzle_40
+; CHECK: pshufb
+; CHECK-NOT: pshufb
+; CHECK-NOT: shufpd
+; CHECK: ret
+
+
+define <8 x i16> @swizzle_41(<8 x i16> %v) {
+ %1 = shufflevector <8 x i16> %v, <8 x i16> undef, <8 x i32> <i32 6, i32 7, i32 5, i32 4, i32 0, i32 1, i32 3, i32 2>
+ %2 = shufflevector <8 x i16> %1, <8 x i16> undef, <8 x i32> <i32 6, i32 7, i32 5, i32 4, i32 0, i32 1, i32 3, i32 2>
+ ret <8 x i16> %2
+}
+; CHECK-LABEL: swizzle_41
+; CHECK: pshufb
+; CHECK-NOT: pshufb
+; CHECK-NOT: shufpd
+; CHECK: ret
+
+
+define <8 x i16> @swizzle_42(<8 x i16> %v) {
+ %1 = shufflevector <8 x i16> %v, <8 x i16> undef, <8 x i32> <i32 0, i32 1, i32 3, i32 2, i32 7, i32 6, i32 4, i32 5>
+ %2 = shufflevector <8 x i16> %1, <8 x i16> undef, <8 x i32> <i32 0, i32 1, i32 3, i32 2, i32 7, i32 6, i32 4, i32 5>
+ ret <8 x i16> %2
+}
+; CHECK-LABEL: swizzle_42
+; Mask: [0,1,2,3,5,4,7,6]
+; CHECK: pshufhw $-79
+; CHECK-NOT: pshufb
+; CHECK: ret
+
+
Index: test/CodeGen/X86/swizzle-avx2.ll
===================================================================
--- test/CodeGen/X86/swizzle-avx2.ll (revision 0)
+++ test/CodeGen/X86/swizzle-avx2.ll (working copy)
@@ -0,0 +1,91 @@
+; RUN: llc < %s -march=x86-64 -mtriple=x86_64-unknown-linux-gnu -mcpu=core-avx2 | FileCheck %s
+
+; Test that we correctly fold a shuffle that performs a swizzle of another
+; shuffle node according to the rule
+; shuffle (shuffle (x, undef, M0), undef, M1) -> shuffle(x, undef, M2)
+;
+; We only do this if the resulting mask is legal to avoid introducing an
+; illegal shuffle that is expanded into a sub-optimal sequence of instructions
+; during lowering stage.
+
+; Check that we produce a single vector permute / shuffle in all cases.
+
+define <8 x i32> @swizzle_1(<8 x i32> %v) {
+ %1 = shufflevector <8 x i32> %v, <8 x i32> undef, <8 x i32> <i32 3, i32 1, i32 2, i32 0, i32 7, i32 5, i32 6, i32 4>
+ %2 = shufflevector <8 x i32> %1, <8 x i32> undef, <8 x i32> <i32 1, i32 0, i32 2, i32 3, i32 7, i32 5, i32 6, i32 4>
+ ret <8 x i32> %2
+}
+; CHECK-LABEL: swizzle_1
+; CHECK: vpermd
+; CHECK-NOT: vpermd
+; CHECK: ret
+
+
+define <8 x i32> @swizzle_2(<8 x i32> %v) {
+ %1 = shufflevector <8 x i32> %v, <8 x i32> undef, <8 x i32> <i32 6, i32 7, i32 4, i32 5, i32 0, i32 1, i32 2, i32 3>
+ %2 = shufflevector <8 x i32> %1, <8 x i32> undef, <8 x i32> <i32 6, i32 7, i32 4, i32 5, i32 0, i32 1, i32 2, i32 3>
+ ret <8 x i32> %2
+}
+; CHECK-LABEL: swizzle_2
+; CHECK: vpshufd $78
+; CHECK-NOT: vpermd
+; CHECK-NOT: vpshufd
+; CHECK: ret
+
+
+define <8 x i32> @swizzle_3(<8 x i32> %v) {
+ %1 = shufflevector <8 x i32> %v, <8 x i32> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 2, i32 3, i32 0, i32 1>
+ %2 = shufflevector <8 x i32> %1, <8 x i32> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 2, i32 3, i32 0, i32 1>
+ ret <8 x i32> %2
+}
+; CHECK-LABEL: swizzle_3
+; CHECK: vpshufd $78
+; CHECK-NOT: vpermd
+; CHECK-NOT: vpshufd
+; CHECK: ret
+
+
+define <8 x i32> @swizzle_4(<8 x i32> %v) {
+ %1 = shufflevector <8 x i32> %v, <8 x i32> undef, <8 x i32> <i32 4, i32 7, i32 5, i32 6, i32 3, i32 2, i32 0, i32 1>
+ %2 = shufflevector <8 x i32> %1, <8 x i32> undef, <8 x i32> <i32 4, i32 7, i32 5, i32 6, i32 3, i32 2, i32 0, i32 1>
+ ret <8 x i32> %2
+}
+; CHECK-LABEL: swizzle_4
+; CHECK: vpermd
+; CHECK-NOT: vpermd
+; CHECK: ret
+
+
+define <8 x i32> @swizzle_5(<8 x i32> %v) {
+ %1 = shufflevector <8 x i32> %v, <8 x i32> undef, <8 x i32> <i32 7, i32 4, i32 6, i32 5, i32 0, i32 2, i32 1, i32 3>
+ %2 = shufflevector <8 x i32> %1, <8 x i32> undef, <8 x i32> <i32 7, i32 4, i32 6, i32 5, i32 0, i32 2, i32 1, i32 3>
+ ret <8 x i32> %2
+}
+; CHECK-LABEL: swizzle_5
+; CHECK: vpermd
+; CHECK-NOT: vpermd
+; CHECK: ret
+
+
+define <8 x i32> @swizzle_6(<8 x i32> %v) {
+ %1 = shufflevector <8 x i32> %v, <8 x i32> undef, <8 x i32> <i32 2, i32 1, i32 3, i32 0, i32 4, i32 7, i32 6, i32 5>
+ %2 = shufflevector <8 x i32> %1, <8 x i32> undef, <8 x i32> <i32 2, i32 1, i32 3, i32 0, i32 4, i32 7, i32 6, i32 5>
+ ret <8 x i32> %2
+}
+; CHECK-LABEL: swizzle_6
+; CHECK: vpermd
+; CHECK-NOT: vpermd
+; CHECK: ret
+
+
+define <8 x i32> @swizzle_7(<8 x i32> %v) {
+ %1 = shufflevector <8 x i32> %v, <8 x i32> undef, <8 x i32> <i32 0, i32 3, i32 1, i32 2, i32 5, i32 4, i32 6, i32 7>
+ %2 = shufflevector <8 x i32> %1, <8 x i32> undef, <8 x i32> <i32 0, i32 3, i32 1, i32 2, i32 5, i32 4, i32 6, i32 7>
+ ret <8 x i32> %2
+}
+; CHECK-LABEL: swizzle_7
+; CHECK: vpermd
+; CHECK-NOT: vpermd
+; CHECK: ret
+
+
More information about the llvm-commits
mailing list