An optimization problem, perhaps arm-elf related

[Tobias Ringstrom <tori at ringstrom dot mine dot nu>]

I have a small problem with gcc 3.0.3 for arm-elf.  The C code looks like
this:  (I know its stupid coding.  It's a trimmed down [broken] version,
but it shows the same problem.  Removing the volatile makes no difference, 
by the way.)

	static __inline set(unsigned mask)
	{
		*(volatile unsigned*)0xffff0030 = mask;
	}
	static __inline clear(unsigned mask)
	{
		*(volatile unsigned*)0xffff0034 = mask;
	}
	void fpga_prog(unsigned x)
	{
		if (x == 0) set(1); else clear(2);
		if (x == 1) set(1); else clear(2);
		if (x == 2) set(1); else clear(2);
		if (x == 3) set(1); else clear(2);
	}
	void fpga_prog_loop(unsigned x)
	{
		int i;
		for (i = 0; i < 4; ++i) {
			if (x == i) set(1); else clear(2);
		}
	}

The problem is that the manual loop unrolling of fpga_prog generates much 
worse code than the -funroll-loops version of fpga_prog_loop.  Compiling 
with "arm-elf-gcc -S -O3 -funroll-loops tmp.c", I get the following nice 
code for fpga_prog_loop:
	@ args = 0, pretend = 0, frame = 0
	@ frame_needed = 0, current_function_anonymous_args = 0
	str	lr, [sp, #-4]!
	mvn	r3, #65280
	mov	r1, #1
	mov	r2, #2
	cmp	r0, #0
	streq	r1, [r3, #-207]
	strne	r2, [r3, #-203]
	sub	lr, r3, #207
	sub	ip, r3, #203
	cmp	r0, #1
	streq	r1, [lr, #0]
	strne	r2, [ip, #0]
	cmp	r0, #2
	streq	r1, [lr, #0]
	strne	r2, [ip, #0]
	cmp	r0, #3
	streq	r1, [lr, #0]
	strne	r2, [ip, #0]
	ldr	pc, [sp], #4

For the manually unrolled loop fpga_prog, I get the following much worse 
code:
	@ args = 0, pretend = 0, frame = 0
	@ frame_needed = 0, current_function_anonymous_args = 0
	@ link register save eliminated.
	cmp	r0, #0
	mvneq	r2, #65280
	mvnne	r2, #65280
	subeq	r1, r2, #207
	subne	r1, r2, #203
	moveq	ip, #1
	movne	ip, #2
	streq	ip, [r1, #0]
	strne	ip, [r1, #0]
	cmp	r0, #1
	mvneq	ip, #65280
	mvnne	r2, #65280
	subeq	r1, ip, #207
	subne	r1, r2, #203
	movne	ip, #2
	strne	ip, [r1, #0]
	streq	r0, [r1, #0]
	cmp	r0, #2
	mvneq	r2, #65280
	mvnne	r2, #65280
	subeq	r1, r2, #207
	subne	r1, r2, #203
	moveq	ip, #1
	movne	ip, #2
	streq	ip, [r1, #0]
	strne	ip, [r1, #0]
	cmp	r0, #3
	mvneq	ip, #65280
	mvnne	r2, #65280
	subeq	r0, ip, #207
	moveq	r1, #1
	subne	r0, r2, #203
	movne	r1, #2
	streq	r1, [r0, #0]
	strne	r1, [r0, #0]
	mov	pc, lr

Is there an optimization that I must use that is not included in O3, or is 
it a bug?  Please let me know if you want me to supply more information, 
and please CC me since I'm not on the list.

/Tobias