llvm-svn: 41029

llvm-svn: 41028

have complex pointer manipulation going on.  This allows us to compile
stuff like this:

__m128i foo(__m128i x){
                static const unsigned int c_0[4] = { 0, 0, 0, 0 };
                __m128i v_Zero = _mm_loadu_si128((__m128i*)c_0);
                x  = _mm_unpacklo_epi8(x,  v_Zero);
                return x;
}

into:

_foo:
        xorps   %xmm1, %xmm1
        punpcklbw       %xmm1, %xmm0
        ret

llvm-svn: 41022

llvm-svn: 40998

llvm-svn: 40997

llvm-svn: 40979

llvm-svn: 40978

not split condition constraints.

llvm-svn: 40977

alloca, increase the alignment of the load, turning it into an aligned load.

This allows us to compile:

#include <xmmintrin.h>
__m128i foo(__m128i x){
 static const unsigned int c_0[4] = { 0, 0, 0, 0 };
	  __m128i v_Zero = _mm_loadu_si128((__m128i*)c_0);
  x  = _mm_unpacklo_epi8(x,  v_Zero);
  return x;
}

into:

_foo:
	punpcklbw	_c_0.5944, %xmm0
	ret
	.data
	.lcomm	_c_0.5944,16,4		# c_0.5944

instead of:

_foo:
	movdqu	_c_0.5944, %xmm1
	punpcklbw	%xmm1, %xmm0
	ret
	.data
	.lcomm	_c_0.5944,16,2		# c_0.5944

llvm-svn: 40971

llvm-svn: 40961

llvm-svn: 40960

llvm-svn: 40952

llvm-svn: 40947

llvm-svn: 40946

llvm-svn: 40944

llvm-svn: 40941

llvm-svn: 40936

and one hack to avoid hitting a bad case when the alias analysis is imprecise.

llvm-svn: 40935

it for potentially undeading pointers.

llvm-svn: 40933

No functionality change.

llvm-svn: 40932

llvm-svn: 40922

llvm-svn: 40919

llvm-svn: 40915

llvm-svn: 40912

llvm-svn: 40909

llvm-svn: 40903

Fix comment typo

llvm-svn: 40898

Fix GLIBCXX_DEBUG error triggered by incrementing erased iterator.

llvm-svn: 40897

llvm-svn: 40883

llvm-svn: 40870

Make this code more efficient.

llvm-svn: 40861

llvm-svn: 40859

actual argument name of the documented function.

llvm-svn: 40851

This shrinks it down to something small.  On the testcase
from PR1432, this speeds up instcombine from 0.7959s to 0.5000s,
(59%)

llvm-svn: 40840

In the old way, we computed and inserted phi nodes for the whole IDF of 
the definitions of the alloca, then computed which ones were dead and
removed them.

In the new method, we first compute the region where the value is live,
and use that information to only insert phi nodes that are live.  This
eliminates the need to compute liveness later, and stops the algorithm
from inserting a bunch of phis which it then later removes.

This speeds up the testcase in PR1432 from 2.00s to 0.15s (14x) in a
release build and 6.84s->0.50s (14x) in a debug build.

llvm-svn: 40825

llvm-svn: 40824

measurable speedup.

llvm-svn: 40823

to the worklist, and handling the last one with a 'tail call'.  This speeds
up PR1432 from 2.0578s to 2.0012s (2.8%)

llvm-svn: 40822

mem2reg from 2.0742->2.0522s on PR1432.

llvm-svn: 40821

llvm-svn: 40820