[LoopRotate] Drop loop dispositions when rotating loops. PR56260

[mkazantsev]

Summary

This is required because if there is a pure loop-invariant instruction, Loop Rotation
may decide to not clone it and just hoist it instead. If SCEV has previously cached
that it was loop-variant (not being smart enough to prove invariance), we may end
up with inconsistent cache state (which may later trigger false-negative assertion
failures checking that something was invariant).

This is a conservative fix that unconditionally drops the dispositions. We could
only drop it if the hoisting has actually happened, but it should take some time
understanding whether it's safe with all other things this function does.

[fhahn]

LGTM, thanks. As mentioned in D134173, it would be good to adjust the API to make this harder to miss.

llvm/lib/Transforms/Utils/LoopRotationUtils.cpp

do {
     if (!OrigPreheader || !L->hasDedicatedExits())
       return Rotated;
 
     // Anything ScalarEvolution may know about this loop or the PHI nodes
     // in its header will soon be invalidated. We should also invalidate
     // all outer loops because insertion and deletion of blocks that happens
     // during the rotation may violate invariants related to backedge taken
     // infos in them.
-    if (SE)
+    if (SE) {
       SE->forgetTopmostLoop(L);
+      // We may hoist some instructions out of loop. In case if they were cached
+      // as "loop variant" or "loop computable", these caches must be dropped.
+      SE->forgetLoopDispositions();
+    }
 
     LLVM_DEBUG(dbgs() << "LoopRotation: rotating "; L->dump());
     if (MSSAU && VerifyMemorySSA)
       MSSAU->getMemorySSA()->verifyMemorySSA();
 
     // Find new Loop header. NewHeader is a Header's one and only successor
     // that is inside loop.  Header's other successor is outside the
     // loop.  Otherwise loop is not suitable for rotation.

llvm/test/Transforms/LoopRotate/pr56260.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -passes='loop(loop-rotate,loop-deletion)' -S | FileCheck %s
+
+target datalayout = "e-m:o-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+
+define void @main() {
+; CHECK-LABEL: @main(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br label [[L0_PREHEADER:%.*]]
+; CHECK:       L0.L0.preheader.loopexit_crit_edge:
+; CHECK-NEXT:    br label [[L0_PREHEADER_LOOPEXIT:%.*]]
+; CHECK:       L0.preheader.loopexit:
+; CHECK-NEXT:    br label [[L0_PREHEADER]]
+; CHECK:       L0.preheader:
+; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i32 0, 0
+; CHECK-NEXT:    [[INC:%.*]] = zext i1 [[CMP]] to i32
+; CHECK-NEXT:    [[SPEC_SELECT1:%.*]] = add nsw i32 0, [[INC]]
+; CHECK-NEXT:    [[TOBOOL3_NOT2:%.*]] = icmp eq i32 [[SPEC_SELECT1]], 0
+; CHECK-NEXT:    br i1 [[TOBOOL3_NOT2]], label [[L0_PREHEADER_LOOPEXIT]], label [[L1_PREHEADER_LR_PH:%.*]]
+; CHECK:       L1.preheader.lr.ph:
+; CHECK-NEXT:    br label [[L1_PREHEADER:%.*]]
+; CHECK:       L1.preheader:
+; CHECK-NEXT:    [[SPEC_SELECT4:%.*]] = phi i32 [ [[SPEC_SELECT1]], [[L1_PREHEADER_LR_PH]] ], [ [[SPEC_SELECT:%.*]], [[L0_LATCH:%.*]] ]
+; CHECK-NEXT:    [[K_03:%.*]] = phi i32 [ 0, [[L1_PREHEADER_LR_PH]] ], [ [[SPEC_SELECT4]], [[L0_LATCH]] ]
+; CHECK-NEXT:    [[TOBOOL8_NOT:%.*]] = icmp eq i32 [[K_03]], 0
+; CHECK-NEXT:    br label [[L0_LATCH]]
+; CHECK:       L0.latch:
+; CHECK-NEXT:    [[SPEC_SELECT]] = add nsw i32 [[SPEC_SELECT4]], [[INC]]
+; CHECK-NEXT:    [[TOBOOL3_NOT:%.*]] = icmp eq i32 [[SPEC_SELECT]], 0
+; CHECK-NEXT:    br i1 [[TOBOOL3_NOT]], label [[L0_L0_PREHEADER_LOOPEXIT_CRIT_EDGE:%.*]], label [[L1_PREHEADER]]
+;
+entry:
+  br label %L0.preheader
+
+L0.preheader:
+  br label %L0
+
+L0:                                               ; preds = %L0.latch, %L0.preheader
+  %k.0 = phi i32 [ 0, %L0.preheader ], [ %spec.select, %L0.latch ]
+  %cmp = icmp slt i32 0, 0
+  %inc = zext i1 %cmp to i32
+  %spec.select = add nsw i32 %k.0, %inc
+  %tobool3.not = icmp eq i32 %spec.select, 0
+  br i1 %tobool3.not, label %L0.preheader, label %L1.preheader
+
+L1.preheader:
+  %tobool8.not = icmp eq i32 %k.0, 0
+  br label %L1
+
+L1:
+  br i1 %tobool8.not, label %L1.latch, label %L0.latch
+
+L1.latch:
+  br i1 false, label %L1, label %L0.latch
+
+L0.latch:
+  br label %L0
+}