PPC vector register break

[paulidale]

Reported by @evanmiller as a comment on the commit:

This commit generates broken PPC code, such as

vsel 0,0,0,0

It appears that vsr2vr1 is being passed a full register name e.g. "v7" but the logic is expecting only the integer portion e.g. "7".

Downstream bug report: https://trac.macports.org/ticket/63847

[evanmiller]

Just to clarify, the flavor in question is osx32. Reproduce with

perl5 crypto/aes/asm/aesp8-ppc.pl "osx32" | grep vsel

[t8m]

@martin-schwenke Could you please help?

[martin-schwenke]

Oh great! Hopefully the last of the PPC platforms broken with that handful of commits... 😃

I've pushed potential fix to https://github.com/martin-schwenke/openssl/tree/fix-osx.

I used the following script to test that it only changes output on OS X platforms:

#!/bin/sh

set -e

dir="$1"
mkdir -p "$dir"

flavours="aix32 aix64 linux32 linux64 linux64le linux64v2 osx32 osx64"
scripts=$(git grep -l ppc-xlate)

for f in $flavours ; do
	for s in $scripts ; do
		# Do not run scripts for 64-bit flavours against 32-bit flavours
		case "$f" in
		*32*)
			case "$s" in
			*64*)
				continue
			esac
		esac
		b=$(basename "$s")
		out="${dir}/${b}-${f}.out"
		perl "$s" "$f" "$out"
	done
done

Run with master branch with argument (say) before and then on this branch with argument (say) after. Then diff -qru before after shows only changes for OS X flavours. Dropping the q from the diff command appears to show useful changes for the OS X flavours.

@evanmiller Can you please verify that my branch fixes the build on osx32?

I'm no longer at IBM, so:

• Perhaps @amitay can please take a look from a linux64le perspective and confirm that I'm not doing anything stupid.
• If this turns into a PR then this will be under personal rather than corporate CLA - I'll try to remember to note that in the PR if we get to that point. 😉

[t8m]

The fix looks good to me - I did not verify it though.

[evanmiller]

I do not know PPC assembly, but the comment over the original function (line 153) makes it sound like the function will change the prefix from vs to v.

Building now on a PPC32 machine.

[martin-schwenke]

I do not know PPC assembly, but the comment over the original function (line 153) makes it sound like the function will change the prefix from vs to v.

The comment is a little confusing but, at the time, I couldn't figure out a way to keep it both useful and short. It is really referring to the 2 different ways that registers are numbered in the ISA. I blame the confusion on whoever decided the vector registers should be numbered in 2 different ways.

[evanmiller]

@martin-schwenke Thanks for the clarification.

The generated assembly now builds successfully with the patch. Will report back when the complete build and test suite are finished.

[evanmiller]

Builds and test suite passes locally 👍

[paulidale]

Our CIs do a linux64le cross compile build. We've no other PPC builds otherwise.