[InstCombine] Avoid folding GEPs across loop boundaries

Folding a GEP from outside to inside a loop will materialize an add where there wasn't an equivalent operation before. Check the containing loops before making this fold.

Reviewed By: lebedev.ri

Differential Revision: https://reviews.llvm.org/D107935

diff --git a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
index 1a80374..1026b9d 100644 (file)
--- a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
@@ -2132,8 +2132,17 @@ Instruction *InstCombinerImpl::visitGetElementPtrInst(GetElementPtrInst &GEP) {
         }
       }
 
+      // Guard the gep(gep) fold so we don't create an add inside a loop
+      // when there wasn't an equivalent instruction there before.
+      bool DifferentLoops = false;
+      if (LI)
+        if (auto *GEPLoop = LI->getLoopFor(GEP.getParent()))
+          if (auto *SrcOpI = dyn_cast<Instruction>(Src))
+            if (LI->getLoopFor(SrcOpI->getParent()) != GEPLoop)
+              DifferentLoops = true;
+
       // Fold (gep(gep(Ptr,Idx0),Idx1) -> gep(Ptr,add(Idx0,Idx1))
-      if (GO1->getType() == SO1->getType()) {
+      if (!DifferentLoops && GO1->getType() == SO1->getType()) {
         bool NewInBounds = GEP.isInBounds() && Src->isInBounds();
         auto *NewIdx =
             Builder.CreateAdd(GO1, SO1, GEP.getName() + ".idx",

diff --git a/llvm/test/Transforms/InstCombine/gep-combine-loop-invariant.ll b/llvm/test/Transforms/InstCombine/gep-combine-loop-invariant.ll
index 1373f86..dfa664f 100644 (file)
--- a/llvm/test/Transforms/InstCombine/gep-combine-loop-invariant.ll
+++ b/llvm/test/Transforms/InstCombine/gep-combine-loop-invariant.ll
@@ -216,6 +216,7 @@ define float @gep_cross_loop(i64* %_arg_, float* %_arg_3, float %_arg_8)
 ; CHECK-LABEL: @gep_cross_loop(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[TMP0:%.*]] = load i64, i64* [[_ARG_:%.*]], align 8
+; CHECK-NEXT:    [[ADD_PTR:%.*]] = getelementptr inbounds float, float* [[_ARG_3:%.*]], i64 [[TMP0]]
 ; CHECK-NEXT:    br label [[FOR_COND_I:%.*]]
 ; CHECK:       for.cond.i:
 ; CHECK-NEXT:    [[IDX:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[ADD11_I:%.*]], [[FOR_BODY_I:%.*]] ]
@@ -225,8 +226,7 @@ define float @gep_cross_loop(i64* %_arg_, float* %_arg_3, float %_arg_8)
 ; CHECK:       for.cond.i.i.i.preheader:
 ; CHECK-NEXT:    ret float [[SUM]]
 ; CHECK:       for.body.i:
-; CHECK-NEXT:    [[ARRAYIDX_I84_I_IDX:%.*]] = add nsw i64 [[IDX]], [[TMP0]]
-; CHECK-NEXT:    [[ARRAYIDX_I84_I:%.*]] = getelementptr inbounds float, float* [[_ARG_3:%.*]], i64 [[ARRAYIDX_I84_I_IDX]]
+; CHECK-NEXT:    [[ARRAYIDX_I84_I:%.*]] = getelementptr inbounds float, float* [[ADD_PTR]], i64 [[IDX]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = load float, float* [[ARRAYIDX_I84_I]], align 4
 ; CHECK-NEXT:    [[ADD_I]] = fadd fast float [[SUM]], [[TMP1]]
 ; CHECK-NEXT:    [[ADD11_I]] = add nuw nsw i64 [[IDX]], 1