@jhibbits, this last change addresses your last comment (moving VSX code to _vsx.S files) and also adds ifunc support.

[PPC64] Optimize bcopy/memcpy/memmove

Commandeering to address issues and use ifunc to decide whether VSX should be used.

• Use ifunc to choose best implementation based on running system

Taking over to add ifunc support, which will make it possible to have the optimized strcpy version and avoid breaking POWER5 and earlier.

In D21131#459452, @jhibbits wrote:

Looks good. Can the same change apply to powerpc?

• [PPC64] strncpy - fix rtld crash

• [PPC64] strncpy - fix rtld-libc build issue

In D20827#458843, @jhibbits wrote:

Thanks for checking. Looks good. I assume you tested on bare metal, too?

This last change fixes the previous "dst not always zeroed" issue and adds ifunc support to strncpy, so that the optimized version is selected on ISAs >= 2.05 while others can use the fallback implementation in C.

• [PPC64] strncpy - fix 'dst' not zeroed issue
• Initial ifunc capable strncpy implementation

• Add missing space

Redo merge:

In D21102#458426, @markj wrote:

In D21102#458391, @luporl wrote:

Now the merge result should match the behavior seen in upstream.
Was this what you had in mind? Or is the plan to revert the local changes and apply this patch on it instead?

I think we should indeed revert the local changes and apply the two patches from upstream, but IMO it's fine to do that in a single commit. I believe that is effectively what you have done here?

LGTM

Now the merge result should match the behavior seen in upstream.

• Match upstream behavior

I guess it is ok now.

• Fix merge issues

In D21102#458061, @markj wrote:

Were there any conflicts in the merge?

In D15369#456889, @jhibbits wrote:

Have you checked performance on this latest version?

• Addressed jhibbits' comments

• Fix cross build detection
• Refactored LIB32LD detection code
• Avoid using short form of -fuse-ld, that avoids issues with native builds and picking up a newer ld in PATH
• Using COMPILER_TYPE instead of LINKER_TYPE in stand/defs.mk, as an installed /usr/local/bin/ld pointing to ld.bfd can confuse LINKER_TYPE detection in this case.
• Aborting if LIB32LD is not defined when building stand. This will tipically happen when trying to build stand alone and without performing a 'make buildenv' first

I've tested @alfredo.junior_eldorado.org.br's patch and I've found a few issues, that I plan to fix in the next diff, as well as address @imp's comments.

I've posted a previous diff over an already patched file, that temporarily screwed the patch/history.

• [PPC64] Implement CAS
• Skip CAS if PSL_HV bit is set on MSR

• Fix inverted tab/newline in asm

Skip CAS if PSL_HV bit is set on MSR

LGTM.

I consider the strncpy implementation complete now.
Adding ifunc to select the optimized implementation, however, is still not possible, as rtld doesn't support it yet.

• Handled misalignment issues
• Improved implementation
• Now using memset() for faster zeroing of dst buffer

Looks good.
But I think MOEA64_STATS should become a kernel config option, to make it easier to enable the statistics and to have this option documented.

Nice catch!

Changed register convention.
Still need to investigate and fix the misaligned access issue.

• Use %r* instead of raw numbers for registers

• Addressed jhibbits' comments

• Address jhibbits comments

Looks good.

Looks good.

Nice, LGTM.

Looks good, thanks for fixing this.
I've tested this on QEMU, both with and without huge pages and it worked fine.

This is pretty old and it seems the issue that this change was supposed to fix was already fixed in a proper way quite some time ago.

The change itself LGTM.

LGTM. Tested on ppcdevref.

@git_bdragon.rtk0.net, I haven't tested your last revision, but Alfredo and I have been using your previous revision for more than a month, both in ELFv1 environments as in ELFv2, with no issues so far, and it is essential for the ELFv2 transition on PPC.