diff mbox series

libgcc, i386, optabs, v2: Add __float{, un}tibf to libgcc and expand BF -> integral through SF intermediate [PR107703]

Message ID Y3TOsnjlLHvSarjl@tucnak
State New
Headers
Series libgcc, i386, optabs, v2: Add __float{, un}tibf to libgcc and expand BF -> integral through SF intermediate [PR107703] |

Commit Message

Jakub Jelinek Nov. 16, 2022, 11:51 a.m. UTC 
  On Wed, Nov 16, 2022 at 10:06:17AM +0100, Jakub Jelinek via Gcc-patches wrote:
> Thoughts on this?  I guess my preference would be the BF -> SF -> TI
> path because we won't need to waste
>     32: 0000000000015e10   321 FUNC    GLOBAL DEFAULT   13 __fixbfti@@GCC_13.0.0
>     89: 0000000000015f60   299 FUNC    GLOBAL DEFAULT   13 __fixunsbfti@@GCC_13.0.0
> If so, I'd need to cut the fix parts of the patch below and
> do something in the middle-end.

Here is adjusted patch that does that.

2022-11-16  Jakub Jelinek  <jakub@redhat.com>

	PR target/107703
	* optabs.cc (expand_fix): For conversions from BFmode to integral,
	use shifts to convert it to SFmode first and then convert SFmode
	to integral.

	* soft-fp/floattibf.c: New file.
	* soft-fp/floatuntibf.c: New file.
	* config/i386/libgcc-glibc.ver: Export __float{,un}tibf @ GCC_13.0.0.
	* config/i386/64/t-softfp (softfp_extras): Add floattibf and
	floatuntibf.
	(CFLAGS-floattibf.c, CFLAGS-floatunstibf.c): Add -msse2.



	Jakub

Comments

Jakub Jelinek March 1, 2023, 12:32 p.m. UTC | #1 
Hi!

On Wed, Nov 16, 2022 at 12:51:14PM +0100, Jakub Jelinek via Gcc-patches wrote:
> On Wed, Nov 16, 2022 at 10:06:17AM +0100, Jakub Jelinek via Gcc-patches wrote:
> > Thoughts on this?  I guess my preference would be the BF -> SF -> TI
> > path because we won't need to waste
> >     32: 0000000000015e10   321 FUNC    GLOBAL DEFAULT   13 __fixbfti@@GCC_13.0.0
> >     89: 0000000000015f60   299 FUNC    GLOBAL DEFAULT   13 __fixunsbfti@@GCC_13.0.0
> > If so, I'd need to cut the fix parts of the patch below and
> > do something in the middle-end.
> 
> Here is adjusted patch that does that.
> 
> 2022-11-16  Jakub Jelinek  <jakub@redhat.com>
> 
> 	PR target/107703
> 	* optabs.cc (expand_fix): For conversions from BFmode to integral,
> 	use shifts to convert it to SFmode first and then convert SFmode
> 	to integral.
> 
> 	* soft-fp/floattibf.c: New file.
> 	* soft-fp/floatuntibf.c: New file.
> 	* config/i386/libgcc-glibc.ver: Export __float{,un}tibf @ GCC_13.0.0.
> 	* config/i386/64/t-softfp (softfp_extras): Add floattibf and
> 	floatuntibf.
> 	(CFLAGS-floattibf.c, CFLAGS-floatunstibf.c): Add -msse2.

I'd like to ping the libgcc non-i386 part of this patch, Uros said the i386
part is ok but that one depends on the generic libgcc changes.
I'll ping the optabs.cc change separately.

https://gcc.gnu.org/pipermail/gcc-patches/2022-November/606398.html
with more info in
https://gcc.gnu.org/pipermail/gcc-patches/2022-November/606382.html

Thanks.

	Jakub
Jakub Jelinek March 10, 2023, 9 a.m. UTC | #2 
Hi!

