JIT: ensure magic divisors are always marked DONT_CSE for all targets

[AndyAyersMS]

fgMorphSmpOpOptional calls CheckDivideByConstOptimized for GT_UDIV/GT_UMOD but the corresponding case GT_DIV: arm only checks for val / 1 and breaks. Signed integer divides with a constant divisor reach lowering without GTF_DONT_CSE on the divisor.

Magic-divide expansion in LowerSignedDivOrMod -> TryLowerConstIntDivOrMod bails at if (!divisor->IsCnsIntOrI()) return false; and falls back to idiv if CSE has rewritten the divisor to a LCL_VAR.

This is latent on x64 because the default JitConstCSE setting disables most integer-constant CSE there, so a full SPMI sweep on x64 shows zero asm diffs. It is observable on ARM64/RISC-V where constant CSE is enabled, and would also surface immediately on x64 if JitConstCSE is ever broadened.

Signed GT_MOD is already covered: fgMorphModToSubMulDiv transforms a % b into a - (a / b) * b and calls CheckDivideByConstOptimized on the inner DIV.

Note

This PR description was drafted with the assistance of GitHub Copilot.

[dotnet-policy-service]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

[Copilot]

Pull request overview

This PR updates JIT morphing so GT_DIV nodes also call CheckDivideByConstOptimized, matching existing handling for GT_UDIV/GT_UMOD. The intent is to keep eligible constant divisors from being constant-CSE’d into locals so later lowering can still recognize and apply “magic divide” transformations.

Changes:

• In fgMorphSmpOpOptional, call CheckDivideByConstOptimized for GT_DIV after the existing val / 1 fast-path.

[AndyAyersMS]

Good catch. Pushed f11b2fa which extends UsesDivideByConstOptimized to accept negative magic divisors (|d| >= 3) and negative pow2 divisors, mirroring what TryLowerConstIntDivOrMod actually lowers.

Verified on the original repro (System.TimeOnly.AddTicks already uses positive divisors so unchanged), and a quick standalone test with x / -3, x / -2, x % -3 confirms the expected magic-divide / shift+neg sequences are emitted with no idiv fallback. Full x64 SPMI sweep still shows zero asm diffs (constant CSE is off by default there), as expected.

Note

Comment drafted with the assistance of GitHub Copilot.

[AndyAyersMS]

@EgorBo PTAL
fyi @dotnet/jit-contrib

future-proofing so we can look into enabling const CSE for more platforms. Should be no diff.