[PATCH] D26952: Fix for lost FastMathFlags in GVN

[Vyacheslav Klochkov via llvm-commits]

v_klochkov created this revision.
v_klochkov added subscribers: llvm-commits, mzolotukhin, majnemer.

Hello,

Please review the fix for lost fast-math flags in GVN (Global Value Numbering) optimization.

Thank you,
Vyacheslav Klochkov


https://reviews.llvm.org/D26952

Files:
  llvm/lib/Transforms/Scalar/GVN.cpp
  llvm/test/Transforms/GVN/propagate-ir-flags.ll


Index: llvm/lib/Transforms/Scalar/GVN.cpp
===================================================================
--- llvm/lib/Transforms/Scalar/GVN.cpp
+++ llvm/lib/Transforms/Scalar/GVN.cpp
@@ -1731,7 +1731,12 @@
 
   // Patch the replacement so that it is not more restrictive than the value
   // being replaced.
-  ReplInst->andIRFlags(I);
+  // Note that if 'I' is a load being replaced by some operation, 
+  // for example, by an arithmetic operation, then andIRFlags()
+  // would just erase all math flags from the original arithmetic
+  // operation, which is clearly not wanted and not needed.
+  if (!isa<LoadInst>(I))
+    ReplInst->andIRFlags(I);
 
   // FIXME: If both the original and replacement value are part of the
   // same control-flow region (meaning that the execution of one
Index: llvm/test/Transforms/GVN/propagate-ir-flags.ll
===================================================================
--- llvm/test/Transforms/GVN/propagate-ir-flags.ll
+++ llvm/test/Transforms/GVN/propagate-ir-flags.ll
@@ -0,0 +1,25 @@
+
+; RUN: opt < %s -gvn -S | FileCheck %s
+
+; CHECK-LABEL: @func_fast(
+; CHECK: fadd fast double
+define double @func_fast(double %a, double %b) {
+entry:
+  %a.addr = alloca double, align 8
+  %add = fadd fast double %b, 3.000000e+00
+  store double %add, double* %a.addr, align 8
+  %load_add = load double, double* %a.addr, align 8
+  ret double %load_add
+}
+
+; CHECK-LABEL: @func_no_fast(
+; CHECK: fadd double
+define double @func_no_fast(double %a, double %b) {
+entry:
+  %a.addr = alloca double, align 8
+  %add = fadd fast double %b, 3.000000e+00
+  store double %add, double* %a.addr, align 8
+  %duplicated_add = fadd double %b, 3.000000e+00
+  ret double %duplicated_add
+}
+


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D26952.78820.patch
Type: text/x-patch
Size: 1731 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20161122/d6573044/attachment.bin>