[LoopSimplifyCFG] Fix SCEV invalidation after removing dead exit (#127536)

Fixes #127534

Author: aleks-tmb

llvm/lib/Transforms/Scalar/LoopSimplifyCFG.cpp

@@ -369,7 +369,7 @@ class ConstantTerminatorFoldingImpl {
         DeadInstructions.emplace_back(LandingPad);
 
       for (Instruction *I : DeadInstructions) {
-        SE.forgetBlockAndLoopDispositions(I);
+        SE.forgetValue(I);
         I->replaceAllUsesWith(PoisonValue::get(I->getType()));
         I->eraseFromParent();
       }

llvm/test/Transforms/LoopSimplifyCFG/pr117537.ll

@@ -0,0 +1,109 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -S -passes='print<scalar-evolution>,loop-mssa(licm,loop-simplifycfg,loop-predication)' -verify-scev < %s 2>/dev/null | FileCheck %s
+
+; Make sure we don't assert due to insufficient SCEV invalidation.
+
+define i64 @"main"(ptr addrspace(1) %p, i1 %check) {
+; CHECK-LABEL: define i64 @main(
+; CHECK-SAME: ptr addrspace(1) [[P:%.*]], i1 [[CHECK:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    switch i32 0, label %[[ENTRY_SPLIT:.*]] [
+; CHECK-NEXT:      i32 1, label %[[LOOP1_PREHEADER_SPLIT_LOOP_EXIT:.*]]
+; CHECK-NEXT:    ]
+; CHECK:       [[ENTRY_SPLIT]]:
+; CHECK-NEXT:    br label %[[LOOP0_PRE:.*]]
+; CHECK:       [[LOOP0_PRE]]:
+; CHECK-NEXT:    br i1 [[CHECK]], label %[[EXIT:.*]], label %[[LOOP0:.*]]
+; CHECK:       [[LOOP0]]:
+; CHECK-NEXT:    [[LENGTH:%.*]] = load atomic i32, ptr addrspace(1) [[P]] unordered, align 4
+; CHECK-NEXT:    [[TMP0:%.*]] = icmp ugt i32 [[LENGTH]], 1
+; CHECK-NEXT:    br i1 [[TMP0]], label %[[LOOP0_OUT:.*]], label %[[LOOP1_PREHEADER_SPLIT_LOOP_EXIT1:.*]]
+; CHECK:       [[LOOP0_OUT]]:
+; CHECK-NEXT:    br label %[[LOOP0_PRE]]
+; CHECK:       [[LOOP1_PREHEADER_SPLIT_LOOP_EXIT]]:
+; CHECK-NEXT:    [[T_LE:%.*]] = add i32 0, 1
+; CHECK-NEXT:    br label %[[LOOP1_PREHEADER:.*]]
+; CHECK:       [[LOOP1_PREHEADER_SPLIT_LOOP_EXIT1]]:
+; CHECK-NEXT:    [[LENGTH_LCSSA_PH2:%.*]] = phi i32 [ [[LENGTH]], %[[LOOP0]] ]
+; CHECK-NEXT:    [[LOCAL_PH3:%.*]] = phi i32 [ 0, %[[LOOP0]] ]
+; CHECK-NEXT:    br label %[[LOOP1_PREHEADER]]
+; CHECK:       [[LOOP1_PREHEADER]]:
+; CHECK-NEXT:    [[LENGTH_LCSSA:%.*]] = phi i32 [ poison, %[[LOOP1_PREHEADER_SPLIT_LOOP_EXIT]] ], [ [[LENGTH_LCSSA_PH2]], %[[LOOP1_PREHEADER_SPLIT_LOOP_EXIT1]] ]
+; CHECK-NEXT:    [[LOCAL:%.*]] = phi i32 [ [[T_LE]], %[[LOOP1_PREHEADER_SPLIT_LOOP_EXIT]] ], [ [[LOCAL_PH3]], %[[LOOP1_PREHEADER_SPLIT_LOOP_EXIT1]] ]
+; CHECK-NEXT:    [[TMP1:%.*]] = add i32 [[LENGTH_LCSSA]], -1
+; CHECK-NEXT:    [[TMP2:%.*]] = icmp ult i32 310, [[TMP1]]
