[PATCH SH] Don't switch mode around fmov (pr61195)

[PATCH SH] Don't switch mode around fmov (pr61195)

[Christian Bruel]

[-- Attachment #1: Type: text/plain, Size: 300 bytes --]

Hi,

This patch reduces the number of unnecessary PR mode switches for single
precision moves. When FPSCR.SZ is not forced. Test illustration in the patch

Regtested for sh-none-elf with -m4, (-m2a still running), bootstrapped
on sh4-linux-elf board. ok for trunk ?

Many thanks,

Christian





 



[-- Attachment #2: fmov.patch --]
[-- Type: text/x-patch, Size: 2051 bytes --]

2014-05-16  Christian Bruel  <christian.bruel@st.com>

	PR target/61195
	* config/sh/sh.c (movsf_ie): Unset fp_mode for fmov.

2014-05-16  Christian Bruel  <christian.bruel@st.com>

	PR target/61195
	* gcc.target/sh/pr61195.c: New test.

Index: config/sh/sh.md
===================================================================
--- config/sh/sh.md	(revision 210475)
+++ config/sh/sh.md	(working copy)
@@ -8357,9 +8357,26 @@ label:
       (const_int 2)
       (const_int 2)
       (const_int 0)])
-   (set (attr "fp_mode") (if_then_else (eq_attr "fmovd" "yes")
-					   (const_string "single")
-					   (const_string "single")))])
+  (set_attr_alternative "fp_mode"
+     [(if_then_else (eq_attr "fmovd" "yes") (const_string "single") (const_string "none"))
+      (const_string "none")
+      (const_string "single")
+      (const_string "single")
+      (const_string "none")
+      (if_then_else (eq_attr "fmovd" "yes") (const_string "single") (const_string "none"))
+      (if_then_else (eq_attr "fmovd" "yes") (const_string "single") (const_string "none"))
+      (const_string "none")
+      (const_string "none")
+      (const_string "none")
+      (const_string "none")
+      (const_string "none")
+      (const_string "none")
+      (const_string "none")
+      (const_string "none")
+      (const_string "none")
+      (const_string "none")
+      (const_string "none")
+      (const_string "none")])])
 
 (define_split
   [(set (match_operand:SF 0 "register_operand" "")
Index: testsuite/gcc.target/sh/pr61195.c
===================================================================
--- testsuite/gcc.target/sh/pr61195.c	(revision 0)
+++ testsuite/gcc.target/sh/pr61195.c	(working copy)
@@ -0,0 +1,19 @@
+/* Verify that we don't switch mode for single moves.  */
+/* { dg-do compile }  */
+/* { dg-require-effective-target hard_float } */
+/* { dg-skip-if "" { *-*-* }  { "mfmovd" } { "" } } */
+/* { dg-final { scan-assembler-not "fpscr" } } */
+
+float *g;
+
+float
+foo(float f)
+{
+  return f;
+}
+
+float
+foo1(void)
+{
+  return *g;
+}

Re: [PATCH SH] Don't switch mode around fmov (pr61195)

[Kaz Kojima]

Christian Bruel <christian.bruel@st.com> wrote:
> This patch reduces the number of unnecessary PR mode switches for single
> precision moves. When FPSCR.SZ is not forced. Test illustration in the patch
> 
> Regtested for sh-none-elf with -m4, (-m2a still running), bootstrapped
> on sh4-linux-elf board. ok for trunk ?

OK.

Regards,
	kaz

Re: [PATCH SH] Don't switch mode around fmov (pr61195)

[Oleg Endo]

On Fri, 2014-05-16 at 11:14 +0200, Christian Bruel wrote:
> Hi,
> 
> This patch reduces the number of unnecessary PR mode switches for single
> precision moves. When FPSCR.SZ is not forced. Test illustration in the patch
> 
> Regtested for sh-none-elf with -m4, (-m2a still running), bootstrapped
> on sh4-linux-elf board. ok for trunk ?

Please wrap the overlong lines ( eq_attr "fmovd" "yes") ... ) and
include a change log for sh.md.
OK with those changes if -m2a doesn't show failures.

Cheers,
Oleg