JIT: Allow strength reducing to GCD of IVs

[jakobbotsch]

This adds support for strength reduction to create a new primary IV that is the GCD of several IVs found in the loop. When the same index is used to access arrays of different sizes we will often see the IV being multiplied by different values; however, it is usually still profitable to strength reduce to the GCD of the step values and then "recover" the final IV by scaling.

Example:

public static void Foo()
{
    string puzzle = "003020600900305001001806400008102900700000008006708200002609500800203009005010300";
    int[] board = new int[81];

    for (int i = 0; i < puzzle.Length; i++)
    {
        board[i] = puzzle[i] - '0';
    }
}

Codegen diff for loop:

        xor      ecx, ecx
+       mov      edx, 81

 G_M24659_IG03:
-       mov      edx, ecx
-       movzx    r8, word  ptr [rbx+2*rdx+0x10]
+       movzx    r8, word  ptr [rbx+rcx+0x10]
        add      r8d, -48
-       mov      dword ptr [rax+4*rdx+0x10], r8d
-       inc      ecx
-       cmp      ecx, 81
-       jl       SHORT G_M24659_IG03
-						;; size=24 bbWeight=3.96 PerfScore 19.80
+       mov      dword ptr [rax+2*rcx+0x10], r8d
+       add      rcx, 2
+       dec      edx
+       jne      SHORT G_M24659_IG03
+						;; size=23 bbWeight=3.96 PerfScore 18.81

A similar diff in System.Linq.Enumerable+EnumerableSorter`2[System.__Canon,System.Decimal]:ComputeKeys(System.__Canon[],int):

+       xor      edx, edx
 G_M57524_IG05:
-       mov      edx, r15d
-       mov      r8, gword ptr [rbx+8*rdx+0x10]
+       mov      r8, gword ptr [rbx+rdx+0x10]
        vmovups  xmm0, xmmword ptr [r8+0x20]
        vmovups  xmmword ptr [rsp+0x28], xmm0
-       shl      rdx, 4
        vmovups  xmm0, xmmword ptr [rsp+0x28]
-       vmovups  xmmword ptr [r14+rdx+0x10], xmm0
-       inc      r15d
-       cmp      r13d, r15d
-       jg       SHORT G_M57524_IG05
-						;; size=45 bbWeight=75.73 PerfScore 1079.10
+       vmovups  xmmword ptr [r14+2*rdx+0x10], xmm0
+       add      rdx, 8
+       dec      r13d
+       jne      SHORT G_M57524_IG05
+						;; size=39 bbWeight=75.73 PerfScore 1022.31

Fix #102068
Fix #105241

[jakobbotsch]

This needs some heuristics for arm64. There strength reducing in this case can break address modes, e.g. before:

G_M51654_IG04:  ;; offset=0x0048
            mov     w4, w1
            ldr     x5, [x2, x4, LSL #3]
            ldr     w5, [x5, #0x28]
            str     w5, [x3, x4, LSL #2]
            add     w1, w1, #1
            cmp     w20, w1
            bgt     G_M51654_IG04
						;; size=28 bbWeight=4.95 PerfScore 47.03

After

G_M51654_IG04:  ;; offset=0x004C
            lsl     x4, x1, #1
            ldr     x4, [x2, x4]
            ldr     w4, [x4, #0x28]
            str     w4, [x3, x1]
            add     x1, x1, #4
            sub     w20, w20, #1
            cbnz    w20, G_M51654_IG04

Before we have two address modes. However, ARM64 cannot represent an address mode with a scale that doesn't correspond to the element size, so in the "after" case it is no longer possible to represent any address modes.

[jakobbotsch]

cc @dotnet/jit-contrib PTAL @AndyAyersMS

Diffs. Gets around a hundred cases in libraries_tests.run and a few dozen in aspnet/benchmarks. Also fixes two motivating cases.

I investigated why we don't see more cases and opened #110315 for one of the reasons, but I think I'll save further investigation for another time.

[AndyAyersMS]

Why does this check iv2->Type when the previous check was against iv1->Type?

[jakobbotsch]

Just a typo, switched it to iv1.