PATCH: new peephole2 in arm.md

[Tamás Gergely <gertom at rgai dot hu>]

Hello All,

This patch extends a peephole2 optimization: in arm, move and comparison to
0 (zero) can be solved in parallel. There is a peephole2 optimization 
defined in arm.md that recognizes one kind of these cases: when two 
instructions follow each other:
  (1) move the value of rs into rd,
  (2) compare rd to 0.
However, if rs is compared instead of rd in the 2nd instruction, the
program  is identical to the first version, but gcc does not
recognizes the pattern, and leaves the two instructions unchanged. This
patch adds a new peephole2 pattern to arm.md that recognizes the second 
case, and makes the two instructions parallel.
--
Testcases:

Example (part of assembly of
gcc/testsuite/gcc.c-torture/execute/20020503-1.c compiled with -O2):

Without the patch:

inttostr:
	@ args = 0, pretend = 0, frame = 0
	@ frame_needed = 0, uses_anonymous_args = 0
	mov	r3, #0
	str	lr, [sp, #-4]!
	cmp	r0, #0;		<-- (2)
	mov	lr, r0;		<-- (1)
	mov	ip, r0;		<-- (3)
	strb	r3, [r1, #127]
	add	r0, r1, #127
	ldr	r1, .L8
	rsblt	ip, lr, #0
.L3:
	umull	r2, r3, r1, ip
	...

With the patch:

inttostr:
	@ args = 0, pretend = 0, frame = 0
	@ frame_needed = 0, uses_anonymous_args = 0
	mov	r3, #0
	str	lr, [sp, #-4]!
	subs	lr, r0, #0;	<-- (1-2 parallel)
	strb	r3, [r1, #127]
	add	r0, r1, #127
	ldr	r1, .L8
	mov	ip, lr;		<-- (3)
	rsblt	ip, lr, #0
.L3:
	umull	r2, r3, r1, ip
	...

--
Change log entry:

2003-02-04 Tamas Gergely <gertom@rgai.hu>

	* gcc/config/arm/arm.md: new peephole2 pattern added for extending
	move and compare parallelization.

--
Testing:

binutils-2.13, gcc-20030127, newlib-1.10.0; HOST=i686-pc-linux-elf, 
TARGET=arm-unknown-elf

``make check-gcc'' produced the same .sum files with and without the patch.
Tested also on some real applications, no regressions.

--
Patch:

gcc-20030127 and gcc-20030203 contains the same arm.md file, so the patch
can be applied directly to the later snapshot.

Patch attached.
---
*** gcc/config/arm/arm.md.orgn	Mon Feb  3 17:05:52 2003
--- gcc/config/arm/arm.md	Mon Feb  3 17:09:22 2003
***************
*** 8534,8539 ****
--- 8534,8553 ----
    ""
  )
  
+ ; The same as the previous, but with exchanged operands
+ 
+ (define_peephole2
+   [(set (match_operand:SI 0 "s_register_operand" "")
+ 	(match_operand:SI 1 "s_register_operand" ""))
+    (set (reg:CC CC_REGNUM)
+ 	(compare:CC (match_dup 0) (const_int 0)))]
+   "TARGET_ARM
+   "
+   [(parallel [(set (reg:CC CC_REGNUM) (compare:CC (match_dup 1) (const_int 0)))
+ 	      (set (match_dup 0) (match_dup 1))])]
+   ""
+ )
+ 
  ; Peepholes to spot possible load- and store-multiples, if the ordering is
  ; reversed, check that the memory references aren't volatile.