[Bug target/11052] [3.3 regression] [arm] ICE (segfault) compiling xfree86

[[EMAIL PROTECTED]]

PLEASE REPLY TO [EMAIL PROTECTED] ONLY, *NOT* [EMAIL PROTECTED]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=11052



------- Additional Comments From [EMAIL PROTECTED]  2003-06-02 17:09 -------
Subject: Re:  [3.3 regression] [arm] ICE (segfault) 
 compiling xfree86

Bizzare.  Your analysis does not match mine at all (though I am getting a 
segfault).

Firstly, I get this on arm-elf as well as on arm-linux (but not if I have 
-mcpu=strongarm on the options list).

What I am seeing happening is that we have the construct

    int quarks[100 + 1];
    register int num_quarks;

        num_quarks = 0;

        for(;;) {
                        quarks[num_quarks++] = something;
        }

And this is being transformed into (roughly):


    int quarks[100 + 1];
    int *q_ptr = quarks;

    for (;;) {
        *q_ptr++ = something;
    }

Or, in rtl:

(insn 193 23 30 0 (nil) (set (reg/v/f:SI 39)
        (reg/f:SI 73)) 175 {*arm_movsi_insn} (insn_list 23 (nil))
    (expr_list:REG_EQUAL (plus:SI (reg/f:SI 25 sfp)
            (const_int -9000 [0xffffdcd8]))
        (nil)))

...

(insn 94 162 95 6 0x4001d7bc (set (mem:SI (pre_inc:SI (reg/v/f:SI 39)) [8 
S4 A32
])
        (reg:SI 54)) 175 {*arm_movsi_insn} (nil)
    (expr_list:REG_INC (reg/v/f:SI 39)
        (expr_list:REG_DEAD (reg:SI 54)
            (nil))))

Now during local register allocation that REG_EQUAL note on insn 193 is 
being transformed into a REQ_EQUIV (which implies that 39 never changes).  
So when reload fails to allocate a register to reg 39 it simply substitues 
in the equiv expression.  That leaves us with a post_inc of something that 
isn't a register and has no stack equivalent.

(insn:HI 211 210 103 5 0x401d52ec (set (mem:SI (pre_inc:SI (plus:SI 
(reg/f:SI 11 fp)
                    (const_int -9040 [0xffffdcb0]))) [8 S4 A32])
        (reg:SI 3 r3 [93])) 124 {*arm_movsi_insn} (insn_list 210 (nil))
    (expr_list:REG_DEAD (reg:SI 3 r3 [93])
        (expr_list:REG_INC (plus:SI (reg/f:SI 11 fp)
                (const_int -9040 [0xffffdcb0]))
            (nil))))

We then get confused in a later pass when we try to kill an auto-inc value 
because an auto-inc should never apply to anything other than a register.

I need to look into this further, but it appears that REG_N_SETS hasn't 
taken into account the auto_inc operation ;-(

R.





------- You are receiving this mail because: -------
You reported the bug, or are watching the reporter.