[llvm] [MCA][X86] Pretend To Have a Stack Engine (PR #153348)

[Andrea Di Biagio via llvm-commits]

================
@@ -36,11 +36,31 @@ void X86InstrPostProcess::setMemBarriers(std::unique_ptr<Instruction> &Inst,
   }
 }
 
+void X86InstrPostProcess::useStackEngine(std::unique_ptr<Instruction> &Inst,
+                                         const MCInst &MCI) {
+  if (X86::isPOP(MCI.getOpcode())) {
+    assert(Inst->getUses().size() == 1 &&
+           "Expected pop instruction to only use stack pointer register");
+    Inst->getUses().clear();
+  }
----------------
adibiagio wrote:

> I think I made a bit of a mountain out of a mole hill.
> 
> <snip>
> 
> This means that in MCA we can just clear writes to rsp in stack operation instructions.

I think it is reasonable. However, I wouldn't always the write always. Otherwise, you won't correctly simulate perf for the case where a sync is required.

Just to be clear: the idea of removing the write makes sense.

I think you should not remove a write to RSP if the next RSP user isn't also a stack operation. That way, you would simulate the sync. You still need to post-process instructions imho.



https://github.com/llvm/llvm-project/pull/153348