[PATCH][ARM] minmax_arithsi for non-canonical operand order with MINUS operator

[PATCH][ARM] minmax_arithsi for non-canonical operand order with MINUS operator

[Kyrylo Tkachov]

[-- Attachment #1: Type: text/plain, Size: 834 bytes --]

Hi all,

This patch adds a splitter variant of the minmax_arithsi pattern for when
the operator
is non-commutative (MINUS) and the ordering of the operands is not
canonical.

That is, it will trigger for:
#define MAX(a, b) (a > b ? a : b)
int
foo (int a, int b, int c)
{
  return c - MAX (a,b);
}

and will generate:
         cmp     r1, r0
         rsbge   r0, r1, r2
         rsblt   r0, r0, r2

instead of the current:
        cmp     r0, r1
        movlt   r0, r1
        rsb     r0, r0, r2

No regressions on arm-none-eabi.

Ok for trunk?

Thanks,
Kyrill

gcc/
2013-03-21  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>

	* config/arm/arm.md (minmax_arithsi_non_canon): New pattern.

gcc/testsuite
2013-03-21  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>

	* gcc.target/arm/minmax_minus.c: New test.

[-- Attachment #2: minmax_arithsi.txt --]
[-- Type: text/plain, Size: 2620 bytes --]

diff --git a/gcc/config/arm/arm.md b/gcc/config/arm/arm.md
index f3c59f3..c2875f9 100644
--- a/gcc/config/arm/arm.md
+++ b/gcc/config/arm/arm.md
@@ -3423,6 +3423,49 @@
 		      (const_int 12)))]
 )
 
+; Reject the frame pointer in operand[1], since reloading this after
+; it has been eliminated can cause carnage.
+(define_insn_and_split "*minmax_arithsi_non_canon"
+  [(set (match_operand:SI 0 "s_register_operand" "=r,r")
+	(minus:SI
+	 (match_operand:SI 1 "s_register_operand" "0,?r")
+	  (match_operator:SI 4 "minmax_operator"
+	   [(match_operand:SI 2 "s_register_operand" "r,r")
+	    (match_operand:SI 3 "arm_rhs_operand" "rI,rI")])))
+   (clobber (reg:CC CC_REGNUM))]
+  "TARGET_32BIT && !arm_eliminable_register (operands[1])"
+  "#"
+  "TARGET_32BIT && !arm_eliminable_register (operands[1])"
+  [(set (reg:CC CC_REGNUM)
+        (compare:CC (match_dup 2) (match_dup 3)))
+   (cond_exec (match_op_dup 4 [(reg:CC CC_REGNUM) (const_int 0)])
+              (set (match_dup 0)
+                   (minus:SI (match_dup 1)
+                             (match_dup 2))))
+   (cond_exec (match_op_dup 5 [(reg:CC CC_REGNUM) (const_int 0)])
+              (set (match_dup 0)
+                   (minus:SI (match_dup 1)
+                             (match_dup 3))))]
+  {
+  enum machine_mode mode = SELECT_CC_MODE (GET_CODE (operands[1]),
+                                           operands[2], operands[3]);
+  enum rtx_code rc = minmax_code (operands[4]);
+  operands[4] = gen_rtx_fmt_ee (rc, SImode,
+                                operands[2], operands[3]);
+
+  if (mode == CCFPmode || mode == CCFPEmode)
+    rc = reverse_condition_maybe_unordered (rc);
+  else
+    rc = reverse_condition (rc);
+  operands[5] = gen_rtx_fmt_ee (rc, SImode, operands[2], operands[3]);
+  }
+  [(set_attr "conds" "clob")
+   (set (attr "length")
+	(if_then_else (eq_attr "is_thumb" "yes")
+		      (const_int 14)
+		      (const_int 12)))]
+)
+
 (define_code_iterator SAT [smin smax])
 (define_code_iterator SATrev [smin smax])
 (define_code_attr SATlo [(smin "1") (smax "2")])
diff --git a/gcc/testsuite/gcc.target/arm/minmax_minus.c b/gcc/testsuite/gcc.target/arm/minmax_minus.c
new file mode 100644
index 0000000..050c847
--- /dev/null
+++ b/gcc/testsuite/gcc.target/arm/minmax_minus.c
@@ -0,0 +1,12 @@
+/* { dg-do compile } */
+/* { dg-options "-O2" } */
+
+#define MAX(a, b) (a > b ? a : b)
+int
+foo (int a, int b, int c)
+{
+  return c - MAX (a, b);
+}
+
+/* { dg-final { scan-assembler "rsbge" } } */
+/* { dg-final { scan-assembler "rsblt" } } */

Re: [PATCH][ARM] minmax_arithsi for non-canonical operand order with MINUS operator

[Richard Earnshaw]

