maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity

[X86][AARCH64] Improve ISD::ABS support 

This patch takes some of the code from D49837 to allow us to enable ISD::ABS support for all SSE vector types.

Differential Revision: https://reviews.llvm.org/D56544

llvm-svn: 350998
This commit is contained in:
Simon Pilgrim 2019-01-12 09:59:32 +00:00
parent 5f393eb5da
commit ca0de0363b
8 changed files with 85 additions and 29 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
llvm/include/llvm/CodeGen/TargetLowering.h
View File

@ -3787,6 +3787,14 @@ public:
/// \returns True, if the expansion was successful, false otherwise
bool expandCTTZ(SDNode *N, SDValue &Result, SelectionDAG &DAG) const;
/// Expand ABS nodes. Expands vector/scalar ABS nodes,
/// vector nodes can only succeed if all operations are legal/custom.
/// (ABS x) -> (XOR (ADD x, (SRA x, type_size)), (SRA x, type_size))
/// \param N Node to expand
/// \param Result output after conversion
/// \returns True, if the expansion was successful, false otherwise
bool expandABS(SDNode *N, SDValue &Result, SelectionDAG &DAG) const;
/// Turn load of vector type into a load of the individual elements.
/// \param LD load to expand
/// \returns MERGE_VALUEs of the scalar loads with their chains.

4
llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
View File

@ -2645,6 +2645,10 @@ bool SelectionDAGLegalize::ExpandNode(SDNode *Node) {
SDValue Tmp1, Tmp2, Tmp3, Tmp4;
bool NeedInvert;
switch (Node->getOpcode()) {
case ISD::ABS:
if (TLI.expandABS(Node, Tmp1, DAG))
Results.push_back(Tmp1);
break;
case ISD::CTPOP:
if (TLI.expandCTPOP(Node, Tmp1, DAG))
Results.push_back(Tmp1);

19
llvm/lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp
View File

@ -117,6 +117,12 @@ class VectorLegalizer {
/// the remaining lanes, finally bitcasting to the proper type.
SDValue ExpandZERO_EXTEND_VECTOR_INREG(SDValue Op);
/// Implement expand-based legalization of ABS vector operations.
/// If following expanding is legal/custom then do it:
/// (ABS x) --> (XOR (ADD x, (SRA x, sizeof(x)-1)), (SRA x, sizeof(x)-1))
/// else unroll the operation.
SDValue ExpandABS(SDValue Op);
/// Expand bswap of vectors into a shuffle if legal.
SDValue ExpandBSWAP(SDValue Op);
@ -355,6 +361,7 @@ SDValue VectorLegalizer::LegalizeOp(SDValue Op) {
case ISD::FSHR:
case ISD::ROTL:
case ISD::ROTR:
case ISD::ABS:
case ISD::BSWAP:
case ISD::BITREVERSE:
case ISD::CTLZ:
@ -749,6 +756,8 @@ SDValue VectorLegalizer::Expand(SDValue Op) {
return ExpandFSUB(Op);
case ISD::SETCC:
return UnrollVSETCC(Op);
case ISD::ABS:
return ExpandABS(Op);
case ISD::BITREVERSE:
return ExpandBITREVERSE(Op);
case ISD::CTPOP:
@ -1064,6 +1073,16 @@ SDValue VectorLegalizer::ExpandVSELECT(SDValue Op) {
return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Val);
}
SDValue VectorLegalizer::ExpandABS(SDValue Op) {
// Attempt to expand using TargetLowering.
SDValue Result;
if (TLI.expandABS(Op.getNode(), Result, DAG))
return Result;
// Otherwise go ahead and unroll.
return DAG.UnrollVectorOp(Op.getNode());
}
SDValue VectorLegalizer::ExpandFP_TO_UINT(SDValue Op) {
// Attempt to expand using TargetLowering.
SDValue Result;

20
llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
View File

@ -4715,6 +4715,26 @@ bool TargetLowering::expandCTTZ(SDNode *Node, SDValue &Result,
return true;
}
bool TargetLowering::expandABS(SDNode *N, SDValue &Result,
SelectionDAG &DAG) const {
SDLoc dl(N);
EVT VT = N->getValueType(0);
SDValue Op = N->getOperand(0);
// Only expand vector types if we have the appropriate vector operations.
if (VT.isVector() && (!isOperationLegalOrCustom(ISD::SRA, VT) ||
!isOperationLegalOrCustom(ISD::ADD, VT) ||
!isOperationLegalOrCustomOrPromote(ISD::XOR, VT)))
return false;
SDValue Shift =
DAG.getNode(ISD::SRA, dl, VT, Op,
DAG.getConstant(VT.getScalarSizeInBits() - 1, dl, VT));
SDValue Add = DAG.getNode(ISD::ADD, dl, VT, Op, Shift);
Result = DAG.getNode(ISD::XOR, dl, VT, Add, Shift);
return true;
}
SDValue TargetLowering::scalarizeVectorLoad(LoadSDNode *LD,
SelectionDAG &DAG) const {
SDLoc SL(LD);

