Bug 27663 - missed-optimization transforming a byte array to unsigned long
Summary: missed-optimization transforming a byte array to unsigned long
Status: NEW
Alias: None
Product: gcc
Classification: Unclassified
Component: middle-end (show other bugs)
Version: 4.2.0
: P3 enhancement
Target Milestone: ---
Assignee: Not yet assigned to anyone
URL:
Keywords: missed-optimization
Depends on:
Blocks:
 
Reported: 2006-05-18 17:58 UTC by berndtrog
Modified: 2014-02-18 08:36 UTC (History)
5 users (show)

See Also:
Host:
Target: avr
Build:
Known to work:
Known to fail: 4.2.0, 4.2.1
Last reconfirmed: 2007-07-25 17:29:16


Attachments

Note You need to log in before you can comment on or make changes to this bug.
Description berndtrog 2006-05-18 17:58:17 UTC
Hi,
this code:

unsigned long f (unsigned char  *P)
{
  unsigned long C;
  C  = ((unsigned long)P[1] << 24)
     | ((unsigned long)P[2] << 16)
     | ((unsigned long)P[3] <<  8)
     | ((unsigned long)P[4] <<  0);
  return C;
}

compiles to this:

00000000 <f>:
   0:   f9 2f           mov     r31, r25
   2:   e8 2f           mov     r30, r24
   4:   61 81           ldd     r22, Z+1        ; 0x01
   6:   77 27           eor     r23, r23
   8:   88 27           eor     r24, r24
   a:   99 27           eor     r25, r25
   c:   96 2f           mov     r25, r22
   e:   88 27           eor     r24, r24
  10:   77 27           eor     r23, r23
  12:   66 27           eor     r22, r22
  14:   22 81           ldd     r18, Z+2        ; 0x02
  16:   33 27           eor     r19, r19
  18:   44 27           eor     r20, r20
  1a:   55 27           eor     r21, r21
  1c:   53 2f           mov     r21, r19
  1e:   42 2f           mov     r20, r18
  20:   33 27           eor     r19, r19
  22:   22 27           eor     r18, r18
  24:   62 2b           or      r22, r18
  26:   73 2b           or      r23, r19
  28:   84 2b           or      r24, r20
  2a:   95 2b           or      r25, r21
  2c:   24 81           ldd     r18, Z+4        ; 0x04
  2e:   33 27           eor     r19, r19
  30:   44 27           eor     r20, r20
  32:   55 27           eor     r21, r21
  34:   62 2b           or      r22, r18
  36:   73 2b           or      r23, r19
  38:   84 2b           or      r24, r20
  3a:   95 2b           or      r25, r21
  3c:   23 81           ldd     r18, Z+3        ; 0x03
  3e:   33 27           eor     r19, r19
  40:   44 27           eor     r20, r20
  42:   55 27           eor     r21, r21
  44:   54 2f           mov     r21, r20
  46:   43 2f           mov     r20, r19
  48:   32 2f           mov     r19, r18
  4a:   22 27           eor     r18, r18
  4c:   62 2b           or      r22, r18
  4e:   73 2b           or      r23, r19
  50:   84 2b           or      r24, r20
  52:   95 2b           or      r25, r21
  54:   08 95           ret

using this cmd line:
avr-gcc -c -Os f.c

IMO, most of the or, eor and mov instructions are unnecessary.
Comment 1 Eric Weddington 2007-04-09 23:33:20 UTC
Bernd, what mcu type was this compiled for?
Comment 2 Eric Weddington 2007-07-25 17:29:16 UTC
Confirmed on 4.2.1.
Comment 3 Urja Rannikko 2009-04-24 20:24:25 UTC
Confirmed on 4.3.2 - it's a bit different and actually worse (longer):
(Please add 4.3.2 to known to fail - i cant)
00000000 <f>:
   0:   e8 2f           mov     r30, r24
   2:   f9 2f           mov     r31, r25
   4:   21 81           ldd     r18, Z+1        ; 0x01
   6:   30 e0           ldi     r19, 0x00       ; 0
   8:   40 e0           ldi     r20, 0x00       ; 0
   a:   50 e0           ldi     r21, 0x00       ; 0
   c:   52 2f           mov     r21, r18
   e:   44 27           eor     r20, r20
  10:   33 27           eor     r19, r19
  12:   22 27           eor     r18, r18
  14:   82 81           ldd     r24, Z+2        ; 0x02
  16:   90 e0           ldi     r25, 0x00       ; 0
  18:   a0 e0           ldi     r26, 0x00       ; 0
  1a:   b0 e0           ldi     r27, 0x00       ; 0
  1c:   a8 2f           mov     r26, r24
  1e:   b9 2f           mov     r27, r25
  20:   99 27           eor     r25, r25
  22:   88 27           eor     r24, r24
  24:   28 2b           or      r18, r24
  26:   39 2b           or      r19, r25
  28:   4a 2b           or      r20, r26
  2a:   5b 2b           or      r21, r27
  2c:   84 81           ldd     r24, Z+4        ; 0x04
  2e:   90 e0           ldi     r25, 0x00       ; 0
  30:   a0 e0           ldi     r26, 0x00       ; 0
  32:   b0 e0           ldi     r27, 0x00       ; 0
  34:   28 2b           or      r18, r24
  36:   39 2b           or      r19, r25
  38:   4a 2b           or      r20, r26
  3a:   5b 2b           or      r21, r27
  3c:   83 81           ldd     r24, Z+3        ; 0x03
  3e:   90 e0           ldi     r25, 0x00       ; 0
  40:   a0 e0           ldi     r26, 0x00       ; 0
  42:   b0 e0           ldi     r27, 0x00       ; 0
  44:   ba 2f           mov     r27, r26
  46:   a9 2f           mov     r26, r25
  48:   98 2f           mov     r25, r24
  4a:   88 27           eor     r24, r24
  4c:   28 2b           or      r18, r24
  4e:   39 2b           or      r19, r25
  50:   4a 2b           or      r20, r26
  52:   5b 2b           or      r21, r27
  54:   62 2f           mov     r22, r18
  56:   73 2f           mov     r23, r19
  58:   84 2f           mov     r24, r20
  5a:   95 2f           mov     r25, r21
  5c:   08 95           ret
Comment 4 Richard Biener 2009-04-24 20:39:18 UTC
It's a target independent issue.  On non-strict alignment targets we can do a
32bit load instead of 4 byte loads.  This is what you ask for, correct?

