Re: [PATCHv2] powerpc64le/power9: guard power9 strcmp against rtld usage [BZ# 25905]

[Adhemerval Zanella <adhemerval.zanella@linaro.org>]

On 04/05/2020 15:02, Paul E Murphy wrote:
> 
> 
> On 5/4/20 11:36 AM, Adhemerval Zanella wrote:
> 
>> As a side note, is strcmp the only string routine that used in lazy
>> resolution?
> 
> It was the only one I encountered while debugging this issue.  I won't claim it is the only one, I have limited experience with that corner of glibc.  Though, I am surprised only one test failed due to the vector register clobbering.

I am wondering which would be best way to avoid such issue to happen
with different string routines since some specific loader code need 
to adhere to more strict ABI requirements.

The powerpc Makefile already adds the expected compiler tools to
avoid this very issue:

---
sysdeps/powerpc/powerpc64/Makefile

 13 # These flags prevent FPU or Altivec registers from being used,
 14 # for code called in contexts that is not allowed to touch those registers.
 15 # Stupid GCC requires us to pass all these ridiculous switches.  We need to
 16 # pass the -mno-* switches as well to prevent the compiler from attempting
 17 # to emit altivec or vsx instructions, especially when the registers aren't
 18 # available.
 19 no-special-regs := $(sort $(foreach n,40 41 50 51 60 61 62 63 \
 20                                       $(foreach m,2 3 4 5 6 7 8 9, \
 21                                                   3$m 4$m 5$m),\
 22                                     -ffixed-$n)) \
 23                    $(sort $(foreach n,$(foreach m,0 1 2 3 4 5 6 7 8 9,\
 24                                                 $m 1$m 2$m) 30 31,\
 25                                     -ffixed-v$n)) \
 26                    -ffixed-vrsave -ffixed-vscr -mno-altivec -mno-vsx
[...]
 36 CFLAGS-dl-runtime.os = $(no-special-regs)
 37 CFLAGS-dl-lookup.os = $(no-special-regs)
 38 CFLAGS-dl-misc.os = $(no-special-regs)
 39 CFLAGS-rtld-mempcpy.os = $(no-special-regs)
 40 CFLAGS-rtld-memmove.os = $(no-special-regs)
 41 CFLAGS-rtld-memchr.os = $(no-special-regs)
 42 CFLAGS-rtld-strnlen.os = $(no-special-regs)
---

However it does not help when an arch-specific implementation might ended being
pulled in rtld build and thus making this ineffective (And it seems that 
the strcmp is missing as well for the powerpc case).

And using __builtin_strcmp not -ffreestanding does really help on dl-runtime,
since strcmp call might still being generated by compiler.  

Maybe one option could to enforce all rtld code to use generic C implementation
instead, not allowing sysdeps folder to override it.  It should scale better
than actually testing all possible permutation and require less hacky solutions
like overriding the loader object for an specific build option as this patch.