[Bug c/16562] New: [ H8300 Target] Wrong code with -O2, -O3 optimizations

[Bug c/16562] New: [ H8300 Target] Wrong code with -O2, -O3 optimizations

[asgarij at kpitcummins dot com]

Following test case generates wrong code with optimization option -O2 or O3.
Details are as under.

h8300-elf-gcc -v
----------------
Configured with: /home/kpit/fsfsrc/downloads/gcc-3.5-20040711/configure --
enable-languages=c,c++ --target=h8300-elf --with-newlib --
prefix=/home/nitins2/intmdt_share/gnuh8300_v0403_elf-1
Thread model: single
gcc version 3.5-GNUH8_v0403

----------------------------------------------
Commandline: h8300-elf-gcc test.c -ms -m2600s -S -O2
//test.c
void theLoop( void )
{
unsigned long int cnt;
cnt = 0L;
do
cnt++;
while (cnt < 1000000L);
}
//test.c

//test.s
	.file	"test.c"
	.h8300s
	.section .text
	.align 2
	.align 1
	.global _theLoop
_theLoop:
	mov.l	er6,@-er7
	mov.l	er7,er6
	mov.l	#999999,er2
.L2:
	sub.l	#50,er2 !This should be 1
	mov.l	er2,er2
	bge	.L2
	mov.l	@er7+,er6
	rts
	.size	_theLoop, .-_theLoop
	.end
//test.s

Additional Info: This bug comes with target options -mh,-ms and -ms2600 only.

-- 
           Summary: [ H8300 Target] Wrong code with -O2, -O3 optimizations
           Product: gcc
           Version: 3.5.0
            Status: UNCONFIRMED
          Severity: critical
          Priority: P2
         Component: c
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: asgarij at kpitcummins dot com
                CC: asgarij at kpitcummins dot com,gcc-bugs at gcc dot gnu
                    dot org
 GCC build triplet: i686-pc-linux-gnu
  GCC host triplet: i686-pc-linux-gnu
GCC target triplet: h8300-unknown-elf


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16562

[Bug target/16562] [ H8300 Target] Wrong code with -O2, -O3 optimizations

[pinskia at gcc dot gnu dot org]

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|critical                    |normal
          Component|c                           |target
           Keywords|                            |wrong-code


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16562

[Bug target/16562] [ H8300 Target] Wrong code with -O2, -O3 optimizations

[anilp1 at kpitcummins dot com]

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |anilp1 at kpitcummins dot
                   |                            |com


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16562

[Bug target/16562] [ H8300 Target] Wrong code with -O2, -O3 optimizations

[anilp1 at kpitcummins dot com]

------- Additional Comments From anilp1 at kpitcummins dot com  2004-08-04 05:20 -------
Hi,

The workaround to avoid bug is to pass either of following option along 
with   -O2 ,
	-fno-loop-optimize	
	-fno-strength-reduce

This bug is not target dependent as was first impression. 
Since counting variable is of type unsigned long, the number of instructions 
in loop for plain H8/300 are 7. So problem doesn't crop up for plain H8/300. 
If we change data type of counting variable to unsigned int (16 bits) and 
value also within range of 16 bits ( less than 65535) then bug can be seen in 
case of plain H8/300. 

Bug is not observed in following program code,
//test.c
void theLoop( void )
{
unsigned long int cnt;
cnt = 0L;
do{
	cnt++;
	asm("nop");
} while (cnt < 1000000L);
}
//test.c
Compiled with -S -O2 -ms options.
The number of instructions in loop becomes 4 in above case. Hence bug is not 
observed.

I tracked the problem to loop.c, line no. 5358 (gcc 3.5 snapshot dated 25 July 
2004).
Following call doesn't take place for insn_count more than 4 
since "unrolled_insn_copies <= insn_count" this condition is dissatisfied. 
		unroll_loop (loop, insn_count, 1);

In case of H8S target and insn_count 3, call takes place and the magic number 
50 appears instead of 1.
This magic number is also dependent on number of iterations of loop. 

Will it be ok to add additional condition as insn_count > 4 while calling 
unroll_loop at line 5358 ?

Please comment.

Regards,
Anil Paranjpe
KPIT Cummins InfoSystems Ltd.
Pune, India

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Free download of GNU based tool-chains for Renesas' SH and H8 Series.
The following site also offers free technical support to its users. 
Visit http://www.kpitgnutools.com for details. 
Latest versions of KPIT GNU tools were released on June 1, 2004.
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16562

[Bug target/16562] [ H8300 Target] Wrong code with -O2, -O3 optimizations

[kazu at cs dot umass dot edu]

------- Additional Comments From kazu at cs dot umass dot edu  2004-08-16 07:07 -------
I doubt there is anything in the C language specification that
prohibits the compiler from removing or shortening these empty loops.

To guarantee the presevation of the loop iterations, you need to use
asm ("nop") as you have demonstrated already.


-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|                            |INVALID


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16562