diff mbox series

x86_64: Avoid rorx rotation instructions with -Os

Message ID 07fe01d7da28$5b78a820$1269f860$@nextmovesoftware.com
State New
Headers show
Series x86_64: Avoid rorx rotation instructions with -Os | expand

Commit Message

Roger Sayle Nov. 15, 2021, 1:54 p.m. UTC 
This patch teaches the i386 backend to avoid using BMI2's rorx
instructions when optimizing for size.  The benefits are shown
with the following example:

unsigned int ror1(unsigned int x) { return (x >> 1) | (x << 31); }
unsigned int ror2(unsigned int x) { return (x >> 2) | (x << 30); }
unsigned int rol2(unsigned int x) { return (x >> 30) | (x << 2); }
unsigned int rol1(unsigned int x) { return (x >> 31) | (x << 1); }

which currently with -Os -march=cascadelake generates:

ror1:   rorx    $1, %edi, %eax          // 6 bytes
        ret
ror2:   rorx    $2, %edi, %eax          // 6 bytes
        ret
rol2:   rorx    $30, %edi, %eax         // 6 bytes
        ret
rol1:   rorx    $31, %edi, %eax         // 6 bytes
        ret

but with this patch now generates:

ror1:   movl    %edi, %eax              // 2 bytes
        rorl    %eax                    // 2 bytes
        ret
ror2:   movl    %edi, %eax              // 2 bytes
        rorl    $2, %eax                // 3 bytes
        ret
rol2:   movl    %edi, %eax              // 2 bytes
        roll    $2, %eax                // 3 bytes
        ret
rol1:   movl    %edi, %eax              // 2 bytes
        roll    %eax                    // 2 bytes
        ret

I've confirmed that this patch is a win on the CSiBE benchmark,
even though rotations are rare, where for example libmspack/test/md5.o
shrinks from 5824 bytes to 5632 bytes.

This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
and make -k check with no new failures.  Ok for mainline?


2021-11-15  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
	* config/i386/i386.md (*bmi2_rorx<mode3>_1): Make conditional
	on !optimize_function_for_size_p.
	(*<any_rotate><mode>3_1): Add preferred_for_size attribute.
	(define_splits): Conditionalize on !optimize_function_for_size_p.
	(*bmi2_rorxsi3_1_zext): Likewise.
	(*<any_rotate>si2_1_zext): Add preferred_for_size attribute.
	(define_splits): Conditionalize on !optimize_function_for_size_p.

Thanks in advance,
Roger
--

Comments

Uros Bizjak Nov. 16, 2021, 8:34 a.m. UTC | #1 
On Mon, Nov 15, 2021 at 2:54 PM Roger Sayle <roger@nextmovesoftware.com> wrote:
>
>
> This patch teaches the i386 backend to avoid using BMI2's rorx
> instructions when optimizing for size.  The benefits are shown
> with the following example:
>
> unsigned int ror1(unsigned int x) { return (x >> 1) | (x << 31); }
> unsigned int ror2(unsigned int x) { return (x >> 2) | (x << 30); }
> unsigned int rol2(unsigned int x) { return (x >> 30) | (x << 2); }
> unsigned int rol1(unsigned int x) { return (x >> 31) | (x << 1); }
>
> which currently with -Os -march=cascadelake generates:
>
> ror1:   rorx    $1, %edi, %eax          // 6 bytes
>         ret
> ror2:   rorx    $2, %edi, %eax          // 6 bytes
>         ret
> rol2:   rorx    $30, %edi, %eax         // 6 bytes
>         ret
> rol1:   rorx    $31, %edi, %eax         // 6 bytes
>         ret
>
> but with this patch now generates:
>
> ror1:   movl    %edi, %eax              // 2 bytes
>         rorl    %eax                    // 2 bytes
>         ret
> ror2:   movl    %edi, %eax              // 2 bytes
>         rorl    $2, %eax                // 3 bytes
>         ret
> rol2:   movl    %edi, %eax              // 2 bytes
>         roll    $2, %eax                // 3 bytes
>         ret
> rol1:   movl    %edi, %eax              // 2 bytes
>         roll    %eax                    // 2 bytes
>         ret
>
> I've confirmed that this patch is a win on the CSiBE benchmark,
> even though rotations are rare, where for example libmspack/test/md5.o
> shrinks from 5824 bytes to 5632 bytes.
>
> This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
> and make -k check with no new failures.  Ok for mainline?
>
>
> 2021-11-15  Roger Sayle  <roger@nextmovesoftware.com>
>
> gcc/ChangeLog
>         * config/i386/i386.md (*bmi2_rorx<mode3>_1): Make conditional
>         on !optimize_function_for_size_p.
>         (*<any_rotate><mode>3_1): Add preferred_for_size attribute.
>         (define_splits): Conditionalize on !optimize_function_for_size_p.
>         (*bmi2_rorxsi3_1_zext): Likewise.
>         (*<any_rotate>si2_1_zext): Add preferred_for_size attribute.
>         (define_splits): Conditionalize on !optimize_function_for_size_p.

