On Sun, Dec 30, 2018 at 12:51 PM Jozef Lawrynowicz
<joze...@mittosystems.com> wrote:
>
> There have been some ICEs in the past for msp430-elf with the large
> memory model (20-bit pointers), caused by SET_{DECL,TYPE}_ALIGN being called
> with an argument which resolves to 20 i.e. POINTER_SIZE,
> or the default value of TARGET_VTABLE_ENTRY_ALIGN (which is POINTER_SIZE).
>
> The attached patch adds an assertion that SET_{DECL,TYPE}_ALIGN is called with
> a value which is a power of 2, or 0, for targets which support a partial int
> mode. This should catch issues in the future with non-power of 2
> alignments being set, which could propagate into serious problems later in the
> compilation process.
>
> If the filtering to only perform this check for targets supporting a partial
> int mode is unnecessary, I can remove that so CHECK_POW2_OR_ZEROP always
> expands to check_pow2_or_zerop.
>
> Successfully bootstrapped and regtested on x86_64-pc-linux-gnu and
> msp430-elf/-mlarge.
>
> Ok for trunk, or does this have to wait for GCC10 Stage 1?
I think the use of ffs_hwi suggests that the current behavior was intended
(or kept exactly because of calls with bougs values). If that's not wanted
why not simply use exact_log2 instead of ffs_hwi in the SET_ macros?
Btw, I'd simply make SET_{TYPE,DECL}_ALIGN inline functions and
then gcc_checking_assert that exact_log2 didn't return -1.
Richard.
