[PATCH] D107935: [InstCombine] Avoid folding GEPs across loop boundaries

[Chang Lin via Phabricator via llvm-commits]

clin1 created this revision.
clin1 added a reviewer: RKSimon.
Herald added a subscriber: hiraditya.
clin1 requested review of this revision.
Herald added a project: LLVM.
Herald added a subscriber: llvm-commits.

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.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D107935

Files:
  llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
  llvm/test/Transforms/InstCombine/gep-combine-loop-invariant.ll


Index: llvm/test/Transforms/InstCombine/gep-combine-loop-invariant.ll
===================================================================
--- llvm/test/Transforms/InstCombine/gep-combine-loop-invariant.ll
+++ llvm/test/Transforms/InstCombine/gep-combine-loop-invariant.ll
@@ -186,3 +186,29 @@
   call void @blackhole(<2 x i8*> %e6)
   br label %loop
 }
+
+; Avoid folding the GEP outside the loop to inside, and increasing loop
+; instruction count.
+define float @gep_cross_loop(i64* %_arg_, float* %_arg_3, float %_arg_8)
+{
+entry:
+  %0 = load i64, i64* %_arg_, align 8
+  %add.ptr = getelementptr inbounds float, float* %_arg_3, i64 %0
+  br label %for.cond.i
+
+for.cond.i:                                       ; preds = %for.body.i, %entry
+  %idx = phi i64 [ 0, %entry ], [ %add11.i, %for.body.i ]
+  %sum = phi float [ 0.000000e+00, %entry ], [ %add.i, %for.body.i ]
+  %cmp = icmp ule i64 %idx, 16
+  br i1 %cmp, label %for.body.i, label %for.cond.i.i.i.preheader
+
+for.cond.i.i.i.preheader:                         ; preds = %for.cond.i
+  ret float %sum
+
+for.body.i:                                       ; preds = %for.cond.i
+  %arrayidx.i84.i = getelementptr inbounds float, float * %add.ptr, i64 %idx
+  %1 = load float, float* %arrayidx.i84.i, align 4
+  %add.i = fadd fast float %sum, %1
+  %add11.i = add nsw i64 %idx, 1
+  br label %for.cond.i
+}
Index: llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
===================================================================
--- llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
+++ llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
@@ -2130,8 +2130,17 @@
         }
       }
 
+      // 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",


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D107935.365867.patch
Type: text/x-patch
Size: 2424 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20210811/91b3895b/attachment.bin>