On 02/04/13 17:06, Kyrylo Tkachov wrote:
>> From: Ramana Radhakrishnan [mailto:ramana.gcc@googlemail.com]
>> Sent: 02 April 2013 11:10
>> To: Kyrylo Tkachov
>> Cc: gcc-patches@gcc.gnu.org; Richard Earnshaw; Ramana Radhakrishnan
>> Subject: Re: [PATCH][ARM] minmax_arithsi for non-canonical operand
>> order with MINUS operator
>>
>> On Thu, Mar 21, 2013 at 6:09 PM, Kyrylo Tkachov
>> <kyrylo.tkachov@arm.com> wrote:
>>> Hi all,
>>>
>>> This patch adds a splitter variant of the minmax_arithsi pattern for
>> when
>>> the operator
>>> is non-commutative (MINUS) and the ordering of the operands is not
>>> canonical.
>>>
>>> That is, it will trigger for:
>>> #define MAX(a, b) (a > b ? a : b)
>>> int
>>> foo (int a, int b, int c)
>>> {
>>>    return c - MAX (a,b);
>>> }
>>>
>>> and will generate:
>>>           cmp     r1, r0
>>>           rsbge   r0, r1, r2
>>>           rsblt   r0, r0, r2
>>>
>>> instead of the current:
>>>          cmp     r0, r1
>>>          movlt   r0, r1
>>>          rsb     r0, r0, r2
>>>
>>> No regressions on arm-none-eabi.
>>>
>>> Ok for trunk?
>>>
>>
>> Split after reload please into cond-exec or use if_then_else instead
>> if you are splitting before reload - I originally thought it to be
>> safe when you asked me, but then have gone back and corrected myself.
>>
>> Read this thread . http://patches.linaro.org/6469/ .
>
> Hi Ramana, thanks for the review.
> How about this? We split after reload now.
> Using if_then_else got me a lot of unrecognisable instruction ICEs and
> delaying the split till after reload seemed like a cleaner approach.
>
> Tested arm-none-eabi on qemu.
>
>
> gcc/
> 2013-04-02  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>
>
>          * config/arm/arm.md (minmax_arithsi_non_canon): New pattern.
>
> gcc/testsuite
> 2013-04-02  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>
>
>          * gcc.target/arm/minmax_minus.c: New test.
>
>

OK.

R.

RE: [PATCH][ARM] minmax_arithsi for non-canonical operand order with MINUS operator

[Kyrylo Tkachov]

[-- Attachment #1: Type: text/plain, Size: 1902 bytes --]

> From: Ramana Radhakrishnan [mailto:ramana.gcc@googlemail.com]
> Sent: 02 April 2013 11:10
> To: Kyrylo Tkachov
> Cc: gcc-patches@gcc.gnu.org; Richard Earnshaw; Ramana Radhakrishnan
> Subject: Re: [PATCH][ARM] minmax_arithsi for non-canonical operand
> order with MINUS operator
> 
> On Thu, Mar 21, 2013 at 6:09 PM, Kyrylo Tkachov
> <kyrylo.tkachov@arm.com> wrote:
> > Hi all,
> >
> > This patch adds a splitter variant of the minmax_arithsi pattern for
> when
> > the operator
> > is non-commutative (MINUS) and the ordering of the operands is not
> > canonical.
> >
> > That is, it will trigger for:
> > #define MAX(a, b) (a > b ? a : b)
> > int
> > foo (int a, int b, int c)
> > {
> >   return c - MAX (a,b);
> > }
> >
> > and will generate:
> >          cmp     r1, r0
> >          rsbge   r0, r1, r2
> >          rsblt   r0, r0, r2
> >
> > instead of the current:
> >         cmp     r0, r1
> >         movlt   r0, r1
> >         rsb     r0, r0, r2
> >
> > No regressions on arm-none-eabi.
> >
> > Ok for trunk?
> >
> 
> Split after reload please into cond-exec or use if_then_else instead
> if you are splitting before reload - I originally thought it to be
> safe when you asked me, but then have gone back and corrected myself.
> 
> Read this thread . http://patches.linaro.org/6469/ .

Hi Ramana, thanks for the review.
How about this? We split after reload now.
Using if_then_else got me a lot of unrecognisable instruction ICEs and
delaying the split till after reload seemed like a cleaner approach.

Tested arm-none-eabi on qemu.


gcc/
2013-04-02  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>

        * config/arm/arm.md (minmax_arithsi_non_canon): New pattern.

gcc/testsuite
2013-04-02  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>

        * gcc.target/arm/minmax_minus.c: New test.


