[IR] Introduce hash_value for FMF, GEPNoWrapFlags

[artagnon]

This patch is motivated by wanting to do CSE at the VPlan-level: as VPlan stores FMF and GEPNoWrapFlags among other things in its recipes, de-duplicating them for common-subexpression-elimination would require hashing the VPlan recipes along with their embedded information.

[llvmbot]

@llvm/pr-subscribers-llvm-ir

Author: Ramkumar Ramachandra (artagnon)

Changes

This patch is motivated by wanting to do CSE at the VPlan-level: as VPlan stores FMF and GEPNoWrapFlags among other things in its recipes, de-duplicating them for common-subexpression-elimination would require hashing the VPlan recipes along with their embedded information.

---

Full diff: https://github.com/llvm/llvm-project/pull/151946.diff

2 Files Affected:

• (modified) llvm/include/llvm/IR/FMF.h (+7)
• (modified) llvm/include/llvm/IR/GEPNoWrapFlags.h (+7)

diff --git a/llvm/include/llvm/IR/FMF.h b/llvm/include/llvm/IR/FMF.h
index 8bd591250a38a..da4d9f564bd5d 100644
--- a/llvm/include/llvm/IR/FMF.h
+++ b/llvm/include/llvm/IR/FMF.h
@@ -13,6 +13,7 @@
 #ifndef LLVM_IR_FMF_H
 #define LLVM_IR_FMF_H
 
+#include "llvm/ADT/Hashing.h"
 #include "llvm/Support/Compiler.h"
 
 namespace llvm {
@@ -121,6 +122,9 @@ class FastMathFlags {
     const unsigned ValueMask = NoNaNs | NoInfs | NoSignedZeros;
     return FastMathFlags(ValueMask & (LHS.Flags | RHS.Flags));
   }
+
+  // Hashing.
+  friend hash_code hash_value(const FastMathFlags &FMF);
 };
 
 inline FastMathFlags operator|(FastMathFlags LHS, FastMathFlags RHS) {
@@ -138,6 +142,9 @@ inline raw_ostream &operator<<(raw_ostream &O, FastMathFlags FMF) {
   return O;
 }
 
+inline hash_code hash_value(const FastMathFlags &FMF) {
+  return hash_value(FMF.Flags);
+}
 } // end namespace llvm
 
 #endif // LLVM_IR_FMF_H
diff --git a/llvm/include/llvm/IR/GEPNoWrapFlags.h b/llvm/include/llvm/IR/GEPNoWrapFlags.h
index 4e6ab0d88bfcf..35974b323d1e1 100644
--- a/llvm/include/llvm/IR/GEPNoWrapFlags.h
+++ b/llvm/include/llvm/IR/GEPNoWrapFlags.h
@@ -13,6 +13,7 @@
 #ifndef LLVM_IR_GEPNOWRAPFLAGS_H
 #define LLVM_IR_GEPNOWRAPFLAGS_H
 
+#include "llvm/ADT/Hashing.h"
 #include <assert.h>
 
 namespace llvm {
@@ -101,8 +102,14 @@ class GEPNoWrapFlags {
     Flags |= Other.Flags;
     return *this;
   }
+
+  // Hashing.
+  friend hash_code hash_value(const GEPNoWrapFlags &NW);
 };
 
+inline hash_code hash_value(const GEPNoWrapFlags &NW) {
+  return hash_value(NW.Flags);
+}
 } // end namespace llvm
 
 #endif // LLVM_IR_GEPNOWRAPFLAGS_H

[fhahn]

I don't think that's needed. Same as IR CSE, we should drop the flags when re-using a value