On Wed, Mar 01, 2023 at 01:32:43PM +0100, Jakub Jelinek via Gcc-patches wrote:
> On Wed, Nov 16, 2022 at 12:51:14PM +0100, Jakub Jelinek via Gcc-patches wrote:
> > On Wed, Nov 16, 2022 at 10:06:17AM +0100, Jakub Jelinek via Gcc-patches wrote:
> > > Thoughts on this?  I guess my preference would be the BF -> SF -> TI
> > > path because we won't need to waste
> > >     32: 0000000000015e10   321 FUNC    GLOBAL DEFAULT   13 __fixbfti@@GCC_13.0.0
> > >     89: 0000000000015f60   299 FUNC    GLOBAL DEFAULT   13 __fixunsbfti@@GCC_13.0.0
> > > If so, I'd need to cut the fix parts of the patch below and
> > > do something in the middle-end.
> > 
> > Here is adjusted patch that does that.
> > 
> > 2022-11-16  Jakub Jelinek  <jakub@redhat.com>
> > 
> > 	PR target/107703
> > 	* optabs.cc (expand_fix): For conversions from BFmode to integral,
> > 	use shifts to convert it to SFmode first and then convert SFmode
> > 	to integral.
> > 
> > 	* soft-fp/floattibf.c: New file.
> > 	* soft-fp/floatuntibf.c: New file.
> > 	* config/i386/libgcc-glibc.ver: Export __float{,un}tibf @ GCC_13.0.0.
> > 	* config/i386/64/t-softfp (softfp_extras): Add floattibf and
> > 	floatuntibf.
> > 	(CFLAGS-floattibf.c, CFLAGS-floatunstibf.c): Add -msse2.
> 
> I'd like to ping the libgcc non-i386 part of this patch, Uros said the i386
> part is ok but that one depends on the generic libgcc changes.
> I'll ping the optabs.cc change separately.
> 
> https://gcc.gnu.org/pipermail/gcc-patches/2022-November/606398.html
> with more info in
> https://gcc.gnu.org/pipermail/gcc-patches/2022-November/606382.html

I'd like to ping this again.  I've posted the previously added
bfloat16 changes as well as the above 2 new files to libc-alpha as well
https://sourceware.org/pipermail/libc-alpha/2023-March/146246.html
if it makes the review easier.

Thanks

	Jakub
Ian Lance Taylor March 10, 2023, 6:10 p.m. UTC | #3 
Jakub Jelinek <jakub@redhat.com> writes:

> On Wed, Mar 01, 2023 at 01:32:43PM +0100, Jakub Jelinek via Gcc-patches wrote:
>> On Wed, Nov 16, 2022 at 12:51:14PM +0100, Jakub Jelinek via Gcc-patches wrote:
>> > On Wed, Nov 16, 2022 at 10:06:17AM +0100, Jakub Jelinek via
>> > Gcc-patches wrote:
>> > > Thoughts on this?  I guess my preference would be the BF -> SF -> TI
>> > > path because we won't need to waste
>> > >     32: 0000000000015e10 321 FUNC GLOBAL DEFAULT 13
>> > > __fixbfti@@GCC_13.0.0
>> > >     89: 0000000000015f60 299 FUNC GLOBAL DEFAULT 13
>> > > __fixunsbfti@@GCC_13.0.0
>> > > If so, I'd need to cut the fix parts of the patch below and
>> > > do something in the middle-end.
>> > 
>> > Here is adjusted patch that does that.
>> > 
>> > 2022-11-16  Jakub Jelinek  <jakub@redhat.com>
>> > 
>> > 	PR target/107703
>> > 	* optabs.cc (expand_fix): For conversions from BFmode to integral,
>> > 	use shifts to convert it to SFmode first and then convert SFmode
>> > 	to integral.
>> > 
>> > 	* soft-fp/floattibf.c: New file.
>> > 	* soft-fp/floatuntibf.c: New file.
>> > 	* config/i386/libgcc-glibc.ver: Export __float{,un}tibf @ GCC_13.0.0.
>> > 	* config/i386/64/t-softfp (softfp_extras): Add floattibf and
>> > 	floatuntibf.
>> > 	(CFLAGS-floattibf.c, CFLAGS-floatunstibf.c): Add -msse2.
>> 
>> I'd like to ping the libgcc non-i386 part of this patch, Uros said the i386
>> part is ok but that one depends on the generic libgcc changes.
>> I'll ping the optabs.cc change separately.
>> 
>> https://gcc.gnu.org/pipermail/gcc-patches/2022-November/606398.html
>> with more info in
>> https://gcc.gnu.org/pipermail/gcc-patches/2022-November/606382.html
>
> I'd like to ping this again.  I've posted the previously added
> bfloat16 changes as well as the above 2 new files to libc-alpha as well
> https://sourceware.org/pipermail/libc-alpha/2023-March/146246.html
> if it makes the review easier.


