[arm64] optimize test for bound checks against a 0 index

[sandreenko]

static int Test(int[] arr) => arr[0];

Current codegen:

G_M50965_IG01:
        A9BF7BFD          stp     fp, lr, [sp,#-16]!
        910003FD          mov     fp, sp
                                                ;; bbWeight=1    PerfScore 1.50
G_M50965_IG02:
        B9400801          ldr     w1, [x0,#8]
        7100003F          cmp     w1, #0
        54000089          bls     G_M50965_IG04
        B9401000          ldr     w0, [x0,#16]
                                                ;; bbWeight=1    PerfScore 7.50
G_M50965_IG03:
        A8C17BFD          ldp     fp, lr, [sp],#16
        D65F03C0          ret     lr
                                                ;; bbWeight=1    PerfScore 2.00
G_M50965_IG04:
        97FF2D9A          bl      CORINFO_HELP_RNGCHKFAIL
        D4200000          brk     #0
                                                ;; bbWeight=0    PerfScore 0.00

this

    cmp     w1, #0
    bls     G_M50965_IG04

can be optimized to:

    cbz     w1, G_M50965_IG04  ;; or CBNZ

branch if (not)zero.

Originally posted by @EgorBo in #42295 (comment)

@nathan-moore's analysis and prototype: #42295 (comment) :

@nathan-moore: I looked into what it will take to emit cbz here for arm. There isn't a way to emit it with genJumpToThrowHlpBlk, which is what everything currently uses. You can add a way, but since cbz is internally represented very differently from a normal jump, I don't see a good way of not duplicating the logic, which is what I did here: nathan-moore/runtime@8500e53. Does that seem like the right approach?

Also, I checked and there's already a peephole for this in the general case.

category:cq
theme:basic-cq
skill-level:beginner
cost:small
impact:small

[sandreenko]

Also, I checked and there's already a peephole for this in the general case.

Do you mean GT_JCMP and genCodeForJumpCompare?

[nathan-moore]

Yes, though the peephole itself is in lowering:

runtime/src/coreclr/src/jit/lower.cpp

Lines 2775 to 2780 in 755f016

	if (relop->OperIs(GT_EQ, GT_NE) && relopOp2->IsIntegralConst(0))
	{
	// Codegen will use cbz or cbnz in codegen which do not affect the flag register
	flags = relop->OperIs(GT_EQ) ? GTF_JCMP_EQ : 0;
	useJCMP = true;
	}

.

[BruceForstall]

@echesakovMSFT arm64 CQ issue