Modify ARM .S code so it can be processed by both gnu-as and llvm-mc.

Modify ARM .S code so it can be processed by both gnu-as and llvm-mc.
NOTE: there is now only one assert keeping us from switchingover to llvm-mc

BUG= http://code.google.com/p/nativeclient/issues/detail?id=1968
Committed: http://src.chromium.org/viewvc/native_client?view=rev&revision=7387

[robertm, 2011-12-07 17:49:41]

PTAL

[jvoung - send to chromium..., 2011-12-07 18:39:02]

http://codereview.chromium.org/8826003/diff/7001/src/trusted/service_runtime/...
File src/trusted/service_runtime/arch/arm/nacl_text_pad_test.S (left):

http://codereview.chromium.org/8826003/diff/7001/src/trusted/service_runtime/...
src/trusted/service_runtime/arch/arm/nacl_text_pad_test.S:17: .cpu cortex-a8
doesn't movw/movt require some cpu baseline?

Is this just something that wasn't added to the llvm arm asm parser yet?

http://codereview.chromium.org/8826003/diff/7001/src/trusted/service_runtime/...
File src/trusted/service_runtime/arch/arm/nacl_text_pad_test.S (right):

http://codereview.chromium.org/8826003/diff/7001/src/trusted/service_runtime/...
src/trusted/service_runtime/arch/arm/nacl_text_pad_test.S:11: /*
space to re-align the /*

http://codereview.chromium.org/8826003/diff/7001/src/trusted/service_runtime/...
src/trusted/service_runtime/arch/arm/nacl_text_pad_test.S:126: movw r1,
:lower16:(end_of_text)
Shouldn't there be a #? This could be a bug in the llvm asm parser?

Do you have an ARM reference that says the # is optional for these special
immediates? Almost every hit I see on google for arm lower16 syntax shows it
with the #. E.g., this one:

http://blogs.arm.com/software-enablement/251-how-to-load-constants-in-assembl...

http://codereview.chromium.org/8826003/diff/7001/src/trusted/service_runtime/...
src/trusted/service_runtime/arch/arm/nacl_text_pad_test.S:224: .fill
(TEXT_EXTEND - (. - start_of_text))/4, 4, halt
is this workaround mostly because the llvm parser doesn't handle the feature
yet, or is it truly illegal?

[robertm, 2011-12-07 19:24:33]

PTAL

http://codereview.chromium.org/8826003/diff/7001/src/trusted/service_runtime/...
File src/trusted/service_runtime/arch/arm/nacl_text_pad_test.S (left):

http://codereview.chromium.org/8826003/diff/7001/src/trusted/service_runtime/...
src/trusted/service_runtime/arch/arm/nacl_text_pad_test.S:17: .cpu cortex-a8
llvm-mc does not like this pseudo and gnu-as does not miss it. We pass in
-mfpu=vfp -march=armv7-a through the command-line anyway.

On 2011/12/07 18:39:02, jvoung wrote:
> doesn't movw/movt require some cpu baseline?
> 
> Is this just something that wasn't added to the llvm arm asm parser yet?

http://codereview.chromium.org/8826003/diff/7001/src/trusted/service_runtime/...
File src/trusted/service_runtime/arch/arm/nacl_text_pad_test.S (right):

http://codereview.chromium.org/8826003/diff/7001/src/trusted/service_runtime/...
src/trusted/service_runtime/arch/arm/nacl_text_pad_test.S:11: /*
On 2011/12/07 18:39:02, jvoung wrote:
> space to re-align the /*

Done.

http://codereview.chromium.org/8826003/diff/7001/src/trusted/service_runtime/...
src/trusted/service_runtime/arch/arm/nacl_text_pad_test.S:126: movw r1,
:lower16:(end_of_text)
gnu-as seems happy without it and llvm-mc dies not like it.
Also, note that :lower16: is not just merely a abbrev for
x & 0xffff but also forces a special relocation in case 
x is a symbol.
Bottom line: :lower16: is special enough that '#' might be implicit/optional


On 2011/12/07 18:39:02, jvoung wrote:
> Shouldn't there be a #? This could be a bug in the llvm asm parser?
> 
> Do you have an ARM reference that says the # is optional for these special
> immediates? Almost every hit I see on google for arm lower16 syntax shows it
> with the #. E.g., this one:
> 
>
http://blogs.arm.com/software-enablement/251-how-to-load-constants-in-assembl...
>

http://codereview.chromium.org/8826003/diff/7001/src/trusted/service_runtime/...
src/trusted/service_runtime/arch/arm/nacl_text_pad_test.S:224: .fill
(TEXT_EXTEND - (. - start_of_text))/4, 4, halt
This is just a NYI in the llvm asm parser/handler for arm.
I added similar hacks to the x86 siblings a while back.
There are few  comments in the llvm code that only a bare minimum has been
implemented, c.f. my updates to the bug tracker.

The assert mentioned in the description is likely also due to one of those NYI.
Apparently, some values given to :lower16: are not what llvm-mc expects. It does
not affect all values, though.



On 2011/12/07 18:39:02, jvoung wrote:
> is this workaround mostly because the llvm parser doesn't handle the feature
> yet, or is it truly illegal?

[jvoung - send to chromium..., 2011-12-07 20:32:11]

lgtm

http://codereview.chromium.org/8826003/diff/7001/src/trusted/service_runtime/...
File src/trusted/service_runtime/arch/arm/nacl_text_pad_test.S (right):

http://codereview.chromium.org/8826003/diff/7001/src/trusted/service_runtime/...
src/trusted/service_runtime/arch/arm/nacl_text_pad_test.S:126: movw r1,
:lower16:(end_of_text)
On 2011/12/07 19:24:33, robertm wrote:
> gnu-as seems happy without it and llvm-mc dies not like it.
> Also, note that :lower16: is not just merely a abbrev for
> x & 0xffff but also forces a special relocation in case 
> x is a symbol.
> Bottom line: :lower16: is special enough that '#' might be implicit/optional
> 
> 
> On 2011/12/07 18:39:02, jvoung wrote:
> > Shouldn't there be a #? This could be a bug in the llvm asm parser?
> > 
> > Do you have an ARM reference that says the # is optional for these special
> > immediates? Almost every hit I see on google for arm lower16 syntax shows it
> > with the #. E.g., this one:
> > 
> >
>
http://blogs.arm.com/software-enablement/251-how-to-load-constants-in-assembl...
> > 
> 

Sure, :lower16: is special and generates a relocation, but this may just be an
NYI on the llvm-mc front. If it is non-standard, perhaps file an LLVM bug -- it
may not take them too long to fix it...