On Sun, Feb 25, 2024 at 09:13:47AM +1300, Thomas Munro wrote:
> On Sun, Feb 25, 2024 at 9:12 AM Thomas Munro <thomas.munro@gmail.com> wrote:
> > On Sun, Feb 25, 2024 at 8:50 AM Noah Misch <noah@leadboat.com> wrote:
> > > On GNU/Linux x64, gcc correctly records alignment=2**12 for the associated
> > > section (.rodata for bulk_write.o zero_buffer, .bss for pg_prewarm.o
> > > blockbuffer). If I'm reading this right, neither AIX gcc nor xlc is marking
> > > the section with sufficient alignment, in bulk_write.o or pg_prewarm.o:
> >
> > Ah, that is a bit of a hazard that we should probably document.
> >
> > I guess the ideas to fix this would be: use smgrzeroextend() instead
> > of this coding, and/or perhaps look at the coding of pg_pwrite_zeros()
> > (function-local static) for any other place that needs such a thing,
> > if it would be satisfied by function-local scope?
True. Alternatively, could arrange for "#define PG_O_DIRECT 0" on AIX, which
disables the alignment assertions (and debug_io_direct).
> Erm, wait, how does that function-local static object work differently?
I don't know specifically, but I expect they're different parts of the gcc
implementation. Aligning an xcoff section may entail some xcoff-specific gcc
component. Aligning a function-local object just changes the early
instructions of the function; it's independent of the object format.