OSchip
/
llvm-project
mirror of https://github.com/THU-DSP-LAB/llvm-project.git
13
0
Fork 6
Code Issues Packages Projects Releases Wiki Activity

[X86][SSE] Add support for combining ANDNP byte masks with target shuffles 

llvm-svn: 293178
This commit is contained in:
Simon Pilgrim 2017-01-26 14:31:12 +00:00
parent 0b034d6f25
commit 027bb453d9
2 changed files with 47 additions and 39 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

30
llvm/lib/Target/X86/X86ISelLowering.cpp
View File

@ -4132,6 +4132,7 @@ static bool isTargetShuffleVariableMask(unsigned Opcode) {
return true; return true;
// 'Faux' Target Shuffles. // 'Faux' Target Shuffles.
case ISD::AND: case ISD::AND:
case X86ISD::ANDNP:
return true; return true;
} }
} }
@ -5742,11 +5743,16 @@ static bool getFauxShuffleMask(SDValue N, SmallVectorImpl<int> &Mask,
unsigned Opcode = N.getOpcode(); unsigned Opcode = N.getOpcode();
switch (Opcode) { switch (Opcode) {
case ISD::AND: { case ISD::AND:
case X86ISD::ANDNP: {
// Attempt to decode as a per-byte mask. // Attempt to decode as a per-byte mask.
SmallBitVector UndefElts; SmallBitVector UndefElts;
SmallVector<APInt, 32> EltBits; SmallVector<APInt, 32> EltBits;
if (!getTargetConstantBitsFromNode(N.getOperand(1), 8, UndefElts, EltBits)) SDValue N0 = N.getOperand(0);
SDValue N1 = N.getOperand(1);
bool IsAndN = (X86ISD::ANDNP == Opcode);
uint64_t ZeroMask = IsAndN ? 255 : 0;
if (!getTargetConstantBitsFromNode(IsAndN ? N0 : N1, 8, UndefElts, EltBits))
return false; return false;
for (int i = 0, e = (int)EltBits.size(); i != e; ++i) { for (int i = 0, e = (int)EltBits.size(); i != e; ++i) {
if (UndefElts[i]) { if (UndefElts[i]) {
@ -5756,9 +5762,9 @@ static bool getFauxShuffleMask(SDValue N, SmallVectorImpl<int> &Mask,
uint64_t ByteBits = EltBits[i].getZExtValue(); uint64_t ByteBits = EltBits[i].getZExtValue();
if (ByteBits != 0 && ByteBits != 255) if (ByteBits != 0 && ByteBits != 255)
return false; return false;
Mask.push_back(ByteBits == 0 ? SM_SentinelZero : i); Mask.push_back(ByteBits == ZeroMask ? SM_SentinelZero : i);
} }
Ops.push_back(N.getOperand(0)); Ops.push_back(IsAndN ? N1 : N0);
return true; return true;
} }
case X86ISD::VSHLI: case X86ISD::VSHLI:
@ -33010,6 +33016,7 @@ static SDValue combineFMinNumFMaxNum(SDNode *N, SelectionDAG &DAG,
/// Do target-specific dag combines on X86ISD::ANDNP nodes. /// Do target-specific dag combines on X86ISD::ANDNP nodes.
static SDValue combineAndnp(SDNode *N, SelectionDAG &DAG, static SDValue combineAndnp(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
const X86Subtarget &Subtarget) { const X86Subtarget &Subtarget) {
// ANDNP(0, x) -> x // ANDNP(0, x) -> x
if (ISD::isBuildVectorAllZeros(N->getOperand(0).getNode())) if (ISD::isBuildVectorAllZeros(N->getOperand(0).getNode()))
@ -33019,6 +33026,19 @@ static SDValue combineAndnp(SDNode *N, SelectionDAG &DAG,
if (ISD::isBuildVectorAllZeros(N->getOperand(1).getNode())) if (ISD::isBuildVectorAllZeros(N->getOperand(1).getNode()))
return getZeroVector(N->getSimpleValueType(0), Subtarget, DAG, SDLoc(N)); return getZeroVector(N->getSimpleValueType(0), Subtarget, DAG, SDLoc(N));
EVT VT = N->getValueType(0);
// Attempt to recursively combine a bitmask ANDNP with shuffles.
if (VT.isVector() && (VT.getScalarSizeInBits() % 8) == 0) {
SDValue Op(N, 0);
SmallVector<int, 1> NonceMask; // Just a placeholder.
NonceMask.push_back(0);
if (combineX86ShufflesRecursively({Op}, 0, Op, NonceMask,
/*Depth*/ 1, /*HasVarMask*/ false, DAG,
DCI, Subtarget))
return SDValue(); // This routine will use CombineTo to replace N.
}
return SDValue(); return SDValue();
} }
@ -34088,7 +34108,7 @@ SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
case ISD::FSUB: return combineFaddFsub(N, DAG, Subtarget); case ISD::FSUB: return combineFaddFsub(N, DAG, Subtarget);
case ISD::FNEG: return combineFneg(N, DAG, Subtarget); case ISD::FNEG: return combineFneg(N, DAG, Subtarget);
case ISD::TRUNCATE: return combineTruncate(N, DAG, Subtarget); case ISD::TRUNCATE: return combineTruncate(N, DAG, Subtarget);
case X86ISD::ANDNP: return combineAndnp(N, DAG, Subtarget); case X86ISD::ANDNP: return combineAndnp(N, DAG, DCI, Subtarget);
case X86ISD::FAND: return combineFAnd(N, DAG, Subtarget); case X86ISD::FAND: return combineFAnd(N, DAG, Subtarget);
case X86ISD::FANDN: return combineFAndn(N, DAG, Subtarget); case X86ISD::FANDN: return combineFAndn(N, DAG, Subtarget);
case X86ISD::FXOR: case X86ISD::FXOR:

