Performance regression

[Dale Johannesen <dalej at apple dot com>]

This patch:

2002-07-10  Roger Sayle  <roger@eyesopen.com>

        PR c/2454
        * combine.c (nonzero_bits): LOAD_EXTEND_OP should only apply
        to SUBREGs of MEMs.  (num_sign_bit_copies): Likewise.

introduces a performance regression on the following code (ppc
[on which LOAD_EXTEND_OP==ZERO_EXTEND and WORD_REGISTER_OPERATIONS]
and probably other RISCs):

struct x { unsigned char c; };
unsigned char foo(struct x* x, unsigned char q) {
  return x->c + (unsigned int)(255 - x->c)*q;
}

An extra and unnecessary zero-extend-byte-to-word insn is now emitted.
Formerly, combine improved this RTL [where reg 120 is also used later]:

(insn 15 7 16 (set (reg:QI 120)
        (mem/s/j:QI (reg/v/f:SI 116) [0 <variable>.c+0 S1 A8])) 314
                                 {*rs6000.md:8046} (insn_list 4 (nil))
    (expr_list:REG_DEAD (reg/v/f:SI 116)
        (nil)))

(insn 16 15 18 (set (reg:SI 119)
        (zero_extend:SI (reg:QI 120))) 18 {*rs6000.md:1247} (insn_list 
15 (nil))
    (nil))

(insn 18 16 20 (set (reg:SI 121)
        (minus:SI (const_int 255 [0xff])
            (reg:SI 119))) 44 {*rs6000.md:1908} (insn_list 16 (nil))
    (expr_list:REG_DEAD (reg:SI 119)
        (nil)))

into this:

(insn 15 7 16 (set (reg:QI 120)
        (mem/s/j:QI (reg:SI 3 r3) [0 <variable>.c+0 S1 A8])) 314 
{*rs6000.md:8046} (nil)
    (expr_list:REG_DEAD (reg:SI 3 r3)
        (nil)))

(note 16 15 18 NOTE_INSN_DELETED)

(insn 18 16 20 (set (reg:SI 121)
        (minus:SI (const_int 255 [0xff])
            (subreg:SI (reg:QI 120) 0))) 44 {*rs6000.md:1908} 
(insn_list 15 (nil))
    (nil))

With the changed semantics of SUBREG, this no longer happens.  And
since the merged instruction 18 above no longer has the same semantics
as the 2 insns, it should not happen.  (Transforming zero_extend
to subreg now loses information; unfortunately disabling this seems
impractical, since the subreg RTL is needed to match a lot of
register_operand constraints.)  I'm not suggesting reverting the patch
because it seems to be necessary for correctness, and that's more
important; but I'd like to get this fixed. Suggestions?