combine unfortunately does not see enough insns to catch it.  Within a single
stmt we can teach fold to do it, otherwise forwprop is our tree level combiner.
Comment 5 Richard Biener 2009-04-24 20:43:04 UTC
Maybe it fits within the byteswap recognition pass (certainly the loads may
be because on-the-fly byteswap is requested).
Comment 6 Georg-Johann Lay 2011-05-16 14:20:25 UTC
Author: gjl
Date: Mon May 16 14:20:19 2011
New Revision: 173792

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=173792
Log:
	PR target/27663
	PR target/41076
	* config/avr/predicates.md (const_8_16_24_operand): New predicate.
	* config/avr/avr.md ("*ior<mode>qi.byte0",
	"*ior<mode>qi.byte1-3"): New define_insn_and_split patterns.


Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/config/avr/avr.md
    trunk/gcc/config/avr/predicates.md
Comment 7 Georg-Johann Lay 2011-05-16 15:05:41 UTC
The patch tries to fix the middle-end flaw in the BE by introducing some combine patterns that recognize byte-insert.

Wouldn't it be possible to recognize such situations in the middle-end and map them to something like (set (zero_extract:QI (reg:SI) ...)) or (set (subreg:QI (reg:SI) ...)?

Even if the bytes inserted do not come from consecutive memory locations, such a recognition would help.

The patch does not lead to optimal code, there is still room for improvement:

With -Os -mmcu=atmega8:

f:
	push r16
	push r17
/* prologue: function */
/* frame size = 0 */
/* stack size = 2 */
.L__stack_usage = 2
	movw r30,r24
	ldd r24,Z+1
	ldd r16,Z+2
	ldi r17,lo8(0)
	ldi r18,lo8(0)
	ldi r19,hi8(0)
	movw r18,r16
	clr r17
	clr r16
	or r19,r24
	ldd r24,Z+4
	or r16,r24
	ldd r24,Z+3
	or r17,r24
	movw r22,r16
	movw r24,r18
/* epilogue start */
	pop r17
	pop r16
	ret

The usage of r16/r17 might be an artifact of IRA because only half of a SI reg is call-saved, the other half is call-used. There is the following comment in ira-color.c:

	/* We need to save/restore the hard register in
	   epilogue/prologue.  Therefore we increase the cost.  */
	{
	  /* ??? If only part is call clobbered.  */

Despite subreg lowering, the call-used r26/r27 are not used.

Maybe you should also try to disable subreg lowering by means of -fno-split-wide-types. For the code in question that gives:

With -Os -mmcu=atmega8 -fno-split-wide-types:

f:
/* prologue: function */
/* frame size = 0 */
/* stack size = 0 */
.L__stack_usage = 0
	movw r30,r24
	ldd r18,Z+1
	ldd r22,Z+2
	mov r24,r22
	ldi r25,lo8(0)
	ldi r26,lo8(0)
	ldi r27,hi8(0)
	clr r23
	clr r22
	or r25,r18
	ldd r18,Z+4
	or r22,r18
	ldd r18,Z+3
	or r23,r18
/* epilogue start */
	ret

What I do not understand are the insns clearing r26/r27 because they are dead (which is not detected). It is an HI insn that looks like that:

; (insn 32 34 42 (set (reg:HI 26 r26 [ MEM[(unsigned char *)P_1(D) + 2B]+2 ])
;         (const_int 0 [0])) insert-byte.c:5 10 {*movhi}
;      (nil))
	ldi r26,lo8(0)	 ;  32	*movhi/1	[length = 2]
	ldi r27,hi8(0)
Comment 8 Jackie Rosen 2014-02-16 13:16:19 UTC Comment hidden (spam)