Crossjumping of tablejumps

[Josef Zlomek]

> >        * cfgcleanup.c (outgoing_edges_match): Compare the jump tables.
> >        (try_crossjump_to_edge): Replace refereces to one jump table by
> >        references to identical jump table.
> 
> This breaks s390, apparently because it doesn't cope correctly with the
> case where the jump table label is stored in the literal pool.

I'll look at it tomorrow.

Josef

> In my test case, this sequence
> 
> ;; Start of basic block 4, registers live: 15 [%r15] 43 57
> (note 185 75 76 4 [bb 4] NOTE_INSN_BASIC_BLOCK)
> 
> (insn 76 185 77 4 0x401ab0c0 (set (reg:SI 3 %r3 [orig:57 kt ] [57])
>         (ashift:SI (reg:SI 3 %r3 [orig:57 kt ] [57])
>             (const_int 2 [0x2]))) 246 {ashlsi3} (nil)
>     (nil))
> 
> (insn 77 76 78 4 0x401ab0c0 (set (reg/f:SI 1 %r1 [58])
>         (mem/u/f:SI (symbol_ref/u:SI ("*.LC2")) [3 S4 A32])) 56 {*movsi} (nil)
>     (insn_list:REG_LABEL 80 (expr_list:REG_EQUIV (label_ref:SI 80)
>             (nil))))
> 
> (insn 78 77 79 4 0x401ab0c0 (set (reg:SI 2 %r2 [59])
>         (mem:SI (plus:SI (reg/f:SI 1 %r1 [58])
>                 (reg:SI 3 %r3 [orig:57 kt ] [57])) [0 S4 A8])) 56 {*movsi} (insn_list 77 (insn_list 76 (nil)))
>     (nil))
> 
> (jump_insn 79 78 80 4 0x401ab0c0 (parallel [
>             (set (pc)
>                 (reg:SI 2 %r2 [59]))
>             (use (label_ref 80))
>         ]) 265 {casesi_jump} (insn_list:REG_DEP_ANTI 76 (insn_list:REG_DEP_ANTI 77 (insn_list 78 (nil))))
>     (nil))
> ;; End of basic block 4, registers live:
>  11 [%r11] 15 [%r15] 32 [%ap] 34 [%fp] 43
> 
> ;; Insn is not within a basic block
> (code_label/s 80 79 81 10 "" [3 uses])
> 
> ;; Insn is not within a basic block
> (jump_insn 81 80 82 (nil) (addr_vec:SI [
>             (label_ref:SI 18)
>             (label_ref:SI 30)
>             (label_ref:SI 40)
>             (label_ref:SI 52)
>         ]) -1 (nil)
>     (nil))
> 
> (barrier 82 81 18)
> 
> 
> is replaced by
> 
> ;; Start of basic block 4, registers live: 3 [%r3] 4 [%r4] 5 [%r5] 15 [%r15]
> (note 185 75 76 4 [bb 4] NOTE_INSN_BASIC_BLOCK)
> 
> (insn 76 185 77 4 0x401ab0c0 (set (reg:SI 3 %r3 [orig:57 kt ] [57])
>         (ashift:SI (reg:SI 3 %r3 [orig:57 kt ] [57])
>             (const_int 2 [0x2]))) 246 {ashlsi3} (nil)
>     (nil))
> 
> (insn 77 76 282 4 0x401ab0c0 (set (reg/f:SI 1 %r1 [58])
>         (mem/u/f:SI (symbol_ref/u:SI ("*.LC2")) [3 S4 A32])) 56 {*movsi} (nil)
>     (insn_list:REG_LABEL 80 (expr_list:REG_EQUIV (label_ref:SI 153)
>             (nil))))
> 
> (jump_insn 282 77 283 4 (nil) (set (pc)
>         (label_ref 281)) -1 (nil)
>     (nil))
> ;; End of basic block 4, registers live:
>  1 [%r1] 3 [%r3] 4 [%r4] 5 [%r5] 15 [%r15]
> 
> (barrier 283 282 80)
> 
> (note/s 80 283 19 "" NOTE_INSN_DELETED_LABEL 10)
> 
> 
> This is broken, because the literal pool entry .LC2 still contains 
> a reference to the now-deleted label 10.  This causes the label to
> be emitted at some random point in the code stream, and the target
> of the redirected jump:
> 
> ;; Start of basic block 31, registers live: 1 [%r1] 3 [%r3] 4 [%r4] 5 [%r5] 15 [%r15]
> (code_label 281 150 279 31 23 "" [1 uses])
> 
> (note 279 281 151 31 [bb 31] NOTE_INSN_BASIC_BLOCK)
> 
> (insn 151 279 152 31 0x401ab0c0 (set (reg:SI 2 %r2 [74])
>         (mem:SI (plus:SI (reg/f:SI 1 %r1 [73])
>                 (reg:SI 3 %r3 [orig:72 kt ] [72])) [0 S4 A8])) 56 {*movsi} (nil)
>     (expr_list:REG_DEAD (reg/f:SI 1 %r1 [73])
>         (expr_list:REG_DEAD (reg:SI 3 %r3 [orig:72 kt ] [72])
>             (nil))))
> 
> (jump_insn 152 151 153 31 0x401ab0c0 (parallel [
>             (set (pc)
>                 (reg:SI 2 %r2 [74]))
>             (use (label_ref 153))
>         ]) 265 {casesi_jump} (insn_list 151 (nil))
>     (expr_list:REG_DEAD (reg:SI 2 %r2 [74])
>         (nil)))
> 
> uses this piece of code as jump table and jumps to some random
> location ...
> 
> 
> Bye,
> Ulrich
> 
> -- 
>   Dr. Ulrich Weigand
>   weigand@informatik.uni-erlangen.de

-- 
      Josef Zlomek
      zlomek@users.sf.net
      zlomj9am@artax.karlin.mff.cuni.cz
      http://zlomek.matfyz.cz/
      http://artax.karlin.mff.cuni.cz/~zlomj9am/