| Issue |
120657
|
| Summary |
In LTO builds, PostOrderFunctionAttrsPass can infer that `__cxa_throw` is `nounwind`
|
| Labels |
|
| Assignees |
|
| Reporter |
ilovepi
|
When trying to enable LTO/FatLTO in the runtimes build, I hit a single test failure in RTSan, ThrowingAnExceptionDiesWhenRealtime(.
After some investigation, I've found that the test, which has a trivial try/catch w/ an explicit `throw` is miscompiled.
Before GlobalOpt the IR has a an invoke to `__cxa_throw` and some code to catch the exception in one of its successors
```llvm
; Function Attrs: cold mustprogress noinline uwtable
define dso_local void @func() #0 personality ptr @__gxx_personality_v0 {
%1 = tail call ptr @__cxa_allocate_exception(i64 8) #6
store ptr getelementptr inbounds nuw inrange(-16, 24) (i8, ptr @_ZTVSt9exception, i64 16), ptr %1, align 8, !tbaa !4478
invoke void @__cxa_throw(ptr nonnull %1, ptr nonnull @_ZTISt9exception, ptr nonnull @_ZNSt9exceptionD1Ev) #7
to label %11 unwind label %2
2: ; preds = %0
%3 = landingpad { ptr, i32 }
catch ptr @_ZTISt9exception
%4 = extractvalue { ptr, i32 } %3, 1
%5 = tail call i32 @llvm.eh.typeid.for.p0(ptr nonnull @_ZTISt9exception) #6
%6 = icmp eq i32 %4, %5
br i1 %6, label %7, label %10
7: ; preds = %2
%8 = extractvalue { ptr, i32 } %3, 0
%9 = tail call ptr @__cxa_begin_catch(ptr %8) #6
tail call void @__cxa_end_catch()
ret void
10: ; preds = %2
resume { ptr, i32 } %3
11: ; preds = %0
unreachable
}
After GlobalOpt runs the catch handling is removed:
```llvm
; Function Attrs: cold mustprogress noinline uwtable
define dso_local void @func() local_unnamed_addr #0 personality ptr @__gxx_personality_v0 {
%1 = tail call ptr @__cxa_allocate_exception(i64 8) #4
store ptr getelementptr inbounds nuw inrange(-16, 24) (i8, ptr @_ZTVSt9exception, i64 16), ptr %1, align 8, !tbaa !4478
call void @__cxa_throw(ptr nonnull %1, ptr nonnull @_ZTISt9exception, ptr nonnull @_ZNSt9exceptionD1Ev) #5
unreachable
}
```
This happens because w/in GlobalOpt the analysis checks if the ivoke instruction's callee is marked `nounwind`, and in this case it is. I've run the change of attributes for `__cxa_throw` back to the attribute inference in FunctionAttrs.cpp. Specifically, it seems that the logic in `IstrBreaksNonThrowing` is incomplete, since its determining that `__cxa_throw` is `nounwind`
```llvm
; *** IR Dump After PGOIndirectCallPromotion on [module] ***
; Function Attrs: mustprogress noreturn uwtable
define dso_local void @__cxa_throw(ptr noundef initializes((-120, -80), (-32, -16)) %0, ptr noundef %1, ptr noundef %2) #24 !dbg !51606 {
#dbg_value(ptr %0, !51610, !DIExpression(), !51615)
#dbg_value(ptr %1, !51611, !DIExpression(), !51615)
#dbg_value(ptr %2, !51612, !DIExpression(), !51615)
%4 = tail call ptr @__cxa_get_globals(), !dbg !51616
#dbg_value(ptr %4, !51613, !DIExpression(), !51615)
%5 = getelementptr inbounds nuw i8, ptr %4, i64 8, !dbg !51617
%6 = load i32, ptr %5, align 8, !dbg !51618, !tbaa !51457
%7 = add i32 %6, 1, !dbg !51618
store i32 %7, ptr %5, align 8, !dbg !51618, !tbaa !51457
%8 = tail call ptr @__cxa_init_primary_exception(ptr noundef %0, ptr noundef %1, ptr noundef %2) #59, !dbg !51619
#dbg_value(ptr %8, !51614, !DIExpression(), !51615)
%9 = getelementptr inbounds nuw i8, ptr %8, i64 8, !dbg !51620
store i64 1, ptr %9, align 8, !dbg !51621, !tbaa !51495
%10 = getelementptr inbounds nuw i8, ptr %8, i64 96, !dbg !51622
%11 = tail call i32 @_Unwind_RaiseException(ptr noundef nonnull %10), !dbg !51623
tail call fastcc void @_ZN10__cxxabiv1L12failed_throwEPNS_15__cxa_exceptionE(ptr noundef nonnull %8) #60, !dbg !51624
unreachable, !dbg !51624
}
; *** IR Dump After PostOrderFunctionAttrsPass on (__cxa_throw) ***
; Function Attrs: cold mustprogress noreturn nounwind uwtable
define dso_local void @__cxa_throw(ptr noundef initializes((-120, -80), (-32, -16)) %0, ptr noundef %1, ptr noundef %2) #36 !dbg !8793 {
#dbg_value(ptr %0, !8798, !DIExpression(), !8803)
#dbg_value(ptr %1, !8799, !DIExpression(), !8803)
#dbg_value(ptr %2, !8800, !DIExpression(), !8803)
%4 = tail call ptr @__cxa_get_globals(), !dbg !8804
#dbg_value(ptr %4, !8801, !DIExpression(), !8803)
%5 = getelementptr inbounds nuw i8, ptr %4, i64 8, !dbg !8805
%6 = load i32, ptr %5, align 8, !dbg !8806, !tbaa !8807
%7 = add i32 %6, 1, !dbg !8806
store i32 %7, ptr %5, align 8, !dbg !8806, !tbaa !8807
%8 = tail call ptr @__cxa_init_primary_exception(ptr noundef %0, ptr noundef %1, ptr noundef %2) #60, !dbg !8814
#dbg_value(ptr %8, !8802, !DIExpression(), !8803)
%9 = getelementptr inbounds nuw i8, ptr %8, i64 8, !dbg !8815
store i64 1, ptr %9, align 8, !dbg !8816, !tbaa !8817
%10 = getelementptr inbounds nuw i8, ptr %8, i64 96, !dbg !8824
%11 = tail call i32 @_Unwind_RaiseException(ptr noundef nonnull %10), !dbg !8825
tail call fastcc void @_ZN10__cxxabiv1L12failed_throwEPNS_15__cxa_exceptionE(ptr noundef nonnull %8) #61, !dbg !8826
unreachable, !dbg !8826
}
```
I would have assumed that the call to` _Unwind_RaiseException` would prevent this from happening.
![]()
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
