[PATCH] D65471: [AMDGPU] Reserve all AGPRs on targets which do not have them

Fri Aug 21 12:17:21 PDT 2020

bjope added inline comments.

================
Comment at: llvm/trunk/test/CodeGen/AMDGPU/spill-vgpr-to-agpr.ll:2
 ; RUN: llc -march=amdgcn -mcpu=gfx908 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX908 %s
-; RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX900 %s
+; RUN: not llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s 2>&1 | FileCheck -check-prefixes=GCN,GFX900 %s

----------------
rampitec wrote:
> bjope wrote:
> > What is the idea behind this test. This single RUN-line takes more than 50 seconds to run, just to discover a failure?
> > 
> > A little bit annoying when `llvm-lit test/CodeGen/AMDGPU/` only takes 11s when removing this particular RUN-line.
> I just have tried it. This run took 4.5 seconds with the debug build. It is still fairly long, but nothing close to 50s.
Ehm, ok, sorry! Seem to be something downstream.

Using latest version from upstream master is fast (0.6s).
But on our development track it seem to spend an awful lot of time in "SI post-RA bundler". It must be some modification that we have done downstream, so I'll try to debug that myself.

Repository:
  rL LLVM

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D65471/new/

https://reviews.llvm.org/D65471