On Thu, 29 Jul 2021 14:57:28 +0000
Charlene Wendling <[email protected]> wrote:
> Hi,
>
> George emitted the hypothesis that libffi may be partly broken on
> macppc. He was right :)
>
> After searching around, i found out that upstream has already a fix for
> that [0]. Tests results have improved a lot:
>
> Before the below diff:
>
> # of expected passes 440
> # of unexpected failures 104
>
> After the below diff:
>
> # of expected passes 530
> # of unexpected failures 14
Thank you for finding this diff!
ok gkoehler@
The patched code is inside #ifndef POWERPC64, so there is no change
on powerpc64. There are 10 "unexpected failures" on powerpc64. Now
I hope that our libffi consumers might get lucky and dodge the few
remaining failures on powerpc*.
macppc and powerpc64:
2 failures in libffi.call/float2.c
-- They fail because our <float.h>'s LBDL_EPSILON is a double, not
a long double, so printf("%Lf", LDBL_EPSILON) warns.
macppc only:
10 failures in libffi.bhaible/test-call.c
2 failures in libffi.call/many.c
-- These pass more than 8 floating-point arguments.
Args after registers f1 .. f8 must go on the stack.
powerpc64 only:
2 failures in libffi.call/va_1.c
2 failures in libffi.call/va_struct1.c
2 failures in libffi.call/va_struct2.c
2 failures in libffi.call/va_struct3.c
-- These do va_arg(3) with a struct.
--George
> I have been able to build gforth on macppc and make it run with that
> patched libffi. Finally. I'll send a diff for gforth after that one,
> some extra bits are required.
>
> I think no further consumer testing is really needed, given the
> spectacular change in tests results, we should assume the libffi we're
> shipping on macppc currently is in a pretty bad shape anyway.
>
> Comments/feedback are welcome,
>
> Charlène.
>
>
> [0] https://github.com/libffi/libffi/commit/4d6d2866
>
>
> Index: Makefile
> ===================================================================
> RCS file: /cvs/ports/devel/libffi/Makefile,v
> retrieving revision 1.43
> diff -u -p -u -p -r1.43 Makefile
> --- Makefile 11 Jun 2021 15:37:14 -0000 1.43
> +++ Makefile 29 Jul 2021 13:54:53 -0000
> @@ -3,7 +3,7 @@
> COMMENT= Foreign Function Interface
>
> DISTNAME= libffi-3.3
> -REVISION= 0
> +REVISION= 1
> SHARED_LIBS += ffi 1.2 # .6.4
> CATEGORIES= devel
>
> Index: patches/patch-src_powerpc_sysv_S
> ===================================================================
> RCS file: patches/patch-src_powerpc_sysv_S
> diff -N patches/patch-src_powerpc_sysv_S
> --- /dev/null 1 Jan 1970 00:00:00 -0000
> +++ patches/patch-src_powerpc_sysv_S 29 Jul 2021 13:54:53 -0000
> @@ -0,0 +1,48 @@
> +$OpenBSD$
> +
> +From 4d6d2866ae43e55325e8ee96561221804602cd7a Mon Sep 17 00:00:00 2001
> +From: Samuel Holland <[email protected]>
> +Date: Fri, 21 Feb 2020 21:06:15 -0600
> +Subject: [PATCH] Update powerpc sysv assembly for ffi_powerpc.h changes
> (#541)
> +
> +Some of the flag bits were moved when adding powerpc64 vector support.
> +
> +Fixes #536
> +
> +Index: src/powerpc/sysv.S
> +--- src/powerpc/sysv.S.orig
> ++++ src/powerpc/sysv.S
> +@@ -104,17 +104,16 @@ ENTRY(ffi_call_SYSV)
> + bctrl
> +
> + /* Now, deal with the return value. */
> +- mtcrf 0x01,%r31 /* cr7 */
> ++ mtcrf 0x03,%r31 /* cr6-cr7 */
> + bt- 31,L(small_struct_return_value)
> + bt- 30,L(done_return_value)
> + #ifndef __NO_FPRS__
> + bt- 29,L(fp_return_value)
> + #endif
> + stw %r3,0(%r30)
> +- bf+ 28,L(done_return_value)
> ++ bf+ 27,L(done_return_value)
> + stw %r4,4(%r30)
> +- mtcrf 0x02,%r31 /* cr6 */
> +- bf 27,L(done_return_value)
> ++ bf 26,L(done_return_value)
> + stw %r5,8(%r30)
> + stw %r6,12(%r30)
> + /* Fall through... */
> +@@ -145,10 +144,9 @@ L(done_return_value):
> + #ifndef __NO_FPRS__
> + L(fp_return_value):
> + .cfi_restore_state
> +- bf 28,L(float_return_value)
> ++ bf 27,L(float_return_value)
> + stfd %f1,0(%r30)
> +- mtcrf 0x02,%r31 /* cr6 */
> +- bf 27,L(done_return_value)
> ++ bf 26,L(done_return_value)
> + stfd %f2,8(%r30)
> + b L(done_return_value)
> + L(float_return_value):
--
George Koehler <[email protected]>
