Patchwork [ARM] AArch32 vmaxnm, vminnm support

login
register
mail settings
Submitter Kyrylo Tkachov
Date Nov. 26, 2012, 2:40 p.m.
Message ID <000601cdcbe3$f1f8edd0$d5eac970$@tkachov@arm.com>
Download mbox | patch
Permalink /patch/201705/
State New
Headers show

Comments

Kyrylo Tkachov - Nov. 26, 2012, 2:40 p.m.
Hi all,
This patch adds support for the AArch32 vmaxnm and vminnm VFP instructions
in that can be used to implement the smax[sf,df]3 and smin[sf,df]3 RTL
patterns.
The patterns are only used by gcc when unsafe math optimisations are turned
on.
Two new values for the type attribute are introduced: f_minmaxs and
f_minmaxd.

New compilation tests are added. They pass and no regressions on
arm-none-eabi.

Ok for trunk?

Thanks,
Kyrill

gcc/ChangeLog

2012-11-26  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>

        * config/arm/arm.md (f_minmaxs, f_minmaxd): New types.
        * config/arm/vfp.md (smax<mode>3): New pattern.
         (smin<mode>3): Likewise.


gcc/testsuite/ChangeLog

2012-11-26  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>

        * gcc.target/arm/vmaxnmdf.c: New test.
        * gcc.target/arm/vmaxnmsf.c: Likewise.
        * gcc.target/arm/vminnmsf.c: Likewise.
        * gcc.target/arm/vminnmdf.c: Likewise.
Kyrylo Tkachov - Dec. 4, 2012, 6 p.m.
Ping.

Thanks,
Kyrill

> -----Original Message-----
> From: gcc-patches-owner@gcc.gnu.org [mailto:gcc-patches-
> owner@gcc.gnu.org] On Behalf Of Kyrylo Tkachov
> Sent: 26 November 2012 14:40
> To: gcc-patches@gcc.gnu.org
> Cc: Ramana Radhakrishnan; Richard Earnshaw
> Subject: [PATCH][ARM] AArch32 vmaxnm, vminnm support
> 
> Hi all,
> This patch adds support for the AArch32 vmaxnm and vminnm VFP
> instructions
> in that can be used to implement the smax[sf,df]3 and smin[sf,df]3 RTL
> patterns.
> The patterns are only used by gcc when unsafe math optimisations are
> turned
> on.
> Two new values for the type attribute are introduced: f_minmaxs and
> f_minmaxd.
> 
> New compilation tests are added. They pass and no regressions on
> arm-none-eabi.
> 
> Ok for trunk?
> 
> Thanks,
> Kyrill
> 
> gcc/ChangeLog
> 
> 2012-11-26  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>
> 
>         * config/arm/arm.md (f_minmaxs, f_minmaxd): New types.
>         * config/arm/vfp.md (smax<mode>3): New pattern.
>          (smin<mode>3): Likewise.
> 
> 
> gcc/testsuite/ChangeLog
> 
> 2012-11-26  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>
> 
>         * gcc.target/arm/vmaxnmdf.c: New test.
>         * gcc.target/arm/vmaxnmsf.c: Likewise.
>         * gcc.target/arm/vminnmsf.c: Likewise.
>         * gcc.target/arm/vminnmdf.c: Likewise.
Kyrylo Tkachov - Dec. 19, 2012, 2:04 p.m.
Ping.

http://gcc.gnu.org/ml/gcc-patches/2012-11/msg02097.html

Thanks,
Kyrill

> -----Original Message-----
> From: gcc-patches-owner@gcc.gnu.org [mailto:gcc-patches-
> owner@gcc.gnu.org] On Behalf Of Kyrylo Tkachov
> Sent: 26 November 2012 14:40
> To: gcc-patches@gcc.gnu.org
> Cc: Ramana Radhakrishnan; Richard Earnshaw
> Subject: [PATCH][ARM] AArch32 vmaxnm, vminnm support
> 
> Hi all,
> This patch adds support for the AArch32 vmaxnm and vminnm VFP
> instructions
> in that can be used to implement the smax[sf,df]3 and smin[sf,df]3 RTL
> patterns.
> The patterns are only used by gcc when unsafe math optimisations are
> turned
> on.
> Two new values for the type attribute are introduced: f_minmaxs and
> f_minmaxd.
> 
> New compilation tests are added. They pass and no regressions on
> arm-none-eabi.
> 
> Ok for trunk?
> 
> Thanks,
> Kyrill
> 
> gcc/ChangeLog
> 
> 2012-11-26  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>
> 
>         * config/arm/arm.md (f_minmaxs, f_minmaxd): New types.
>         * config/arm/vfp.md (smax<mode>3): New pattern.
>          (smin<mode>3): Likewise.
> 
> 
> gcc/testsuite/ChangeLog
> 
> 2012-11-26  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>
> 
>         * gcc.target/arm/vmaxnmdf.c: New test.
>         * gcc.target/arm/vmaxnmsf.c: Likewise.
>         * gcc.target/arm/vminnmsf.c: Likewise.
>         * gcc.target/arm/vminnmdf.c: Likewise.
Richard Earnshaw - Dec. 19, 2012, 3:43 p.m.
On 26/11/12 14:40, Kyrylo Tkachov wrote:
> Hi all,
> This patch adds support for the AArch32 vmaxnm and vminnm VFP instructions
> in that can be used to implement the smax[sf,df]3 and smin[sf,df]3 RTL
> patterns.
> The patterns are only used by gcc when unsafe math optimisations are turned
> on.
> Two new values for the type attribute are introduced: f_minmaxs and
> f_minmaxd.
>
> New compilation tests are added. They pass and no regressions on
> arm-none-eabi.
>
> Ok for trunk?
>
OK.

