Date: Tue, 27 Apr 1999 20:41:32 +0100
From: cmatraki AT ee DOT ucl DOT ac DOT uk (Chris Matrakidis)
Message-Id: <199904271941.UAA10783@melchard.ee.ucl.ac.uk>
To: djgpp-workers AT delorie DOT com
Subject: Re: [dma AT hpesdma DOT fc DOT hp DOT com: Performance Observation]
X-Sun-Charset: US-ASCII
Reply-To: djgpp-workers AT delorie DOT com

> > The performance drop seems to be an alignment problem.
> 
> Details, please.  What are the differences, as far as alignment is 
> concerned, between 2.7.2 and 2.8.x?
> 

There are no significant differences between the code generated by the two
versions.

The different can be seen in alignment of the global arrays in the .bss section
of the executable.

gcc 2.7.2:
0000be78 B _src_i
0000de78 B _co
0000fe78 B _rev
00010e78 B _si
00012e78 B _a_real
00014e78 B _a_imag
00016e78 B _src_r

gcc 2.8.1:
0000d694 B _src_i
0000f694 B _co
00011694 B _rev
00012694 B _si
00014694 B _a_real
00016694 B _a_imag
00018694 B _src_r

It just happens that a new global variable is defined in libgcc.a for version
2.8.1, which moves those arrays out of doubleword alignment, which seriously
affects the performance. (doubles should be doubleword aligned for better 
performance).

I thought that one way around the problem was to use -fno-common -malign-double
which should put those arrays in the .data section properly aligned. However,
no even that works in all cases:

gcc 2.8.1:
0000b6d0 D _co
0000d6d0 D _si
0000f6d0 D _rev
000106d0 D _a_real
000126d0 D _a_imag
000146d0 D _src_r
000166d0 D _src_i

gcc 2.7.2:
0000a4cc D _co
0000c4cc D _si
0000e4cc D _rev
0000f4cc D _a_real
000114cc D _a_imag
000134cc D _src_r
000154cc D _src_i

Even though the object file has the arrays correctly aligned, ld has AFAIK no
way to identify what the required alignment is. I can't think of any easy way
to overcome this.