[llvm-commits] [llvm] r135992 - in /llvm/trunk: include/llvm/Support/BlockFrequency.h lib/Support/BlockFrequency.cpp lib/Support/CMakeLists.txt

[Jakob Stoklund Olesen]

On Jul 25, 2011, at 3:24 PM, Jakub Staszak wrote:

> --- llvm/trunk/include/llvm/Support/BlockFrequency.h (added)
> +++ llvm/trunk/include/llvm/Support/BlockFrequency.h Mon Jul 25 17:24:51 2011
> @@ -0,0 +1,64 @@
> +  static void mult96bit(uint64_t freq, uint32_t N, uint64_t W[2]);
> +  static uint64_t div96bit(uint64_t W[2], uint32_t D);

These functions should not be in the header, they can be static functions in the .cpp file.

> +/// div96bit - Divide 96-bit value stored in W array by D. Return 64-bit frequency.
> +uint64_t BlockFrequency::div96bit(uint64_t W[2], uint32_t D) {
> +  uint64_t y = W[0];
> +  uint64_t x = W[1];
> +
> +  for (int i = 1; i <= 64; ++i) {
> +    uint32_t t = (int)x >> 31;
> +    x = (x << 1) | (y >> 63);
> +    y = y << 1;
> +    if ((x | t) >= D) {
> +      x -= D;
> +      ++y;
> +    }
> +  }
> +
> +  return y;
> +}

There has to be a faster way to divide two numbers than this.

Please also compare the performance of these routines to the same computations in floating point.

/jakob