On 25 April 2017 at 18:39, Matthias Kaehlcke <mka@xxxxxxxxxxxx> wrote:
> Hi,
>
> El Tue, Apr 18, 2017 at 04:35:02PM +0100 Ard Biesheuvel ha dit:
>
>> On 18 April 2017 at 15:47, Paul Gortmaker <paul.gortmaker@xxxxxxxxxxxxx> wrote:
>> > On Wed, Apr 5, 2017 at 2:34 PM, Matthias Kaehlcke <mka@xxxxxxxxxxxx> wrote:
>> >> The operand is an integer constant, make the constness explicit by
>> >> adding the modifier. This is needed for clang to generate valid code
>> >> and also works with gcc.
>> >
>> > Actually it doesn't work with all gcc. I've got an older arm64 toolchain that I
>> > only use for syntax checking (and hence I don't care if it is the latest and
>> > greatest) and this commit breaks it:
>> >
>> > arch/arm64/crypto/sha1-ce-glue.c:21:2: error: invalid 'asm': invalid
>> > operand prefix '%c'
>> > asm(".globl " #sym "; .set " #sym ", %c0" :: "i"(val));
>> >
>> > I'm currently reverting this change locally so I can continue to use the old
>> > toolchain:
>> >
>> > $ aarch64-linux-gnu-gcc --version
>> > aarch64-linux-gnu-gcc (crosstool-NG linaro-1.13.1-4.8-2013.12 - Linaro
>> > GCC 2013.11) 4.8.3 20131202 (prerelease)
>> > Copyright (C) 2013 Free Software Foundation, Inc.
>> >
>> > $ aarch64-linux-gnu-as --version
>> > GNU assembler (crosstool-NG linaro-1.13.1-4.8-2013.12 - Linaro GCC
>> > 2013.11) 2.24.0.20131220
>> > Copyright 2013 Free Software Foundation, Inc.
>> >
>> > Maybe it is finally too old and nobody cares, but I thought it worth a mention.
>> >
>>
>> Thanks for the report. I think we care more about GCC 4.8 than about
>> Clang, which argues for reverting this patch.
>>
>> I understand these issues must be frustrating if you are working on
>> this stuff, but to me, it is not entirely obvious why we want to
>> support Clang in the first place (i.e., what does it buy you if your
>> distro/environment is not already using Clang for userland), and why
>> the burden is on Linux to make modifications to support Clang,
>> especially when it comes to GCC extensions such as inline assembly
>> syntax.
>>
>> It is ultimately up to the maintainers to decide what to do with this
>> patch, but my vote would be to revert it, especially given that the %c
>> placeholder prefix is not documented anywhere, and appears to simply
>> trigger some GCC internals that happen to do the right thing in this
>> case.
>>
>> However, the I -> i change is arguably an improvement, and considering
>> that the following
>>
>> asm("foo: .long %0" :: "i"(some value))
>>
>> doesn't compile with clang either, I suggest you (Matthias) file a bug
>> against Clang to get this fixed, and we can propose another patch just
>> for the I->i change.
>
> I consulted with folks with more expertise in this area than myself.
> This is their analysis of the situation:
>
> "The ARM ARM specifies that the correct AArch64 instruction assembly
> syntax is to have a hash sign (#) before an immediate.
>
It does not specify that at all:
"""
The A64 assembly language does not require the # character to
introduce constant immediate operands, but an assembler must allow
immediate values introduced with or without the # character. ARM
recommends that an A64 disassembler outputs a # before an immediate
operand.
"""
(ARM DDI 0487A.g page C1-121)
IOW, it only /recommends/ the # sign for *dis*assemblers. Big difference.
> Therefore, every time an inline assembly constraint is used that
> specifies to print an immediate (like 'i' or 'I'), the immediate
> (e.g. 42) should be printed with the hash (e.g. #42).
>
> Therefore, if you're using an immediate constraint where the hash sign
> must not be printed, you have to use the "c" operand modifier. The "c"
> operand modifier apparently got introduced to gcc after the 4.8
> release.
>
My problem with the %c modifier is that it is completely undocumented,
and appears in an internal GCC code generation code path. IOW, the GCC
developers could also remove it at any time (although this is highly
unlikely, of course)
> The binutils assembler and the clang integrated assembler accept
> immediates without the hash sign as a non-official extension.
Nope. *That* is mandated by the ARM ARM, see above.
> Some of
> the immediate constraints on gcc seem to not print out the hash sign
> either; which is why the variant in the linux kernel works with gcc.
>
Yes, and since it is perfectly legal for the "i" constraint not to
have a #, I don't understand what the big deal is tbh.
> In summary, it seems to me that the inline assembly with the %c0
> operand is the correct one and gcc 4.8 is simply too old to support
> this."
>
OK, so we're back to having to choose between GCC 4.8 and Clang.
> If the above is correct it seems that the solution is not to "fix"
> clang, but to use different instructions for gcc<=4.8 and newer
> compilers. I am aware that this is not a popular option.
>
> What do you think?
>
Perhaps I should just rework the code not to rely on inline asm at all.
I will take a look,
