Don't do all the lowering stuff for 2-wide build_vector's. Also, minor optimization for shuffle of undef.

llvm-svn: 27946

Fix a performance regression. Use {p}shuf* when there are only two distinct elements in a build_vector.

llvm-svn: 27945

movd always clear the top 96 bits and movss does so when it's loading the
value from memory.
The net result is codegen for 4-wide shuffles is much improved. It is near
optimal if one or more elements is a zero. e.g.

__m128i test(int a, int b) {
  return _mm_set_epi32(0, 0, b, a);
}

compiles to

_test:
	movd 8(%esp), %xmm1
	movd 4(%esp), %xmm0
	punpckldq %xmm1, %xmm0
	ret

compare to gcc:

_test:
	subl	$12, %esp
	movd	20(%esp), %xmm0
	movd	16(%esp), %xmm1
	punpckldq	%xmm0, %xmm1
	movq	%xmm1, %xmm0
	movhps	LC0, %xmm0
	addl	$12, %esp
	ret

or icc:

_test:
        movd      4(%esp), %xmm0                                #5.10
        movd      8(%esp), %xmm3                                #5.10
        xorl      %eax, %eax                                    #5.10
        movd      %eax, %xmm1                                   #5.10
        punpckldq %xmm1, %xmm0                                  #5.10
        movd      %eax, %xmm2                                   #5.10
        punpckldq %xmm2, %xmm3                                  #5.10
        punpckldq %xmm3, %xmm0                                  #5.10
        ret                                                     #5.10

There are still room for improvement, for example the FP variant of the above example:

__m128 test(float a, float b) {
  return _mm_set_ps(0.0, 0.0, b, a);
}

_test:
	movss 8(%esp), %xmm1
	movss 4(%esp), %xmm0
	unpcklps %xmm1, %xmm0
	xorps %xmm1, %xmm1
	movlhps %xmm1, %xmm0
	ret

The xorps and movlhps are unnecessary. This will require post legalizer optimization to handle.

llvm-svn: 27939

scalar value.

e.g.
        _mm_set_epi32(0, a, 0, 0);
==>
	movd 4(%esp), %xmm0
	pshufd $69, %xmm0, %xmm0

        _mm_set_epi8(0, 0, 0, 0, 0, a, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
==>
	movzbw 4(%esp), %ax
	movzwl %ax, %eax
	pxor %xmm0, %xmm0
	pinsrw $5, %eax, %xmm0

llvm-svn: 27923

to a vector shuffle.
- VECTOR_SHUFFLE lowering change in preparation for more efficient codegen
of vector shuffle with zero (or any splat) vector.

llvm-svn: 27875

but i64 is not. If possible, change a i64 op to a f64 (e.g. load, constant)
and then cast it back.

llvm-svn: 27849

llvm-svn: 27847

  instructions.
- Fixed a commute vector_shuff bug.

llvm-svn: 27845

llvm-svn: 27840

llvm-svn: 27782

llvm-svn: 27779

llvm-svn: 27768

llvm-svn: 27767

llvm-svn: 27755

llvm-svn: 27719

llvm-svn: 27718

llvm-svn: 27711

llvm-svn: 27644

Clean up and fix various logical ops issues.

llvm-svn: 27633

llvm-svn: 27612

llvm-svn: 27575

llvm-svn: 27556

llvm-svn: 27501

- Normalize shuffle nodes so result vector lower half elements come from the
  first vector, the rest come from the second vector. (Except for the
  exceptions :-).
- Other minor fixes.

llvm-svn: 27474

llvm-svn: 27444

vector_shuffle v1, v1, <0, 4, 1, 5, 2, 6, 3, 7>

This is turned into
vector_shuffle v1, <undef>, <0, 0, 1, 1, 2, 2, 3, 3>
by dag combiner.

It would match a {p}unpckl on x86.

llvm-svn: 27437

llvm-svn: 27434

llvm-svn: 27433

llvm-svn: 27427

Use movhpd to: store upper f64 extracted from v2f64.

llvm-svn: 27382

- Some bug fixes and naming inconsistency fixes.

llvm-svn: 27377

INSERT_VECTOR_ELT to insert a 16-bit value in a 128-bit vector.

llvm-svn: 27314

from a 128-bit vector.

llvm-svn: 27304

llvm-svn: 27275

llvm-svn: 27272

llvm-svn: 27271

Use pshufd, pshuhw, and pshulw to shuffle v4f32 if shufps doesn't match.
Use shufps to shuffle v4f32 if pshufd, pshuhw, and pshulw don't match.

llvm-svn: 27259

- Added SSE2 128-bit integer pack with signed saturation ops.
- Added pshufhw and pshuflw ops.

llvm-svn: 27252

vector_shuffle undef.

llvm-svn: 27250

- Whenever possible use ops of the right packed types for vector shuffles /
  splats.

llvm-svn: 27246