[WinAVR-users] [ winavr-Bugs-2902772 ] Optimisation regression from 4.2 to 4.3

[SourceForge.net <no...@so...>]

Bugs item #2902772, was opened at 2009-11-23 23:11
Message generated for change (Comment added) made by wutje
You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=520074&aid=2902772&group_id=68108

Please note that this message will contain a full copy of the comment thread,
including the initial issue submission, for this request,
not just the latest update.
Category: GCC
Group: None
Status: Open
Resolution: None
Priority: 5
Private: No
Submitted By: David Brown (davidbrown)
Assigned to: Nobody/Anonymous (nobody)
Summary: Optimisation regression from 4.2 to 4.3

Initial Comment:
Functions that do arithmetic on an expression, then return a 16-bit value, use an extra register pair r18:r19 for the arithmetic instead of using r24:r25 directly.  This leads to extra register moves, and limits other optimisations (such as disallowing the use of adiw and similar instructions).  This is a regression since gcc 4.2.2 in earlier WinAVR releases, and applies to gcc 4.3.x in various newer WinAVR releases.  Compile options are -mmcu=atmega128 -Os, though it also applies to higher optimisations.

extern int ext3 (void);
int foo6(void) {
	return ext3() + 1;
}

avr-gcc 4.2 generates optimal code:
foo6:
	call ext3
	adiw r24,1
	ret

avrgcc 4.3.2 generates much worse code:
foo6:
	call ext3
	movw r18,r24
	subi r18,lo8(-(1))
	sbci r19,hi8(-(1))
	movw r24,r18
	ret

Many other similar functions exhibit the same behaviour, but the function above is perhaps the simplest example.


----------------------------------------------------------------------

Comment By: Wouter (wutje)
Date: 2009-11-24 12:57

Message:
This problem is solved when using -fno-split-wide-types

----------------------------------------------------------------------

You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=520074&aid=2902772&group_id=68108