OK.

Thanks,
Uros.
diff mbox series

Patch

diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
index 6eb9de8..7394906 100644
--- a/gcc/config/i386/i386.md
+++ b/gcc/config/i386/i386.md
@@ -12775,7 +12775,7 @@ 
 	(rotatert:SWI48
 	  (match_operand:SWI48 1 "nonimmediate_operand" "rm")
 	  (match_operand:QI 2 "<rorx_immediate_operand>" "<S>")))]
-  "TARGET_BMI2"
+  "TARGET_BMI2 && !optimize_function_for_size_p (cfun)"
   "rorx\t{%2, %1, %0|%0, %1, %2}"
   [(set_attr "type" "rotatex")
    (set_attr "mode" "<MODE>")])
@@ -12803,6 +12803,10 @@ 
 }
   [(set_attr "isa" "*,bmi2")
    (set_attr "type" "rotate,rotatex")
+   (set (attr "preferred_for_size")
+     (cond [(eq_attr "alternative" "0")
+	      (symbol_ref "true")]
+	   (symbol_ref "false")))
    (set (attr "length_immediate")
      (if_then_else
        (and (eq_attr "type" "rotate")
@@ -12819,7 +12823,7 @@ 
 	(rotate:SWI48 (match_operand:SWI48 1 "nonimmediate_operand")
 		      (match_operand:QI 2 "const_int_operand")))
    (clobber (reg:CC FLAGS_REG))]
-  "TARGET_BMI2 && reload_completed"
+  "TARGET_BMI2 && reload_completed && !optimize_function_for_size_p (cfun)"
   [(set (match_dup 0)
 	(rotatert:SWI48 (match_dup 1) (match_dup 2)))]
 {
@@ -12833,7 +12837,7 @@ 
 	(rotatert:SWI48 (match_operand:SWI48 1 "nonimmediate_operand")
 			(match_operand:QI 2 "const_int_operand")))
    (clobber (reg:CC FLAGS_REG))]
-  "TARGET_BMI2 && reload_completed"
+  "TARGET_BMI2 && reload_completed && !optimize_function_for_size_p (cfun)"
   [(set (match_dup 0)
 	(rotatert:SWI48 (match_dup 1) (match_dup 2)))])
 
@@ -12842,7 +12846,7 @@ 
 	(zero_extend:DI
 	  (rotatert:SI (match_operand:SI 1 "nonimmediate_operand" "rm")
 		       (match_operand:QI 2 "const_0_to_31_operand" "I"))))]
-  "TARGET_64BIT && TARGET_BMI2"
+  "TARGET_64BIT && TARGET_BMI2 && !optimize_function_for_size_p (cfun)"
   "rorx\t{%2, %1, %k0|%k0, %1, %2}"
   [(set_attr "type" "rotatex")
    (set_attr "mode" "SI")])
@@ -12870,6 +12874,10 @@ 
 }
   [(set_attr "isa" "*,bmi2")
    (set_attr "type" "rotate,rotatex")
+   (set (attr "preferred_for_size")
+     (cond [(eq_attr "alternative" "0")
+	      (symbol_ref "true")]
+	   (symbol_ref "false")))
    (set (attr "length_immediate")
      (if_then_else
        (and (eq_attr "type" "rotate")
@@ -12887,7 +12895,8 @@ 
 	  (rotate:SI (match_operand:SI 1 "nonimmediate_operand")
 		     (match_operand:QI 2 "const_int_operand"))))
    (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && TARGET_BMI2 && reload_completed"
+  "TARGET_64BIT && TARGET_BMI2 && reload_completed
+   && !optimize_function_for_size_p (cfun)"
   [(set (match_dup 0)
 	(zero_extend:DI (rotatert:SI (match_dup 1) (match_dup 2))))]
 {
@@ -12902,7 +12911,8 @@ 
 	  (rotatert:SI (match_operand:SI 1 "nonimmediate_operand")
 		       (match_operand:QI 2 "const_int_operand"))))
    (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && TARGET_BMI2 && reload_completed"
+  "TARGET_64BIT && TARGET_BMI2 && reload_completed
+   && !optimize_function_for_size_p (cfun)"
   [(set (match_dup 0)
 	(zero_extend:DI (rotatert:SI (match_dup 1) (match_dup 2))))])