This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: Try harder to preserve operand ties in maybe_legitimize_operands
- From: Jeff Law <law at redhat dot com>
- To: gcc-patches at gcc dot gnu dot org, richard dot sandiford at linaro dot org
- Date: Thu, 24 May 2018 16:48:04 -0600
- Subject: Re: Try harder to preserve operand ties in maybe_legitimize_operands
- References: <87in7elvfn.fsf@linaro.org>
On 05/23/2018 12:47 AM, Richard Sandiford wrote:
> maybe_legitimize_operands normally goes through each operand in turn
> and legitimises it in isolation. For example, if two operands to
> an instruction initially have constant value C, and the instruction
> requires both operands to be registers, the function ends up forcing
> C into a register twice and passing two different registers to the
> instruction.
>
> I think we should try a bit harder to preserve the rtx_equal_p
> property, if it's easy to do. Some targets can optimise that
> case better than they would the general case of all operands
> being different. This is particularly true for SVE after the
> upcoming changes to the IFN_COND_* routines.
>
> This is hard to test on its own, but is covered by the upcoming
> IFN_COND_* patches.
>
> Tested on aarch64-linux-gnu (with and without SLP, and with and without
> the follow-on patch that needs it), aarch64_be-elf and x86_64-linux-gnu.
> OK to install?
>
> Richard
>
>
> 2018-05-23 Richard Sandiford <richard.sandiford@linaro.org>
>
> gcc/
> * optabs.c (can_reuse_operands_p): New function.
> (maybe_legitimize_operands): Try to reuse the results for
> earlier operands.
OK.
jeff