The libgcc parts of this are fine.  Thanks.

Ian
Uros Bizjak March 11, 2023, 9:17 a.m. UTC | #4 
On Fri, Mar 10, 2023 at 7:11 PM Ian Lance Taylor <ian@airs.com> wrote:
>
> Jakub Jelinek <jakub@redhat.com> writes:
>
> > On Wed, Mar 01, 2023 at 01:32:43PM +0100, Jakub Jelinek via Gcc-patches wrote:
> >> On Wed, Nov 16, 2022 at 12:51:14PM +0100, Jakub Jelinek via Gcc-patches wrote:
> >> > On Wed, Nov 16, 2022 at 10:06:17AM +0100, Jakub Jelinek via
> >> > Gcc-patches wrote:
> >> > > Thoughts on this?  I guess my preference would be the BF -> SF -> TI
> >> > > path because we won't need to waste
> >> > >     32: 0000000000015e10 321 FUNC GLOBAL DEFAULT 13
> >> > > __fixbfti@@GCC_13.0.0
> >> > >     89: 0000000000015f60 299 FUNC GLOBAL DEFAULT 13
> >> > > __fixunsbfti@@GCC_13.0.0
> >> > > If so, I'd need to cut the fix parts of the patch below and
> >> > > do something in the middle-end.
> >> >
> >> > Here is adjusted patch that does that.
> >> >
> >> > 2022-11-16  Jakub Jelinek  <jakub@redhat.com>
> >> >
> >> >    PR target/107703
> >> >    * optabs.cc (expand_fix): For conversions from BFmode to integral,
> >> >    use shifts to convert it to SFmode first and then convert SFmode
> >> >    to integral.
> >> >
> >> >    * soft-fp/floattibf.c: New file.
> >> >    * soft-fp/floatuntibf.c: New file.
> >> >    * config/i386/libgcc-glibc.ver: Export __float{,un}tibf @ GCC_13.0.0.
> >> >    * config/i386/64/t-softfp (softfp_extras): Add floattibf and
> >> >    floatuntibf.
> >> >    (CFLAGS-floattibf.c, CFLAGS-floatunstibf.c): Add -msse2.
> >>
> >> I'd like to ping the libgcc non-i386 part of this patch, Uros said the i386
> >> part is ok but that one depends on the generic libgcc changes.
> >> I'll ping the optabs.cc change separately.
> >>
> >> https://gcc.gnu.org/pipermail/gcc-patches/2022-November/606398.html
> >> with more info in
> >> https://gcc.gnu.org/pipermail/gcc-patches/2022-November/606382.html
> >
> > I'd like to ping this again.  I've posted the previously added
> > bfloat16 changes as well as the above 2 new files to libc-alpha as well
> > https://sourceware.org/pipermail/libc-alpha/2023-March/146246.html
> > if it makes the review easier.
>
>
> The libgcc parts of this are fine.  Thanks.

Also OK for the x86 part.

