Added more insertps optimizations

[oota-llvm.git] / lib / Target / X86 / X86RegisterInfo.td

diff --git a/lib/Target/X86/X86RegisterInfo.td b/lib/Target/X86/X86RegisterInfo.td
index b8027283cc1fd37d18af5a1ef5c007438a2073cb..33c402b69a4aa493110bd7935cbc2ef1e102b099 100644 (file)
--- a/lib/Target/X86/X86RegisterInfo.td
+++ b/lib/Target/X86/X86RegisterInfo.td
@@ -463,9 +463,13 @@ def VR128X : RegisterClass<"X86", [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64],
 def VR256X : RegisterClass<"X86", [v32i8, v16i16, v8i32, v4i64, v8f32, v4f64],
                           256, (sequence "YMM%u", 0, 31)>;
 
-def VK8     : RegisterClass<"X86", [v8i1],   8, (sequence "K%u", 0, 7)>;
-def VK16    : RegisterClass<"X86", [v16i1], 16, (add VK8)>;
-
-def VK8WM   : RegisterClass<"X86", [v8i1],   8, (sub VK8, K0)>;
+// The size of the all masked registers is 16 bit because we have only one
+// KMOVW istruction that can store this register in memory, and it writes 2 bytes
+def VK1     : RegisterClass<"X86", [i1],    16, (sequence "K%u", 0, 7)>;
+def VK8     : RegisterClass<"X86", [v8i1],  16, (add VK1)> {let Size = 16;}
+def VK16    : RegisterClass<"X86", [v16i1], 16, (add VK8)> {let Size = 16;}
+
+def VK1WM   : RegisterClass<"X86", [i1],    16, (sub VK1, K0)> {let Size = 16;}
+def VK8WM   : RegisterClass<"X86", [v8i1],  16, (sub VK8, K0)> {let Size = 16;}
 def VK16WM  : RegisterClass<"X86", [v16i1], 16, (add VK8WM)>;