Remove more filters from the disassembler. Mark some AVX512 instructions as CodeGenOnly.

[oota-llvm.git] / lib / Target / X86 / README-SSE.txt

diff --git a/lib/Target/X86/README-SSE.txt b/lib/Target/X86/README-SSE.txt
index a581993c3c61bcf65d45c7a32d4ee58136ada49c..adfa7fa1232c2d6ac973cc639b741e7ad4cf436a 100644 (file)
--- a/lib/Target/X86/README-SSE.txt
+++ b/lib/Target/X86/README-SSE.txt
@@ -517,37 +517,6 @@ to <2 x i64> ops being so bad.
 
 //===---------------------------------------------------------------------===//
 
 
 //===---------------------------------------------------------------------===//
 

-'select' on vectors and scalars could be a whole lot better.  We currently 
-lower them to conditional branches.  On x86-64 for example, we compile this:
-
-double test(double a, double b, double c, double d) { return a<b ? c : d; }
-
-to:
-
-_test:
-       ucomisd %xmm0, %xmm1
-       ja      LBB1_2  # entry
-LBB1_1:        # entry
-       movapd  %xmm3, %xmm2
-LBB1_2:        # entry
-       movapd  %xmm2, %xmm0
-       ret
-
-instead of:
-
-_test:
-       cmpltsd %xmm1, %xmm0
-       andpd   %xmm0, %xmm2
-       andnpd  %xmm3, %xmm0
-       orpd    %xmm2, %xmm0
-       ret
-
-For unpredictable branches, the later is much more efficient.  This should
-just be a matter of having scalar sse map to SELECT_CC and custom expanding
-or iseling it.
-
-//===---------------------------------------------------------------------===//
-

 LLVM currently generates stack realignment code, when it is not necessary
 needed. The problem is that we need to know about stack alignment too early,
 before RA runs.
 LLVM currently generates stack realignment code, when it is not necessary
 needed. The problem is that we need to know about stack alignment too early,
 before RA runs.


@@ -922,3 +891,43 @@ _test2:                                 ## @test2
 The insertps's of $0 are pointless complex copies.
 
 //===---------------------------------------------------------------------===//
 The insertps's of $0 are pointless complex copies.
 
 //===---------------------------------------------------------------------===//

+
+[UNSAFE FP]
+
+void foo(double, double, double);
+void norm(double x, double y, double z) {
+  double scale = __builtin_sqrt(x*x + y*y + z*z);
+  foo(x/scale, y/scale, z/scale);
+}
+
+We currently generate an sqrtsd and 3 divsd instructions. This is bad, fp div is
+slow and not pipelined. In -ffast-math mode we could compute "1.0/scale" first
+and emit 3 mulsd in place of the divs. This can be done as a target-independent
+transform.
+
+If we're dealing with floats instead of doubles we could even replace the sqrtss
+and inversion with an rsqrtss instruction, which computes 1/sqrt faster at the
+cost of reduced accuracy.
+
+//===---------------------------------------------------------------------===//
+
+This function should be matched to haddpd when the appropriate CPU is enabled:
+
+#include <x86intrin.h>
+double f (__m128d p) {
+  return p[0] + p[1];
+}
+
+similarly, v[0]-v[1] should match to hsubpd, and {v[0]-v[1], w[0]-w[1]} should
+turn into hsubpd also.
+
+//===---------------------------------------------------------------------===//
+
+define <2 x i32> @foo(<2 x double> %in) {
+  %x = fptosi <2 x double> %in to <2 x i32>
+  ret <2 x i32> %x
+}
+
+Should compile into cvttpd2dq instead of being scalarized into 2 cvttsd2si.
+
+//===---------------------------------------------------------------------===//