Add a fast path to alignTo.

[ruiu]

Summary

If an alignment is a power of two, we can avoid integer division,
and it is very likely to happen.

[davide]

Not sure I really like this one for two reasons.

1. I profiled llvm/lld to death and this pretty much never shows up in a profile.
2. I think after inlining instcombine/instsimplify should handle this case.

[davide]

In other words, I propose abandoning this one and opening an llvm bug to get this fixed (if you still care).

[ruiu]

Sorry not to explicitly abandon. Closing it now.

[davide]

In D25344#621824, @ruiu wrote:

Sorry not to explicitly abandon. Closing it now.

No problem at all.
I opened https://llvm.org/bugs/show_bug.cgi?id=31368 to keep track of the missing transform

include/llvm/Support/MathExtras.h

 ///
 ///   alignTo(5, 8, 7) = 7
 ///   alignTo(17, 8, 1) = 17
 ///   alignTo(~0LL, 8, 3) = 3
 ///   alignTo(321, 255, 42) = 552
 /// \endcode
 inline uint64_t alignTo(uint64_t Value, uint64_t Align, uint64_t Skew = 0) {
   assert(Align != 0u && "Align can't be 0.");
+  if (LLVM_LIKELY(isPowerOf2_64(Align))) {
+    Skew &= Align - 1;
+    return ((Value + Align - 1 - Skew) & -Align) + Skew;
+  }
   Skew %= Align;
   return (Value + Align - 1 - Skew) / Align * Align + Skew;
 }
 
 /// Returns the next integer (mod 2**64) that is greater than or equal to
 /// \p Value and is a multiple of \c Align. \c Align must be non-zero.
 template <uint64_t Align>
 LLVM_CONSTEXPR inline uint64_t alignTo(uint64_t Value) {