54
llvm/test/CodeGen/X86/clear_upper_vector_element_bits.ll
View File

@ -356,43 +356,31 @@ define <4 x i32> @_clearupper4xi32b(<4 x i32>) nounwind {
define <8 x i16> @_clearupper8xi16b(<8 x i16>) nounwind { define <8 x i16> @_clearupper8xi16b(<8 x i16>) nounwind {
; SSE-LABEL: _clearupper8xi16b: ; SSE-LABEL: _clearupper8xi16b:
; SSE: # BB#0: ; SSE: # BB#0:
; SSE-NEXT: movdqa {{.*#+}} xmm2 = [255,0,255,255,255,255,255,255,255,255,255,255,255,255,255,255] ; SSE-NEXT: movdqa {{.*#+}} xmm1 = [255,0,255,255,255,255,255,255,255,255,255,255,255,255,255,255]
; SSE-NEXT: pand %xmm2, %xmm0 ; SSE-NEXT: pand %xmm1, %xmm0
; SSE-NEXT: xorl %eax, %eax ; SSE-NEXT: xorl %eax, %eax
; SSE-NEXT: movd %eax, %xmm1 ; SSE-NEXT: movd %eax, %xmm2
; SSE-NEXT: movdqa %xmm1, %xmm3 ; SSE-NEXT: movdqa %xmm2, %xmm3
; SSE-NEXT: psllw $8, %xmm3 ; SSE-NEXT: psllw $8, %xmm3
; SSE-NEXT: pandn %xmm3, %xmm2 ; SSE-NEXT: pandn %xmm3, %xmm1
; SSE-NEXT: por %xmm2, %xmm0 ; SSE-NEXT: por %xmm1, %xmm0
; SSE-NEXT: movdqa {{.*#+}} xmm2 = [255,255,255,0,255,255,255,255,255,255,255,255,255,255,255,255] ; SSE-NEXT: movdqa {{.*#+}} xmm1 = [255,255,255,0,255,255,255,255,255,255,255,255,255,255,255,255]
; SSE-NEXT: pand %xmm2, %xmm0 ; SSE-NEXT: pand %xmm1, %xmm0
; SSE-NEXT: movdqa %xmm1, %xmm3 ; SSE-NEXT: movdqa %xmm2, %xmm3
; SSE-NEXT: pslld $24, %xmm3 ; SSE-NEXT: pslld $24, %xmm3
; SSE-NEXT: pandn %xmm3, %xmm2 ; SSE-NEXT: pandn %xmm3, %xmm1
; SSE-NEXT: por %xmm2, %xmm0 ; SSE-NEXT: por %xmm1, %xmm0
; SSE-NEXT: movdqa {{.*#+}} xmm2 = [255,255,255,255,255,0,255,255,255,255,255,255,255,255,255,255] ; SSE-NEXT: movdqa {{.*#+}} xmm1 = [255,255,255,255,255,0,255,255,255,255,255,255,255,255,255,255]
; SSE-NEXT: pand %xmm2, %xmm0 ; SSE-NEXT: pand %xmm1, %xmm0
; SSE-NEXT: movdqa %xmm1, %xmm3 ; SSE-NEXT: movdqa %xmm2, %xmm3
; SSE-NEXT: psllq $40, %xmm3 ; SSE-NEXT: psllq $40, %xmm3
; SSE-NEXT: pandn %xmm3, %xmm2 ; SSE-NEXT: pandn %xmm3, %xmm1
; SSE-NEXT: por %xmm2, %xmm0 ; SSE-NEXT: por %xmm1, %xmm0
; SSE-NEXT: movdqa {{.*#+}} xmm2 = [255,255,255,255,255,255,255,0,255,255,255,255,255,255,255,255] ; SSE-NEXT: movdqa {{.*#+}} xmm1 = [255,255,255,255,255,255,255,0,255,255,255,255,255,255,255,255]
; SSE-NEXT: pand %xmm2, %xmm0 ; SSE-NEXT: pand %xmm1, %xmm0
; SSE-NEXT: movdqa %xmm1, %xmm3 ; SSE-NEXT: psllq $56, %xmm2
; SSE-NEXT: psllq $56, %xmm3 ; SSE-NEXT: pandn %xmm2, %xmm1
; SSE-NEXT: pandn %xmm3, %xmm2 ; SSE-NEXT: por %xmm1, %xmm0
; SSE-NEXT: por %xmm2, %xmm0
; SSE-NEXT: movdqa {{.*#+}} xmm2 = [255,255,255,255,255,255,255,255,255,0,255,255,255,255,255,255]
; SSE-NEXT: pand %xmm2, %xmm0
; SSE-NEXT: movdqa %xmm1, %xmm3
; SSE-NEXT: pslldq {{.*#+}} xmm3 = zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm3[0,1,2,3,4,5,6]
; SSE-NEXT: pandn %xmm3, %xmm2
; SSE-NEXT: por %xmm2, %xmm0
; SSE-NEXT: movdqa {{.*#+}} xmm2 = [255,255,255,255,255,255,255,255,255,255,255,0,255,255,255,255]
; SSE-NEXT: pand %xmm2, %xmm0
; SSE-NEXT: pslldq {{.*#+}} xmm1 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm1[0,1,2,3,4]
; SSE-NEXT: pandn %xmm1, %xmm2
; SSE-NEXT: por %xmm2, %xmm0
; SSE-NEXT: pand {{.*}}(%rip), %xmm0 ; SSE-NEXT: pand {{.*}}(%rip), %xmm0
; SSE-NEXT: retq ; SSE-NEXT: retq
; ;