[PATCH] D116928: [LoopVectorize] Support epilogue vectorisation of loops with reductions

[Florian Hahn via Phabricator via llvm-commits]

fhahn added inline comments.


================
Comment at: llvm/lib/Transforms/Vectorize/LoopVectorize.cpp:10642
+                    ReductionPhi->getRecurrenceDescriptor())) {
+              VPValue *StartVal = new VPValue(Resume);
+              ReductionPhi->setOperand(0, StartVal);
----------------
I think you might have to add this as external def to the VPlan (using `addExternalDef`), otherwise the allocated memory won't get freed.


================
Comment at: llvm/test/Transforms/LoopVectorize/X86/pr35432.ll:92
 ; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i8 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ [[CONV3]], [[FOR_BODY8_LR_PH]] ], [ [[CONV3]], [[VECTOR_SCEVCHECK]] ]
-; CHECK-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i32 [ [[DOTPROMOTED]], [[FOR_BODY8_LR_PH]] ], [ [[DOTPROMOTED]], [[VECTOR_SCEVCHECK]] ], [ [[TMP31]], [[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i32 [ [[DOTPROMOTED]], [[VECTOR_SCEVCHECK]] ], [ [[DOTPROMOTED]], [[FOR_BODY8_LR_PH]] ], [ [[TMP27]], [[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    br label [[FOR_BODY8:%.*]]
----------------
lots of spurrious changes here due to auto generating the check lines. Can you just reorder the incoming values here without auto-generating the check lines?


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D116928/new/

https://reviews.llvm.org/D116928