llvm dev - Jun 2015 - [LLVMdev] Debug info: test crashes at -O0

Hi Duncan,

For one of my tasks I am checking the LLVM tests with -O0 passed to llc.

There are some llc crashes and one of them happens on the 
CodeGen/ARM/debug-info-d16-reg.ll test. It has two DILocalVariable's 
with the same arg number:

!50 = !DILocalVariable(tag: DW_TAG_arg_variable, name: "val", line: 4,
arg: 2, scope: !9, file: !1, type: !7)
!51 = !DILocalVariable(tag: DW_TAG_arg_variable, name: "c", line: 4, 
arg: 2, scope: !9, file: !1, type: !8).

This breaks the assert condition "assert(V.Var == Var &&
"conflicting
variable")" in DbgVariable::addMMIEntry, as the two entries
("c" and
"val") with different DILocalVariable* pointers are added to the same 
DbgVariable. In release build it crashes later on a null pointer 
dereference.

Although this is an obvious mistake as the IR is handwritten, it might 
be a good idea to adjust the code so that it doesn't crash and shows a 
comprehensive error message instead.

Would you share you thoughts whether this makes sense, please?

Thanks.
Oleg

> On 2015-Jun-10, at 12:42, Oleg Ranevskyy <llvm.mail.list at
gmail.com> wrote:
> 
> Hi Duncan,
> 
> For one of my tasks I am checking the LLVM tests with -O0 passed to llc.
> 
> There are some llc crashes and one of them happens on the
CodeGen/ARM/debug-info-d16-reg.ll test. It has two DILocalVariable's with
the same arg number:
> 
> !50 = !DILocalVariable(tag: DW_TAG_arg_variable, name: "val",
line: 4, arg: 2, scope: !9, file: !1, type: !7)
> !51 = !DILocalVariable(tag: DW_TAG_arg_variable, name: "c", line:
4, arg: 2, scope: !9, file: !1, type: !8).
> 
> This breaks the assert condition "assert(V.Var == Var &&
"conflicting variable")" in DbgVariable::addMMIEntry, as the two
entries ("c" and "val") with different DILocalVariable*
pointers are added to the same DbgVariable. In release build it crashes later on
a null pointer dereference.
> 
> Although this is an obvious mistake as the IR is handwritten,
Looking at `git log`, I doubt this was handwritten.  Just old (r129934),
and likely suffering from bitrot.
> it might be a good idea to adjust the code so that it doesn't crash and
shows a comprehensive error message instead.
> 
> Would you share you thoughts whether this makes sense, please?
Yup, a misbehaving frontend could easily generate this IR.  I think the
`Verifier` is the right place to catch this.  Thanks for working on it!