[llvm] [SimplifyCFG] Treat umul + extract pattern as cheap single instruction. (PR #124933)

[Gábor Spaits via llvm-commits]

================
@@ -3286,7 +3286,21 @@ bool SimplifyCFGOpt::speculativelyExecuteBB(BranchInst *BI,
 
   SmallVector<Instruction *, 4> SpeculatedDbgIntrinsics;
 
+  // The number of already examined instructions. Debug instructions don't
+  // count!
   unsigned SpeculatedInstructions = 0;
+  // By default the number of instructions that may be speculatevly executed is
+  // one. Whenever a pattern is found in the basic block, that is cheap for sure
+  // we increase this number to the size of the pattern (how many instructions
+  // are there in that pattern).
+  unsigned MaxSpeculatedInstructionsToHoist = 1;
+  // In case we have found a cheap pattern, we don't want to do cost checking
+  // anymore. We are sure we want to hoist the pattern. To know, that we are
+  // only hoisting the cheap pattern only and not other expensive instructions
+  // too, we have the `MaxSpeculatedInstructionsToHoist` variable to track that
+  // the basic block truly only contains that pattern.
+  bool PartialInst = false;
----------------
spaits wrote:

We got rid of ` MaxSpeculatedInstructionsToHoist` and now we are only using ` MaxSpeculatedInstructionsToHoist`. I think now this can also be resolved.

https://github.com/llvm/llvm-project/pull/124933