[PATCH 0/3] stdlib, x86: Add internal stdc_rotate_left and use it in x86 pointer (de-)mangling

[Uros Bizjak]

x86 pointer (de-)mangling macros are currently implemented using
inline rol/ror assembly. This prevents the compiler from selecting
more efficient instructions on newer microarchitectures. In particular,
x86-64-v3 targets with BMI2 support can use rorx, which avoids
flag dependecies.

This series introduces an internal stdc_rotate_left helper, modelled
after Adhemerval Zanella's earlier stdc_rotate_right implementation.
It allows the compiler to emit optimal rotate sequences (including
rorx when available) instead of relying on inline assembly.

Tests are imported from gnulib and adapted to glibc's libsupport.

With these changes, pointer demangling improves from:

    33d8:	48 8b 40 20          	mov    0x20(%rax),%rax
    33dc:	48 c1 c8 11          	ror    $0x11,%rax
    33e0:	64 48 33 04 25 30 00 	xor    %fs:0x30,%rax
    33e7:	00 00 

to:

    3678:	c4 e3 fb f0 40 20 11 	rorx   $0x11,0x20(%rax),%rax
    367f:	64 48 33 04 25 30 00 	xor    %fs:0x30,%rax
    3686:	00 00 

The series has been tested on {x86_64,i686}-pc-linux-gnu, both with
and without GCC rotate builtins used.

Cc: Adhemerval Zanella  <adhemerval.zanella@linaro.org>

Uros Bizjak (3):
  stdlib: Add internal stdc_rotate_left implementation
  x86_64: Replace inline asm rotates in pointer_guard with
    stdc_rotate_{left,right}
  i386: Replace inline asm rotates in pointer_guard with
    stdc_rotate_{left,right}

 SHARED-FILES                                  |   1 +
 include/stdbit.h                              |  62 ++-
 stdlib/Makefile                               |   1 +
 stdlib/tst-stdc_rotate_left.c                 | 457 ++++++++++++++++++
 sysdeps/unix/sysv/linux/i386/pointer_guard.h  |   7 +-
 .../unix/sysv/linux/x86_64/pointer_guard.h    |  14 +-
 6 files changed, 535 insertions(+), 7 deletions(-)
 create mode 100644 stdlib/tst-stdc_rotate_left.c

-- 
2.54.0