Six wins surfaced by the inner-loop analysis. This file gets deleted when everything's shipped — anchors the work + the after-each-change checklist while it's in flight. https://claude.ai/code/session_01FRmSBruVWCufm3LsUVMs8v
5 KiB
Performance work
Six performance wins surfaced by the SHA-256 example's inner-loop analysis. Working through them on this branch; this file is a scratchpad for the milestone and gets deleted in the final commit once everything is shipped.
The numbers are per 64-byte SHA-256 block on the example, where the baseline is ~550K cycles ≈ 18 NTSC frames per block.
#1 — Skip parameter spill in leaf functions
Status: TODO
Where: src/codegen/ir_codegen.rs
Estimate: ~66K cycles/block, ~2.2 frames
Every function currently opens with LDA $04 / STA <local> to
spill the param transport slots into a per-function RAM slot,
defending against nested calls that would re-clobber $04-$07.
Functions that never JSR from inside their body are leaves —
the spill is dead work for them.
Fix: in IrCodeGen::new, scan each function for an IrOp::Call.
Functions with none get their parameters mapped directly to the
transport slots $04-$07, and gen_function skips emitting the
spill prologue.
#2 — Direct-branch comparisons (drop bool materialization)
Status: TODO
Where: src/codegen/ir_codegen.rs (Lt/Gt/Eq/etc. lowering)
Estimate: ~9K cycles/block, ~0.3 frames
if x < N currently lowers to "compute the result as 0 or 1 in A,
then BEQ on it":
LDA x ; 3
CMP #N ; 2
BCC cmp_t ; 2/3
LDA #0 ; 2
JMP cmp_e ; 3
cmp_t: LDA #1 ; 2
cmp_e: BEQ end ; 3
The canonical 6502 idiom is one CMP + one branch (8 cycles vs.
~16). The IR already gives us the false target label — we just
need to teach gen_op for the comparison ops to branch directly
when their result feeds straight into a conditional.
#3 — Drop dead LDA #imm before INC/DEC
Status: TODO
Where: src/codegen/peephole.rs
Estimate: ~5K cycles/block, ~0.2 frames
i += 1 currently emits:
LDA #1 ; 2 — A is overwritten by the next op
INC ZeroPage(i) ; 5
The LDA #1 is dead. A peephole rule "drop LDA #imm if A is
re-written or never read before the next LDA/branch/RTS" should
catch this. The same rule fires elsewhere (any += const /
-= const that strength-reduced to INC/DEC).
#4 — Specialize rotr_wk per amount (.ne refactor)
Status: TODO
Where: examples/sha256/sha_core.ne
Estimate: ~45K cycles/block, ~1.5 frames
rotr_wk(dst, n) is a generic loop wrapper. Every SHA-256
rotation amount is a compile-time constant (2, 6, 7, 11, 13, 17,
18, 19, 22, 25), so the loop body is wasted work — the compiler
can't see through the runtime n.
Fix: declare one rotr_wk_<N>(dst) per used amount, each calling
the appropriate sequence of byte_rotr_wk and rotr1_wk. The
sigma helpers swap rotr_wk(SIG, 6) → rotr_wk_6(SIG). The
loop wrapper stays available for any future caller that needs a
runtime amount.
#5 — Inline-asm {param} substitution after inline fun splice
Status: DEFERRED
Where: src/ir/lowering.rs::try_inline_call_stmt +
src/codegen/ir_codegen.rs::substitute_asm_vars
Estimate: ~45K cycles/block (potential), ~1.5 frames
Marking primitives inline fun would eliminate JSR + RTS + the
rest of the call apparatus — but the inline-asm {param}
substitution today resolves names against the analyzer's per-
function allocation table, which doesn't see the inline frame.
A spliced cp_wk(32, 28) ends up emitting LDX {dst} against
the caller's scope where dst doesn't exist.
Properly fixing this is non-trivial: substitution needs to be
addressing-mode-aware (immediate #$20 vs. zero-page $27 vs.
absolute) and depends on whether the inline arg is a constant or
a runtime value. Documenting the design here so a future pass can
take it.
The simpler half of the win: at inline expansion time, build a
per-frame map from param-name → arg-temp. Pass that map down to
gen_op's asm handler, which substitutes {name} with the
arg-temp's allocated slot instead of the param's address. For
constant args, fold further to immediates. This is ~150 lines of
Rust; deferring to a follow-up.
#6 — Const-fold r << 2 style index math
Status: TODO
Where: src/optimizer/mod.rs
Estimate: <1K cycles/block, negligible
round_one(r << 2) recomputes the shift on every iteration of the
phased compression driver. The optimizer already folds shifts when
both operands are constant; extending it to fold "shift a constant
by a constant" inside the IR would catch this case. Trivial in
cycles but worth the cleanliness.
After-each-change checklist
For every codegen change, verify:
cargo fmt --checkcargo clippy --all-targets -- -D warningson rustc 1.95.0cargo test --all-targetspasses- Rebuild every committed
examples/*.nes - Run the emulator harness — if any golden drifts, eyeball the
diff and update with
UPDATE_GOLDENS=1 - Regenerate
docs/{platformer,war,pong}.gifif any of those examples' captured frames changed
For each .ne change:
- Just the example-side checks (rebuild ROM, harness, gif).