[PATCH] D118807: [IRSim] Make sure that commutative intrinsics are treated as function calls without commutativity

[Andrew Litteken via Phabricator via llvm-commits]

AndrewLitteken created this revision.
AndrewLitteken added reviewers: paquette, uabelho.
Herald added a subscriber: hiraditya.
AndrewLitteken requested review of this revision.
Herald added a project: LLVM.
Herald added a subscriber: llvm-commits.

Created to fix: https://github.com/llvm/llvm-project/issues/53537

Some intrinsics functions are considered commutative since they are performing operations like addition or multiplication.  Some of these have extra parameters to provide extra information that are not part of the operation itself and are not commutative.  This makes sure that if an instruction that is an intrinsic takes the non commutative path to handle this case.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D118807

Files:
  llvm/lib/Analysis/IRSimilarityIdentifier.cpp
  llvm/unittests/Analysis/IRSimilarityIdentifierTest.cpp


Index: llvm/unittests/Analysis/IRSimilarityIdentifierTest.cpp
===================================================================
--- llvm/unittests/Analysis/IRSimilarityIdentifierTest.cpp
+++ llvm/unittests/Analysis/IRSimilarityIdentifierTest.cpp
@@ -2608,6 +2608,39 @@
   }
 }
 
+// This test makes sure that intrinsic functions that are marked commutative
+// are still treated as non-commutative since they are function calls.
+TEST(IRSimilarityIdentifier, InstrinsicCommutative) {
+  // If treated as commutative, we will fail to find a valid mapping, causing
+  // an assertion error.
+  StringRef ModuleString = R"(
+  define void @foo() {
+    entry:
+      %0 = call i16 @llvm.smul.fix.i16(i16 16384, i16 16384, i32 15)
+      store i16 %0, i16* undef, align 1
+      %1 = icmp eq i16 undef, 8192
+      call void @bar()
+      %2 = call i16 @llvm.smul.fix.i16(i16 -16384, i16 16384, i32 15)
+      store i16 %2, i16* undef, align 1
+      %3 = icmp eq i16 undef, -8192
+      call void @bar()
+      %4 = call i16 @llvm.smul.fix.i16(i16 -16384, i16 -16384, i32 15)
+      ret void
+  }
+
+  declare void @bar()
+
+  ; Function Attrs: nofree nosync nounwind readnone speculatable willreturn
+  declare i16 @llvm.smul.fix.i16(i16, i16, i32 immarg))";
+  LLVMContext Context;
+  std::unique_ptr<Module> M = makeLLVMModule(Context, ModuleString);
+
+  std::vector<std::vector<IRSimilarityCandidate>> SimilarityCandidates;
+  getSimilarities(*M, SimilarityCandidates);
+
+  ASSERT_TRUE(SimilarityCandidates.size() == 0);
+}
+
 // This test checks to see whether we can detect different structure in
 // commutative instructions.  In this case, the second operand in the second
 // add is different.
Index: llvm/lib/Analysis/IRSimilarityIdentifier.cpp
===================================================================
--- llvm/lib/Analysis/IRSimilarityIdentifier.cpp
+++ llvm/lib/Analysis/IRSimilarityIdentifier.cpp
@@ -791,7 +791,8 @@
     // We have different paths for commutative instructions and non-commutative
     // instructions since commutative instructions could allow multiple mappings
     // to certain values.
-    if (IA->isCommutative() && !isa<FPMathOperator>(IA)) {
+    if (IA->isCommutative() && !isa<FPMathOperator>(IA) &&
+        !isa<IntrinsicInst>(IA)) {
       if (!compareCommutativeOperandMapping(
               {A, OperValsA, ValueNumberMappingA},
               {B, OperValsB, ValueNumberMappingB}))


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D118807.405306.patch
Type: text/x-patch
Size: 2439 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20220202/508642ce/attachment.bin>