Andy Hutchinson wrote:
I have been testing GCC 4.0 for avr wrt loop optimisations. I need to
compare results with earlier release to see if this is new or old
limitation.
Attached is short makefile and source. Only important thing is to keep
optimisation options the same.
Can somebody help me out here? Ive got way too many versions of gcc
live right now to deal with more. The comments explain the effects in
4.0.
3.3.1 or 3.3.5 plus 3.4.3 would be great.
Attached are the results for GCC 3.4.3. Hot off the press. ;-)
HTH
Eric
------------------------------------------------------------------------
.file "testcase.c"
.arch atmega169
__SREG__ = 0x3f
__SP_H__ = 0x3e
__SP_L__ = 0x3d
__tmp_reg__ = 0
__zero_reg__ = 1
.global __do_copy_data
.global __do_clear_bss
.text
.global testloop1
.type testloop1, @function
testloop1:
/* prologue: frame size=0 */
push r28
push r29
/* prologue end (size=2) */
ldi r28,lo8(9)
ldi r29,hi8(9)
.L5:
call foo
sbiw r28,1
sbrs r29,7
rjmp .L5
/* epilogue: frame size=0 */
pop r29
pop r28
ret
/* epilogue end (size=3) */
/* function testloop1 size 12 (7) */
.size testloop1, .-testloop1
.global testloop2
.type testloop2, @function
testloop2:
/* prologue: frame size=0 */
push r28
push r29
/* prologue end (size=2) */
ldi r28,lo8(0)
ldi r29,hi8(0)
.L20:
lds r24,value
tst r24
breq .L20
call foo
adiw r28,1
cpi r28,10
cpc r29,__zero_reg__
brlt .L20
/* epilogue: frame size=0 */
pop r29
pop r28
ret
/* epilogue end (size=3) */
/* function testloop2 size 17 (12) */
.size testloop2, .-testloop2
.global testloop3
.type testloop3, @function
testloop3:
/* prologue: frame size=0 */
push r28
push r29
/* prologue end (size=2) */
ldi r28,lo8(9)
ldi r29,hi8(9)
.L33:
lds r24,value
tst r24
brne .L32
call foo
rjmp .L33
.L32:
sbiw r28,1
sbrs r29,7
rjmp .L33
/* epilogue: frame size=0 */
pop r29
pop r28
ret
/* epilogue end (size=3) */
/* function testloop3 size 17 (12) */
.size testloop3, .-testloop3
.global testloop4
.type testloop4, @function
testloop4:
/* prologue: frame size=0 */
push r28
push r29
/* prologue end (size=2) */
ldi r28,lo8(0)
ldi r29,hi8(0)
.L47:
call foo
lds r24,value
tst r24
breq .L47
adiw r28,1
cpi r28,10
cpc r29,__zero_reg__
brlt .L47
/* epilogue: frame size=0 */
pop r29
pop r28
ret
/* epilogue end (size=3) */
/* function testloop4 size 17 (12) */
.size testloop4, .-testloop4
.global testloop5
.type testloop5, @function
testloop5:
/* prologue: frame size=0 */
push r28
push r29
/* prologue end (size=2) */
ldi r28,lo8(9)
ldi r29,hi8(9)
.L53:
lds r24,value
tst r24
brne .L51
call foo
.L51:
sbiw r28,1
sbrs r29,7
rjmp .L53
/* epilogue: frame size=0 */
pop r29
pop r28
ret
/* epilogue end (size=3) */
/* function testloop5 size 16 (11) */
.size testloop5, .-testloop5
.comm value,1,1
/* File "testcase.c": code 79 = 0x004f ( 54), prologues 10, epilogues 15 */
------------------------------------------------------------------------
OPT = s
MCU = atmega169
CC = m:/WinAVR/bin/avr-gcc
CFLAGS = -mmcu=$(MCU) -W -Wall -v -funsigned-char -funsigned-bitfields
-fpack-struct -fshort-enums -O$(OPT)
.c.s:
$(CC) $(CFLAGS) -c -S $<
all: testcase.s
clean:
rm -f *.o
rm -f testcase.s
------------------------------------------------------------------------
extern int foo(void);
volatile char value;
/* Test ablity of gcc to optimise simple loops
* Expectations - all loops should reduce to simple decrement, test and branch
* Assembler code being
*
* Looptop:
* <body>
* sbiw Rx,1;
* brlt looptop (or similar)
*/
/* TESTCASE 1 the simple loop.
