[Patch, AArch64, Testsuite] Check for expected MOVI vectorization.

[Tejas Belagod <tejas dot belagod at arm dot com>]

Hi,

This change:

+2014-12-05 Martin Jambor mjambor@suse.cz
+

    PR ipa/64192
        ipa-prop.c (ipa_compute_jump_functions_for_edge): Convert alignment
    from bits to bytes after checking they are byte-aligned.
    +

causes this regression on AArch64.

FAIL: gcc.target/aarch64/vect-movi.c scan-assembler movi\\tv[0-9]+.4s, 
0xab, msl 16
FAIL: gcc.target/aarch64/vect-movi.c scan-assembler movi\\tv[0-9]+.4s, 
0xab, msl 8
FAIL: gcc.target/aarch64/vect-movi.c scan-assembler mvni\\tv[0-9]+.4s, 
0xab, msl 16
FAIL: gcc.target/aarch64/vect-movi.c scan-assembler mvni\\tv[0-9]+.4s, 
0xab, msl 8

It causes AArch64 vector cost model to vectorize the loops in the test 
case with a VF = 2 on A53/default and VF = 4 for A57.

A53/default:
movi v0.2s, 0xab, msl 8
str d0, [x0]
str d0, [x0, 8]
str d0, [x0, 16]
str d0, [x0, 24]
str d0, [x0, 32]
str d0, [x0, 40]
str d0, [x0, 48]
str d0, [x0, 56]

vs. A57

    movi v0.4s, 0xab, msl 8
    str q0, [x0]
    str q0, [x0, 16]
    str q0, [x0, 32]
    str q0, [x0, 48]


But the test case isn't checking for a per-core optimized code, just 
whether we vectorize for MOVI or not. So, this patch improves reg exp to 
make sure the compiler vectorizes the code for either vectorization factor.

OK for trunk?

Thanks,
Tejas.

Changelog:

gcc/testsuite:

	* gcc.target/aarch64/vect-movi.c: Check for vectorization for
	64-bit and 128-bit.

Attachment: movi.txt
Description: Text document