[PATCH] D59778: AMDGPU: Make exec mask optimzations more resistant to block splits

Mon Mar 25 10:29:28 PDT 2019

rampitec added inline comments.

================
Comment at: lib/Target/AMDGPU/SIOptimizeExecMaskingPreRA.cpp:144
+
+    // TODO: Is this really necessary?
+    if (!MBB->isLayoutSuccessor(Succ))
----------------
Yes, in case block is only preceded bit itself. We do not generate this way, but I can write such IR manually. Anyway layout successor means there will be no branch.

================
Comment at: test/CodeGen/AMDGPU/collapse-endcf.mir:146
   ; GCN:   [[V_CMP_NE_U32_e64_:%[0-9]+]]:sreg_64 = V_CMP_NE_U32_e64 2, [[COPY1]], implicit $exec
-  ; GCN:   [[COPY4:%[0-9]+]]:sreg_64 = COPY $exec, implicit-def $exec
-  ; GCN:   [[S_AND_B64_1:%[0-9]+]]:sreg_64 = S_AND_B64 [[COPY4]], [[V_CMP_NE_U32_e64_]], implicit-def dead $scc
+  ; GCN:   [[S_AND_B64_1:%[0-9]+]]:sreg_64 = S_AND_B64 $exec, [[V_CMP_NE_U32_e64_]], implicit-def dead $scc
   ; GCN:   $exec = S_MOV_B64_term [[S_AND_B64_1]]
----------------
I am not sure autogenerated test really tests anything, as there is no GCN-NEXT. The copy may easily remain and be untested.

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D59778/new/

https://reviews.llvm.org/D59778