target/6039: ARM interrupt attribute generats wrong code

[tori@unhappy.mine.nu]

>Number:         6039
>Category:       target
>Synopsis:       ARM interrupt attribute generats wrong code
>Confidential:   no
>Severity:       critical
>Priority:       high
>Responsible:    unassigned
>State:          open
>Class:          wrong-code
>Submitter-Id:   net
>Arrival-Date:   Fri Mar 22 07:56:00 PST 2002
>Closed-Date:
>Last-Modified:
>Originator:     Tobias Ringstrom
>Release:        gcc-3.0.4
>Organization:
>Environment:
Crosscompiler for arm-elf (cygwin and Linux host)
>Description:
Compile the following code with arm-elf-gcc -S tmp.c:

void foo(void) __attribute__((__interrupt__));
void bar(void);
void foo()
{
        int a=0, b=0, c=a+b;
        bar();
}

gives you the following:

@ Generated by gcc 3.0.4 for ARM/elf
        .file   "tmp.c"
        .text
        .align  2
        .global foo
        .type   foo,function
foo:
        @ Interrupt Service Routine.
        @ args = 0, pretend = 0, frame = 12
        @ frame_needed = 1, current_function_anonymous_args = 0
        mov     ip, sp
        stmfd   sp!, {r2, r3, fp, ip, lr, pc}
        sub     fp, ip, #4
        sub     sp, sp, #12
        mov     r3, #0
        str     r3, [fp, #-16]
        str     r3, [fp, #-20]
        ldr     r2, [fp, #-16]
        ldr     r3, [fp, #-20]
        add     r3, r2, r3
        str     r3, [fp, #-24]
        bl      bar
        ldmea   fp, {r2, r3, fp, sp, lr}
        subs    pc, lr, #4
.Lfe1:
        .size   foo,.Lfe1-foo

Major problem 1:
The ip register is overwritten and not restored.

Major problem 2:
The local variables overlap the pushed registers.  In this case variable a occupies the same space as the stored lr.

Major problem 3:
A function is called which has every right to alter r0 and r1, and still those registers are not saved/restored in the prologue/epilogue.

Question:
Why is the register pc stored?

If this bug is not fixed for the next gcc-3.0.x, I strongly recommend that you disable the interrupt attribute since it is unusable in its current form.

I have tried the later pre-3.1 CVS version as well, and the above bugs seem to be fixed there, but unfortunately there is another (actually much worse) showstopper there.

I will generate ASM stubs for my ISRs for now, but it would be very nice to use the interrupt attribute in the future.

/Tobias
>How-To-Repeat:
Compile the program in the Description above.  Note that no optimization is neccessary.  It may even hide the real problems.
>Fix:

>Release-Note:
>Audit-Trail:
>Unformatted: