Re: [musl] busybox problem on powerpc PPC/32bit (hardware TP-Link-WDR-4900-v1)

[Rich Felker <dalias@libc.org>]

[-- Attachment #1: Type: text/plain, Size: 1277 bytes --]

On Sat, Feb 11, 2023 at 12:51:47PM +0000, Bastian Bittorf wrote:
> With OpenWRT i recognized a strange behavior of ash-shell scripts.
> It happens only on target PowerPC, e.g. m68k, arm, mips, x86 are
> unaffected.
> 
> The visible strange behavior in 'ash' is:
> 
> $ test A -gt 5 && echo OK
> ash: A: out of range
> OK
> 
> Ofcourse this wrong returncode leads to all sorts of things...
> 
> I tested several busybox releases (1.33.2, 1.35.0, 1.36.0) and
> crosscompiled with musl-git-b76f37f (from musl.cc) and 
> crosscompiled with glibc: powerpc-linux-gnu-gcc (Debian-12.2.0)
> (it's the same for all versions)
> 
> The resulting linux + busybox images ready for QEMU are here:
> http://intercity-vpn.de/mpc85xx/
> 
> If needed i can provide build instructions.
> Maybe somebody with more powerpc assembly knowledge can help here.
> 
> The underlying code is here:  
> https://git.busybox.net/busybox/tree/coreutils/test.c#n488
> But i can not spot the error:
> 
> static number_t getn(const char *s)
> {
>   char *p;
>   errno = 0;
>   r = strtol(s, &p, 10);
>   if (errno != 0)
>     syntax(s, "out of range");
>   return r;
> }
> 
> Best Greetings,
> Bastian Bittorf

It looks like the powerpc spe longjmp code is clobbering the value
argument. Try the attached patch.

[-- Attachment #2: ppc-spe-longjmp-fix.diff --]
[-- Type: text/plain, Size: 446 bytes --]

diff --git a/src/setjmp/powerpc/longjmp.S b/src/setjmp/powerpc/longjmp.S
index 611389fe..465e4cd7 100644
--- a/src/setjmp/powerpc/longjmp.S
+++ b/src/setjmp/powerpc/longjmp.S
@@ -42,10 +42,10 @@ longjmp:
 	bl 1f
 	.hidden __hwcap
 	.long __hwcap-.
-1:	mflr 4
-	lwz 5, 0(4)
-	lwzx 4, 4, 5
-	andis. 4, 4, 0x80
+1:	mflr 6
+	lwz 5, 0(6)
+	lwzx 6, 6, 5
+	andis. 6, 6, 0x80
 	beq 1f
 	.long 0x11c35b01 /* evldd 14,88(3) */
 	.long 0x11e36301 /* ... */