[Top] [All Lists]

Confused assembler

Subject: Confused assembler
From: Peter Fuerst <pf@Indigo2.Peter>
Date: Fri, 31 Dec 2004 02:06:02 +0100 (CET)
In-reply-to: <>
Original-recipient: rfc822;
References: <> <>

Hello !

When building 2.6.10, the assembler ( gets confused by a "b target2"
(compiler generated) immediately following a "beqzl target1" (inline assembly
macro), and reorders these instructions (with wrong address calculation too)
to an infinite loop.
(Well, not really infinite - after about ten minutes some timer interrupt
makes an end to it by a null-pointer dereference in __queue_work()  ;-)
This bug is rather unreliable though, going away with minimal, unintentional
code changes, unexpectedly reappearing after some later rebuild...

Was this behaviour already observed elsewhere ?  Is it fixed in some newer
assembler version ?  Or should i just be content with it and work around
with appropriate "nop"s in the concerned inline-assembly macros ? ... ?

as -EB -G 0 -mips4 -O2 -g0 -64 -mcpu=r8000 -v -64 -non_shared -64 -march=r8000 
-mips4 --trap -o kernel/.tmp_fork.o
GNU assembler version (mips64-ip28-linux-gnu) using BFD version
GNU assembler version (mips64-linux) using BFD version

Here are examples of what the assembler receives from the compiler,
and what it produces:

a80000002003a6c0 <copy_process>:

        .set    macro
        .set    reorder
        cache 0x14,0($sp)       # Cache Barrier
        1:      lld     $3, 256($4)     # atomic64_add
        addu    $3, $2
        scd     $3, 256($4)
        beqzl   $3, 1b
        b       $L5881
         # Cache Barrier omitted.
        move    $5,$0
a80000002003a944:       bfb40000        cache   0x14,0(sp)
a80000002003a948:       d0830100        lld     v1,256(a0)
a80000002003a94c:       00621821        addu    v1,v1,v0
a80000002003a950:       f0830100        scd     v1,256(a0)
a80000002003a954 <!>:   1000ffff        b       a80000002003a954 <$L6939+0xe4>
a80000002003a958:       00000000        nop
a80000002003a95c:       50600000        beqzl   v1,a80000002003a960 <$L5887>
a80000002003a960 <$L5887>:
a80000002003a960:       0000282d        move    a1,zero
a80000002003a964 <$L5881>:

        1:      ll      $5, 0($2)       # atomic_add
        addu    $5, $3
        sc      $5, 0($2)
        beqzl   $5, 1b
        b       $L6092
a80000002003abe4:       c0450000        ll      a1,0(v0)
a80000002003abe8 <!>:   00a32821        addu    a1,a1,v1
a80000002003abec:       e0450000        sc      a1,0(v0)
a80000002003abf0:       1000fffd        b       a80000002003abe8 <$L6045+0x70>
a80000002003abf4:       00000000        nop
a80000002003abf8:       50a00000        beqzl   a1,a80000002003abfc <$L6057>
a80000002003abfc <$L6057>:

        1:      ll      $5, 4($2)       # atomic_add
        addu    $5, $3
        sc      $5, 4($2)
        beqzl   $5, 1b
        b       $L6480
a80000002003aedc:       c0450004        ll      a1,4(v0)
a80000002003aee0:       00a32821        addu    a1,a1,v1
a80000002003aee4 <!>:   e0450004        sc      a1,4(v0)
a80000002003aee8:       1000fffe        b       a80000002003aee4 <$L6401+0x54>
a80000002003aeec:       00000000        nop
a80000002003aef0:       50a00000        beqzl   a1,a80000002003aef4 <$L6411>
a80000002003aef4 <$L6411>:

with kind regards


<Prev in Thread] Current Thread [Next in Thread>