Module Name: src
Committed By: rin
Date: Thu Aug 17 06:38:08 UTC 2023
Modified Files:
src/external/gpl3/gdb/dist/gdbsupport: common-defs.h
Log Message:
gdb: Clean up duplicate ATTRIBUTE_NONNULL. No binary changes.
To generate a diff of this commit:
cvs rdiff -u -r1.3 -r1.4 src/external/gpl3/gdb/dist/gdbsupport/common-defs.h
Please note that diffs are not public domain; they are subject to the
copyright notices on the relevant files.
Modified files:
Index: src/external/gpl3/gdb/dist/gdbsupport/common-defs.h
diff -u src/external/gpl3/gdb/dist/gdbsupport/common-defs.h:1.3 src/external/gpl3/gdb/dist/gdbsupport/common-defs.h:1.4
--- src/external/gpl3/gdb/dist/gdbsupport/common-defs.h:1.3 Mon Jul 31 17:00:52 2023
+++ src/external/gpl3/gdb/dist/gdbsupport/common-defs.h Thu Aug 17 06:38:08 2023
@@ -187,81 +187,6 @@
#undef ATTRIBUTE_NONNULL
#define ATTRIBUTE_NONNULL(m)
-/* This is defined by ansidecl.h, but we disable the attribute.
-
- Say a developer starts out with:
- ...
- extern void foo (void *ptr) __atttribute__((nonnull (1)));
- void foo (void *ptr) {}
- ...
- with the idea in mind to catch:
- ...
- foo (nullptr);
- ...
- at compile time with -Werror=nonnull, and then adds:
- ...
- void foo (void *ptr) {
- + gdb_assert (ptr != nullptr);
- }
- ...
- to catch:
- ...
- foo (variable_with_nullptr_value);
- ...
- at runtime as well.
-
- Said developer then verifies that the assert works (using -O0), and commits
- the code.
-
- Some other developer then checks out the code and accidentally writes some
- variant of:
- ...
- foo (variable_with_nullptr_value);
- ...
- and builds with -O2, and ... the assert doesn't trigger, because it's
- optimized away by gcc.
-
- There's no suppported recipe to prevent the assertion from being optimized
- away (other than: build with -O0, or remove the nonnull attribute). Note
- that -fno-delete-null-pointer-checks does not help. A patch was submitted
- to improve gcc documentation to point this out more clearly (
- https://gcc.gnu.org/pipermail/gcc-patches/2021-July/576218.html ). The
- patch also mentions a possible workaround that obfuscates the pointer
- using:
- ...
- void foo (void *ptr) {
- + asm ("" : "+r"(ptr));
- gdb_assert (ptr != nullptr);
- }
- ...
- but that still requires the developer to manually add this in all cases
- where that's necessary.
-
- A warning was added to detect the situation: -Wnonnull-compare, which does
- help in detecting those cases, but each new gcc release may indicate a new
- batch of locations that needs fixing, which means we've added a maintenance
- burden.
-
- We could try to deal with the problem more proactively by introducing a
- gdb_assert variant like:
- ...
- void gdb_assert_non_null (void *ptr) {
- asm ("" : "+r"(ptr));
- gdb_assert (ptr != nullptr);
- }
- void foo (void *ptr) {
- gdb_assert_nonnull (ptr);
- }
- ...
- and make it a coding style to use it everywhere, but again, maintenance
- burden.
-
- With all these things considered, for now we go with the solution with the
- least maintenance burden: disable the attribute, such that we reliably deal
- with it everywhere. */
-#undef ATTRIBUTE_NONNULL
-#define ATTRIBUTE_NONNULL(m)
-
#if GCC_VERSION >= 3004
#define ATTRIBUTE_UNUSED_RESULT __attribute__ ((__warn_unused_result__))
#else
