Alignment of FP Context Storage

Alignment of FP Context Storage

[Kevin D. Kissell]

I'm cleaning up some old Linux kernel sandboxes, and
came across a patch which I had long ago made in a
local copy of include/asm-mips/processor.h but which
does not seem to have been propagated more widely.
I had added "__attribute__((aligned(8))))" to the
declarations of the mips_fpu_hard_struct and
mips_fpu_soft_struct data structures, presumably
because there was a need to ensure 64-bit alignment
of the elements so that LDC1 instructions would work.
We don't generally have a problem here, presumably
because either the previous data declarations naturally
align things to 64-bits, or because we've ensured things 
at a higher level of makfile compiler directives.  Are we 
in fact guarnateed to be safe without the source code 
directive, or should those __attribute__ directives be 
added as insurance?

            Kevin K.

Alignment of FP Context Storage

[Kevin D. Kissell]

I'm cleaning up some old Linux kernel sandboxes, and
came across a patch which I had long ago made in a
local copy of include/asm-mips/processor.h but which
does not seem to have been propagated more widely.
I had added "__attribute__((aligned(8))))" to the
declarations of the mips_fpu_hard_struct and
mips_fpu_soft_struct data structures, presumably
because there was a need to ensure 64-bit alignment
of the elements so that LDC1 instructions would work.
We don't generally have a problem here, presumably
because either the previous data declarations naturally
align things to 64-bits, or because we've ensured things 
at a higher level of makfile compiler directives.  Are we 
in fact guarnateed to be safe without the source code 
directive, or should those __attribute__ directives be 
added as insurance?

            Kevin K.

Re: Alignment of FP Context Storage

[Ralf Baechle]

On Mon, Nov 18, 2002 at 02:24:06PM +0100, Kevin D. Kissell wrote:

> I'm cleaning up some old Linux kernel sandboxes, and
> came across a patch which I had long ago made in a
> local copy of include/asm-mips/processor.h but which
> does not seem to have been propagated more widely.
> I had added "__attribute__((aligned(8))))" to the
> declarations of the mips_fpu_hard_struct and
> mips_fpu_soft_struct data structures, presumably
> because there was a need to ensure 64-bit alignment
> of the elements so that LDC1 instructions would work.
> We don't generally have a problem here, presumably
> because either the previous data declarations naturally
> align things to 64-bits, or because we've ensured things 
> at a higher level of makfile compiler directives.  Are we 
> in fact guarnateed to be safe without the source code 
> directive, or should those __attribute__ directives be 
> added as insurance?

The definition of mips_cpu_hard_struct uses doubles which is ensuring
mips_fpu_hard_struct will have 64-bit alignment.  This is actually part
of thread_struct which itself has 8kB alignment on 32-bit and 16kB
alignment on 64-bit.

  Ralf