[PATCH] D93891: [DAGCombine] Respect the unsafe-fp-math when perform the transformation of fpow()

[Qing Shan Zhang via Phabricator via llvm-commits]

steven.zhang created this revision.
steven.zhang added reviewers: spatel, RKSimon, dmgreen, arsenm, PowerPC.
Herald added subscribers: hiraditya, nemanjai.
steven.zhang requested review of this revision.
Herald added a subscriber: wdng.
Herald added a project: LLVM.

Happen to see this issue. We miss to check the unsafe-fp-math when perform the transform for fpow.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D93891

Files:
  llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
  llvm/test/CodeGen/PowerPC/pow.75.ll


Index: llvm/test/CodeGen/PowerPC/pow.75.ll
===================================================================
--- llvm/test/CodeGen/PowerPC/pow.75.ll
+++ llvm/test/CodeGen/PowerPC/pow.75.ll
@@ -1,4 +1,3 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=powerpc64le-unknown-unknown -debug 2>&1 | FileCheck %s
 ; REQUIRES: asserts
 
@@ -46,3 +45,15 @@
   %r = call ninf afn <2 x double> @llvm.pow.v2f64(<2 x double> %x, <2 x double> <double 7.5e-1, double 7.5e-1>)
   ret <2 x double> %r
 }
+
+define float @pow_f32_three_fourth_fmf_unsafe(float %x) nounwind #0 {
+; CHECK: Combining: {{.*}}: f32 = fpow ninf [[X:t[0-9]+]], ConstantFP:f32<7.500000e-01>
+; CHECK-NEXT: Creating new node: [[SQRT:t[0-9]+]]: f32 = fsqrt ninf [[X]]
+; CHECK-NEXT: Creating new node: [[SQRTSQRT:t[0-9]+]]: f32 = fsqrt ninf [[SQRT]]
+; CHECK-NEXT: Creating new node: [[R:t[0-9]+]]: f32 = fmul ninf [[SQRT]], [[SQRTSQRT]]
+; CHECK-NEXT:  ... into: [[R]]: f32 = fmul ninf [[SQRT]], [[SQRTSQRT]]
+  %r = call ninf float @llvm.pow.f32(float %x, float 7.5e-01)
+  ret float %r
+}
+
+attributes #0 = { "unsafe-fp-math"="true" }
Index: llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
===================================================================
--- llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -13995,6 +13995,7 @@
   // TODO: Since we're approximating, we don't need an exact 1/3 exponent.
   //       Some range near 1/3 should be fine.
   EVT VT = N->getValueType(0);
+  const TargetOptions &Options = DAG.getTarget().Options;
   if ((VT == MVT::f32 && ExponentC->getValueAPF().isExactlyValue(1.0f/3.0f)) ||
       (VT == MVT::f64 && ExponentC->getValueAPF().isExactlyValue(1.0/3.0))) {
     // pow(-0.0, 1/3) = +0.0; cbrt(-0.0) = -0.0.
@@ -14005,7 +14006,7 @@
     // TODO: We could select out the special cases if we don't have nsz/ninf.
     SDNodeFlags Flags = N->getFlags();
     if (!Flags.hasNoSignedZeros() || !Flags.hasNoInfs() || !Flags.hasNoNaNs() ||
-        !Flags.hasApproximateFuncs())
+        (!Flags.hasApproximateFuncs() && !Options.UnsafeFPMath))
       return SDValue();
 
     // Do not create a cbrt() libcall if the target does not have it, and do not
@@ -14036,7 +14037,7 @@
 
     // We only need no signed zeros for the 0.25 case.
     if ((!Flags.hasNoSignedZeros() && ExponentIs025) || !Flags.hasNoInfs() ||
-        !Flags.hasApproximateFuncs())
+        (!Flags.hasApproximateFuncs() && !Options.UnsafeFPMath))
       return SDValue();
 
     // Don't double the number of libcalls. We are trying to inline fast code.


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D93891.313941.patch
Type: text/x-patch
Size: 2638 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20201229/f6b8211c/attachment.bin>