Please advise on dealing with compiler issue

Please advise on dealing with compiler issue

[Khalid Aziz]

I am adding code to the sparc64 kernel that would require 
reading/writing one of the new registers on M7 processor, namely MCDPER. 
There are three ways I know of to do this:

1. __asm__ __volatile__("rd %%mcdper, %0\n\t": "=r" (tmp)) - This is the 
best way to do it but the support for %%mcdper is available in binutils 
2.26 and newer only. Compiling the kernel with older binutils will cause 
build failure.

2. __asm__ __volatile__("rd %%asr14, %0\n\t": "=r" (tmp)) - This should 
work independent of binutils version but it does not due to two bugs in 
assembler that have been fixed recently but not in a released version 
yet - <https://sourceware.org/ml/binutils/2016-03/msg00302.html>, 
<https://sourceware.org/ml/binutils/2016-03/msg00303.html>.

3. .byte 0x83,0x43,0x80,0x00 - This is byte coded equivalent of "rd 
%asr14, %g1". This works independent of binutils version but is ugly and 
requires pre-assigning a register to read MCDPER into, which does not 
help compiler with optimization.

Of the three options, only option 3 can work without requiring everyone 
compiling kernel to update to latest binutils.

Please advise on best way to proceed.

Thanks,
Khalid

Re: Please advise on dealing with compiler issue

[David Miller]

From: Khalid Aziz <khalid.aziz@oracle.com>
Date: Wed, 20 Apr 2016 09:08:02 -0600

> 3. .byte 0x83,0x43,0x80,0x00 - This is byte coded equivalent of "rd
> %asr14, %g1". This works independent of binutils version but is ugly
> and requires pre-assigning a register to read MCDPER into, which does
> not help compiler with optimization.

The register read is so expensive relative to executing other
instructions, that I don't think compiler optimizations are a valid
consideration at all.

This is the only valid option given the constraints, and you should
use ".word xxxxx" just like we do for other similar hard coded
instruction sequences.

Thanks.