[PATCH] D136169: [AMDGPU] Avoid SCC clobbering before S_CSELECT_B32

[Stanislav Mekhanoshin via Phabricator via llvm-commits]

rampitec added inline comments.


================
Comment at: llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp:2219
+          while (I != MBB->end()) {
+            if ((*I != MI && I->readsRegister(AMDGPU::SCC)) ||
+                (++Depth == getSCCLivenessScanMaxLength())) {
----------------
I think you need to skip debug instructions without counting so that debug build does not differ from release.


================
Comment at: llvm/lib/Target/AMDGPU/SIRegisterInfo.h:96
+  unsigned getSCCLivenessScanMaxLength() const {
+     return 100; 
+  }
----------------
100 seems to large number to me. Compiling with spills is already sufficiently and painfully slow. I would not go more than 10 here. In a worst case you are now burning just 2 instructions and probably 8 cycles.

Besides I do not think it deserves a special function, just a constant threshold right inside the code. Just because it is only used once.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D136169/new/

https://reviews.llvm.org/D136169