[PATCH] D72737: [AMDGPU] Bundle loads before post-RA scheduler

[Stanislav Mekhanoshin via Phabricator via llvm-commits]

rampitec marked an inline comment as done.
rampitec added inline comments.


================
Comment at: llvm/lib/Target/AMDGPU/SIMemoryLegalizer.cpp:1300
+            if (MO.isReg())
+              MO.setIsInternalRead(false);
+        }
----------------
foad wrote:
> rampitec wrote:
> > vpykhtin wrote:
> > > just out of curiosity: why is this needed?
> > We need to set it when we forming bundle to model the fact a single bundle instruction reads a register which is defined inside the same instruction. Therefore, when we unpack the bundle this flag needs to be reversed.
> You mean when the address of one load depends on the result of another load? Surely we shouldn't be bundling loads that are dependent like that?
Yes, we do not. But I really cannot rely on that when unpacking.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D72737/new/

https://reviews.llvm.org/D72737