[PATCH] D20263: X86: Avoid using _chkstk when lowering WIN_ALLOCA instructions

[Michael Kuperstein via llvm-commits]

mkuper added inline comments.

================
Comment at: lib/Target/X86/X86WinAllocaExpander.cpp:181
@@ +180,3 @@
+                 MI.getOpcode() == X86::ADJCALLSTACKUP64) {
+        Offset -= MI.getOperand(0).getImm();
+      } else if (MI.getOpcode() == X86::ADJCALLSTACKDOWN32 ||
----------------
This looks a bit scary to me. It can cause an offset to be temporary negative, and negative offsets are special-cased now, right? (Well, we'll probably never hit "-1" due to alignment anyway, but...)

I think I can live with this, though, especially since, as Dave's example shows, allocations within this region are probably broken as is, so it's not like this would make anything worse. But if we keep it as is, then I think we need at least a FIXME.

================
Comment at: lib/Target/X86/X86WinAllocaExpander.cpp:264
@@ +263,3 @@
+    MachineInstr *AmountDef = MRI->getUniqueVRegDef(AmountReg);
+    if (!AmountDef)
+      break;
----------------
Never mind, I misread, sorry for the noise.


http://reviews.llvm.org/D20263