[llvm] [RegisterCoalescer] Prefer copy over rematerialization when smaller (PR #182279)

[Matt Arsenault via llvm-commits]

================
@@ -1319,6 +1319,15 @@ bool RegisterCoalescer::reMaterializeDef(const CoalescerPair &CP,
   if (!TII->isReMaterializable(*DefMI))
     return false;
 
+  // For multiple uses of a cheap instruction, prefer to copy if it affects code
+  // size.
+  if (!MRI->hasOneNonDBGUse(SrcReg)) {
+    unsigned RematSize = TII->getInstSizeInBytes(*DefMI);
----------------
arsenm wrote:

Does getInstSizeInBytes even work correctly for pseudoinstructions? I'm pretty sure the AMDGPU implementation will get the wrong answer, at least.

I'm not sure this make sense unless maybe you're using minsize? The spill code you're increasing the chances of emitting may be larger 

https://github.com/llvm/llvm-project/pull/182279