replica
/
hanchenye-llvm-project
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[X86][SSE] Lower 128-bit MOVDDUP with existing VBROADCAST mechanisms 

We have a number of useful lowering strategies for VBROADCAST instructions (both from memory and register element 0) which the 128-bit form of the MOVDDUP instruction can make use of.

This patch tweaks lowerVectorShuffleAsBroadcast to enable it to broadcast 2f64 args using MOVDDUP as well.

It does require a slight tweak to the lowerVectorShuffleAsBroadcast mechanism as the existing MOVDDUP lowering uses isShuffleEquivalent which can match binary shuffles that can lower to (unary) broadcasts.

Differential Revision: http://reviews.llvm.org/D17680

llvm-svn: 262478
This commit is contained in:
Simon Pilgrim 2016-03-02 11:43:05 +00:00
parent f2fbabe9c1
commit c02b72627a
6 changed files with 61 additions and 55 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

90
llvm/lib/Target/X86/X86ISelLowering.cpp
View File

@ -8359,29 +8359,40 @@ static SDValue lowerVectorShuffleAsTruncBroadcast(SDLoc DL, MVT VT, SDValue V0,
/// filtering. While a little annoying to re-dispatch on type here, there isn't
/// a convenient way to factor it out.
/// FIXME: This is very similar to LowerVectorBroadcast - can we merge them?
static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
ArrayRef<int> Mask,
static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V1,
SDValue V2, ArrayRef<int> Mask,
const X86Subtarget &Subtarget,
SelectionDAG &DAG) {
if (!Subtarget.hasAVX())
return SDValue();
if (VT.isInteger() && !Subtarget.hasAVX2())
if (!((Subtarget.hasSSE3() && VT == MVT::v2f64) ||
(Subtarget.hasAVX() && VT.isFloatingPoint()) ||
(Subtarget.hasAVX2() && VT.isInteger())))
return SDValue();
// With MOVDDUP (v2f64) we can broadcast from a register or a load, otherwise
// we can only broadcast from a register with AVX2.
unsigned NumElts = Mask.size();
unsigned Opcode = VT == MVT::v2f64 ? X86ISD::MOVDDUP : X86ISD::VBROADCAST;
bool BroadcastFromReg = (Opcode == X86ISD::MOVDDUP) || Subtarget.hasAVX2();
// Check that the mask is a broadcast.
int BroadcastIdx = -1;
for (int M : Mask)
if (M >= 0 && BroadcastIdx == -1)
BroadcastIdx = M;
else if (M >= 0 && M != BroadcastIdx)
return SDValue();
for (int i = 0; i != (int)NumElts; ++i) {
SmallVector<int, 8> BroadcastMask(NumElts, i);
if (isShuffleEquivalent(V1, V2, Mask, BroadcastMask)) {
BroadcastIdx = i;
break;
}
}
if (BroadcastIdx < 0)
return SDValue();
assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
"a sorted mask where the broadcast "
"comes from V1.");
// Go up the chain of (vector) values to find a scalar load that we can
// combine with the broadcast.
SDValue V = V1;
for (;;) {
switch (V.getOpcode()) {
case ISD::CONCAT_VECTORS: {
@ -8434,9 +8445,8 @@ static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
(V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
V = V.getOperand(BroadcastIdx);
// If the scalar isn't a load, we can't broadcast from it in AVX1.
// Only AVX2 has register broadcasts.
if (!Subtarget.hasAVX2() && !isShuffleFoldableLoad(V))
// If we can't broadcast from a register, check that the input is a load.
if (!BroadcastFromReg && !isShuffleFoldableLoad(V))
return SDValue();
} else if (MayFoldLoad(BC) && !cast<LoadSDNode>(BC)->isVolatile()) {
// 32-bit targets need to load i64 as a f64 and then bitcast the result.
@ -8453,8 +8463,8 @@ static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
V = DAG.getLoad(SVT, DL, Ld->getChain(), NewAddr,
DAG.getMachineFunction().getMachineMemOperand(
Ld->getMemOperand(), Offset, SVT.getStoreSize()));
} else if (!Subtarget.hasAVX2()) {
// We can't broadcast from a vector register without AVX2.
} else if (!BroadcastFromReg) {
// We can't broadcast from a vector register.
return SDValue();
} else if (BroadcastIdx != 0) {
// We can only broadcast from the zero-element of a vector register,
@ -8477,8 +8487,10 @@ static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
DAG.getIntPtrConstant(BroadcastIdx, DL));
}
V = DAG.getNode(X86ISD::VBROADCAST, DL, BroadcastVT, V);
return DAG.getBitcast(VT, V);
if (Opcode == X86ISD::MOVDDUP && !V.getValueType().isVector())
V = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V);
return DAG.getBitcast(VT, DAG.getNode(Opcode, DL, BroadcastVT, V));
}
// Check for whether we can use INSERTPS to perform the shuffle. We only use
@ -8694,10 +8706,10 @@ static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
if (isSingleInputShuffleMask(Mask)) {
// Use low duplicate instructions for masks that match their pattern.
if (Subtarget.hasSSE3())
if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
// Check for being able to broadcast a single element.
if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(
DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
return Broadcast;
// Straight shuffle of a single input vector. Simulate this by using the
// single input as both of the "inputs" to this instruction..
@ -8776,8 +8788,8 @@ static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
if (isSingleInputShuffleMask(Mask)) {
// Check for being able to broadcast a single element.
if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
Mask, Subtarget, DAG))
if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(
DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
return Broadcast;
// Straight shuffle of a single input vector. For everything from SSE2
@ -8998,8 +9010,8 @@ static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
if (NumV2Elements == 0) {
// Check for being able to broadcast a single element.
if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
Mask, Subtarget, DAG))
if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(
DL, MVT::v4f32, V1, V2, Mask, Subtarget, DAG))
return Broadcast;
// Use even/odd duplicate instructions for masks that match their pattern.
@ -9090,8 +9102,8 @@ static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
if (NumV2Elements == 0) {
// Check for being able to broadcast a single element.
if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
Mask, Subtarget, DAG))
if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(
DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
return Broadcast;
// Straight shuffle of a single input vector. For everything from SSE2
@ -9730,8 +9742,8 @@ static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
if (NumV2Inputs == 0) {
// Check for being able to broadcast a single element.
if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
Mask, Subtarget, DAG))
if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(
DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
return Broadcast;
// Try to use shift instructions.
@ -9929,8 +9941,8 @@ static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
// For single-input shuffles, there are some nicer lowering tricks we can use.
if (NumV2Elements == 0) {
// Check for being able to broadcast a single element.
if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
Mask, Subtarget, DAG))
if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(
DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
return Broadcast;
// Check whether we can widen this to an i16 shuffle by duplicating bytes.
@ -10982,8 +10994,8 @@ static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
if (isSingleInputShuffleMask(Mask)) {
// Check for being able to broadcast a single element.
if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
Mask, Subtarget, DAG))
if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(
DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
return Broadcast;
// Use low duplicate instructions for masks that match their pattern.
@ -11080,7 +11092,7 @@ static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
return Blend;
// Check for being able to broadcast a single element.
if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1, V2,
Mask, Subtarget, DAG))
return Broadcast;
@ -11153,7 +11165,7 @@ static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
return Blend;
// Check for being able to broadcast a single element.
if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1, V2,
Mask, Subtarget, DAG))
return Broadcast;
@ -11259,7 +11271,7 @@ static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
return Blend;
// Check for being able to broadcast a single element.
if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1, V2,
Mask, Subtarget, DAG))
return Broadcast;
@ -11341,7 +11353,7 @@ static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
return ZExt;
// Check for being able to broadcast a single element.
if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1, V2,
Mask, Subtarget, DAG))
return Broadcast;
@ -11438,7 +11450,7 @@ static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
return ZExt;
// Check for being able to broadcast a single element.
if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1, V2,
Mask, Subtarget, DAG))
return Broadcast;
@ -11748,7 +11760,7 @@ static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
// Check for being able to broadcast a single element.
if (SDValue Broadcast =
lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
lowerVectorShuffleAsBroadcast(DL, VT, V1, V2, Mask, Subtarget, DAG))
return Broadcast;
// Dispatch to each element type for lowering. If we don't have supprot for

3
llvm/test/CodeGen/X86/avx-splat.ll
View File

@ -135,8 +135,7 @@ entry:
define <2 x double> @splat_load_2f64_11(<2 x double>* %ptr) {
; CHECK-LABEL: splat_load_2f64_11:
; CHECK: ## BB#0:
; CHECK-NEXT: vmovaps (%rdi), %xmm0
; CHECK-NEXT: vmovhlps {{.*#+}} xmm0 = xmm0[1,1]
; CHECK-NEXT: vmovddup {{.*#+}} xmm0 = mem[0,0]
; CHECK-NEXT: retq
%x = load <2 x double>, <2 x double>* %ptr
%x1 = shufflevector <2 x double> %x, <2 x double> undef, <2 x i32> <i32 1, i32 1>

