[PATCH] D80163: [X86][VARARG] Avoid spilling xmm registers for va_start.

[Alexey Lapshin via Phabricator via llvm-commits]

avl added inline comments.


================
Comment at: llvm/lib/Target/X86/X86ExpandPseudo.cpp:531
+
+    GuardedBlock->addLiveIn(
+        SaveVarargXmmRegsInstr->getOperand(OpndIdx).getReg());
----------------
avl wrote:
> craig.topper wrote:
> > Why do we only add them ass liveins this late? Shouldn't they have been livein when the block was created in the custom inserter?
> Things were done that way(fill liveins when the block was created 
> in the custom inserter) in the first version of the fix.
> But it creates new liveness rule for physical registers
> (newly added block should have physical registers in its liveins 
> but it is not allowed before regalloc):
> 
> https://reviews.llvm.org/D69372?id=227838#change-TJhqJRPUjpDM
> 
> So there was a request to do things in other way, not requiring 
> to change liveness rule:
> 
> https://reviews.llvm.org/D69372#1734756
> 
> The solution done in that patch creates a block with SAVE_VARARG_XMM_REGS
> pseudo instruction. That pseudo instruction defines xmm registers 
> (used by varargs). After register allocator, SAVE_VARARG_XMM_REGS pseudo 
> instruction is expanded to the real copy instructions and put xmm registers 
> into the block liveins. 
But... Reading this recent comment https://reviews.llvm.org/D80163#2224926 I think I implemented not exactly what was requested. Implementation from this patch does not hide control flow. Would do implementation which will hide control flow.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D80163/new/

https://reviews.llvm.org/D80163