[llvm] r251450 - Use the 'arcp' fast-math-flag when combining repeated FP divisors
Sanjay Patel via llvm-commits
llvm-commits at lists.llvm.org
Tue Oct 27 13:27:25 PDT 2015
Author: spatel
Date: Tue Oct 27 15:27:25 2015
New Revision: 251450
URL: http://llvm.org/viewvc/llvm-project?rev=251450&view=rev
Log:
Use the 'arcp' fast-math-flag when combining repeated FP divisors
This is a usage of the IR-level fast-math-flags now that they are propagated to SDNodes.
This was originally part of D8900.
Removing the global 'enable-unsafe-fp-math' checks will require auto-upgrade and
possibly other changes.
Differential Revision: http://reviews.llvm.org/D9708
Modified:
llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
llvm/trunk/test/CodeGen/X86/fdiv-combine.ll
Modified: llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp?rev=251450&r1=251449&r2=251450&view=diff
==============================================================================
--- llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp (original)
+++ llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp Tue Oct 27 15:27:25 2015
@@ -8477,7 +8477,9 @@ SDValue DAGCombiner::visitFMA(SDNode *N)
// FDIVs may be lower than the cost of one FDIV and two FMULs. Another reason
// is the critical path is increased from "one FDIV" to "one FDIV + one FMUL".
SDValue DAGCombiner::combineRepeatedFPDivisors(SDNode *N) {
- if (!DAG.getTarget().Options.UnsafeFPMath)
+ bool UnsafeMath = DAG.getTarget().Options.UnsafeFPMath;
+ const SDNodeFlags *Flags = N->getFlags();
+ if (!UnsafeMath && !Flags->hasAllowReciprocal())
return SDValue();
// Skip if current node is a reciprocal.
@@ -8496,9 +8498,14 @@ SDValue DAGCombiner::combineRepeatedFPDi
// Find all FDIV users of the same divisor.
// Use a set because duplicates may be present in the user list.
SetVector<SDNode *> Users;
- for (auto *U : N1->uses())
- if (U->getOpcode() == ISD::FDIV && U->getOperand(1) == N1)
- Users.insert(U);
+ for (auto *U : N1->uses()) {
+ if (U->getOpcode() == ISD::FDIV && U->getOperand(1) == N1) {
+ // This division is eligible for optimization only if global unsafe math
+ // is enabled or if this division allows reciprocal formation.
+ if (UnsafeMath || U->getFlags()->hasAllowReciprocal())
+ Users.insert(U);
+ }
+ }
// Now that we have the actual number of divisor uses, make sure it meets
// the minimum threshold specified by the target.
@@ -8508,7 +8515,6 @@ SDValue DAGCombiner::combineRepeatedFPDi
EVT VT = N->getValueType(0);
SDLoc DL(N);
SDValue FPOne = DAG.getConstantFP(1.0, DL, VT);
- const SDNodeFlags *Flags = &cast<BinaryWithFlagsSDNode>(N)->Flags;
SDValue Reciprocal = DAG.getNode(ISD::FDIV, DL, VT, FPOne, N1, Flags);
// Dividend / Divisor -> Dividend * Reciprocal
Modified: llvm/trunk/test/CodeGen/X86/fdiv-combine.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/fdiv-combine.ll?rev=251450&r1=251449&r2=251450&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/fdiv-combine.ll (original)
+++ llvm/trunk/test/CodeGen/X86/fdiv-combine.ll Tue Oct 27 15:27:25 2015
@@ -1,9 +1,11 @@
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=x86-64 | FileCheck %s
-; Anything more than one division using a single divisor operand
+; More than one 'arcp' division using a single divisor operand
; should be converted into a reciprocal and multiplication.
-define float @div1_arcp(float %x, float %y, float %z) #0 {
+; Don't do anything for just one division.
+
+define float @div1_arcp(float %x, float %y, float %z) {
; CHECK-LABEL: div1_arcp:
; CHECK: # BB#0:
; CHECK-NEXT: divss %xmm1, %xmm0
@@ -12,13 +14,15 @@ define float @div1_arcp(float %x, float
ret float %div1
}
-define float @div2_arcp(float %x, float %y, float %z) #0 {
-; CHECK-LABEL: div2_arcp:
+; All math instructions are 'arcp', so optimize.
