arc patch from glibc repo

Vineet Gupta Vineet.Gupta1 at synopsys.com
Thu Nov 9 11:09:05 PST 2017


On 11/09/2017 05:18 AM, Cupertino Miranda wrote:
> Hi Waldemar,
>
> I personally made this tst-cancel test start to pass in the past,
> related to unwinding issues we had in our tools.
> However, I have noticed that some of the tst-cancel tests would have
> some non-deterministic behavior.
>
> This was related to some pipes not blocking, when writing some data, but
> not blocking.
> If I increased the written elements to the pipe, those would block
> always, and the test would pass.
>
> It is possible that there was some change in the kernel that might have
> fixed this.
> Vineet: Do you know anything about this ?

Not in ARC port per se.

> Regarding your patch I would advise not to apply it until we find the
> real reason for blink not to be restored.

Right. Although in SNPS github uclibc repo we carry an equivalent patch (from 
Cuper) for fixing some of the cancelation tests. That is the only out of tree 
patch we have in uClibc just because we (I) didn't have time to analyze things 
fully. However I'm planning to clean this up and prove that we need that patch 
(and submit it for merge) or prove it is useless - which is quite unlikely - since 
Cuper is generally NOT smoking pot when coding ;-)

OTOH, while we do need this analysis, the change itself seemed relatively benign, 
but was kept on hold due to reasons needing a historic detour (read below if 
interested)

In 2010 ? I rewrote the uClibc syscall generators. To reduce generated code bloat 
and stack spills for prologue/epilogue etc ernro handling was moved into an 
out-of-line function. For fast path case of a simpler successful syscalls (no 
errno), no stack operations were needed.

    0000c244 <__libc_nanosleep>:
         c244:    mov       r8,162
         c248:    swi
         c24c:    brge       r0,0,c25c
         c250:    st.a        blink, [sp,-4]     # this blink save is from inline
    asm - not gcc
         c254:    bl           ad5c                 # this is errno setter
         c258:    ld.ab      blink, [sp,4]
         c25c:    j_s        [blink]
         c25e:    nop_s


The proposed change adds blink to clobber list, making BLINK saved twice now (once 
by gcc and other due to above) and also undoes all the hand crafted beauty above.

-Vineet



More information about the linux-snps-arc mailing list