diff --git a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp index 23821e6f603..a83203b6635 100644 --- a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp +++ b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp @@ -5729,6 +5729,18 @@ SDValue DAGCombiner::visitFADD(SDNode *N) { DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0.getOperand(1), N1)); + // If allow, fold (fadd (fneg x), x) -> 0.0 + if (DAG.getTarget().Options.UnsafeFPMath && + N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1) { + return DAG.getConstantFP(0.0, VT); + } + + // If allow, fold (fadd x, (fneg x)) -> 0.0 + if (DAG.getTarget().Options.UnsafeFPMath && + N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0) { + return DAG.getConstantFP(0.0, VT); + } + // In unsafe math mode, we can fold chains of FADD's of the same value // into multiplications. This transform is not safe in general because // we are reducing the number of rounding steps. @@ -6038,6 +6050,12 @@ SDValue DAGCombiner::visitFMA(SDNode *N) { EVT VT = N->getValueType(0); DebugLoc dl = N->getDebugLoc(); + if (DAG.getTarget().Options.UnsafeFPMath) { + if (N0CFP && N0CFP->isZero()) + return N2; + if (N1CFP && N1CFP->isZero()) + return N2; + } if (N0CFP && N0CFP->isExactlyValue(1.0)) return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N1, N2); if (N1CFP && N1CFP->isExactlyValue(1.0)) diff --git a/test/CodeGen/X86/fp-fast.ll b/test/CodeGen/X86/fp-fast.ll index 091f0de9308..192a1273580 100644 --- a/test/CodeGen/X86/fp-fast.ll +++ b/test/CodeGen/X86/fp-fast.ll @@ -35,3 +35,23 @@ define float @test3(float %a) { ret float %r } +; CHECK: test4 +define float @test4(float %a) { +; CHECK-NOT: fma +; CHECK-NOT mul +; CHECK-NOT: add +; CHECK: ret + %t1 = fmul float %a, 0.0 + %t2 = fadd float %a, %t1 + ret float %t2 +} + +; CHECK: test5 +define float @test5(float %a) { +; CHECK-NOT: add +; CHECK: vxorps +; CHECK: ret + %t1 = fsub float -0.0, %a + %t2 = fadd float %a, %t1 + ret float %t2 +}