Currently, LLVM lowers a cttz8 on x86_64 to these instructions:

```asm
    movzbl      %dil, %eax
    bsfl        %eax, %eax
    movl        $32, %ecx
    cmovnel     %eax, %ecx
    cmpl        $32, %ecx
    movl        $8, %eax
    cmovnel     %ecx, %eax
```

which has some unnecessary overhead, having two conditional moves.

To improve the codegen, we can zero extend the 8 bit integer, then set
bit 8 and perform a cttz operation on the extended value. That way
there's no conditional operation involved at all.

diff --git a/src/libcore/num/mod.rs b/src/libcore/num/mod.rs
index 44d5333ce1f46bb582a11b97e2fe08718c52029e..b8638c5b09be257ea87e9c7db993ea0c930328d5 100644 (file)
--- a/src/libcore/num/mod.rs
+++ b/src/libcore/num/mod.rs
@@ -745,7 +745,20 @@ pub fn leading_zeros(self) -> u32 {
         #[stable(feature = "rust1", since = "1.0.0")]
         #[inline]
         pub fn trailing_zeros(self) -> u32 {
-            unsafe { $cttz(self as $ActualT) as u32 }
+            // As of LLVM 3.6 the codegen for the zero-safe cttz8 intrinsic
+            // emits two conditional moves on x86_64. By promoting the value to
+            // u16 and setting bit 8, we get better code without any conditional
+            // operations.
+            // FIXME: There's a LLVM patch (http://reviews.llvm.org/D9284)
+            // pending, remove this workaround once LLVM generates better code
+            // for cttz8.
+            unsafe {
+                if $BITS == 8 {
+                    intrinsics::cttz16(self as u16 | 0x100) as u32
+                } else {
+                    $cttz(self as $ActualT) as u32
+                }
+            }
         }
 
         /// Shifts the bits to the left by a specified amount, `n`,