| Message ID | CAFULd4YGx4G0n_mB3vYW8kZVkuDcdLXRQV=poW47MVp_QOxy8g@mail.gmail.com |
|---|---|
| State | New |
| Headers | show |
On Sun, Sep 2, 2012 at 1:22 PM, Uros Bizjak <ubizjak@gmail.com> wrote: > Hello! > > Attached patch prevents compute_bb_for_insn to calculate BB for > barrier RTXes. This is in fact the same approach all other > *_bb_for_insn use. > > The patch is bordering on obvious. It is anything _but_ obvious. The code looks like this: void compute_bb_for_insn (void) { basic_block bb; FOR_EACH_BB (bb) { rtx end = BB_END (bb); rtx insn; for (insn = BB_HEAD (bb); ; insn = NEXT_INSN (insn)) { BLOCK_FOR_INSN (insn) = bb; if (insn == end) break; } } } This could (&should) actually be written as: void compute_bb_for_insn (void) { basic_block bb; rtx insn; FOR_EACH_BB (bb) FOR_BB_INSNS (bb, insn) BLOCK_FOR_INSN (insn) = bb; } What is happening for you, is that you're seeing a BARRIER between BB_HEAD(bb) and BB_END(bb), which is not possible. The barrier is mis-placed. The patch is not OK. Ciao! Steven
On Sun, Sep 2, 2012 at 1:38 PM, Steven Bosscher <stevenb.gcc@gmail.com> wrote: > On Sun, Sep 2, 2012 at 1:22 PM, Uros Bizjak <ubizjak@gmail.com> wrote: >> Hello! >> >> Attached patch prevents compute_bb_for_insn to calculate BB for >> barrier RTXes. This is in fact the same approach all other >> *_bb_for_insn use. > What is happening for you, is that you're seeing a BARRIER between > BB_HEAD(bb) and BB_END(bb), which is not possible. The barrier is > mis-placed. Please see how the dump looks: (insn 12 11 28 3 (set (reg:SI 1 dx [63]) (const_int 2 [0x2])) pr54455.c:5 65 {*movsi_internal} (nil)) (note 28 12 29 3 NOTE_INSN_EPILOGUE_BEG) (insn/f 29 28 30 3 (set (reg/f:DI 6 bp) (mem:DI (post_inc:DI (reg/f:DI 7 sp)) [0 S8 A8])) pr54455.c:13 -1 (expr_list:REG_CFA_DEF_CFA (plus:DI (reg/f:DI 7 sp) (const_int 8 [0x8])) (nil))) (jump_insn 30 29 31 3 (simple_return) pr54455.c:13 -1 (nil) -> simple_return) (barrier 31 30 15) (note 15 31 21 ("lab") NOTE_INSN_DELETED_LABEL 3) (note 21 15 24 NOTE_INSN_DELETED) (note 24 21 0 NOTE_INSN_DELETED) Please note that we have a bunch of notes at the end, and barrier above them. There are no active insns between barrer and BB_END. Where should this RTX go? Uros.
On Sun, Sep 2, 2012 at 1:46 PM, Uros Bizjak <ubizjak@gmail.com> wrote: > On Sun, Sep 2, 2012 at 1:38 PM, Steven Bosscher <stevenb.gcc@gmail.com> wrote: >> On Sun, Sep 2, 2012 at 1:22 PM, Uros Bizjak <ubizjak@gmail.com> wrote: >>> Hello! >>> >>> Attached patch prevents compute_bb_for_insn to calculate BB for >>> barrier RTXes. This is in fact the same approach all other >>> *_bb_for_insn use. > >> What is happening for you, is that you're seeing a BARRIER between >> BB_HEAD(bb) and BB_END(bb), which is not possible. The barrier is >> mis-placed. > > Please see how the dump looks: > > (insn 12 11 28 3 (set (reg:SI 1 dx [63]) > (const_int 2 [0x2])) pr54455.c:5 65 {*movsi_internal} > (nil)) > (note 28 12 29 3 NOTE_INSN_EPILOGUE_BEG) > (insn/f 29 28 30 3 (set (reg/f:DI 6 bp) > (mem:DI (post_inc:DI (reg/f:DI 7 sp)) [0 S8 A8])) pr54455.c:13 -1 > (expr_list:REG_CFA_DEF_CFA (plus:DI (reg/f:DI 7 sp) > (const_int 8 [0x8])) > (nil))) > (jump_insn 30 29 31 3 (simple_return) pr54455.c:13 -1 > (nil) > -> simple_return) > (barrier 31 30 15) > (note 15 31 21 ("lab") NOTE_INSN_DELETED_LABEL 3) > (note 21 15 24 NOTE_INSN_DELETED) > (note 24 21 0 NOTE_INSN_DELETED) > > Please note that we have a bunch of notes at the end, and barrier > above them. There are no active insns between barrer and BB_END. Where > should this RTX go? It doesn't have to be the barrier that has to move. It could also be that the barrier is inserted in the right place but BB_END wasn't updated to point to the jump insn. Ciao! Steven
