[PATCH] D120346: [AMDGPU] Extend SILoadStoreOptimizer to handle global stores

[Stanislav Mekhanoshin via Phabricator via llvm-commits]

rampitec added inline comments.


================
Comment at: llvm/test/CodeGen/AMDGPU/soft-clause-exceeds-register-budget.ll:7
+; CHECK:         s_load_dwordx16
+; CHECK-NEXT:    s_load_dwordx16
+
----------------
foad wrote:
> What happened here? Autogenerating the checks for this file might make it clearer what actually changed.
There is a bunch of global stores at the end of the test. We have combined 2 more of them, which has changed register pressure. The estimate for the last block changed from 40 to 41. Consequentially that has changed the scheduling and a pack of 4 global loads was scheduled lower, while couple of scalar x16 loads was scheduled higher.

I actually do not think that generating checks for ~400 instructions is beneficial. We will have to regenerate it often and it will obscure actual clauses.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D120346/new/

https://reviews.llvm.org/D120346