[PATCH v2] x86: Disallow GOT memory as vector memory operand

[Jan Beulich]

On 07.02.2025 07:32, Jan Beulich wrote:
> On 07.02.2025 06:22, H.J. Lu wrote:
>> On Thu, Feb 6, 2025 at 9:02 PM Jan Beulich <jbeulich@suse.com> wrote:
>>> On 06.02.2025 04:32, H.J. Lu wrote:
>>>> GOT memory is 4-byte for i386 and 8-byte for x86-64.  Disallow GOT memory
>>>> as vector memory operand for vector instructions which don't support 32-bit
>>>> nor 64-bit memory.
>>>>
>>>> PR gas/32624
>>>> * config/tc-i386.c (output_disp): Disallow GOT memory as vector
>>>> memory operand.
>>>> * testsuite/gas/i386/got.s: Add vcvtph2qq with GOT memory.
>>>> * testsuite/gas/i386/x86-64-gotpcrel-2.s: Add vcvtph2qq and
>>>> vcvtps2pd with GOT memory.
>>>> * testsuite/gas/i386/got-no-relax.d: Updated.
>>>> * testsuite/gas/i386/got.d: Likewise.
>>>> * testsuite/gas/i386/x86-64-gotpcrel-2.d: Likewise.
>>>> * testsuite/gas/i386/i386.exp: Run inval-got.
>>>> * testsuite/gas/i386/inval-got.l: New file.
>>>> * testsuite/gas/i386/inval-got.s: Likewise.
>>>> * testsuite/gas/i386/x86-64-inval-got.l: Likewise.
>>>> * testsuite/gas/i386/x86-64-inval-got.l: Likewise.
>>>> * testsuite/gas/i386/x86-64.exp: Run x86-64-inval-got.
>>>
>>> Hmm, I'm further puzzled: On what basis would VCVTPS2DQ and VCVTPH2QQ be
>>> okay to use with an operand in GOT?
>>>
>>> And then, if I'm reading the added logic right,
>>>
>>>         VADDPS foo@GOTPCREL(%rip), %zmm0, %zmm1
>>>
>>> would be permitted despite having a 512-bit memory operand, just because
>>> it also permits embedded broadcast (which isn't in use here). Whereas
>>
>> Only for broadcast.
> 
> Hmm, perhaps I'm reading the code change incorrectly then. By checking
> the template rather than i.types[] (as I had asked you to do) I would
> assume templates permitting for broadcast would allow all forms of their
> insns through.

Now that I've tried out your patch, behavior is exactly as predicted: Of

	vaddps foo@GOTPCREL(%rip), %zmm0, %zmm1
	vaddps foo@GOTPCREL(%rip){1to16}, %zmm0, %zmm1
	vaddps foo@GOTPCREL(%rip), %ymm0, %ymm1

only the last one triggers the new error.

Jan