*
* loop reversed (i=10....1) - OK
*/
void testloop1(void)
{
int i;
/*a simple loop thats ok*/
for (i=0;i<10;i++)
{
foo();
}
}
/*TESTCASE 2 - add inner loop - directed at itself (neither forward nor
backward)
*
* doloop - loop reversed (i=9....0) OK
* decrement uses *addhi pattern with -1 as subtract - OK
* uses sbrs - not OK!
* Intention was for peephole (L 2347 avr.md) to combine subtract, test and
branch (3 insn)
* by finding compare NE 65535 after subtract. This pattern no longer applies.
* We now have GE 0 - which no longer matches!
* RTL ends up with "*sbrx_and_branchhi" pattern as equivalent x & 32768 sign
test (2 insn)
* Hack - add ANOTHER peephole to use GE 0
* Better maybe - define decrement,test and jump insn and let combiner figure it
out.
* Also - why is gcc inconsistent in loop reversal bounds????
* Need more info to design a robust fix!!
*/
void testloop2(void)
{
int i;
for (i=0;i<10;i++)
{
while (!value);
foo();
}
}
/* TESTCASE 3 add inner loop - forward jump 'while'
*
* doloop not reversed - not OK
* Cause is forward jump inside loop setting maybe_multiple flag that stops
* optimisation. Very poor the loop counter is never used, and there is no way
it can
* get set mutiple times regardless of jumps inside loop. Need to file bug report
*
* Loop increment using *addhi pattern OK
* Test and jump using compare OK
*/
void testloop3(void)
{
int i;
for (i=0;i<10;i++)
{
while (!value)
{
foo();
}
}
}
/* TESTCASE 4
*
* Same as above but with inner while loop
*
* Loop reversed (i=9.....0) - OK
* Same code pattern matching probelm as testcase 2
*/
void testloop4(void)
{
int i;
for (i=0;i<10;i++)
{
do
{
foo();
} while (!value);
}
}
/* TESTCASE 4
*
* Same as above but no inner loop - just 'if'
*
* Loop reversed - (i=10.....1) - OK
* Again why is gcc using 10...1 as opposed to 9....0.
*/
void testloop5(void)
{
int i;
for (i=0;i<10;i++)
{
if (!value)
{
foo();
}
}
}
------------------------------------------------------------------------
m:/WinAVR/bin/avr-gcc -mmcu=atmega169 -W -Wall -v -funsigned-char
-funsigned-bitfields -fpack-struct -fshort-enums -Os -c -S testcase.c
Reading specs from m:/WinAVR/bin/../lib/gcc/avr/3.4.3/specs
Configured with: ../gcc-3.4.3/configure --prefix=m:/WinAVR --build=mingw32
--host=mingw32 --target=avr --enable-languages=c,c++ --with-dwarf2
Thread model: single
gcc version 3.4.3
m:/WinAVR/bin/../libexec/gcc/avr/3.4.3/cc1.exe -quiet -v -iprefix
m:\WinAVR\bin\../lib/gcc/avr/3.4.3/ testcase.c -quiet -dumpbase testcase.c
-mmcu=atmega169 -auxbase testcase -Os -W -Wall -version -funsigned-char
-funsigned-bitfields -fpack-struct -fshort-enums -o testcase.s
ignoring nonexistent directory
"m:/WinAVR/bin/../lib/gcc/avr/3.4.3/../../../../avr/sys-include"
ignoring nonexistent directory
"m:/WinAVR/lib/gcc/avr/3.4.3/../../../../avr/sys-include"
#include "..." search starts here:
#include <...> search starts here:
m:/WinAVR/bin/../lib/gcc/avr/3.4.3/include
m:/WinAVR/bin/../lib/gcc/avr/3.4.3/../../../../avr/include
m:/WinAVR/lib/gcc/avr/3.4.3/include
m:/WinAVR/lib/gcc/avr/3.4.3/../../../../avr/include
End of search list.
GNU C version 3.4.3 (avr)
compiled by GNU C version 3.3.3 (cygwin special).
GGC heuristics: --param ggc-min-expand=46 --param ggc-min-heapsize=31613