Don't use a potentially expensive shift if all we want is one set bit.

[oota-llvm.git] / lib / Target / R600 / SIRegisterInfo.td

diff --git a/lib/Target/R600/SIRegisterInfo.td b/lib/Target/R600/SIRegisterInfo.td
index 4f14931a9c481f4d12d450d53b6e77fd1c62a068..244d4c00348d3b5a45532811194ab1e99f2247dd 100644 (file)
--- a/lib/Target/R600/SIRegisterInfo.td
+++ b/lib/Target/R600/SIRegisterInfo.td
@@ -94,6 +94,12 @@ def VGPR_64 : RegisterTuples<[sub0, sub1],
                              [(add (trunc VGPR_32, 255)),
                               (add (shl VGPR_32, 1))]>;
 
+// VGPR 96-bit registers
+def VGPR_96 : RegisterTuples<[sub0, sub1, sub2],
+                             [(add (trunc VGPR_32, 254)),
+                              (add (shl VGPR_32, 1)),
+                              (add (shl VGPR_32, 2))]>;
+
 // VGPR 128-bit registers
 def VGPR_128 : RegisterTuples<[sub0, sub1, sub2, sub3],
                               [(add (trunc VGPR_32, 253)),
@@ -151,7 +157,7 @@ def SReg_64 : RegisterClass<"AMDGPU", [i64, i1], 64,
   (add SGPR_64, VCCReg, EXECReg)
 >;
 
-def SReg_128 : RegisterClass<"AMDGPU", [v16i8], 128, (add SGPR_128)>;
+def SReg_128 : RegisterClass<"AMDGPU", [v16i8, i128], 128, (add SGPR_128)>;
 
 def SReg_256 : RegisterClass<"AMDGPU", [v32i8], 256, (add SGPR_256)>;
 
@@ -162,6 +168,10 @@ def VReg_32 : RegisterClass<"AMDGPU", [i32, f32, v1i32], 32, (add VGPR_32)>;
 
 def VReg_64 : RegisterClass<"AMDGPU", [i64, f64, v2i32, v2f32], 64, (add VGPR_64)>;
 
+def VReg_96 : RegisterClass<"AMDGPU", [untyped], 96, (add VGPR_96)> {
+  let Size = 96;
+}
+
 def VReg_128 : RegisterClass<"AMDGPU", [v4i32, v4f32], 128, (add VGPR_128)>;
 
 def VReg_256 : RegisterClass<"AMDGPU", [v8i32, v8f32], 256, (add VGPR_256)>;