[llvm] [InstCombine] Preserve profile data with select instructions and binary operators (PR #158375)

[Mircea Trofin via llvm-commits]

================
@@ -1381,24 +1384,28 @@ Value *InstCombinerImpl::SimplifySelectsFeedingBinaryOp(BinaryOperator &I,
     // We need an 'add' and exactly 1 arm of the select to have been simplified.
     if (Opcode != Instruction::Add || (!True && !False) || (True && False))
       return nullptr;
-
-    Value *N;
+    Value *N, *SI = nullptr;
     if (True && match(FVal, m_Neg(m_Value(N)))) {
       Value *Sub = Builder.CreateSub(Z, N);
-      return Builder.CreateSelect(Cond, True, Sub, I.getName());
+      SI = Builder.CreateSelect(Cond, True, Sub, I.getName());
     }
     if (False && match(TVal, m_Neg(m_Value(N)))) {
       Value *Sub = Builder.CreateSub(Z, N);
-      return Builder.CreateSelect(Cond, Sub, False, I.getName());
+      SI = Builder.CreateSelect(Cond, Sub, False, I.getName());
     }
-    return nullptr;
+    if (!ProfcheckDisableMetadataFixes && SI)
+      cast<SelectInst>(SI)->setMetadata(LLVMContext::MD_prof, ProfileData);
+    return SI;
   };
 
   if (LHSIsSelect && RHSIsSelect && A == D) {
     // (A ? B : C) op (A ? E : F) -> A ? (B op E) : (C op F)
     Cond = A;
     True = simplifyBinOp(Opcode, B, E, FMF, Q);
     False = simplifyBinOp(Opcode, C, F, FMF, Q);
+    // Profile weights for both LHS and RHS should be the same because they have
----------------
mtrofin wrote:

there are 2 cases where the pointers won't be the same:

1) the branch weights are e.g. 2, 4 and 1, 2. So same probability, different expression
2) the metadata was added as unique, and then you'd have 2 different objects, content-wise identical

IDK, may be safer to LLVM_DEBUG. The assert may get someone on a wild goose chase in the cases above. Or add a comment above it that we don't expect those cases to happen, but if they do, safe to replace the assert w a LLVM_DEBUG or something of the sort.

https://github.com/llvm/llvm-project/pull/158375