1480594 - make the jit compilable under aarch64 windows

Status: RESOLVED FIXED

(Core :: JavaScript Engine: JIT, enhancement, P2)

Description

[Nathan Froyd [:froydnj]]

Trying to compile any of the JIT code under aarch64 windows results in a mess of compilation errors.  It appears that Microsoft has defined macros for every aarch64 instruction, or something like that?  You get errors about "not enough arguments for function-like macro invocation mvn" or "move32".

There's also some weird warnings about undeclared identifiers like opeqneoneg.  We can sort those out later.

Note: this is not about actually providing JIT support for aarch64 windows; this is just getting the code to compile.  I suspect we'll want to turn the JIT off initially, or something like that.

Comment 1

[Nathan Froyd [:froydnj]]

Attachment: undefine a few macros for aarch64 windows's benefit

Various assembler methods don't play nicely with Windows headers and
their constant macro definitions, so we have to #undef a few things
along the way.

I'm not sure what the Right Way is to patch vixl, so posting this as a sort of
starting point.

Comment 2

[Jan de Mooij [:jandem]]

You could also add this to js/src/util/Windows.h maybe?

https://searchfox.org/mozilla-central/source/js/src/util/Windows.h

Comment 3

[Nathan Froyd [:froydnj]]

(In reply to Jan de Mooij [:jandem] from comment #2)
> You could also add this to js/src/util/Windows.h maybe?
> 
> https://searchfox.org/mozilla-central/source/js/src/util/Windows.h

Empirical evidence says that this is insufficient.  The macros are defined in arm64_neon.h, and if I understand the include paths correctly, we have something like:

vm/Time.h -> intrin.h -> arm64_neon.h

and vm/Time.h gets pulled in through various sources (vm/JSScript.h, vm/Realm.h, vm/DateTime.h), which I assume somehow get pulled into the macro assembler headers in some fashion?

Comment 4

[Nathan Froyd [:froydnj]]

Hm, no, intrin.h only gets pulled in on x86-ish machines.  So maybe I really do have to struggle through /showIncludes to figure out what's going on.

Comment 5

[Nathan Froyd [:froydnj]]

OK, so the intrin.h include comes via mozilla/MathAlgorithms.h -> cmath -> intrin.h, which is probably quite difficult to fix "properly". =/  I guess we could include the undefs in MathAlgorithms.h, but that feels like a hack.

Comment 6

[Jan de Mooij [:jandem]]

Ah fair enough, thanks for checking. I thought this came in via <windows.h>

Comment 7

[Sean Stangl [:sstangl]]

Comment on attachment 9005004 [details] [diff] [review]
undefine a few macros for aarch64 windows's benefit

Review of attachment 9005004 [details] [diff] [review]:
-----------------------------------------------------------------

It is an odd move on Microsoft's part to effectively reserve keywords with such short names. This patch is perfectly OK for the moment if it helps get things compiling.

Comment 8

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/mozilla-inbound/rev/2ab5c5c57528
undefine a few macros for aarch64 windows's benefit; r=sstangl

Comment 9

[Arthur Iakab [arthur_iakab]]

https://hg.mozilla.org/mozilla-central/rev/2ab5c5c57528