[PATCH] D41187: Ignore metainstructions during the shrink wrap analysis

[Matthias Braun via Phabricator via llvm-commits]

MatzeB added inline comments.


================
Comment at: llvm/trunk/lib/CodeGen/ShrinkWrap.cpp:244
+  // Ignore DBG_VALUE and other meta instructions that must not affect codegen.
+  if (MI.isMetaInstruction())
+    return false;
----------------
MatzeB wrote:
> thegameg wrote:
> > MatzeB wrote:
> > > thegameg wrote:
> > > > Do we also want to ignore
> > > > ```
> > > > %csr = IMPLICIT_DEF
> > > > ```
> > > > ?
> > > > 
> > > > I remember @MatzeB suggested this approach to me to write cleaner tests.
> > > Should probably do `MO.isDef() || MO.readsReg()` below instead of this check.
> > That would do it for register operands, but seems that `DBG_VALUE` instructions also have `FrameIndex` operands. Is `MI.isDebugValue()` not enough here?
> the operands of DBG_VALUE have MachineOperand::isDebug() set, so readsReg() will always return false for them. The good thing about MachineOperand::readsReg() is that it also checks some other conditions that are often forgotten (such as undef operands or internals reads in bundles, which both are forgotten here as well it seems).
And you are right you probably need an extra MO.isDebugValue() check for the frame index case, didn't realize the code was checking MachineOperand::isFI() here. But in this case I'd move the isDebugValue() check closer to the isFI() check.


Repository:
  rL LLVM

https://reviews.llvm.org/D41187