Re: DECL_RTL problem

[Mark Mitchell <mark at codesourcery dot com>]

More details.  After loop, we have:

  (insn 2747 2745 2749 (set (reg/f:SI 648)
	  (reg/f:SI 521)) 175 {*arm_movsi_insn} (nil)
      (nil))

  (insn 2749 2747 2754 (set (reg/f:SI 647)
	  (mem/s:SI (plus:SI (reg/f:SI 648)
		  (const_int 80 [0x50])) 0)) 175 {*arm_movsi_insn} (nil)
      (nil))

  (insn 2754 2749 2755 (set (reg:CC 24 cc)
	  (compare:CC (reg/f:SI 647)
	      (const_int 0 [0x0]))) 217 {*arm_cmpsi_insn} (nil)
	  (nil))

  (jump_insn 2755 2754 3963 (set (pc)
	  (if_then_else (ne (reg:CC 24 cc)
		  (const_int 0 [0x0]))
	      (label_ref 2778)
	      (pc))) 234 {*arm_cond_branch} (nil)
	  (nil))

After CSE2, we have:

  (insn 2747 2745 2749 (set (reg/f:SI 648)
	  (reg/f:SI 810)) 175 {*arm_movsi_insn} (nil)
      (nil))

  (insn 2749 2747 2754 (set (reg/f:SI 647)
	  (reg:SI 522)) 175 {*arm_movsi_insn} (nil)
      (expr_list:REG_EQUAL (const_int 0 [0x0])
	  (nil)))

  (insn 2754 2749 2755 (set (reg:CC 24 cc)
	  (compare:CC (reg:SI 522)
		  (const_int 0 [0x0]))) 217 {*arm_cmpsi_insn} (nil)
      (nil))

  (note 2755 2754 3963 NOTE_INSN_DELETED 0)

The conditional branch is gone.  Tsk, tsk.

It's provable at this point that reg:SI 522 is non-zero so CSE should
know that the jump will always be taken, not never be taken.

Hmm.

--
Mark Mitchell                   mark@codesourcery.com
CodeSourcery, LLC               http://www.codesourcery.com