[llvm] [AMDGPU] Move INIT_EXEC lowering from SILowerControlFlow to SIWholeQuadMode (PR #94452)

[Jay Foad via llvm-commits]

================
@@ -1605,11 +1703,14 @@ bool SIWholeQuadMode::runOnMachineFunction(MachineFunction &MF) {
 
   // Shader is simple does not need any state changes or any complex lowering
   if (!(GlobalFlags & (StateWQM | StateStrict)) && LowerToCopyInstrs.empty() &&
-      LowerToMovInstrs.empty() && KillInstrs.empty()) {
+      LowerToMovInstrs.empty() && KillInstrs.empty() &&
+      InitExecInstrs.empty()) {
----------------
jayfoad wrote:

Done. But really, as a future cleanup, I'd prefer to remove this fast path. Instead we should make sure that the "slow" path below is actually fast when all the `Instrs` vectors are empty, and returns an accurate "Changed" flag.

https://github.com/llvm/llvm-project/pull/94452