On Mon, Mar 12, 2012 at 11:13 AM, Paolo Bonzini bonz...@gnu.org wrote:
Il 12/03/2012 09:52, Uros Bizjak ha scritto:
+(define_peephole2
+ [(parallel [(set (reg FLAGS_REG) (match_operand 0 ))
+ (match_operand 4 )])
+ (set (match_operand:QI 1 register_operand )
+ (match_operator:QI 2 ix86_comparison_operator
+ [(reg FLAGS_REG) (const_int 0)]))
+ (set (match_operand 3 q_regs_operand )
+ (zero_extend (match_dup 1)))]
+ (peep2_reg_dead_p (3, operands[1])
+ || operands_match_p (operands[1], operands[3]))
+ ! reg_overlap_mentioned_p (operands[3], operands[0])
I understand that you're assuming the shape of operands[4] to be the
same as operands[3], but would it be preferrable to add another overlap
check on operands[4]?
For example the transformation is invalid if you had an overlap between
operands[3] and the destination of operands[4].
The destination of operands[4] _always_ matches one of operands inside
operand[0]. All arithmetic insn that set flags are destructive on x86.
Uros.