Re: RFC: [MIPS] Add an option to disable ldc1/sdc1

[Richard Sandiford <rdsandiford at googlemail dot com>]

Chao-Ying Fu <Chao-Ying.Fu@imgtec.com> writes:
> Hello All,
>
>   Once in a while we got reports about programs (ex: WebKit, FireFox)
> crash due to ldc1/sdc1 unaligned accesses on MIPS targets.  The root
> cause is that programmers neglect the alignment issue and cast
> arbitrary pointers to point to double variables.
>
>   Although the correct solution is to fix application source code to
> fulfill alignment requirements, we want to add a GCC option to disable
> ldc1 and sdc1 (for the testing purpose or for workaround).  On 32-bit
> MIPS targets, GCC generates lwc1 and swc1 when -mno-ldc1-sdc1 is used,
> so that the memory address can be just 4-byte aligned to avoid
> ldc1/sdc1 address exceptions.

Sounds OK to me, given that the impact of the option is so low.

Bikeshed time, but I'd prefer the option to be named after the thing
that we're guaranteeing, rather than an instruction.  E.g. if the
problem is that doubles might only be 32-bit aligned, we could have
-mmisaligned-double (better suggestions welcome).

What about 64-bit targets?  We can sometimes access doubles using GPRs,
so on 64-bit targets we could end up using LD and SD to access a double
even when this option disables LDC1 and SDC1.  I think we'd need to
patch the move patterns as well.

If you only see the problem on 32-bit targets, then maybe it would be
better to have an option that prevents instructions that require greater
than word alignment.  Say (for sake of argument) -mno-superword-align.
Then it would be:

#define ISA_HAS_LDC1_SDC1 \
  (!ISA_MIPS1 && !TARGET_MIPS16 && (TARGET_64BIT || TARGET_SUPERWORD_ALIGN))

Thanks,
Richard