reload question

[Chris Lattner]

I'm fixing some bugs in a test backend I wrote for GCC, and am hitting a
few snags in the reload phase.  Specifically, GCC isn't able to reload:

(insn 29 28 31 (set (reg:SI 8 %o0 [113])
        (zero_extend:SI (subreg:QI (reg/v:SI 108) 0))) 19 {zero_extendqisi2} (nil)
    (expr_list:REG_EQUIV (mem/f:SI (reg/f:PDI 14 %sp) 0)
        (expr_list:REG_DEAD (reg/v:SI 108)
            (nil))))

and a small set of other cases, involving (zero|sign) extending subregs of
registers.

Can anyone give me a hint as to what could cause this?  I have the
following instruction pattern:

(define_insn "zero_extendqisi2"
  [(set (match_operand:SI 0 "register_operand" "=r")
	(zero_extend:SI (match_operand:QI 1 "register_operand" "")))]
  ""
  "zero_extendqisi2 %1 -> %0")

Which is obviously matched by the labeler, and combined by the combiner...
but reload isn't able to split the combined instruction into something
like this:

(set (reg:QI %reg:QI 1000)
        (subreg:QI (reg/v:SI 108)))

(set (reg:SI 8 %o0 [113])
        (zero_extend:SI (reg:QI 1000)))

Is there something that I'm missing that enables such transformations to
take place?  Is there someplace in the manual that I should reread to
understand this phase of GCC?  Do subreg's ever get turned into
truncate's?

Thanks for the help,

-Chris

http://www.nondot.org/sabre/os/