[Neon intrinsics] Literal vector construction through vcombine is poor

[Michael Collison <Michael dot Collison at arm dot com>]

This patch improves code generation for literal vector construction by expanding and exposing the pattern to rtl optimization earlier. The current implementation delays splitting the pattern until after reload which results in poor code generation for the following code:


#include "arm_neon.h"

int16x8_t
foo ()
{
  return vcombine_s16 (vdup_n_s16 (0), vdup_n_s16 (8));
}

Trunk generates:

foo:
	movi	v1.2s, 0
	movi	v0.4h, 0x8
	dup	d2, v1.d[0]
	ins	v2.d[1], v0.d[0]
	orr	v0.16b, v2.16b, v2.16b
	ret

With the patch we now generate:

foo:
	movi	v1.4h, 0x8
	movi	v0.4s, 0
	ins	v0.d[1], v1.d[0]
	ret

Bootstrapped and tested on aarch64-linux-gnu. Okay for trunk.

2017-06-15  Michael Collison  <michael.collison@arm.com>

	* config/aarch64/aarch64-simd.md(aarch64_combine_internal<mode>):
	Convert from define_insn_and_split into define_expand
	* config/aarch64/aarch64.c(aarch64_split_simd_combine):
	Allow register and subreg operands.

Attachment: pr7057.patch
Description: pr7057.patch