lgm@ODDJOB.UCHICAGO.EDU (11/26/88)
SYNOPSIS:
The '-fstrength-reduce' flag can produce incorrect code when
it attempts to reduce the strength of a nontrivial (i.e., function-
requiring) multiplication more than once within the same loop.
The example below shows incorrect assembly language generated by
'-O -fstrength-reduce'; '-O' alone generates correct code.
The example is simplified from "cccp.c".
MACHINE AND OS:
AT&T UNIX PC 3B1 (based on Motorola 68010 microprocessor) running
OS Version 3.5 (compatible with UNIX System V Release 2).
CONFIGURATION FILES:
(usual for the UNIX PC)
"tm-3b1.h"
"xm-3b1.h" <== By the way, the reference to "config-m68k.h"
in this file must be changed to "xm-m68k.h".
"m68k.md"
"output-m68k.c"
TRANSCRIPT (INPUT FILE, COMMAND LINE, OUTPUT):
_______________________________________
$ cat badstr.c
struct S { /* structure is 14 bytes long */
int dummy1;
short dummy2;
char *field1;
char *field2;
};
extern void g();
void
f( nels, arr )
int nels;
struct S *arr;
{
int i;
for ( i = 0; i < nels; ++i ) {
if ( arr[ i ].field1 )
g( arr[ i ].field1 );
if ( arr[ i ].field2 )
g( arr[ i ].field2 );
}
}
$ gcc -v -S -O -fstrength-reduce badstr.c
gcc version 1.31
/usr/local/lib/gcc-cpp -v -undef -D__GNU__ -D__GNUC__ -Dmc68k -Dunix -Dunixpc -D__OPTIMIZE__ badstr.c /tmp/cca03041.cpp
GNU CPP version 1.31
/usr/local/lib/gcc-cc1 /tmp/cca03041.cpp -quiet -dumpbase badstr.c -fstrength-reduce -O -version -o badstr.s
GNU C version 1.31 (68k, SGS/AT&T unixpc syntax) compiled by GNU C version 1.31.
$ cat badstr.s
file "badstr.c"
text
even
global f
f:
link.w %a6,&0
movm.l &0x3f20,-(%sp)
mov.l 8(%fp),%d4
mov.l 12(%fp),%a2
clr.l %d2
cmp.l %d4,%d2
ble.w L%8
clr.l %d3
L%7:
tst.l 6(%a2,%d3.l) # %d3 is correct here
beq.w L%5
mov.l 6(%a2,%d5.l),-(%sp) # %d5 is garbage here
jsr g
addq.w &4,%sp
L%5:
tst.l 10(%a2,%d6.l) # %d6 is garbage here
beq.w L%4
mov.l 10(%a2,%d7.l),-(%sp) # %d7 is garbage here
jsr g
addq.w &4,%sp
L%4:
mov.l &14,%d0
add.l %d0,%d3 # %d3 is correctly incremented by 14
addq.l &1,%d2
cmp.l %d4,%d2
bgt.w L%7
L%8:
movm.l -28(%a6),&0x4fc
unlk %a6
rts
$
_______________________________________
EXPLANATION OF TRANSCRIPT:
The generated assembly language above *uses* %d5 , %d6 , and
%d7 without ever assigning them meaningful values. The correct
register to use in all three places would have been %d3 .
Lawrence G. Mayka
Aurora, Illinois
chinet!lmayk!lgmjames@bigtex.cactus.org (James Van Artsdalen) (11/27/88)
> Lawrence G. Mayka In <8811260051.aa00576@gamma.eecs.nwu.edu>, Lawrence G. Mayka (lmayk!lgm@ODDJOB.UCHICAGO.EDU) submitted a bug report about -fstrength-reduce. I tried the example on my SysVr3 gcc 1.31 and did not see the failure mode. But I did see some strange, sub-optimal code: This: --- struct S { /* structure is 14 bytes long */ int dummy1; short dummy2; char *field1; char *field2; }; extern void g(); void f( nels, arr ) int nels; struct S *arr; { int i; for ( i = 0; i < nels; ++i ) { if ( arr[ i ].field1 ) g( arr[ i ].field1 ); if ( arr[ i ].field2 ) g( arr[ i ].field2 ); } } --- gives (via gcc -O -fstrength-reduce -S) this: --- .file "strength1.c" .text .align 4 .globl f f: pushl %ebp movl %esp,%ebp subl $4,%esp ; we don't need this temp. space pushl %ebx pushl %esi pushl %edi xorl %ebx,%ebx ; i cmpl 8(%ebp),%ebx ; cmp nels ,i jge .L8 xorl %edi,%edi ; expected nifty array-index follower .L7: movl 12(%ebp),%esi ; arr cmpl $0,8(%edi,%esi) ; why not copy this to ax for push? je .L5 pushl 8(%edi,%esi) call g leal 4(%esp),%esp .L5: movl %edi,-4(%ebp) ; huh? -4(%ebp) is never read! movl 12(%ebp),%esi ; arr cmpl $0,12(%edi,%esi) je .L4 pushl 12(%edi,%esi) call g leal 4(%esp),%esp .L4: leal 16(%edi),%edi ; "add" is slightly faster than lea. incl %ebx ; i cmpl 8(%ebp),%ebx ; cmp nels, i jl .L7 .L8: leal -16(%ebp),%esp popl %edi popl %esi popl %ebx leave ret -- James R. Van Artsdalen james@bigtex.cactus.org "Live Free or Die" Home: 512-346-2444 Work: 338-8789 9505 Arboretum Blvd Austin TX 78759