9deee6bffa
D141386 changed the semantics of !range metadata to return poison
on violation. If !range is combined with !noundef, violation is
immediate UB instead, matching the old semantics.
In theory, these IR semantics should also carry over into SDAG.
In practice, DAGCombine has at least one key transform that is
invalid in the presence of poison, namely the conversion of logical
and/or to bitwise and/or (c7b537bf09/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp (L11252)).
Ideally, we would fix this transform, but this will require
substantial work to avoid codegen regressions.
In the meantime, avoid transferring !range metadata without
!noundef, effectively restoring the old !range metadata semantics
on the SDAG layer.
Fixes https://github.com/llvm/llvm-project/issues/64589.
Differential Revision: https://reviews.llvm.org/D157685
37 lines
1.0 KiB
LLVM
37 lines
1.0 KiB
LLVM
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
|
|
; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu | FileCheck %s
|
|
|
|
define void @test(ptr %p) nounwind {
|
|
; CHECK-LABEL: test:
|
|
; CHECK: # %bb.0: # %bb
|
|
; CHECK-NEXT: movq (%rdi), %rax
|
|
; CHECK-NEXT: andl $-2, %eax
|
|
; CHECK-NEXT: cmpq $2, %rax
|
|
; CHECK-NEXT: cmpl $2, %eax
|
|
; CHECK-NEXT: retq
|
|
bb:
|
|
%i = load i64, ptr %p, align 8, !range !0, !noundef !{}
|
|
%i1 = and i64 %i, 6
|
|
%i2 = icmp eq i64 %i1, 2
|
|
br i1 %i2, label %bb3, label %bb5
|
|
|
|
bb3: ; preds = %bb
|
|
%i4 = icmp ne ptr undef, null
|
|
br label %bb5
|
|
|
|
bb5: ; preds = %bb3, %bb
|
|
br label %bb6
|
|
|
|
bb6: ; preds = %bb5
|
|
br i1 %i2, label %bb7, label %bb9
|
|
|
|
bb7: ; preds = %bb6
|
|
%i8 = getelementptr inbounds i64, ptr undef, i64 5
|
|
br label %bb9
|
|
|
|
bb9: ; preds = %bb7, %bb6
|
|
ret void
|
|
}
|
|
|
|
!0 = !{i64 0, i64 5}
|