+; CHECK-NEXT:    [[TMP3:%.*]] = icmp ult i32 4, [[LENGTH_LCSSA]]
+; CHECK-NEXT:    [[TMP4:%.*]] = and i1 [[TMP3]], [[TMP2]]
+; CHECK-NEXT:    [[TMP5:%.*]] = freeze i1 [[TMP4]]
+; CHECK-NEXT:    br label %[[LOOP1:.*]]
+; CHECK:       [[LOOP1]]:
+; CHECK-NEXT:    [[IV1:%.*]] = phi i32 [ 4, %[[LOOP1_PREHEADER]] ], [ [[IV1_NEXT:%.*]], %[[LOOP1_GUARDED:.*]] ]
+; CHECK-NEXT:    [[TMP6:%.*]] = icmp ult i32 [[IV1]], [[LENGTH_LCSSA]]
+; CHECK-NEXT:    [[WC:%.*]] = call i1 @llvm.experimental.widenable.condition()
+; CHECK-NEXT:    [[TMP7:%.*]] = and i1 [[TMP5]], [[WC]]
+; CHECK-NEXT:    br i1 [[TMP7]], label %[[LOOP1_GUARDED]], label %[[DEOPT_EXIT:.*]]
+; CHECK:       [[LOOP1_GUARDED]]:
+; CHECK-NEXT:    call void @llvm.assume(i1 [[TMP6]])
+; CHECK-NEXT:    [[IV1_NEXT]] = add nuw nsw i32 [[IV1]], 1
+; CHECK-NEXT:    [[CHK:%.*]] = icmp ugt i32 [[IV1]], 310
+; CHECK-NEXT:    br i1 [[CHK]], label %[[LOOP1_EXIT:.*]], label %[[LOOP1]]
+; CHECK:       [[DEOPT_EXIT]]:
+; CHECK-NEXT:    [[TMP8:%.*]] = call i64 (...) @llvm.experimental.deoptimize.i64(i32 13) [ "deopt"() ]
+; CHECK-NEXT:    ret i64 [[TMP8]]
+; CHECK:       [[LOOP1_EXIT]]:
+; CHECK-NEXT:    ret i64 0
+; CHECK:       [[EXIT]]:
+; CHECK-NEXT:    ret i64 0
+;
+entry:
+  br label %loop0.pre
+
+loop0.pre:
+  br i1 %check, label %exit, label %loop0
+
+loop0:
+  %length = load atomic i32, ptr addrspace(1) %p unordered, align 4
+  %28 = icmp ugt i32 %length, 1
+  br i1 %28, label %loop0.out, label %loop1.preheader
+
+loop0.out:
+  %t = add i32 0, 1
+  br i1 false, label %loop1.preheader, label %mid
+
+loop1.preheader:
+  %length.lcssa = phi i32 [ %length, %loop0.out ], [ %length, %loop0 ]
+  %local = phi i32 [ 0, %loop0 ], [ %t, %loop0.out ]
+  br label %loop1
+
+loop1:
+  %iv1 = phi i32 [ 4, %loop1.preheader ], [ %iv1.next, %loop1.guarded ]
+  %82 = icmp ult i32 %iv1, %length.lcssa
+  %wc = call i1 @llvm.experimental.widenable.condition()
+  %guard.chk = and i1 %82, %wc
+  br i1 %guard.chk, label %loop1.guarded, label %deopt-exit
+
+loop1.guarded:
+  %iv1.next = add nuw nsw i32 %iv1, 1
+  %chk = icmp ugt i32 %iv1, 310
+  br i1 %chk, label %loop1.exit, label %loop1
+
+deopt-exit:
+  %100 = call i64 (...) @llvm.experimental.deoptimize.i64(i32 13) [ "deopt"() ]
+  ret i64 %100
+
+loop1.exit:
+  ret i64 0
+
+mid:
+  br label %loop0.pre
+
+exit:
+  ret i64 0
+}
+
+declare i64 @foo()
+
+declare i64 @llvm.experimental.deoptimize.i64(...)
+
+; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(inaccessiblemem: readwrite)
+declare noundef i1 @llvm.experimental.widenable.condition()