diff --git a/src/libextra/ebml.rs b/src/libextra/ebml.rs
index b63f7e495b90..4a47af46afe0 100644
--- a/src/libextra/ebml.rs
+++ b/src/libextra/ebml.rs
@@ -960,3 +960,87 @@ mod tests {
         test_v(Some(3));
     }
 }
+
+#[cfg(test)]
+mod bench {
+    use ebml::reader;
+    use test::BenchHarness;
+
+    #[bench]
+    pub fn vuint_at_A_aligned(bh: &mut BenchHarness) {
+        use std::vec;
+        let data = vec::from_fn(4*100, |i| {
+            match (i % 2) {
+              0 => 0x80u8,
+              _ => i as u8,
+            }
+        });
+        let mut sum = 0u;
+        bh.iter(|| {
+            let mut i = 0;
+            while (i < data.len()) {
+                sum += reader::vuint_at(data, i).val;
+                i += 4;
+            }
+        });
+    }
+
+    #[bench]
+    pub fn vuint_at_A_unaligned(bh: &mut BenchHarness) {
+        use std::vec;
+        let data = vec::from_fn(4*100+1, |i| {
+            match (i % 2) {
+              1 => 0x80u8,
+              _ => i as u8
+            }
+        });
+        let mut sum = 0u;
+        bh.iter(|| {
+            let mut i = 1;
+            while (i < data.len()) {
+                sum += reader::vuint_at(data, i).val;
+                i += 4;
+            }
+        });
+    }
+
+    #[bench]
+    pub fn vuint_at_D_aligned(bh: &mut BenchHarness) {
+        use std::vec;
+        let data = vec::from_fn(4*100, |i| {
+            match (i % 4) {
+              0 => 0x10u8,
+              3 => i as u8,
+              _ => 0u8
+            }
+        });
+        let mut sum = 0u;
+        bh.iter(|| {
+            let mut i = 0;
+            while (i < data.len()) {
+                sum += reader::vuint_at(data, i).val;
+                i += 4;
+            }
+        });
+    }
+
+    #[bench]
+    pub fn vuint_at_D_unaligned(bh: &mut BenchHarness) {
+        use std::vec;
+        let data = vec::from_fn(4*100+1, |i| {
+            match (i % 4) {
+              1 => 0x10u8,
+              0 => i as u8,
+              _ => 0u8
+            }
+        });
+        let mut sum = 0u;
+        bh.iter(|| {
+            let mut i = 1;
+            while (i < data.len()) {
+                sum += reader::vuint_at(data, i).val;
+                i += 4;
+            }
+        });
+    }
+}
diff --git a/src/libstd/unstable/intrinsics.rs b/src/libstd/unstable/intrinsics.rs
index e0089b599c5b..9acc09dfcc68 100644
--- a/src/libstd/unstable/intrinsics.rs
+++ b/src/libstd/unstable/intrinsics.rs
@@ -486,33 +486,33 @@ extern "rust-intrinsic" {
     pub fn u64_mul_with_overflow(x: u64, y: u64) -> (u64, bool);
 }
 
-#[cfg(target_endian = "little")] pub fn to_le16(x: i16) -> i16 { x }
-#[cfg(target_endian = "big")]    pub fn to_le16(x: i16) -> i16 { unsafe { bswap16(x) } }
-#[cfg(target_endian = "little")] pub fn to_le32(x: i32) -> i32 { x }
-#[cfg(target_endian = "big")]    pub fn to_le32(x: i32) -> i32 { unsafe { bswap32(x) } }
-#[cfg(target_endian = "little")] pub fn to_le64(x: i64) -> i64 { x }
-#[cfg(target_endian = "big")]    pub fn to_le64(x: i64) -> i64 { unsafe { bswap64(x) } }
+#[cfg(target_endian = "little")] #[inline] pub fn to_le16(x: i16) -> i16 { x }
+#[cfg(target_endian = "big")]    #[inline] pub fn to_le16(x: i16) -> i16 { unsafe { bswap16(x) } }
+#[cfg(target_endian = "little")] #[inline] pub fn to_le32(x: i32) -> i32 { x }
+#[cfg(target_endian = "big")]    #[inline] pub fn to_le32(x: i32) -> i32 { unsafe { bswap32(x) } }
+#[cfg(target_endian = "little")] #[inline] pub fn to_le64(x: i64) -> i64 { x }
+#[cfg(target_endian = "big")]    #[inline] pub fn to_le64(x: i64) -> i64 { unsafe { bswap64(x) } }
 
