[llvm] [X86][APX] Reuse EFLAGS across multi-predecessor blocks via NF in optimizeCompareInstr (PR #208184)

[Phoebe Wang via llvm-commits]

================
@@ -5474,10 +5560,35 @@ bool X86InstrInfo::optimizeCompareInstr(MachineInstr &CmpInstr, Register SrcReg,
     if (MI || Sub)
       break;
 
-    // Reached begin of basic block. Continue in predecessor if there is
-    // exactly one.
-    if (MBB->pred_size() != 1)
-      return false;
+    // Reached the begin of the basic block. If it has exactly one predecessor,
+    // continue the backward scan there. Otherwise (multiple predecessors), try
+    // to reuse EFLAGS from a dominating producer (handled below).
+    if (MBB->pred_size() != 1) {
+      // The block has multiple predecessors. We can still reuse EFLAGS from an
+      // equivalent flag producer that dominates CmpInstr, provided every path
+      // from that producer to CmpInstr only clobbers EFLAGS via instructions
+      // that have an NF (no-flags) variant (which requires APX). This handles
+      // patterns like (CMP duplicated by CodeGenPrepare across a diamond):
+      //   entry:  cmp %x, C   ; br
+      //   bb1:    imul ...     ; clobbers EFLAGS  ->  {nf} imul
+      //   bb2:    ...
+      //   bb3:    cmp %x, C    ; <-- redundant, reuse EFLAGS from entry
+      //           cmovcc ...
+      //
+      // Only attempt this when no NF conversion has been collected yet. An NF
+      // conversion grows code size, and any already collected lies on the
+      // single-predecessor chain to CmpInstr, i.e. the unconditional path the
+      // compare removal saves, so converting it is pure added cost. Conversions
+      // the dominating walk collects instead lie on the mutually exclusive
+      // diamond arms and are only conditionally executed.
+      if (HasNF && InstsToUpdate.empty())
----------------
phoebewang wrote:

Why InstsToUpdate must be empty?

https://github.com/llvm/llvm-project/pull/208184