Drew T
2.7 KiB
§399 ★★★ — FOUR LEVERS FROM THE FINAL S69 ROUND (P31 S69; each byte-proven, none previously in the cookbook)
(a) A fence BETWEEN two prologue loads, where source order does nothing. Two loads at the top of
a function emitted in the wrong order, and reordering the C statements had no effect — the order
is fixed before statement order matters. A __asm__ __volatile__("") placed between the two
loads (§194-A) swapped them and closed the last 2 instructions.
(byte-proven md_MAIN_013/func_800CB56C, 94 ins)
(b) SINK A CALL INTO BOTH ARMS and let cross_jump keep only the jal suffix. A warm start sat at
88 ins / closeness 86. Sinking func_8017E3E8 into both arms of a >= 0x400 test (so
cross_jump merges only the shared jal tail, §193-C) took it to 92 ins / closeness 13; then §3-T2
field-order on the two zero-stores — deliberately NOT the neighbour's order — let each sh $zero
fill an lhu load-delay slot and closed it. Note the direction: duplicate the call in source so
the compiler merges it, rather than writing the merged form yourself.
(byte-proven ov_SC07_001/func_8017EDC0, 92 ins)
(c) A $v0 → $a0 → $s3 DOUBLE COPY IS A TWO-PSEUDO TELL. Not a scheduling artifact — it means
the source has two distinct variables: an SImode temp holding the call result (living in a
caller-saved register for the first compare) and a separate HImode variable for the tail:
s32 t = call(); /* SImode, caller-saved, feeds the compare */
s16 ret = t; /* HImode, $s3, survives to the tail */
That split alone took 70 → 37. And §195-N precondition 5: nesting return 1 inside the if-chain
with exactly ONE trailing return 0; blocks jump.c's store-flag transform (killing
sltiu $v0,$v0,1), so reorg fills both delay slots with addu $v0,$zero,$zero — 18 → 0.
(byte-proven ov_SC02_017/func_8018347C, 94 ins)
(d) THE RE-TIE AS A BIV KILLER. __asm__("" : "=r"(q) : "0"(q)) on a halfword walker gives q a
second set, so n_times_set > 1 and loop.c refuses it as a biv — which kills the combined
address giv that had biased the base to +2 and had sunk $s0's init into the preheader.
combine_givs picks the last-discovered / lowest-offset giv, so denying it the biv is the whole
lever. Measured alternatives: a volatile q was WORSE (8), a dead q[0] read stayed at 6.
This is the third distinct use of the zero-byte re-tie in one session — §380 (un-hoist a
move_movables invariant), §393 (kill the scheduler's birthing boost), and now §399d (deny a biv).
One line, three passes: when a single-set pseudo is being treated specially, give it a second set.
(byte-proven ov_SC07_001/func_8017E4DC, 92 ins)