+
+define float @div2_arcp_all(float %x, float %y, float %z) {
+; CHECK-LABEL: div2_arcp_all:
; CHECK: # BB#0:
; CHECK-NEXT: movss {{.*#+}} xmm3 = mem[0],zero,zero,zero
; CHECK-NEXT: divss %xmm2, %xmm3
-; CHECK-NEXT: mulss %xmm1, %xmm0
; CHECK-NEXT: mulss %xmm3, %xmm0
+; CHECK-NEXT: mulss %xmm1, %xmm0
; CHECK-NEXT: mulss %xmm3, %xmm0
; CHECK-NEXT: retq
%div1 = fdiv arcp float %x, %z
@@ -27,10 +31,57 @@ define float @div2_arcp(float %x, float
ret float %div2
}
+; The first division is not 'arcp', so do not optimize.
+
+define float @div2_arcp_partial1(float %x, float %y, float %z) {
+; CHECK-LABEL: div2_arcp_partial1:
+; CHECK: # BB#0:
+; CHECK-NEXT: divss %xmm2, %xmm0
+; CHECK-NEXT: mulss %xmm1, %xmm0
+; CHECK-NEXT: divss %xmm2, %xmm0
+; CHECK-NEXT: retq
+ %div1 = fdiv float %x, %z
+ %mul = fmul arcp float %div1, %y
+ %div2 = fdiv arcp float %mul, %z
+ ret float %div2
+}
+
+; The second division is not 'arcp', so do not optimize.
+
+define float @div2_arcp_partial2(float %x, float %y, float %z) {
+; CHECK-LABEL: div2_arcp_partial2:
+; CHECK: # BB#0:
+; CHECK-NEXT: divss %xmm2, %xmm0
+; CHECK-NEXT: mulss %xmm1, %xmm0
+; CHECK-NEXT: divss %xmm2, %xmm0
+; CHECK-NEXT: retq
+ %div1 = fdiv arcp float %x, %z
+ %mul = fmul arcp float %div1, %y
+ %div2 = fdiv float %mul, %z
+ ret float %div2
+}
+
+; The multiply is not 'arcp', but that does not prevent optimizing the divisions.
+
+define float @div2_arcp_partial3(float %x, float %y, float %z) {
+; CHECK-LABEL: div2_arcp_partial3:
+; CHECK: # BB#0:
+; CHECK-NEXT: movss {{.*#+}} xmm3 = mem[0],zero,zero,zero
+; CHECK-NEXT: divss %xmm2, %xmm3
+; CHECK-NEXT: mulss %xmm3, %xmm0
+; CHECK-NEXT: mulss %xmm1, %xmm0
+; CHECK-NEXT: mulss %xmm3, %xmm0
+; CHECK-NEXT: retq
+ %div1 = fdiv arcp float %x, %z
+ %mul = fmul float %div1, %y
+ %div2 = fdiv arcp float %mul, %z
+ ret float %div2
+}
+
; If the reciprocal is already calculated, we should not
; generate an extra multiplication by 1.0.
-define double @div3_arcp(double %x, double %y, double %z) #0 {
+define double @div3_arcp(double %x, double %y, double %z) {
; CHECK-LABEL: div3_arcp:
; CHECK: # BB#0:
; CHECK-NEXT: movsd{{.*#+}} xmm2 = mem[0],zero
@@ -44,7 +95,7 @@ define double @div3_arcp(double %x, doub
ret double %ret
}
-define void @PR24141() #0 {
+define void @PR24141() {
; CHECK-LABEL: PR24141:
; CHECK: callq
; CHECK-NEXT: divsd
@@ -57,11 +108,9 @@ while.body:
%call = call { double, double } @g(double %x.0)
%xv0 = extractvalue { double, double } %call, 0
%xv1 = extractvalue { double, double } %call, 1
- %div = fdiv double %xv0, %xv1
+ %div = fdiv arcp double %xv0, %xv1
br label %while.body
}
declare { double, double } @g(double)
-; FIXME: If the backend understands 'arcp', then this attribute is unnecessary.
-attributes #0 = { "unsafe-fp-math"="true" }
More information about the llvm-commits
mailing list