RFR: 8316234: AArch64: fastdebug build failure with GCC 12 and 13 due to -Wnonnull warning in vm_version_linux_aarch64.cpp

[Jie Fu]

On Mon, 27 Nov 2023 03:35:31 GMT, Hao Sun <haosun at openjdk.org> wrote:

> I tested {release, fastdebug, slowdebug} JDK builds with GCC 11/12/13 on AArch64, and fastdebug build with GCC 12 and 13 failed due to -Wnonnull warning.
> 
> The warning is raised because GCC 12 and 13 still think that `buf` can be null even if we have the assertion `assert(buf != nullptr, ...);`. Note that GCC 12 and 13 correctly determine the value range of `buflen`, i.e. `1 <= buflen <= 0xffff ffff ffff ffff`.
> 
> Note-1: Since it involves the analysis of `assert` statements, that's why release build can pass.
> 
> Note-2: I suppose that the incorrect optimization result of `nonnull/nullability` for variable `buf` only occurs at high GCC optimization levels. That's why slowdebug build can pass.
> 
> Note-3: To be honest, I didn't fully understand why this warning is raised by GCC 12 and 13. Recently we struggled with GCC 12 false positive warnings many times. See [1][2].
> 
> To suppress this warning, we try to fool GCC compiler via minor code change in this patch.
> 
> [1] https://bugs.openjdk.org/browse/JDK-8299580
> [2] https://bugs.openjdk.org/browse/JDK-8294031

src/hotspot/os_cpu/linux_aarch64/vm_version_linux_aarch64.cpp line 184:

> 182: 
> 183: static bool read_fully(const char *fname, char *buf, size_t buflen) {
> 184:   assert(buf != nullptr && buflen >= 1, "invalid arguments");

Is it possible to dump the values of `buf` and `buflen` when the assert is hit?
It would be more helpful for debugging.

-------------

PR Review Comment: https://git.openjdk.org/jdk/pull/16819#discussion_r1405664223