Thanks,
Uros.
Jeff Law March 11, 2023, 3:03 p.m. UTC | #5 
On 3/1/23 05:32, Jakub Jelinek via Gcc-patches wrote:
> Hi!
> 
> On Wed, Nov 16, 2022 at 12:51:14PM +0100, Jakub Jelinek via Gcc-patches wrote:
>> On Wed, Nov 16, 2022 at 10:06:17AM +0100, Jakub Jelinek via Gcc-patches wrote:
>>> Thoughts on this?  I guess my preference would be the BF -> SF -> TI
>>> path because we won't need to waste
>>>      32: 0000000000015e10   321 FUNC    GLOBAL DEFAULT   13 __fixbfti@@GCC_13.0.0
>>>      89: 0000000000015f60   299 FUNC    GLOBAL DEFAULT   13 __fixunsbfti@@GCC_13.0.0
>>> If so, I'd need to cut the fix parts of the patch below and
>>> do something in the middle-end.
>>
>> Here is adjusted patch that does that.
>>
>> 2022-11-16  Jakub Jelinek  <jakub@redhat.com>
>>
>> 	PR target/107703
>> 	* optabs.cc (expand_fix): For conversions from BFmode to integral,
>> 	use shifts to convert it to SFmode first and then convert SFmode
>> 	to integral.
>>
>> 	* soft-fp/floattibf.c: New file.
>> 	* soft-fp/floatuntibf.c: New file.
>> 	* config/i386/libgcc-glibc.ver: Export __float{,un}tibf @ GCC_13.0.0.
>> 	* config/i386/64/t-softfp (softfp_extras): Add floattibf and
>> 	floatuntibf.
>> 	(CFLAGS-floattibf.c, CFLAGS-floatunstibf.c): Add -msse2.
> 
> I'd like to ping the libgcc non-i386 part of this patch, Uros said the i386
> part is ok but that one depends on the generic libgcc changes.
> I'll ping the optabs.cc change separately.
> 
> https://gcc.gnu.org/pipermail/gcc-patches/2022-November/606398.html
> with more info in
> https://gcc.gnu.org/pipermail/gcc-patches/2022-November/606382.html
The optab changes are OK too.
jeff
diff mbox series

Patch

--- gcc/optabs.cc.jj	2022-11-16 07:29:11.665715915 +0100
+++ gcc/optabs.cc	2022-11-16 12:46:25.523281210 +0100
@@ -5574,7 +5574,21 @@  expand_fix (rtx to, rtx from, int unsign
 	    rtx_insn *last = get_last_insn ();
 	    rtx from1 = from;
 	    if (fmode != GET_MODE (from))
-	      from1 = convert_to_mode (fmode, from, 0);
+	      {
+		if (REAL_MODE_FORMAT (GET_MODE (from))
+		    == &arm_bfloat_half_format
+		    && REAL_MODE_FORMAT (fmode) == &ieee_single_format)
+		  /* The BF -> SF conversions can be just a shift, doesn't
+		     need to handle sNANs.  */
+		  {
+		    int save_flag_finite_math_only = flag_finite_math_only;
+		    flag_finite_math_only = true;
+		    from1 = convert_to_mode (fmode, from, 0);
+		    flag_finite_math_only = save_flag_finite_math_only;
+		  }
+		else
+		  from1 = convert_to_mode (fmode, from, 0);
+	      }
 
 	    if (must_trunc)
 	      {
@@ -5646,7 +5660,21 @@  expand_fix (rtx to, rtx from, int unsign
 	    lab2 = gen_label_rtx ();
 
 	    if (fmode != GET_MODE (from))
-	      from = convert_to_mode (fmode, from, 0);
+	      {
+		if (REAL_MODE_FORMAT (GET_MODE (from))
+		    == &arm_bfloat_half_format
+		    && REAL_MODE_FORMAT (fmode) == &ieee_single_format)
+		  /* The BF -> SF conversions can be just a shift, doesn't
+		     need to handle sNANs.  */
+		  {
+		    int save_flag_finite_math_only = flag_finite_math_only;
+		    flag_finite_math_only = true;
+		    from = convert_to_mode (fmode, from, 0);
+		    flag_finite_math_only = save_flag_finite_math_only;
+		  }
+		else
+		  from = convert_to_mode (fmode, from, 0);
+	      }
 
 	    /* See if we need to do the subtraction.  */
 	    do_pending_stack_adjust ();
@@ -5690,6 +5718,22 @@  expand_fix (rtx to, rtx from, int unsign
 	  }
       }
 
