Re: Combine performance regression (was: Fix PR target/18701)

[Richard Henderson <rth at redhat dot com>]

On Thu, Jan 20, 2005 at 08:11:06PM +0100, Ulrich Weigand wrote:
> Unfortunately I cannot use an immediate operand for the constant 255,
> so I have to reload it into a register (or use memory), which means
> I have to expose this detail before reload ...

Or do either

(define_insn "zero_extendqisi2"
  [(set (match_operand:SI 0 "register_operand" "=r,r")
        (zero_extend:QI (match_operand:QI 1 "nonimmediate_operand" "m,r")))
   (clobber (match_scratch:SI 2 "=X,r"))]
  ""
  "@
   load %0 = %1
   #")

(define_split
  [(set (match_operand:SI 0 "register_operand" "")
        (zero_extend:QI (match_operand:QI 1 "register_operand" "")))
   (clobber (match_scratch:SI 2 ""))]
  "reload_completed"
  [(set (match_dup 2) (const_int 255))
   (set (match_dup 0) (and:SI (match_dup 1) (match_dup 2)))]
{
  operands[1] = gen_lowpart (SImode, operands[1]);
})

or

(define_expand "*zero_extendqisi2"
  [(set (match_operand:SI 0 "register_operand" "")
        (zero_extend:QI (match_operand:QI 1 "nonimmediate_operand" "")))
   (use (match_dup 2))]
  ""
{
  operands[2] = force_reg (SImode, GEN_INT (255));
})

(define_insn "*zero_extendqisi2"
  [(set (match_operand:SI 0 "register_operand" "=r,r")
        (zero_extend:QI (match_operand:QI 1 "nonimmediate_operand" "m,r")))
   (use (match_operand:SI 2 "register_operand" "=X,r"))
  ""
  "@
   load %0 = %1
   and %0 = %1, %2")

Mostly depending on whether you've got two-address insns or three.
The later form will work harder to CSE the immediate.


r~