This is a blanket issue meant to track code quality improvements to size, startup, and runtime performance for the new RyuJIT Wasm Backend. Our current priorities are size-focused and so far, size improvements seem highly correlated with startup performance as well. Size improvements are often correlated with steady-state runtime performance improvements, though we shouldn't necessarily assume that will always be the case.

Perf Categories

Size / Startup

Regalloc

• More efficient use of locals: [wasm][r2r] Jit generates code with very large numbers of wasm locals #135139
• Allocate unexposed GC refs to locals: [wasm] Regalloc: Allow Non-exposed GC Ref Locals to Live in Wasm Locals #135543
• More general GC Ref Spilling accounting for liveness across safepoints (see [Wasm][CoreCLR] Enable GC-correct reference/byref locals to avoid hot-loop frame traffic #135980)

Inlining

• Tune inlining policy for Wasm to reduce code size

Morph

• Reduce arithmetic expression expansions: [Wasm][CoreCLR] Preserve native remainder instead of expanding constant MOD to SUB/MUL/DIV #135982

Prolog/Epilog

• Drop stack frame setup and teardown for methods which don't need stack space and don't need unwinding

Loops

• Investigate disabling loop cloning/unrolling on Wasm

SIMD

• [RyuJIT Wasm] CQ: Improve StoreSelectedScalar/LoadScalarAndInsert Codegen #131003

Tail Calls

• [wasm] CQ: Lack of Implicit Tail Calls Leads to Size-Expensive GDV Expansion #136246

Null Checks

• Propagate non-null this assertion: [NO MERGE] JIT: Experiment with non-null this for Wasm R2R #134677
• Teach fgAddrCouldBeNull to recognize more non-null cases

Steady-state runtime perf

Loops

• [Wasm][CoreCLR] Loop-hoisting profitability uses a negative native-register budget #135979 (loop hoisting profitability in a wasm context)

Cross-cutting

Opt-Repeat

• Investigate the impact of running opt repeat in an AOT Wasm context