19
llvm/lib/Target/X86/X86ISelLowering.cpp
View File

@ -867,6 +867,7 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64 }) {
setOperationAction(ISD::SETCC, VT, Custom);
setOperationAction(ISD::CTPOP, VT, Custom);
setOperationAction(ISD::ABS, VT, Custom);
// The condition codes aren't legal in SSE/AVX and under AVX512 we use
// setcc all the way to isel and prefer SETGT in some isel patterns.
@ -1207,6 +1208,7 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
setOperationAction(ISD::MULHU, MVT::v32i8, Custom);
setOperationAction(ISD::MULHS, MVT::v32i8, Custom);
setOperationAction(ISD::ABS, MVT::v4i64, Custom);
setOperationAction(ISD::SMAX, MVT::v4i64, Custom);
setOperationAction(ISD::UMAX, MVT::v4i64, Custom);
setOperationAction(ISD::SMIN, MVT::v4i64, Custom);
@ -23585,7 +23587,8 @@ static SDValue LowerADDSAT_SUBSAT(SDValue Op, SelectionDAG &DAG) {
return split256IntArith(Op, DAG);
}
static SDValue LowerABS(SDValue Op, SelectionDAG &DAG) {
static SDValue LowerABS(SDValue Op, const X86Subtarget &Subtarget,
SelectionDAG &DAG) {
MVT VT = Op.getSimpleValueType();
if (VT == MVT::i16 || VT == MVT::i32 || VT == MVT::i64) {
// Since X86 does not have CMOV for 8-bit integer, we don't convert
@ -23599,10 +23602,14 @@ static SDValue LowerABS(SDValue Op, SelectionDAG &DAG) {
return DAG.getNode(X86ISD::CMOV, DL, VT, Ops);
}
assert(Op.getSimpleValueType().is256BitVector() &&
Op.getSimpleValueType().isInteger() &&
"Only handle AVX 256-bit vector integer operation");
return Lower256IntUnary(Op, DAG);
if (VT.is256BitVector() && !Subtarget.hasInt256()) {
assert(VT.isInteger() &&
"Only handle AVX 256-bit vector integer operation");
return Lower256IntUnary(Op, DAG);
}
// Default to expand.
return SDValue();
}
static SDValue LowerMINMAX(SDValue Op, SelectionDAG &DAG) {
@ -26287,7 +26294,7 @@ SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
case ISD::SMIN:
case ISD::UMAX:
case ISD::UMIN: return LowerMINMAX(Op, DAG);
case ISD::ABS: return LowerABS(Op, DAG);
case ISD::ABS: return LowerABS(Op, Subtarget, DAG);
case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
case ISD::MLOAD: return LowerMLOAD(Op, Subtarget, DAG);
case ISD::MSTORE: return LowerMSTORE(Op, Subtarget, DAG);

3
llvm/test/CodeGen/AArch64/arm64-vabs.ll
View File

@ -542,7 +542,8 @@ define <1 x i64> @abs_1d(<1 x i64> %A) nounwind {
define i64 @abs_1d_honestly(i64 %A) nounwind {
; CHECK-LABEL: abs_1d_honestly:
; CHECK: abs d0, d0
; CHECK: cmp x0, #0
; CHECK-NEXT: cneg x0, x0, mi
%abs = call i64 @llvm.aarch64.neon.abs.i64(i64 %A)
ret i64 %abs
}

3
llvm/test/CodeGen/X86/combine-abs.ll
View File

@ -67,9 +67,6 @@ define <4 x i64> @combine_v4i64_abs_abs(<4 x i64> %a) {
; AVX2-LABEL: combine_v4i64_abs_abs:
; AVX2: # %bb.0:
; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX2-NEXT: vpcmpgtq %ymm0, %ymm1, %ymm2
; AVX2-NEXT: vpaddq %ymm2, %ymm0, %ymm0
; AVX2-NEXT: vpxor %ymm2, %ymm0, %ymm0
; AVX2-NEXT: vpcmpgtq %ymm0, %ymm1, %ymm1
; AVX2-NEXT: vpaddq %ymm1, %ymm0, %ymm0
; AVX2-NEXT: vpxor %ymm1, %ymm0, %ymm0