6
llvm/test/CodeGen/X86/avx-vbroadcast.ll
View File

@ -315,14 +315,12 @@ define <2 x double> @load_splat_2f64_2f64_1111(<2 x double>* %ptr) nounwind uwta
; X32-LABEL: load_splat_2f64_2f64_1111:
; X32: ## BB#0: ## %entry
; X32-NEXT: movl {{[0-9]+}}(%esp), %eax
; X32-NEXT: vmovaps (%eax), %xmm0
; X32-NEXT: vmovhlps {{.*#+}} xmm0 = xmm0[1,1]
; X32-NEXT: vmovddup {{.*#+}} xmm0 = mem[0,0]
; X32-NEXT: retl
;
; X64-LABEL: load_splat_2f64_2f64_1111:
; X64: ## BB#0: ## %entry
; X64-NEXT: vmovaps (%rdi), %xmm0
; X64-NEXT: vmovhlps {{.*#+}} xmm0 = xmm0[1,1]
; X64-NEXT: vmovddup {{.*#+}} xmm0 = mem[0,0]
; X64-NEXT: retq
entry:
%ld = load <2 x double>, <2 x double>* %ptr

6
llvm/test/CodeGen/X86/avx2-vbroadcast.ll
View File

@ -494,14 +494,12 @@ define <2 x double> @load_splat_2f64_2f64_1111(<2 x double>* %ptr) nounwind uwta
; X32-LABEL: load_splat_2f64_2f64_1111:
; X32: ## BB#0: ## %entry
; X32-NEXT: movl {{[0-9]+}}(%esp), %eax
; X32-NEXT: vmovaps (%eax), %xmm0
; X32-NEXT: vmovhlps {{.*#+}} xmm0 = xmm0[1,1]
; X32-NEXT: vmovddup {{.*#+}} xmm0 = mem[0,0]
; X32-NEXT: retl
;
; X64-LABEL: load_splat_2f64_2f64_1111:
; X64: ## BB#0: ## %entry
; X64-NEXT: vmovaps (%rdi), %xmm0
; X64-NEXT: vmovhlps {{.*#+}} xmm0 = xmm0[1,1]
; X64-NEXT: vmovddup {{.*#+}} xmm0 = mem[0,0]
; X64-NEXT: retq
entry:
%ld = load <2 x double>, <2 x double>* %ptr

8
llvm/test/CodeGen/X86/extractelement-load.ll
View File

@ -63,13 +63,13 @@ define void @t3() {
;
; X64-SSSE3-LABEL: t3:
; X64-SSSE3: # BB#0: # %bb
; X64-SSSE3-NEXT: movupd (%rax), %xmm0
; X64-SSSE3-NEXT: movhpd %xmm0, (%rax)
; X64-SSSE3-NEXT: movddup {{.*#+}} xmm0 = mem[0,0]
; X64-SSSE3-NEXT: movlpd %xmm0, (%rax)
;
; X64-AVX-LABEL: t3:
; X64-AVX: # BB#0: # %bb
; X64-AVX-NEXT: vmovupd (%rax), %xmm0
; X64-AVX-NEXT: vmovhpd %xmm0, (%rax)
; X64-AVX-NEXT: vmovddup {{.*#+}} xmm0 = mem[0,0]
; X64-AVX-NEXT: vmovlpd %xmm0, (%rax)
bb:
%tmp13 = load <2 x double>, <2 x double>* undef, align 1
%.sroa.3.24.vec.extract = extractelement <2 x double> %tmp13, i32 1

3
llvm/test/CodeGen/X86/vec_extract-sse4.ll
View File

@ -20,8 +20,7 @@ define float @t2(<4 x float>* %P1) nounwind {
; CHECK: # BB#0:
; CHECK-NEXT: pushl %eax
; CHECK-NEXT: movl {{[0-9]+}}(%esp), %eax
; CHECK-NEXT: movapd (%eax), %xmm0
; CHECK-NEXT: shufpd {{.*#+}} xmm0 = xmm0[1,0]
; CHECK-NEXT: movddup {{.*#+}} xmm0 = mem[0,0]
; CHECK-NEXT: movss %xmm0, (%esp)
; CHECK-NEXT: flds (%esp)
; CHECK-NEXT: popl %eax