[PATCH, alpha]: Fix libgomp.c/autopar-1.c execution test failure

[Uros Bizjak <ubizjak at gmail dot com>]

Hello!

This one was surprisingly hard and the fact that alpha thread debugging 
is currently FUBAR on Ubuntu 2.6.30.5 was not helpful at all :(

At the end, the problem was found in LL/SC sequence that implements 
sync_compare_and_swap functionality. Looking at the Alpha Archicecture 
Handbook [1], Notes for Section 4.2.4, LL/SC link will always fail for 
various things, such as FP insn, memory access or branch in between LL 
and SC.

So, it happens that various optimization passes after split-after-reload 
mangled LL/SC sequence into:

?> 120000c7c:   00 40 00 60     mb
   120000c80:   0b 14 4b 59     addt    $f10,$f11,$f11
   120000c84:   02 0e 7f 71     ftoit   $f11,t1
-> 120000c88:   00 00 70 ac     ldq_l   t2,0(a0)
   120000c8c:   a4 05 61 40     cmpeq   t2,t0,t3
*>   120000c90:   06 00 80 f4     bne     t3,120000cac <foo.$loopfn.0+0xfc>
   120000c94:   a1 05 61 40     cmpeq   t2,t0,t0
   120000c98:   02 04 e3 47     mov     t2,t1
   120000c9c:   08 00 20 e4     beq     t0,120000cc0 <foo.$loopfn.0+0x110>
   120000ca0:   00 00 5e a7     ldq     ra,0(sp)
   120000ca4:   20 00 de 23     lda     sp,32(sp)
   120000ca8:   01 80 fa 6b     ret
   120000cac:   04 04 e2 47     mov     t1,t3
-> 120000cb0:   00 00 90 bc     stq_c   t3,0(a0)
   120000cb4:   f4 ff 9f e4     beq     t3,120000c88 <foo.$loopfn.0+0xd8>
   120000cb8:   00 40 00 60     mb

And yeah... the branch at 0x120000c94 to 0x120000cac in LL/SC sequence 
broke the link, since it jumped to SC when compare-and-swap arguments 
were equal.

Patched gcc creates following sequence:

?   mb
    addt $f10,$f11,$f11
    ftoit $f11,$2
$L12:
>   ldq_l $3,0($16)
    cmpeq $3,$1,$4
    beq $4,$L13
    mov $2,$4
>   stq_c $4,0($16)
    beq $4,$L12
    mb

So, there is no _TAKEN_ branch inside SC/LL sequence. This sequence 
relases lock as expected (and is AFAICS the same as 64bit cmpxchg 
sequence in linux-2.6/arch/alpha/include/asm/xchg.h, modulo wrongly 
predicted backward branch at the end of our sequence.).

The actual gcc problem is in bbro pass that for some reason takes hot BB 
out of the main sequence. The proposed solution is to limit compiler 
creativity by delaying insn splitting after the bbro pass, 
conditionalizing split on epilogue_completed instead of 
reload_completed. There is no problem with follow-up scheduling pass, 
since both LL and SC insns are unspec_volatiles, effectively a 
scheduling barriers.

Attached patch uses the same approach to all other sync LL/SC sequences, 
so generated sequences will more or less stay in the original expanded form.

Yes, the comment in sync.md was wrong. Taken branches clear lock, at 
least on

cpu            : Alpha
cpu model        : EV68AL
cpu variation        : 7
cpu revision        : 0

BTW: Linux does not need memory barrier in front of the LL/SC sequence, 
I will propose a follow-up patch that removes it.

2009-08-25  Uros Bizjak <ubizjak@gmail.com>

    * config/alpha/sync.md: Update comment about unpredictable LL/SC lock
    clearing by a taken branch.
    (sync_<fetchop_name><mode>): Split when epilogue_completed is set,
    effectively after bbro pass.
    (sync_nand<mode>): Ditto.
    (sync_old_<fetchop_name><mode>): Ditto.
    (sync_old_nand<mode>): Ditto.
    (sync_new_<fetchop_name><mode>): Dito.
    (sync_new_nand<mode>): Ditto.
    (sync_compare_and_swap<mode>_1): Ditto.
    (*sync_compare_and_swap<mode>): Ditto.
    (sync_lock_test_and_set<mode>_1): Ditto.
    ("sync_lock_test_and_set<mode>): Ditto.

The patch was tested on 4-processor alphaev68-unknown-linux-gnu. Patch 
fixes:

WARNING: program timed out.
FAIL: libgomp.c/autopar-1.c execution test

in libgomp testsuite.

OK for mainline and after a week or two for 4.4 and 4.3?

[1]http://www.comms.scitech.susx.ac.uk/fft/programming/alphaahb.pdf

Uros.

Attachment: a.diff.txt
Description: Text document