> regards
> Ramana
> 
Thanks,
Kyrill

[-- Attachment #2: minmax_arithsi.txt --]
[-- Type: text/plain, Size: 2645 bytes --]

diff --git a/gcc/config/arm/arm.md b/gcc/config/arm/arm.md
index 8895080..7ff066e 100644
--- a/gcc/config/arm/arm.md
+++ b/gcc/config/arm/arm.md
@@ -3417,6 +3417,50 @@
 		      (const_int 12)))]
 )
 
+; Reject the frame pointer in operand[1], since reloading this after
+; it has been eliminated can cause carnage.
+(define_insn_and_split "*minmax_arithsi_non_canon"
+  [(set (match_operand:SI 0 "s_register_operand" "=r,r")
+	(minus:SI
+	 (match_operand:SI 1 "s_register_operand" "0,?r")
+	  (match_operator:SI 4 "minmax_operator"
+	   [(match_operand:SI 2 "s_register_operand" "r,r")
+	    (match_operand:SI 3 "arm_rhs_operand" "rI,rI")])))
+   (clobber (reg:CC CC_REGNUM))]
+  "TARGET_32BIT && !arm_eliminable_register (operands[1])"
+  "#"
+  "TARGET_32BIT && !arm_eliminable_register (operands[1]) && reload_completed"
+  [(set (reg:CC CC_REGNUM)
+        (compare:CC (match_dup 2) (match_dup 3)))
+
+   (cond_exec (match_op_dup 4 [(reg:CC CC_REGNUM) (const_int 0)])
+              (set (match_dup 0)
+                   (minus:SI (match_dup 1)
+                             (match_dup 2))))
+   (cond_exec (match_op_dup 5 [(reg:CC CC_REGNUM) (const_int 0)])
+              (set (match_dup 0)
+                   (minus:SI (match_dup 1)
+                             (match_dup 3))))]
+  {
+  enum machine_mode mode = SELECT_CC_MODE (GET_CODE (operands[1]),
+                                           operands[2], operands[3]);
+  enum rtx_code rc = minmax_code (operands[4]);
+  operands[4] = gen_rtx_fmt_ee (rc, VOIDmode,
+                                operands[2], operands[3]);
+
+  if (mode == CCFPmode || mode == CCFPEmode)
+    rc = reverse_condition_maybe_unordered (rc);
+  else
+    rc = reverse_condition (rc);
+  operands[5] = gen_rtx_fmt_ee (rc, SImode, operands[2], operands[3]);
+  }
+  [(set_attr "conds" "clob")
+   (set (attr "length")
+	(if_then_else (eq_attr "is_thumb" "yes")
+		      (const_int 14)
+		      (const_int 12)))]
+)
+
 (define_code_iterator SAT [smin smax])
 (define_code_iterator SATrev [smin smax])
 (define_code_attr SATlo [(smin "1") (smax "2")])
diff --git a/gcc/testsuite/gcc.target/arm/minmax_minus.c b/gcc/testsuite/gcc.target/arm/minmax_minus.c
new file mode 100644
index 0000000..050c847
--- /dev/null
+++ b/gcc/testsuite/gcc.target/arm/minmax_minus.c
@@ -0,0 +1,12 @@
+/* { dg-do compile } */
+/* { dg-options "-O2" } */
+
+#define MAX(a, b) (a > b ? a : b)
+int
+foo (int a, int b, int c)
+{
+  return c - MAX (a, b);
+}
+
+/* { dg-final { scan-assembler "rsbge" } } */
+/* { dg-final { scan-assembler "rsblt" } } */

Re: [PATCH][ARM] minmax_arithsi for non-canonical operand order with MINUS operator

[Ramana Radhakrishnan]

On Thu, Mar 21, 2013 at 6:09 PM, Kyrylo Tkachov <kyrylo.tkachov@arm.com> wrote:
> Hi all,
>
> This patch adds a splitter variant of the minmax_arithsi pattern for when
> the operator
> is non-commutative (MINUS) and the ordering of the operands is not
> canonical.
>
> That is, it will trigger for:
> #define MAX(a, b) (a > b ? a : b)
> int
> foo (int a, int b, int c)
> {
>   return c - MAX (a,b);
> }
>
> and will generate:
>          cmp     r1, r0
>          rsbge   r0, r1, r2
>          rsblt   r0, r0, r2
>
> instead of the current:
>         cmp     r0, r1
>         movlt   r0, r1
>         rsb     r0, r0, r2
>
> No regressions on arm-none-eabi.
>
> Ok for trunk?
>

Split after reload please into cond-exec or use if_then_else instead
if you are splitting before reload - I originally thought it to be
safe when you asked me, but then have gone back and corrected myself.

Read this thread . http://patches.linaro.org/6469/ .


regards
Ramana

> Thanks,
> Kyrill
>
> gcc/
> 2013-03-21  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>
>
>         * config/arm/arm.md (minmax_arithsi_non_canon): New pattern.
>
> gcc/testsuite
> 2013-03-21  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>
>
>         * gcc.target/arm/minmax_minus.c: New test.