R.

Patch

--- a/gcc/config/arm/arm.md
+++ b/gcc/config/arm/arm.md
@@ -274,6 +274,8 @@ 
   fmacd,\
   f_rints,\
   f_rintd,\
+  f_minmaxs,\
+  f_minmaxd,\
   f_flag,\
   f_loads,\
   f_loadd,\
diff --git a/gcc/config/arm/vfp.md b/gcc/config/arm/vfp.md
index 480a88d..82b277a 100644
--- a/gcc/config/arm/vfp.md
+++ b/gcc/config/arm/vfp.md
@@ -1263,6 +1263,31 @@ 
    (set_attr "type" "f_rint<vfp_type>")]
 )
 
+;; MIN_EXPR and MAX_EXPR eventually map to 'smin' and 'smax' in RTL.
+;; The 'smax' and 'smin' RTL standard pattern names do not specify which
+;; operand will be returned when both operands are zero (i.e. they may not
+;; honour signed zeroes), or when either operand is NaN.  Therefore GCC
+;; only introduces MIN_EXPR/MAX_EXPR in fast math mode or when not honouring
+;; NaNs.
+
+(define_insn "smax<mode>3"
+  [(set (match_operand:SDF 0 "register_operand" "=<F_constraint>")
+        (smax:SDF (match_operand:SDF 1 "register_operand" "<F_constraint>")
+		  (match_operand:SDF 2 "register_operand" "<F_constraint>")))]
+  "TARGET_HARD_FLOAT && TARGET_FPU_ARMV8 <vfp_double_cond>"
+  "vmaxnm.<V_if_elem>\\t%<V_reg>0, %<V_reg>1, %<V_reg>2"
+  [(set_attr "type" "f_minmax<vfp_type>")]
+)
+
+(define_insn "smin<mode>3"
+  [(set (match_operand:SDF 0 "register_operand" "=<F_constraint>")
+        (smin:SDF (match_operand:SDF 1 "register_operand" "<F_constraint>")
+		  (match_operand:SDF 2 "register_operand" "<F_constraint>")))]
+  "TARGET_HARD_FLOAT && TARGET_FPU_ARMV8 <vfp_double_cond>"
+  "vminnm.<V_if_elem>\\t%<V_reg>0, %<V_reg>1, %<V_reg>2"
+  [(set_attr "type" "f_minmax<vfp_type>")]
+)
+
 ;; Unimplemented insns:
 ;; fldm*
 ;; fstm*
diff --git a/gcc/testsuite/gcc.target/arm/vmaxnmdf.c b/gcc/testsuite/gcc.target/arm/vmaxnmdf.c
new file mode 100644
index 0000000..1a172b8
--- /dev/null
+++ b/gcc/testsuite/gcc.target/arm/vmaxnmdf.c
@@ -0,0 +1,12 @@ 
+/* { dg-do compile } */
+/* { dg-require-effective-target arm_v8_vfp_ok } */
+/* { dg-options "-ffast-math" } */
+/* { dg-add-options arm_v8_vfp } */
+
+double
+foo (double x, double y)
+{
+  return __builtin_fmax (x, y);
+}
+
+/* { dg-final { scan-assembler-times "vmaxnm.f64\td\[0-9\]+" 1 } } */
diff --git a/gcc/testsuite/gcc.target/arm/vmaxnmsf.c b/gcc/testsuite/gcc.target/arm/vmaxnmsf.c
new file mode 100644
index 0000000..bc23261
--- /dev/null
+++ b/gcc/testsuite/gcc.target/arm/vmaxnmsf.c
@@ -0,0 +1,12 @@ 
+/* { dg-do compile } */
+/* { dg-require-effective-target arm_v8_vfp_ok } */
+/* { dg-options "-ffast-math" } */
+/* { dg-add-options arm_v8_vfp } */
+
+float
+foo (float x, float y)
+{
+  return __builtin_fmaxf (x, y);
+}
+
+/* { dg-final { scan-assembler-times "vmaxnm.f32\ts\[0-9\]+" 1 } } */
diff --git a/gcc/testsuite/gcc.target/arm/vminnmdf.c b/gcc/testsuite/gcc.target/arm/vminnmdf.c
new file mode 100644
index 0000000..c2a6915
--- /dev/null
+++ b/gcc/testsuite/gcc.target/arm/vminnmdf.c
@@ -0,0 +1,12 @@ 
+/* { dg-do compile } */
+/* { dg-require-effective-target arm_v8_vfp_ok } */
+/* { dg-options "-ffast-math" } */
+/* { dg-add-options arm_v8_vfp } */
+
+double
+foo (double x, double y)
+{
+  return __builtin_fmin (x, y);
+}
+
+/* { dg-final { scan-assembler-times "vminnm.f64\td\[0-9\]+" 1 } } */
diff --git a/gcc/testsuite/gcc.target/arm/vminnmsf.c b/gcc/testsuite/gcc.target/arm/vminnmsf.c
new file mode 100644
index 0000000..eee43bc
--- /dev/null
+++ b/gcc/testsuite/gcc.target/arm/vminnmsf.c
@@ -0,0 +1,12 @@ 
+/* { dg-do compile } */
+/* { dg-require-effective-target arm_v8_vfp_ok } */
+/* { dg-options "-ffast-math" } */
+/* { dg-add-options arm_v8_vfp } */
+
+float
+foo (float x, float y)
+{
+  return __builtin_fminf (x, y);
+}
+
+/* { dg-final { scan-assembler-times "vminnm.f32\ts\[0-9\]+" 1 } } */