38
llvm/test/CodeGen/X86/viabs.ll
View File

@ -583,12 +583,12 @@ define <4 x i64> @test_abs_gt_v4i64(<4 x i64> %a) nounwind {
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
; AVX1-NEXT: vpxor %xmm2, %xmm2, %xmm2
; AVX1-NEXT: vpcmpgtq %xmm1, %xmm2, %xmm3
; AVX1-NEXT: vpcmpgtq %xmm0, %xmm2, %xmm2
; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm4
; AVX1-NEXT: vpaddq %xmm3, %xmm1, %xmm1
; AVX1-NEXT: vpxor %xmm3, %xmm1, %xmm1
; AVX1-NEXT: vpcmpgtq %xmm0, %xmm2, %xmm2
; AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
; AVX1-NEXT: vpxor %xmm2, %xmm0, %xmm0
; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVX1-NEXT: vxorps %ymm4, %ymm0, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: test_abs_gt_v4i64:
@ -639,20 +639,20 @@ define <8 x i64> @test_abs_le_v8i64(<8 x i64> %a) nounwind {
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm2
; AVX1-NEXT: vpxor %xmm3, %xmm3, %xmm3
; AVX1-NEXT: vpcmpgtq %xmm2, %xmm3, %xmm4
; AVX1-NEXT: vpcmpgtq %xmm0, %xmm3, %xmm5
; AVX1-NEXT: vinsertf128 $1, %xmm4, %ymm5, %ymm6
; AVX1-NEXT: vpaddq %xmm4, %xmm2, %xmm2
; AVX1-NEXT: vpaddq %xmm5, %xmm0, %xmm0
; AVX1-NEXT: vpxor %xmm4, %xmm2, %xmm2
; AVX1-NEXT: vpcmpgtq %xmm0, %xmm3, %xmm4
; AVX1-NEXT: vpaddq %xmm4, %xmm0, %xmm0
; AVX1-NEXT: vpxor %xmm4, %xmm0, %xmm0
; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0
; AVX1-NEXT: vxorps %ymm6, %ymm0, %ymm0
; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
; AVX1-NEXT: vpcmpgtq %xmm2, %xmm3, %xmm4
; AVX1-NEXT: vpcmpgtq %xmm1, %xmm3, %xmm3
; AVX1-NEXT: vinsertf128 $1, %xmm4, %ymm3, %ymm5
; AVX1-NEXT: vpaddq %xmm4, %xmm2, %xmm2
; AVX1-NEXT: vpxor %xmm4, %xmm2, %xmm2
; AVX1-NEXT: vpcmpgtq %xmm1, %xmm3, %xmm3
; AVX1-NEXT: vpaddq %xmm3, %xmm1, %xmm1
; AVX1-NEXT: vpxor %xmm3, %xmm1, %xmm1
; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm1
; AVX1-NEXT: vxorps %ymm5, %ymm1, %ymm1
; AVX1-NEXT: retq
;
; AVX2-LABEL: test_abs_le_v8i64:
@ -713,19 +713,19 @@ define <8 x i64> @test_abs_le_v8i64_fold(<8 x i64>* %a.ptr) nounwind {
; AVX1-NEXT: vmovdqu 48(%rdi), %xmm3
; AVX1-NEXT: vpxor %xmm4, %xmm4, %xmm4
; AVX1-NEXT: vpcmpgtq %xmm1, %xmm4, %xmm5
; AVX1-NEXT: vpcmpgtq %xmm0, %xmm4, %xmm6
; AVX1-NEXT: vinsertf128 $1, %xmm5, %ymm6, %ymm7
; AVX1-NEXT: vpaddq %xmm5, %xmm1, %xmm1
; AVX1-NEXT: vpaddq %xmm6, %xmm0, %xmm0
; AVX1-NEXT: vpxor %xmm5, %xmm1, %xmm1
; AVX1-NEXT: vpcmpgtq %xmm0, %xmm4, %xmm5
; AVX1-NEXT: vpaddq %xmm5, %xmm0, %xmm0
; AVX1-NEXT: vpxor %xmm5, %xmm0, %xmm0
; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVX1-NEXT: vxorps %ymm7, %ymm0, %ymm0
; AVX1-NEXT: vpcmpgtq %xmm3, %xmm4, %xmm1
; AVX1-NEXT: vpcmpgtq %xmm2, %xmm4, %xmm4
; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm4, %ymm5
; AVX1-NEXT: vpaddq %xmm1, %xmm3, %xmm1
; AVX1-NEXT: vpaddq %xmm4, %xmm2, %xmm2
; AVX1-NEXT: vpaddq %xmm1, %xmm3, %xmm3
; AVX1-NEXT: vpxor %xmm1, %xmm3, %xmm1
; AVX1-NEXT: vpcmpgtq %xmm2, %xmm4, %xmm3
; AVX1-NEXT: vpaddq %xmm3, %xmm2, %xmm2
; AVX1-NEXT: vpxor %xmm3, %xmm2, %xmm2
; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm2, %ymm1
; AVX1-NEXT: vxorps %ymm5, %ymm1, %ymm1
; AVX1-NEXT: retq
;
; AVX2-LABEL: test_abs_le_v8i64_fold: