Add DAG-combines for aggressive FMA formation.
This patch adds DAG combines to form FMAs from pairs of FADD + FMUL or
FSUB + FMUL. The combines are performed when:
(a) Either
AllowExcessFPPrecision option (-enable-excess-fp-precision for llc)
OR
UnsafeFPMath option (-enable-unsafe-fp-math)
are set, and
(b) TargetLoweringInfo::isFMAFasterThanMulAndAdd(VT) is true for the type of
the FADD/FSUB, and
(c) The FMUL only has one user (the FADD/FSUB).
If your target has fast FMA instructions you can make use of these combines by
overriding TargetLoweringInfo::isFMAFasterThanMulAndAdd(VT) to return true for
types supported by your FMA instruction, and adding patterns to match ISD::FMA
to your FMA instructions.
llvm-svn: 158757
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 369d927..3517b7c 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -5633,6 +5633,26 @@
DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
N0.getOperand(1), N1));
+ // FADD -> FMA combines:
+ if ((DAG.getTarget().Options.AllowExcessFPPrecision ||
+ DAG.getTarget().Options.UnsafeFPMath) &&
+ DAG.getTarget().getTargetLowering()->isFMAFasterThanMulAndAdd(VT) &&
+ TLI.isOperationLegal(ISD::FMA, VT)) {
+
+ // fold (fadd (fmul x, y), z) -> (fma x, y, z)
+ if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse()) {
+ return DAG.getNode(ISD::FMA, N->getDebugLoc(), VT,
+ N0.getOperand(0), N0.getOperand(1), N1);
+ }
+
+ // fold (fadd x, (fmul y, z)) -> (fma x, y, z)
+ // Note: Commutes FADD operands.
+ if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse()) {
+ return DAG.getNode(ISD::FMA, N->getDebugLoc(), VT,
+ N1.getOperand(0), N1.getOperand(1), N0);
+ }
+ }
+
return SDValue();
}
@@ -5690,6 +5710,29 @@
}
}
+ // FSUB -> FMA combines:
+ if ((DAG.getTarget().Options.AllowExcessFPPrecision ||
+ DAG.getTarget().Options.UnsafeFPMath) &&
+ DAG.getTarget().getTargetLowering()->isFMAFasterThanMulAndAdd(VT) &&
+ TLI.isOperationLegal(ISD::FMA, VT)) {
+
+ // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
+ if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse()) {
+ return DAG.getNode(ISD::FMA, N->getDebugLoc(), VT,
+ N0.getOperand(0), N0.getOperand(1),
+ DAG.getNode(ISD::FNEG, N1->getDebugLoc(), VT, N1));
+ }
+
+ // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
+ // Note: Commutes FSUB operands.
+ if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse()) {
+ return DAG.getNode(ISD::FMA, N->getDebugLoc(), VT,
+ DAG.getNode(ISD::FNEG, N1->getDebugLoc(), VT,
+ N1.getOperand(0)),
+ N1.getOperand(1), N0);
+ }
+ }
+
return SDValue();
}
diff --git a/llvm/lib/Target/ARM/ARMInstrInfo.td b/llvm/lib/Target/ARM/ARMInstrInfo.td
index 5131152..81e3527 100644
--- a/llvm/lib/Target/ARM/ARMInstrInfo.td
+++ b/llvm/lib/Target/ARM/ARMInstrInfo.td
@@ -236,7 +236,7 @@
// Prefer fused MAC for fp mul + add over fp VMLA / VMLS if they are available.
// But only select them if more precision in FP computation is allowed.
// Do not use them for Darwin platforms.
-def UseFusedMAC : Predicate<"!TM.Options.NoExcessFPPrecision && "
+def UseFusedMAC : Predicate<"TM.Options.AllowExcessFPPrecision && "
"!Subtarget->isTargetDarwin()">;
def DontUseFusedMAC : Predicate<"!Subtarget->hasVFP4() || "
"Subtarget->isTargetDarwin()">;
diff --git a/llvm/lib/Target/PowerPC/PPCInstrInfo.td b/llvm/lib/Target/PowerPC/PPCInstrInfo.td
index c9cdd5c..25b6dc7 100644
--- a/llvm/lib/Target/PowerPC/PPCInstrInfo.td
+++ b/llvm/lib/Target/PowerPC/PPCInstrInfo.td
@@ -353,7 +353,7 @@
//===----------------------------------------------------------------------===//
// PowerPC Instruction Predicate Definitions.
-def FPContractions : Predicate<"!TM.Options.NoExcessFPPrecision">;
+def FPContractions : Predicate<"TM.Options.AllowExcessFPPrecision">;
def In32BitMode : Predicate<"!PPCSubTarget.isPPC64()">;
def In64BitMode : Predicate<"PPCSubTarget.isPPC64()">;
def IsBookE : Predicate<"PPCSubTarget.isBookE()">;