On Sun, Sep 2, 2012 at 1:49 PM, Steven Bosscher <stevenb.gcc@gmail.com> wrote: >>>> Attached patch prevents compute_bb_for_insn to calculate BB for >>>> barrier RTXes. This is in fact the same approach all other >>>> *_bb_for_insn use. >> >>> What is happening for you, is that you're seeing a BARRIER between >>> BB_HEAD(bb) and BB_END(bb), which is not possible. The barrier is >>> mis-placed. >> >> Please see how the dump looks: >> >> (insn 12 11 28 3 (set (reg:SI 1 dx [63]) >> (const_int 2 [0x2])) pr54455.c:5 65 {*movsi_internal} >> (nil)) >> (note 28 12 29 3 NOTE_INSN_EPILOGUE_BEG) >> (insn/f 29 28 30 3 (set (reg/f:DI 6 bp) >> (mem:DI (post_inc:DI (reg/f:DI 7 sp)) [0 S8 A8])) pr54455.c:13 -1 >> (expr_list:REG_CFA_DEF_CFA (plus:DI (reg/f:DI 7 sp) >> (const_int 8 [0x8])) >> (nil))) >> (jump_insn 30 29 31 3 (simple_return) pr54455.c:13 -1 >> (nil) >> -> simple_return) >> (barrier 31 30 15) >> (note 15 31 21 ("lab") NOTE_INSN_DELETED_LABEL 3) >> (note 21 15 24 NOTE_INSN_DELETED) >> (note 24 21 0 NOTE_INSN_DELETED) >> >> Please note that we have a bunch of notes at the end, and barrier >> above them. There are no active insns between barrer and BB_END. Where >> should this RTX go? > > It doesn't have to be the barrier that has to move. It could also be > that the barrier is inserted in the right place but BB_END wasn't > updated to point to the jump insn. It looks that NOTE_INSN_DELETED_LABEL interferes with barrier. Uros.
Index: cfgrtl.c =================================================================== --- cfgrtl.c (revision 190857) +++ cfgrtl.c (working copy) @@ -415,7 +415,8 @@ compute_bb_for_insn (void) for (insn = BB_HEAD (bb); ; insn = NEXT_INSN (insn)) { - BLOCK_FOR_INSN (insn) = bb; + if (!BARRIER_P (insn)) + BLOCK_FOR_INSN (insn) = bb; if (insn == end) break; } Index: testsuite/gcc.target/i386/pr54455.c =================================================================== --- testsuite/gcc.target/i386/pr54455.c (revision 0) +++ testsuite/gcc.target/i386/pr54455.c (working copy) @@ -0,0 +1,16 @@ +/* { dg-do compile } */ +/* { dg-options "-fschedule-insns -fselective-scheduling --param max-sched-extend-regions-iters=8" } */ + +static inline void +__attribute__ ((always_inline)) +foo (void) +{ + asm goto ("" : : "r" (1), "r" (2) : "memory" : lab); + lab:; +} + +void bar (void) +{ + foo (); + foo (); +}
Hello! Attached patch prevents compute_bb_for_insn to calculate BB for barrier RTXes. This is in fact the same approach all other *_bb_for_insn use. The patch is bordering on obvious. 2012-09-02 Uros Bizjak <ubizjak@gmail.com> PR rtl-optimization/54455 * cfgrtl.c (compute_bb_for_insn): Do not compute BB for barrier RTXes. testsuite/ChangeLog: 2012-09-02 Uros Bizjak <ubizjak@gmail.com> PR rtl-optimization/54455 * gcc.target/i386/pr54455.c: New test. Bootstrapped and regression tested on x86_64-pc-linux-gnu {-m32}. OK for mainline and 4.7? Uros.