X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTarget%2FX86%2FREADME-SSE.txt;h=b2116e03b148ea222532c956e9dd78f49fea3a06;hb=1292c226458b68a119d3a387a0527f453b2065c2;hp=85cdd17781985cc78558af9a88b3a78e07d58839;hpb=b7f243a638c60e32c4f576d13b03687b0f90cfee;p=oota-llvm.git

diff --git a/lib/Target/X86/README-SSE.txt b/lib/Target/X86/README-SSE.txt
index 85cdd177819..b2116e03b14 100644
--- a/lib/Target/X86/README-SSE.txt
+++ b/lib/Target/X86/README-SSE.txt
@@ -17,6 +17,51 @@ __m128i shift_right(__m128i value, unsigned long offset) {
                _mm_loadu_si128((__m128 *) (___m128i_shift_right + offset)));
 }
 
+//===---------------------------------------------------------------------===//
+
+SSE has instructions for doing operations on complex numbers, we should pattern
+match them.   For example, this should turn into a horizontal add:
+
+typedef float __attribute__((vector_size(16))) v4f32;
+float f32(v4f32 A) {
+  return A[0]+A[1]+A[2]+A[3];
+}
+
+Instead we get this:
+
+_f32:                                   ## @f32
+	pshufd	$1, %xmm0, %xmm1        ## xmm1 = xmm0[1,0,0,0]
+	addss	%xmm0, %xmm1
+	pshufd	$3, %xmm0, %xmm2        ## xmm2 = xmm0[3,0,0,0]
+	movhlps	%xmm0, %xmm0            ## xmm0 = xmm0[1,1]
+	movaps	%xmm0, %xmm3
+	addss	%xmm1, %xmm3
+	movdqa	%xmm2, %xmm0
+	addss	%xmm3, %xmm0
+	ret
+
+Also, there are cases where some simple local SLP would improve codegen a bit.
+compiling this:
+
+_Complex float f32(_Complex float A, _Complex float B) {
+  return A+B;
+}
+
+into:
+
+_f32:                                   ## @f32
+	movdqa	%xmm0, %xmm2
+	addss	%xmm1, %xmm2
+	pshufd	$1, %xmm1, %xmm1        ## xmm1 = xmm1[1,0,0,0]
+	pshufd	$1, %xmm0, %xmm3        ## xmm3 = xmm0[1,0,0,0]
+	addss	%xmm1, %xmm3
+	movaps	%xmm2, %xmm0
+	unpcklps	%xmm3, %xmm0    ## xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]
+	ret
+
+seems silly when it could just be one addps.
+
+
 //===---------------------------------------------------------------------===//
 
 Expand libm rounding functions inline:  Significant speedups possible.