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

[X86] Stop changing f128 fand/for/fxor to v2i64. 

The additional patterns don't cost us much and it seems better than changing element widths.

llvm-svn: 345564
This commit is contained in:
Craig Topper 2018-10-30 03:27:11 +00:00
parent d8e14a5901
commit 676d7a7a43
2 changed files with 37 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
llvm/lib/Target/X86/X86ISelLowering.cpp
View File

@ -37745,10 +37745,12 @@ static SDValue lowerX86FPLogicOp(SDNode *N, SelectionDAG &DAG,
const X86Subtarget &Subtarget) { const X86Subtarget &Subtarget) {
MVT VT = N->getSimpleValueType(0); MVT VT = N->getSimpleValueType(0);
// If we have integer vector types available, use the integer opcodes. // If we have integer vector types available, use the integer opcodes.
if ((VT.isVector() || VT == MVT::f128) && Subtarget.hasSSE2()) { if (!VT.isVector() || !Subtarget.hasSSE2())
return SDValue();
SDLoc dl(N); SDLoc dl(N);
unsigned IntBits = std::min(VT.getScalarSizeInBits(), 64U); unsigned IntBits = VT.getScalarSizeInBits();
MVT IntSVT = MVT::getIntegerVT(IntBits); MVT IntSVT = MVT::getIntegerVT(IntBits);
MVT IntVT = MVT::getVectorVT(IntSVT, VT.getSizeInBits() / IntBits); MVT IntVT = MVT::getVectorVT(IntSVT, VT.getSizeInBits() / IntBits);
@ -37765,8 +37767,6 @@ static SDValue lowerX86FPLogicOp(SDNode *N, SelectionDAG &DAG,
SDValue IntOp = DAG.getNode(IntOpcode, dl, IntVT, Op0, Op1); SDValue IntOp = DAG.getNode(IntOpcode, dl, IntVT, Op0, Op1);
return DAG.getBitcast(VT, IntOp); return DAG.getBitcast(VT, IntOp);
} }
return SDValue();
}
/// Fold a xor(setcc cond, val), 1 --> setcc (inverted(cond), val) /// Fold a xor(setcc cond, val), 1 --> setcc (inverted(cond), val)

21
llvm/lib/Target/X86/X86InstrVecCompiler.td
View File

@ -466,8 +466,6 @@ def : Pat<(loadf128 addr:$src),
(VMOVUPSZ128rm addr:$src)>; (VMOVUPSZ128rm addr:$src)>;
} }
// With SSE2 the DAG combiner converts fp logic ops to integer logic ops to
// reduce patterns.
let Predicates = [UseSSE1] in { let Predicates = [UseSSE1] in {
// andps is shorter than andpd or pand. andps is SSE and andpd/pand are in SSE2 // andps is shorter than andpd or pand. andps is SSE and andpd/pand are in SSE2
def : Pat<(f128 (X86fand VR128:$src1, (memopf128 addr:$src2))), def : Pat<(f128 (X86fand VR128:$src1, (memopf128 addr:$src2))),
@ -489,4 +487,23 @@ def : Pat<(f128 (X86fxor VR128:$src1, VR128:$src2)),
(XORPSrr VR128:$src1, VR128:$src2)>; (XORPSrr VR128:$src1, VR128:$src2)>;
} }
let Predicates = [HasAVX] in {
// andps is shorter than andpd or pand. andps is SSE and andpd/pand are in SSE2
def : Pat<(f128 (X86fand VR128:$src1, (loadf128 addr:$src2))),
(VANDPSrm VR128:$src1, f128mem:$src2)>;
def : Pat<(f128 (X86fand VR128:$src1, VR128:$src2)),
(VANDPSrr VR128:$src1, VR128:$src2)>;
def : Pat<(f128 (X86for VR128:$src1, (loadf128 addr:$src2))),
(VORPSrm VR128:$src1, f128mem:$src2)>;
def : Pat<(f128 (X86for VR128:$src1, VR128:$src2)),
(VORPSrr VR128:$src1, VR128:$src2)>;
def : Pat<(f128 (X86fxor VR128:$src1, (loadf128 addr:$src2))),
(VXORPSrm VR128:$src1, f128mem:$src2)>;
def : Pat<(f128 (X86fxor VR128:$src1, VR128:$src2)),
(VXORPSrr VR128:$src1, VR128:$src2)>;
}