home blog log atom rss about contact

gcc-14 bugs, pile 3

This week GCC 14 entered stage 3. Most major features are already pushed to master branch and the main focus now is to stabilize the result.

It’s a good time to look at bugs I noticed over past two months. This time I saw 19 of those. That almost twice as much than over previous 2 months.

summary

Bugs (or patches) in the order I observed them:

• tree-optimization/111435: ICE on gcc code (-m32) due to infinite recursion in type conversion rule
• driver/111527: gcc hits environment size limit early due to an internal COLLECT_GCC_OPTIONS variable
• rtl-optimization/111619: make profiledbootstrap is very slow to build in unoptimized builds
• other/111629: make profiledbootstrap SIGSEGVs gcc on shutdown due to a ggc bug
• bootstrap/111642: make profiledbootstrap fails to type check gcc’s own poly_int64 constructor
• c++/111647: -fchecking=0/2 disagree on validity of IFNDR c++ handling
• bootstrap/111653: -fchecking=0/2 generate different code on the same input
• libgcc trampoline build fix: libgcc build failure in libc-less mode
• rtl-optimization/112107: bootstrap failure on i686-linux: enabling debug changed register allocator output
• middle-end/112321: gcc SIGSEGV in debug mode as it generated invalid objects
• target/112332: ICE in gcc when it attempted to use SIMD instruction for stack access
• c/112347: ICE on jemalloc in newly added -Walloc-size analysis
• bootstrap/112379: bootstrap builds failure: unused function when asserts are disabled (code under #ifdef)
• libstdc++/112467: __assume__ in libstdc++ broke clang usage of that library
• target/112523: mpfr, libsodium and unbound tests were failing for an invalid shrd instruction use
• target/112540: gstreamer ICE in RTL (invalid addressing mode for SIMD)
• target/112567: linux ICE on RTL due to gcc generating invalid objects
• ipa/112601: ICE on llvm-17.0.5 code in -fchecking=2 mode related to function attribute inference
• target/112613: bad code in comparison code when AVX2 registers are present in generated code

fun discovery

I found a few new things as part of poking at those bugs:

IFNDR “Ill-Formed, No Diagnostic Required” is the known invalid code from type checking standpoint that is allowed to be compiled. In this case the whole program has an undefined behaviour.

-Walloc-size added in PR71219 detects interesting cases of T * p = malloc(sz) calls when it’s clear that sz is smaller than sizeof(T). It also works on calloc() and already found a few benign instances in elfutils, waypipe, sway and swaybg.

histograms

Looking at the bug categories:

• target: 5
• bootstrap: 3
• rtl-optimization: 2
• tree-optimization: 1
• driver: 1
• other: 1
• c++: 1
• middle-end: 1
• ipa: 1
• c: 1
• libstdc++: 1
• libgcc: 1

This cycle was very unusual: it has more bugs than I expected, it spans over 12 categories of compiler components, most of bugs are in i386 target.

parting words

This cycle felt very busy: about 2-3 bugs per week.

I had a lot of joy fixing PGO bootstrap and submitting a few trivial fixes upstream.

I encountered only two bugs related to wrong code generated by gcc. Most of the bugs were compiler crashes which are a lot easier to detect, report and fix. Chances are I did not get to running enough test suites as there were so many obvious bugs that required some attention.

Let’s see what stage 3 will bring us.

Have fun!