The test case below shows that suppressing -Wuninitialized by #pragma GCC diagnostic doesn't work the same as it does for other warnings (-Wrray-bounds in this instance, but most other warnings behave like it). This makes it difficult for users to control -Wuninitialized (and -Wmaybe-uninitialized), compounding their frustration when they run into one of its false positives. The reason for the difference is that tree-ssa-uninit.c calls linemap_resolve_location (line_table, location, LRK_SPELLING_LOCATION, NULL) before using location, which other warnings don't do. The call was introduced in r186971 along with a test for its effect, gcc.dg/cpp/pragma-diagnostic-2.c, with the goal to improve the macro expansion output printed by GCC with -ftrack-macro-expansion as well as its interaction with on #pragma GCC diagnostic. It seems to me that the change was ill thought out: I can think of no reason why -Wunitialized should be treated differently from all other warnings. $ cat a.c && gcc -O2 -S -Wall -Werror a.c #define X(i) f (a[i]); int f (int); int g (void) { int a[3]; #pragma GCC diagnostic push #pragma GCC diagnostic warning "-Wuninitialized" return X (1); #pragma GCC diagnostic pop } int h (void) { int a[] = { 1, 2, 3 }; #pragma GCC diagnostic push #pragma GCC diagnostic warning "-Warray-bounds" return X (3); #pragma GCC diagnostic pop } a.c: In function ‘g’: a.c:1:14: error: ‘a’ is used uninitialized [-Werror=uninitialized] 1 | #define X(i) f (a[i]); | ^~~~~~~~ a.c:10:10: note: in expansion of macro ‘X’ 10 | return X (1); | ^ a.c:7:7: note: ‘a’ declared here 7 | int a[3]; | ^ a.c: In function ‘h’: a.c:1:14: warning: array subscript 3 is above array bounds of ‘int[3]’ [-Warray-bounds] 1 | #define X(i) f (a[i]); | ^~~~~~~~ a.c:19:10: note: in expansion of macro ‘X’ 19 | return X (3); | ^ a.c:16:7: note: while referencing ‘a’ 16 | int a[] = { 1, 2, 3 }; | ^ cc1: all warnings being treated as errors
See also pr90400 which is about using a _Pragma to suppress -Wmaybe-uninitialized in macros. Removing the linemap_resolve_location() calls from tree-ssa-uninit.c isn't enough to fix that bug but it's prerequisite for a consistent behavior for the suppression of all warnings.
This was fixed by r13-2994. Sorry for not tagging this PR, I came upon the issue via PR69543 comment 9 instead. Regarding PR90400, I think that is about the way the token streamer class works for gcc -E, and it needs to handle _Pragma distinctly from #pragma. I will look at it too sometime.