Execute-prefetch repair throttle · librelane-playground

P144 applies P143’s audit result. Execute-stage prefetch fills still get one adjacent repair word, but no longer get a second repair word by default.

check	result
Verilator build	PASS
BusyBox shell workload reaches `P144-FILE-OK`	PASS
Aux-load queue full drops	PASS
Aux response errors/cancels	PASS
Hardened layout	NOT RUN

metric	P142	P143	P144
post-load cycles	218,856,863	219,787,362	219,161,535
shell window cycles	63,926,691	64,756,820	64,192,833
retired instructions	85,799,045	86,097,554	85,905,855
CPI	2.5508	2.5528	2.5512
S_FETCH cycles	7,598,133	7,607,357	7,601,896
S_MEM cycles	29,199,830	29,346,505	29,215,063

P144 is 563,987 cycles faster than P143 and 28,809 cycles faster than P134, but 266,142 cycles slower than P142. That makes it a useful policy experiment, not the new best shell result.

counter	P143	P144	delta
background repair fills	42,994,352	32,979,525	-10,014,827
first later fetch hits	1,575,088	1,383,118	-191,970
repeat later fetch hits	663,574	606,411	-57,163
demand second-word grants	2,641,195	2,728,297	+87,102
prefetch second-word grants	49,052,366	22,115,553	-26,936,813

The throttle did cut the low-payback traffic, but it also removed enough useful hits to fall behind P142.

class	repair fills	first hits	repeat hits	first+repeat ratio
demand fetch	1,021,344	482,207	331,819	79.70%
execute prefetch	29,021,605	558,785	180,904	2.55%
load prefetch	2,419,665	134,871	15,575	6.22%
writeback prefetch	407,551	170,820	53,952	55.15%
steer prefetch	0	0	0	0.00%
aux prefetch	109,360	36,435	24,161	55.41%
unknown	0	0	0	0.00%

The next rung should bring back execute-prefetch second-word repair only under a tighter condition, rather than enabling it for the whole class.

shell phases label P144 shell workload cycles 219,161,535 cpi 2.55

kernel banner to /init 117,345,818 53.7%
/init to shell banner 1,095,952 0.5%
shell banner to first command 35,898,113 16.4%
echo command 1,649 0%
uname -a 2,449,446 1.1%
ls /bin /usr/share 31,650,054 14.5%
cat sample file 3,003,495 1.4%
touch/write/cat/rm /tmp file 10,595,352 4.9%
8x ash loop with file I/O 16,492,157 7.6%
final marker 680 0%

state breakdown label P144 execute-prefetch repair throttle workload cycles 219,161,535 cpi 2.55

fetch 3.5% 7,601,896
execute 39.2% 85,930,489
mem 13.5% 29,493,185
walker 1.2% 2,682,490
writeback 39.2% 85,905,855
mul/div 3.4% 7,545,904