On Thu, Aug 13, 2026 at 1:51 PM Jeffrey Law <[email protected]> wrote: > > > > On 8/13/2026 2:45 PM, Andrea Pinski wrote: > > On Thu, Aug 13, 2026 at 1:39 PM Jeffrey Law > > <[email protected]> wrote: > >> > >> > >> On 8/12/2026 10:27 PM, Sam Price wrote: > >>> insert_value_copy_on_edge declares 'unsignedp' uninitialized and sets it > >>> only as a side effect of an argument to gcc_assert. When src_mode and > >>> dest_mode differ it is then passed to convert_modes, which uses it to > >>> choose between zero and sign extension. > >>> > >>> gcc_assert does not always evaluate its argument. With GCC_VERSION > >>> < 4005 and without assert checking, gcc/system.h defines > >>> > >>> #define gcc_assert(EXPR) ((void)(0 && (EXPR))) > >>> > >>> so promote_ssa_mode is never called and unsignedp is passed to > >>> convert_modes indeterminate, selecting zero or sign extension at random. > >>> GCC_VERSION is __GNUC__ * 1000 + __GNUC_MINOR__, so any host compiler > >>> reporting less than 4.5 takes this branch. > >>> > >>> Relying on a side effect in gcc_assert is against the coding > >>> conventions regardless of which expansion is in use. Call > >>> promote_ssa_mode outside the assert, as the same file already does in > >>> get_temp_reg, and keep the assert for the invariant alone. unsignedp is > >>> initialized for the !REG_P path, where promote_ssa_mode is not called. > >>> > >>> Signed-off-by: Sam Price<[email protected]> > >>> Assisted-by: Claude (Anthropic) > >>> --- > >>> > >>> Found while getting gcc.c-torture running against a MicroBlaze port to > >>> LLVM, comparing the two compilers over the suite. > >> So I don't think is necessarily as simple as it might appear. In > >> particular what is the right value for the !REG case. I would suggest > >> one of two paths to make sure we're safe WRT the unsignedp initialization. > >> > >> First, make the call to promote_ssa_mode unconditionally. The worry > >> here is that it may not be ready to handle non-REG cases. > >> > >> Second, in the src_mode != dest_mode code, assert REG_P (dest). > >> > >> Given the change affects generic code it would need to be bootstrapped > >> and regression tested on one of the primary targets (x86, aarch64 linux > >> being the most common). > > So the problem will only show up with --disable-checking and a non GCC > > build. > > Since GCC before 5.4.0 can't build GCC at this point. So maybe the > > best way forward is either removing assert checking or just changing > > the non-GCC gcc_assert case into the assert checking case. > > I filed https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126838 and > > Richard B. sugguests we remove --disable-checking. I tend to agree > > with him. > I could live with that as well. But I don't think that resolves the > question of what value to use for unsignedp unless you pair the removal > with making the call to promote_ssa_mode unconditional.
if there is no call to promote_ssa_mode, then the 2 modes won't be different and unsignedp won't be used. So it does the solve issue there too. > > > > jeff
