[PATCH] D18593: [PowerPC] Front end improvements for vec_splat

[amehsan via llvm-commits]

amehsan added a comment.

Thanks for using target-independent builtin and IR.


================
Comment at: lib/Headers/altivec.h:7726-7736
@@ -7725,7 +7725,13 @@
 vec_splat(vector signed int __a, unsigned const int __b) {
-  unsigned char b0 = (__b & 0x03) * 4;
-  unsigned char b1 = b0 + 1, b2 = b0 + 2, b3 = b0 + 3;
-  return vec_perm(__a, __a,
-                  (vector unsigned char)(b0, b1, b2, b3, b0, b1, b2, b3, b0, b1,
-                                         b2, b3, b0, b1, b2, b3));
+  const unsigned __elem = __b & 0x3;
+  switch(__elem) {
+  case 0:
+    return __builtin_shufflevector(__a, __a, 0, 0, 0, 0);
+  case 1:
+    return __builtin_shufflevector(__a, __a, 1, 1, 1, 1);
+  case 2:
+    return __builtin_shufflevector(__a, __a, 2, 2, 2, 2);
+  case 3:
+    return __builtin_shufflevector(__a, __a, 3, 3, 3, 3);
+  }
 }
----------------
Why do we need a switch case here (and in the ones below) instead of just returning __builtin_shufflevector(__a, __a, __elem, __elem, __elem, __elem)? 


Repository:
  rL LLVM

http://reviews.llvm.org/D18593