SIGILL in detect_arm_hwf_by_toolchain under FreeBSD arm64 on a Raspberry Pi 4B

[Jussi Kivilinna]

On 10.7.2024 19.49, Naram Qashat via Gcrypt-devel wrote:
> Hello,
> 
> On 2024-07-10 12:39, Jussi Kivilinna wrote:
>> Hello,
>>
>> On 9.7.2024 20.49, Naram Qashat via Gcrypt-devel wrote:
>>> I don't know if this started to happen after I updated to FreeBSD 14.1 on my RPi4B, but libgcrypt crashes with a SIGILL in detect_arm_hwf_by_toolchain on line 527 when trying to run the assembly code there. As a result, I cannot use gnupg either. I asked about this in one of the FreeBSD IRC channels and this was the conversation about it:
>>
>> Looking at documentation for __ARM_FEATURE_CRYPTO (https://github.com/ARM-software/acle/releases/download/r2024Q2/acle-2024Q2.pdf):
>>
>> "__ARM_FEATURE_CRYPTO is defined to 1 if the Armv8-A Crypto instructions are supported and intrinsics targeting them are available. These instructions include AES{E, D}, SHA1{C, P, M} and others. This also implies __ARM_FEATURE_AES and __ARM_FEATURE_SHA2."
>>
>> And as __ARM_FEATURE_AES is implied:
>>
>> "__ARM_FEATURE_AES is defined to 1 if there is hardware support for the Advanced SIMD AES Crypto instructions from Armv8-A ..."
>>
>> Based on this, libgcrypt has been compiled using CPU target that supports crypto-extensions, but actually HW does not support. Did you use "-mcpu=cortex-a72" for compiler flags to select CPU type? If compiler is enabling crypto-extensions for "-mcpu=cortex-a72" and crypto-extensions are actually optional for Cortex-A72 then bug must be in compiler... "-mcpu=cortex-a72" setting should not be selecting optional crypto-extensions by default. Which compiler are you using? GCC or clang or something else? Which version?
> 
> I am using -mcpu=cortex-a72 in my compiler flags. From what I've read recently, the Raspberry Pi 4B, despite running a Cortex-A72, the SoC being used done not have the crypto extensions at all. I don't think it is a compiler bug, but rather just the hardware excluding the feature.

I think this LLVM issue and comments there give some background on CPU feature selection for clang on ARM:
  https://github.com/llvm/llvm-project/issues/90365
  https://github.com/llvm/llvm-project/issues/90365#issuecomment-2117644294

> 
> I am using clang version 18.1.5 from FreeBSD's base system.
> 
> libcrypt was specifically built through FreeBSD's ports tree instead of the poudriere package builder, but I don't think that has anything to do with this as the file with the relevant function is still the same as upstream's 1.11.0.
> 
> Another thing I read is that code doing this should be attempting to catch the SIGILL signal when trying to check for the crypto extensions that don't exist. In a way, I agree with jhibbits that the code in libgcrypt needs fixing.
> 
No need for SIGILL checks, toolchain configuration (mcpu setting in this case) needs to be such that compiler does not indicate crypto-extension instruction support. Inline assembly code in "src/hwf-arm.c" is exactly for detecting these kind of configuration errors, having compiler indicate support for instruction sets that are not actually supported by target HW.

Anyway, you need to use "-mcpu=cortex-a72+nocrypto" with clang when targeting build for RPi 4B.

-Jussi