[llvm] [AMDGPU][AMDGPULateCodeGenPrepare] Combine scalarized selects back into vector selects (PR #173990)

[Matt Arsenault via llvm-commits]

================
@@ -225,6 +253,20 @@ bool AMDGPULateCodeGenPrepare::run() {
       Changed |= LRO.optimizeLiveType(&I, DeadInsts);
     }
 
+  // Combine scalarized selects back into vector selects.
+  // This uses a top-down approach: iterate over bitcasts (i32 vec -> i8 vec)
+  // and collect all select instructions that use extracted elements with a
+  // zero false value. By starting from the bitcast, we process each source
+  // exactly once, avoiding redundant work when multiple selects share a source.
+  if (CombineScalarSelects) {
----------------
arsenm wrote:

This shouldn't be a separate pass over blocks 

https://github.com/llvm/llvm-project/pull/173990