This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] Allow fpmath differences in inlining when callee doesn't use FP math
- From: Uros Bizjak <ubizjak at gmail dot com>
- To: Richard Biener <rguenther at suse dot de>
- Cc: "gcc-patches at gcc dot gnu dot org" <gcc-patches at gcc dot gnu dot org>, Jan Hubicka <jh at suse dot de>
- Date: Thu, 24 Aug 2017 13:16:44 +0200
- Subject: Re: [PATCH] Allow fpmath differences in inlining when callee doesn't use FP math
- Authentication-results: sourceware.org; auth=none
- References: <alpine.LSU.2.20.1708241307560.14191@zhemvz.fhfr.qr>
On Thu, Aug 24, 2017 at 1:10 PM, Richard Biener <rguenther@suse.de> wrote:
>
> This adjusts the x86 backend to allow -mfpmath differences when
> deciding whether to allow inlining. -mfpmath doesn't really
> matter for functions not containing FP operations.
>
> It appears that the can_inline_p target hook is called from the
> C++ FE for multi-versioning, thus the ! ipa_fn_summaries check.
>
> Bootstrapped and tested on x86_64-unknown-linux-gnu, ok for trunk?
>
> Incidentically this fixes the bootstrap issue I have when enabling
> free-lang-data without LTO.
>
> It doesn't fully solve the fpmath difference issue we have with
> intrinsics but is an optimization.
>
> Thanks,
> Richard.
>
> 2017-08-24 Richard Biener <rguenther@suse.de>
>
> * config/i386/i386.c: Include symbol-summary.h, ipa-prop.h
> and ipa-fnsummary.h.
> (ix86_can_inline_p): When ix86_fpmath flags do not match
> check whether the callee uses FP math at all.
LGTM for branches, but for trunk, I'd still like to remove fpmath
processing from outside ix86_option_override_internal. As mentioned in
the comment in i_o_o_i, blindly switching fpmath for TARGET_SSE can
violate ABI for some environments.
Uros.
> Index: gcc/config/i386/i386.c
> ===================================================================
> --- gcc/config/i386/i386.c (revision 251307)
> +++ gcc/config/i386/i386.c (working copy)
> @@ -85,6 +85,9 @@ along with GCC; see the file COPYING3.
> #include "print-rtl.h"
> #include "intl.h"
> #include "ifcvt.h"
> +#include "symbol-summary.h"
> +#include "ipa-prop.h"
> +#include "ipa-fnsummary.h"
>
> /* This file should be included last. */
> #include "target-def.h"
> @@ -7544,7 +7547,14 @@ ix86_can_inline_p (tree caller, tree cal
> else if (caller_opts->tune != callee_opts->tune)
> ret = false;
>
> - else if (caller_opts->x_ix86_fpmath != callee_opts->x_ix86_fpmath)
> + else if (caller_opts->x_ix86_fpmath != callee_opts->x_ix86_fpmath
> + /* If the calle doesn't use FP expressions differences in
> + ix86_fpmath can be ignored. We are called from FEs
> + for multi-versioning call optimization, so beware of
> + ipa_fn_summaries not available. */
> + && (! ipa_fn_summaries
> + || ipa_fn_summaries->get
> + (cgraph_node::get (callee))->fp_expressions))
> ret = false;
>
> else if (caller_opts->branch_cost != callee_opts->branch_cost)