[llvm] [profcheck] Fix profile metadata in IntegerDivision/ExpandIRinsts (PR #173114)

[Jin Huang via llvm-commits]

================
@@ -235,11 +244,53 @@ static Value *generateUnsignedDivisionCode(Value *Dividend, Value *Divisor,
   Value *Tmp1 = Builder.CreateCall(CTLZ, {Dividend, True});
   Value *SR          = Builder.CreateSub(Tmp0, Tmp1);
   Value *Ret0_4      = Builder.CreateICmpUGT(SR, MSB);
+
+  // Add 'unlikely' branch weights to the select instruction ('Ret0') generated
+  // by 'CreateLogicalOr'. We assume the 'true' path, where either the divisor
+  // or dividend is zero ('Ret0_3'), is less likely than the 'false' path,
+  // which corresponds to the |divisor| > |dividend| ('Ret0_4').
   Value *Ret0        = Builder.CreateLogicalOr(Ret0_3, Ret0_4);
+  if (!ProfcheckDisableMetadataFixes) {
+    if (Instruction *Ret0SelectI = dyn_cast<Instruction>(Ret0)) {
+      Ret0SelectI->setMetadata(
+          LLVMContext::MD_prof,
+          MDBuilder(Ret0SelectI->getContext()).createUnlikelyBranchWeights());
+    }
+  }
   Value *RetDividend = Builder.CreateICmpEQ(SR, MSB);
-  Value *RetVal      = Builder.CreateSelect(Ret0, Zero, Dividend);
-  Value *EarlyRet    = Builder.CreateLogicalOr(Ret0, RetDividend);
-  Builder.CreateCondBr(EarlyRet, End, BB1);
+
+  // This select instruction (RetVal) immediately follows the Ret0 check. Since
+  // Ret0_4 determines if the divisor is greater then the dividend so the
+  // unknown (50/50) branch weights are fairly accurate for RetVal.
----------------
jinhuang1102 wrote:

> both the dividend and the divisor are being randomly sampled from the same distribution...

Yes, this was my assumptoion. 

> I would expect the common case to be that the dividend is greater than the divisor.

I agree. That makes sense.

https://github.com/llvm/llvm-project/pull/173114