-#[cfg(target_endian = "little")] pub fn to_be16(x: i16) -> i16 { unsafe { bswap16(x) } }
-#[cfg(target_endian = "big")]    pub fn to_be16(x: i16) -> i16 { x }
-#[cfg(target_endian = "little")] pub fn to_be32(x: i32) -> i32 { unsafe { bswap32(x) } }
-#[cfg(target_endian = "big")]    pub fn to_be32(x: i32) -> i32 { x }
-#[cfg(target_endian = "little")] pub fn to_be64(x: i64) -> i64 { unsafe { bswap64(x) } }
-#[cfg(target_endian = "big")]    pub fn to_be64(x: i64) -> i64 { x }
+#[cfg(target_endian = "little")] #[inline] pub fn to_be16(x: i16) -> i16 { unsafe { bswap16(x) } }
+#[cfg(target_endian = "big")]    #[inline] pub fn to_be16(x: i16) -> i16 { x }
+#[cfg(target_endian = "little")] #[inline] pub fn to_be32(x: i32) -> i32 { unsafe { bswap32(x) } }
+#[cfg(target_endian = "big")]    #[inline] pub fn to_be32(x: i32) -> i32 { x }
+#[cfg(target_endian = "little")] #[inline] pub fn to_be64(x: i64) -> i64 { unsafe { bswap64(x) } }
+#[cfg(target_endian = "big")]    #[inline] pub fn to_be64(x: i64) -> i64 { x }
 
-#[cfg(target_endian = "little")] pub fn from_le16(x: i16) -> i16 { x }
-#[cfg(target_endian = "big")]    pub fn from_le16(x: i16) -> i16 { unsafe { bswap16(x) } }
-#[cfg(target_endian = "little")] pub fn from_le32(x: i32) -> i32 { x }
-#[cfg(target_endian = "big")]    pub fn from_le32(x: i32) -> i32 { unsafe { bswap32(x) } }
-#[cfg(target_endian = "little")] pub fn from_le64(x: i64) -> i64 { x }
-#[cfg(target_endian = "big")]    pub fn from_le64(x: i64) -> i64 { unsafe { bswap64(x) } }
+#[cfg(target_endian = "little")] #[inline] pub fn from_le16(x: i16) -> i16 { x }
+#[cfg(target_endian = "big")]    #[inline] pub fn from_le16(x: i16) -> i16 { unsafe { bswap16(x) } }
+#[cfg(target_endian = "little")] #[inline] pub fn from_le32(x: i32) -> i32 { x }
+#[cfg(target_endian = "big")]    #[inline] pub fn from_le32(x: i32) -> i32 { unsafe { bswap32(x) } }
+#[cfg(target_endian = "little")] #[inline] pub fn from_le64(x: i64) -> i64 { x }
+#[cfg(target_endian = "big")]    #[inline] pub fn from_le64(x: i64) -> i64 { unsafe { bswap64(x) } }
 
-#[cfg(target_endian = "little")] pub fn from_be16(x: i16) -> i16 { unsafe { bswap16(x) } }
-#[cfg(target_endian = "big")]    pub fn from_be16(x: i16) -> i16 { x }
-#[cfg(target_endian = "little")] pub fn from_be32(x: i32) -> i32 { unsafe { bswap32(x) } }
-#[cfg(target_endian = "big")]    pub fn from_be32(x: i32) -> i32 { x }
-#[cfg(target_endian = "little")] pub fn from_be64(x: i64) -> i64 { unsafe { bswap64(x) } }
-#[cfg(target_endian = "big")]    pub fn from_be64(x: i64) -> i64 { x }
+#[cfg(target_endian = "little")] #[inline] pub fn from_be16(x: i16) -> i16 { unsafe { bswap16(x) } }
+#[cfg(target_endian = "big")]    #[inline] pub fn from_be16(x: i16) -> i16 { x }
+#[cfg(target_endian = "little")] #[inline] pub fn from_be32(x: i32) -> i32 { unsafe { bswap32(x) } }
+#[cfg(target_endian = "big")]    #[inline] pub fn from_be32(x: i32) -> i32 { x }
+#[cfg(target_endian = "little")] #[inline] pub fn from_be64(x: i64) -> i64 { unsafe { bswap64(x) } }
+#[cfg(target_endian = "big")]    #[inline] pub fn from_be64(x: i64) -> i64 { x }
 
 /// `TypeId` represents a globally unique identifier for a type
 #[lang="type_id"] // This needs to be kept in lockstep with the code in trans/intrinsic.rs and