This is the mail archive of the
binutils@sourceware.org
mailing list for the binutils project.
RE: MIPS JAL/JALR to BAL transformation for Linux (o32 ABI)
- From: "Fu, Chao-Ying" <fu at mips dot com>
- To: "Richard Sandiford" <rdsandiford at googlemail dot com>
- Cc: "Adam Nemet" <anemet at caviumnetworks dot com>, <binutils at sourceware dot org>, "Lau, David" <davidlau at mips dot com>, "Garbacea, Ilie" <ilie at mips dot com>
- Date: Wed, 5 Aug 2009 14:38:16 -0700
- Subject: RE: MIPS JAL/JALR to BAL transformation for Linux (o32 ABI)
Richard Sandiford wrote:
> "Fu, Chao-Ying" <fu@mips.com> writes:
> > @@ -5590,11 +5601,12 @@ mips_elf_perform_relocation (struct bfd_
> > prediction hardware. If we are linking for the RM9000, and we
> > see jal, and bal fits, use it instead. Note that this
> > transformation should be safe for all architectures. */
>
> You need to update this comment too. With the new macros, it can
> be a lot simpler, such as:
>
> /* Try converting JAL and JALR to BAL, if the target is in
> range. */
Yes.
>
> > - if (bfd_get_mach (input_bfd) == bfd_mach_mips9000
> > - && !info->relocatable
> > + if (!info->relocatable
> > && !require_jalx
> > - && ((r_type == R_MIPS_26 && (x >> 26) == 0x3)
> /* jal addr */
> > - || (r_type == R_MIPS_JALR && x == 0x0320f809))) /*
> jalr t9 */
> > + && ((JAL_TO_BAL_P (input_bfd)
> > + && (r_type == R_MIPS_26 && (x >> 26) == 0x3))
> /* jal addr */
> > + || (JALR_TO_BAL_P (input_bfd) && (r_type == R_MIPS_JALR
> > + && x == 0x0320f809)))) /* jalr t9 */
>
> Odd formatting. I think it should be:
>
> if (!info->relocatable
> && !require_jalx
> && ((JAL_TO_BAL_P (input_bfd)
> && r_type == R_MIPS_26
> && (x >> 26) == 0x3) /* jal addr */
> || (JALR_TO_BAL_P (input_bfd)
> && r_type == R_MIPS_JALR
> && x == 0x0320f809))) /* jalr t9 */
Yes.
> More importantly, I think we should be checking the output_bfd
> rather than the input bfd. E.g. if you were linking legacy MIPS 3
> n32 objects with MIPS64 objects, you'd want the MIPS64-related
> optimisations to be applied to both.
>
> Looks good to me with those changes.
Yes, I checked in the patch with the first two changes.
To check output_bfd, we need to add a new parameter of output_bfd to
mips_elf_perform_relocation. But now because JALR_TO_BAL_P is true for all
and JAL_TO_BAL_P is true for RM9000, checking input_bfd should be the same
as checking output_bfd for non-RM9000 objects. We still can change to check
output_bfd later. Thanks a lot!
Regards,
Chao-ying