https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127990
Bug ID: 127990
Summary: -g changes the source location of a TARGET_EXPR
initializer in a statement expression
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Keywords: compare-debug-failure
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: yqtian668 at gmail dot com
Target Milestone: ---
Host: x86_64-pc-linux-gnu
Target: x86_64-pc-linux-gnu
Build: x86_64-pc-linux-gnu
Created attachment 65888
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=65888&action=edit
patch and test
The following valid C++ program fails `-fcompare-debug` at `-O1` and above:
```c++
struct D { int i; ~D (); };
D g ();
int
f ()
{
return ({ g (); }).i;
}
```
Command:
```text
g++ -O1 -g -fcompare-debug -S source.C
```
Tested with GCC 17.0.0 20261018 (experimental), revision
2d1f1bc18a59c0f106c551304ac459ad7781a711, configured for
x86_64-pc-linux-gnu.
The generated instructions are identical with `-g` and `-g0`, but the source
location attached to the call and the load of its return slot differs. In the
reduced input it is column 15 with `-g` and column 10 with `-g0`. This makes
`-fcompare-debug` fail and can make a later diagnostic refer to a different
column.
With `-g`, statement-frontier markers turn the `TARGET_EXPR` initializer into
a `STATEMENT_LIST`. `gimplify_target_expr` builds a location-less `INIT_EXPR`
with `build2`, so the two tree shapes cause it to inherit different
`input_location` values.
A candidate patch builds the `INIT_EXPR` with
`rexpr_location (init, EXPR_LOCATION (targ))`. This existing helper skips the
leading debug marker and selects the real inner-expression location, while
keeping the `TARGET_EXPR` location as a fallback. The regression test checks
both `-fcompare-debug` and the precise location of a multi-line statement
expression.
The focused test reports 18 expected passes. An x86_64-pc-linux-gnu C/C++
bootstrap completed successfully, and complete regression testing found no new
failures. `git diff --check` and `contrib/check_GNU_style.sh` are clean.
Note that this bug is found by our non-AI-based automated testing technique,
and then verfied by me with the assistnat of Chatgpt.
Any suggestion would be apprecaited. If it looks good to you, I am happy submit
the patch to gcc-patches
