[Committed] Optimize "popcount(x) eq/ne 0" as "x eq/ne 0"

["Roger Sayle" <roger at eyesopen dot com>]

Some of the new AMDFAM10 patterns recently added to i386.md by Harsha
contained the interesting RTL:

   (set (reg FLAGS_REG)
        (compare
          (popcount:SI (match_operand:SI 1 "nonimmediate_operand" "rm"))
          (const_int 0)))

What's most interesting about this to the RTL optimizers, is that with a
little thought it should be possible to simplify or optimize away all
comparisons of __builtin_popcount against zero as follows:

(eq  (popcount x) 0)  ->  (eq x 0)
(ne  (popcount x) 0)  ->  (ne x 0)
(lt  (popcount x) 0)  ->  false
(ltu (popcount x) 0)  ->  false
(le  (popcount x) 0)  ->  (eq x 0)
(leu (popcount x) 0)  ->  (eq x 0)
(gt  (popcount x) 0)  ->  (ne x 0)
(gtu (popcount x) 0)  ->  (ne x 0)
(ge  (popcount x) 0)  ->  true
(geu (popcount x) 0)  ->  true

The following patch implements the transformations above not already
performed by the compiler.

The following patch has been tested on both i686-pc-linux-gnu and
ia64-unknown-linux-gnu with a full "make bootstrap", all default languages
(including Ada on IA-32), and regression tested with a top-level "make -k
check" with no new failures.  The new testcases confirm that we emit the
Itanium's "popcnt" instruction for __builtin_popcount, but that we can
optimize it away when used as "__builtin_popcount(x) == 0".

Committed to mainline as revision 121838.


2007-02-11  Roger Sayle  <roger@eyesopen.com>

        * simplify-rtx.c (simplify_relational_operation_1): Optimize
        comparisons of POPCOUNT against zero.
        (simplify_const_relational_operation): Likewise.

        * gcc.target/ia64/builtin-popcount-1.c: New test case.
        * gcc.target/ia64/builtin-popcount-2.c: Likewise.

Roger
--

Attachment: patchp.txt
Description: Text document

Attachment: builtin-popcount-1.c
Description: Text document

Attachment: builtin-popcount-2.c
Description: Text document