JIT: inefficient multidimensional array access vis Jit32 (and in general)

[AndyAyersMS]

RyuJit fails to optimize multidimensional array indices and misses some cases that Jit32 is able to optimize. In the following code example,

class M
{
    public static int Example(int[,] a, int b)
    {
        int r = 0;
        for (int i = 0; i < a.GetLength(1); i++)
        {
            r += a[b, i] + a[b, i];
        }
        return r;
    }

    public static int Main()
    {
        int[,] a = new int[10,10];
        a[3,3] = 3;
        a[3,7] = 7;
        int r = Example(a, 3);
        return r == 20 ? 100 : 0;
    }
}

Jit32 is able to CSE the addressing done in the a[b,i] references in Example (though it does not CSE the array fetch itself). RyuJit does not CSE anything in the loop.

Neither jit recognizes a.GetLength as being a loop invariant, or that the initial bounds check r[b, ] is likewise a loop invariant, or inline and decompose GetLength despite inlining the very similar bounds check and element addressing logic. Neither jit is able to optimize away the bounds checks in Main where the array size and indices are known.

category:cq
theme:md-arrays
skill-level:expert
cost:large

[AndyAyersMS]

See also the codegen in Bytemark/BenchNumericSortRectangular's NumSift.

[BruceForstall]

Related: #5481

[BruceForstall]

This is addressed in #70271. There are follow-up items there to continue to improve MD array performance.