+#ifdef HAVE_SFmode
+  if (REAL_MODE_FORMAT (GET_MODE (from)) == &arm_bfloat_half_format
+      && REAL_MODE_FORMAT (SFmode) == &ieee_single_format)
+    /* We don't have BF -> TI library functions, use BF -> SF -> TI
+       instead but the BF -> SF conversion can be just a shift, doesn't
+       need to handle sNANs.  */
+    {
+      int save_flag_finite_math_only = flag_finite_math_only;
+      flag_finite_math_only = true;
+      from = convert_to_mode (SFmode, from, 0);
+      flag_finite_math_only = save_flag_finite_math_only;
+      expand_fix (to, from, unsignedp);
+      return;
+    }
+#endif
+
   /* We can't do it with an insn, so use a library call.  But first ensure
      that the mode of TO is at least as wide as SImode, since those are the
      only library calls we know about.  */
--- libgcc/soft-fp/floattibf.c.jj	2022-11-15 19:10:13.147834226 +0100
+++ libgcc/soft-fp/floattibf.c	2022-11-15 19:13:40.939996482 +0100
@@ -0,0 +1,45 @@ 
+/* Software floating-point emulation.
+   Convert a 128bit signed integer to bfloat16
+   Copyright (C) 2007-2022 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   In addition to the permissions in the GNU Lesser General Public
+   License, the Free Software Foundation gives you unlimited
+   permission to link the compiled version of this file into
+   combinations with other programs, and to distribute those
+   combinations without any restriction coming from the use of this
+   file.  (The Lesser General Public License restrictions do apply in
+   other respects; for example, they cover modification of the file,
+   and distribution when not linked into a combine executable.)
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#include "soft-fp.h"
+#include "brain.h"
+
+BFtype
+__floattibf (TItype i)
+{
+  FP_DECL_EX;
+  FP_DECL_B (A);
+  BFtype a;
+
+  FP_INIT_ROUNDMODE;
+  FP_FROM_INT_B (A, i, TI_BITS, UTItype);
+  FP_PACK_RAW_B (a, A);
+  FP_HANDLE_EXCEPTIONS;
+
+  return a;
+}
--- libgcc/soft-fp/floatuntibf.c.jj	2022-11-15 19:10:53.363285018 +0100
+++ libgcc/soft-fp/floatuntibf.c	2022-11-15 19:13:31.399126765 +0100
@@ -0,0 +1,45 @@ 
+/* Software floating-point emulation.
+   Convert a 128bit unsigned integer to bfloat16
+   Copyright (C) 2007-2022 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   In addition to the permissions in the GNU Lesser General Public
+   License, the Free Software Foundation gives you unlimited
+   permission to link the compiled version of this file into
+   combinations with other programs, and to distribute those
+   combinations without any restriction coming from the use of this
+   file.  (The Lesser General Public License restrictions do apply in
+   other respects; for example, they cover modification of the file,
+   and distribution when not linked into a combine executable.)
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#include "soft-fp.h"
+#include "brain.h"
+
+BFtype
+__floatuntibf (UTItype i)
+{
+  FP_DECL_EX;
+  FP_DECL_B (A);
+  BFtype a;
+
+  FP_INIT_ROUNDMODE;
+  FP_FROM_INT_B (A, i, TI_BITS, UTItype);
+  FP_PACK_RAW_B (a, A);
+  FP_HANDLE_EXCEPTIONS;
+
+  return a;
+}
--- libgcc/config/i386/libgcc-glibc.ver.jj	2022-10-14 09:35:56.269989297 +0200
+++ libgcc/config/i386/libgcc-glibc.ver	2022-11-15 19:26:37.125389987 +0100
@@ -218,6 +218,8 @@  GCC_12.0.0 {
 %inherit GCC_13.0.0 GCC_12.0.0
 GCC_13.0.0 {
   __extendbfsf2
+  __floattibf
+  __floatuntibf
   __truncdfbf2
   __truncsfbf2
   __trunctfbf2
--- libgcc/config/i386/64/t-softfp.jj	2021-12-30 15:12:44.111138056 +0100
+++ libgcc/config/i386/64/t-softfp	2022-11-15 19:26:07.184799287 +0100
@@ -1,6 +1,9 @@ 
-softfp_extras := fixhfti fixunshfti floattihf floatuntihf
+softfp_extras := fixhfti fixunshfti floattihf floatuntihf \
+		 floattibf floatuntibf
 
 CFLAGS-fixhfti.c += -msse2
 CFLAGS-fixunshfti.c += -msse2
 CFLAGS-floattihf.c += -msse2
 CFLAGS-floatunstihf.c += -msse2
+CFLAGS-floattibf.c += -msse2
+CFLAGS-floatunstibf.c += -msse2