3.2 bug or something really odd.

[Dmitry <diwil at eis dot ru>]

Hi Fellows,

I faced with the following gcc3.2 tricky behaviour.

I define following patterns:

(define_expand "extzv"
 [(set (match_operand:QI 0 "nonimmediate_operand_msp430" "")
       (zero_extract:QI (match_operand:QI 1 "nonimmediate_operand_msp430" "")
                        (match_operand 2 "const_int_operand" "")
                        (match_operand 3 "const_int_operand" "")))]
""
"{
    if(INTVAL(operands[2]) != 1 || INTVAL(operands[3]) <= 0)
      FAIL;
    operands[3] = GEN_INT(1<<INTVAL(operands[3]));
}")

(define_insn "*extzv"
 [(set (match_operand:QI 0 "nonimmediate_operand_msp430" "=r,r,r,r,m,m,m,m")
       (zero_extract:QI (match_operand:QI 1 "nonimmediate_operand_msp430" 
"r,m,r,m,r,m,r,m")
                        (match_operand 2 "const_int_operand" 
"i,i,i,i,i,i,i,i")
                        (match_operand 3 "const_int_operand" 
"P,P,i,i,P,P,i,i")))]
 ""
"bit.b\\t%3, %1
\\tclr.b\\t%0
\\tadc.b\\t%0"
  [(set_attr "length" "3,4,4,5,5,6,6,7")
     (set_attr "cc" "clobber")])


And for the code given:
	unsigned char doit(unsigned char x) {return (x >> 1) & 7; }

The compiler generates:
 ; (insn 14 13 15 (set (reg:QI 15 r15 [26])
 ;         (zero_extract:QI (reg/v:QI 15 r15 [23])
 ;             (const_int 3 [0x3])
 ;             (const_int 1 [0x1]))) 156 {*extzv} (nil)
 ;     (nil))
        bit.b   #llo(1), r15     ;  14  *extzv/1        [length = 3]
        clr.b   r15
        adc.b   r15
 ; (insn 15 14 19 (set (reg:HI 15 r15 [24])
 ;         (zero_extend:HI (reg:QI 15 r15 [26]))) 149 {zero_extendqihi2} 
(insn_list 14 (nil))
 ;     (nil))
        and.b   #-1,r15  ;  15  zero_extendqihi2/1      [length = 1]
        ret


As follows from the expander defenition, if the operands[2] not equal to 1 the 
extzv parrern must be ignored.
However, as follows from the rtl dump, gcc ignores 'FAIL' directive and issues 
tricky code.
The extzv pattern first appears after 'combine' pass.

GCC-current lacks this 'feature'.

Should I put an another condition in insn pattern '*extzv' like 
	'operands[2] == 1' ?
Or this is just something I cannot spot?


Thanks,
Dmitry.