[PATCH] D110869: [X86] Implement -fzero-call-used-regs option

[Craig Topper via Phabricator via cfe-commits]

craig.topper added inline comments.


================
Comment at: llvm/test/CodeGen/X86/zero-call-used-regs-fmod.ll:15
+; CHECK-NEXT:    xorps %xmm1, %xmm1
+; CHECK-NEXT:    popq %rax
+; CHECK-NEXT:    .cfi_def_cfa_offset 8
----------------
void wrote:
> craig.topper wrote:
> > Is it ok that this popq popped garbage into the %rax after you cleared it? It's not the return value of the function. That's xmm0. This is just a popq because its a shorter encoding than `sub 8, %esp`
> It's getting the value back from the `pushq %rax` instruction, isn't it? We shouldn't be clearing it, though. I'll see if I can fix that.
There was a mistake in my previous comment. The popq is being used in place of `add 8, %rsp`

In this case it probably is the value from the pushq. But %rax isn't a function argument so it is effectively garbage data. This is X86FrameLowering::emitSPUpdate scavenging %rax to use a pushq/popq instead of subtracting or adding to %rsp.

I think this is being done to align the stack for the call to fmod, but I'm not sure.

I think we can use a pushq like this to allocate stack space that gets overwritten before the popq.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D110869/new/

https://reviews.llvm.org/D110869