Re: Issues with native build of -10 on VAX

[Kalvis Duckmanton <kalvisd%gmail.com@localhost>]

Hi Hans,

On 7/11/23 00:58, Hans Rosenfeld wrote:
> Hi,
>
> has anyone already succeeded in doing a native build of -10 on the VAX?
>
> I'm stuck at the tools build, with gcc failing in a way that I don't
> really understand:
>
> /usr/src/tools/gcc/../../external/gpl3/gcc/dist/gcc/sancov.c: In instantiation of 'const pass_data {anonymous}::pass_sancov<false>::data':
> /usr/src/tools/gcc/../../external/gpl3/gcc/dist/gcc/sancov.c:307:55:   required from '{anonymous}::pass_sancov<O0>::pass_sancov(gcc::context*) [with bool O0 = false]'
> /usr/src/tools/gcc/../../external/gpl3/gcc/dist/gcc/sancov.c:345:38:   required from here
> /usr/src/tools/gcc/../../external/gpl3/gcc/dist/gcc/sancov.c:328:17: error: 'constinit' variable '{anonymous}::pass_sancov<false>::data' does not have a constant initializer
>    328 | const pass_data pass_sancov<O0>::data = {
>        |                 ^~~~~~~~~~~~~~~
> /usr/src/tools/gcc/../../external/gpl3/gcc/dist/gcc/sancov.c: In instantiation of 'const pass_data {anonymous}::pass_sancov<true>::data':
> /usr/src/tools/gcc/../../external/gpl3/gcc/dist/gcc/sancov.c:307:55:   required from '{anonymous}::pass_sancov<O0>::pass_sancov(gcc::context*) [with bool O0 = true]'
> /usr/src/tools/gcc/../../external/gpl3/gcc/dist/gcc/sancov.c:351:37:   required from here
> /usr/src/tools/gcc/../../external/gpl3/gcc/dist/gcc/sancov.c:328:17: error: 'constinit' variable '{anonymous}::pass_sancov<true>::data' does not have a constant initializer
>
> The code in question looks sane, and so does the preprocessor
> output. The same code built correctly when cross-built on amd64 (on
> illumos, though).
>
> I assume this has once worked on -current? Perhaps there something got
> inadvertently broken in the pullups to -10?

Hmm.  I'm not entirely sure but I *think* this is a manifestation of GCC 
bug c++/105491.  Native GCC 10 for VAX from -current also fails in the 
same way, but the GCC 10 cross-compiler from -current (on amd64 on 
Linux), can compile the same code without issue. It's weird.  I'd have 
expected the cross-compiler to also fail. For what it's worth, the 
(patched) native GCC 12.2 that I've built from pkgsrc can compile 
sancov.c from GCC 10 without complaint. (Bug 105491 is fixed in GCC 12.2 
and later.)

I have an inelegant workaround for GCC 10 but it needs some more testing.

kalvis