RISCV can't build contribs

[Douglas Katzman]

The 'cat' command in make-contrib.lisp doesn't work and that's because run-program doesn't work and that seems to be due to broken alien calls for certain call signatures.

I built at this change from last July which must have worked a little bit better:
commit 9d1c91edc2462177d4070a4202480e12ae5ae919 (HEAD)
Author: Stas Boukarev <email address hidden>
Date: Tue Jul 5 16:32:41 2022 +0300
    Workaround a bootstrapping problem on ppc64, riscv.

and observed that many 'run-program' tests failed back then as well (per below). Now that the build depends on run-program, it's much worse off. We should figure out what the support level is because it's only going to deteriorate further if contribs can't be built.

::: UNEXPECTED-FAILURE (RUN-PROGRAM :CAT 1) due to SB-INT:BROKEN-PIPE:
        "Couldn't write to #<SB-SYS:FD-STREAM for "descriptor 6" {526F3393}>:
           Broken pipe"
::: Running (RUN-PROGRAM :CAT 2)
::: UNEXPECTED-FAILURE (RUN-PROGRAM :CAT 2) due to END-OF-FILE:
        "end of file on #<SB-SYS:FD-STREAM for "descriptor 7" {52886E63}>"
::: Running (RUN-PROGRAM :CAT 3)
Unhandled SB-INT:BROKEN-PIPE in thread #<SB-THREAD:THREAD RUNNING {528882D3}>:
  Couldn't write to #<SB-SYS:FD-STREAM for "descriptor 6" {52886BE3}>:
    Broken pipe

// Running run-program.test.sh in COMPILE evaluator mode
foo bar
Unhandled SIMPLE-ERROR in thread #<SB-THREAD:THREAD "main thread" RUNNING
                                    {514F4003}>:
  The assertion
  (STRING= STRING "foo bar
")
  failed with STRING = "".

[Stas Boukarev (stassats)]

I didn't actually build it on riscv back then (or ever).

[Charles (karlosz)]

Somehow I suspect it's not actually broken, because we probably would've gotten a report by the suse folk, and I didn't observe these failures when I build in around February last year. Probably whatever risc-v environment you have is borked?

[Douglas Katzman (dougk)]

are there different ABIs? I used to use whatever gcc farm machine was an amd64 with user mode qemu for riscv. Now they have SiFive physical hardware. So if anything it should be better/newer.
I added some printing into spawn() - all the strings come through as total garbage.

[Douglas Katzman (dougk)]

Must be an unreported regression. I went back even further, to

commit 147d56602ee275c3371d7f1a294c4d1594d8d39a (HEAD)
Author: Douglas Katzman <email address hidden>
Date: Thu Feb 24 00:44:51 2022 -0500
    riscv: workaround QEMU user mode emulator issue

which, with absolute certainty, was able to build, though not pass regressions (for me anyway) because as it remarked "the regression suite crashes a ton in user mode emulation."
Nonetheless when built on the SiFive hardware this *does* pass run-program.impure.lisp

[Douglas Katzman (dougk)]

git bisect says:

e50eda73cf65d593c01221fad6905b076df358be is the first bad commit
commit e50eda73cf65d593c01221fad6905b076df358be
Author: Charles Zhang <email address hidden>
Date: Wed Mar 9 03:20:12 2022 -0800

    riscv: Make SB-CONCURRENCY work again.

[Moody Liu (moodyhunter)]

Digging it further I found it's actually the 'pty_name' (the 8th) argument of `spawn` being mistakenly set to something else (it's set the address of spawn function itself on my side)
By the C calling convention this should be register A7.

Hope it helps :)

[Moody Liu (moodyhunter)]

It.... seems that `#define LINKAGE_TEMP_REG reg_NL3` is actually A7?

[Charles (karlosz)]

Thanks for the tip. It's fixed now.