Assign locations to TFORLOOP loops.

[vext01]

FORLOOP (which yklua already recognises) is for numeric looping, e.g.:

for i = 1, 10; do ...

TFORLOOP is used for other loops, e.g.:

for line in io.lines("fasta1000000.txt") do ...

Peformance impact (vs. main, based on 3 invocations of 10 iterations):

 Benchmark                  Datum0 (ms)  Datum1 (ms)  Ratio  Summary
 revcomp/YkLua/default      2458         1850         0.75   24.75% faster
 knucleotide/YkLua/default  1846         1594         0.86   13.66% faster
 Richards/YkLua/100         4471         4147         0.93   7.24% faster
 LuLPeg/YkLua/default       3357         3252         0.97   3.10% faster
 Json/YkLua/100             2549         2489         0.98   2.37% faster
 Queens/YkLua/1000          504          493          0.98   2.24% faster
 HashIds/YkLua/6000         2873         2829         0.98   1.56% faster
 binarytrees/YkLua/15       3993         3979         1.00   0.36% faster
 Storage/YkLua/1000         9815         9791         1.00   0.24% faster
 CD/YkLua/250               8969         8955         1.00   0.16% faster
 Heightmap/YkLua/2000       749          749          1.00   0.01% faster
 fasta/YkLua/500000         841          841          1.00   0.00% slower
 NBody/YkLua/250000         488          488          1.00   0.04% slower
 spectralnorm/YkLua/1000    902          907          1.01   0.51% slower
 Mandelbrot/YkLua/500       129          130          1.01   0.65% slower
 Sieve/YkLua/3000           468          472          1.01   0.84% slower
 Towers/YkLua/600           908          918          1.01   1.12% slower
 BigLoop/YkLua/1000000000   1617         1641         1.02   1.50% slower
 Havlak/YkLua/1500          18926        19245        1.02   1.69% slower
 Permute/YkLua/1000         790          804          1.02   1.75% slower
 fannkuchredux/YkLua/10     1223         1250         1.02   2.24% slower
 DeltaBlue/YkLua/12000      1949         2000         1.03   2.61% slower
 List/YkLua/1500            832          859          1.03   3.33% slower
 Bounce/YkLua/1500          1142         1186         1.04   3.88% slower

[ltratt]

Why are the slower ones slower?

[vext01]

FWIW, the CI failure is a todo!() in the trace optimiser:

14:21:21 + YKD_SERIALISE_COMPILATION=1 ../src/lua -e_U=true all.lua
14:21:21 
14:21:21 	Starting Tests
14:21:21 random seeds: 1754313681, 574726824
14:21:21 current path:
14:21:21 ****/usr/local/share/lua/5.4/?.lua;/usr/local/share/lua/5.4/?/init.lua;/usr/local/lib/lua/5.4/?.lua;/usr/local/lib/lua/5.4/?/init.lua;./?.lua;./?/init.lua****
14:21:21 
14:21:21     ---- total memory: 50.6K, max memory: 50.6K ----
14:21:21 
14:21:21 time: 5.8e-05 (+5.8e-05)
14:21:21 
14:21:21 ***** FILE 'main.lua'*****
14:21:21 
14:21:21 ***** FILE 'gc.lua'*****
14:21:21 ..testing incremental garbage collection
14:21:21 ..................................................................................................................creating many objects
14:21:21 ....................................................functions with errors
14:21:21 long strings
14:21:21 ..........................steps
14:21:21 steps (2)
14:21:21 ....clearing tables
14:21:21 .................weak tables
14:21:22 ..............................+
14:21:22 .........self-referenced threads
14:21:22 ..........OK
14:21:22     ---- total memory: 1.4M, max memory: 1.4M ----
14:21:22 
14:21:22 time: 0.758009 (+0.757951)
14:21:22 
14:21:22 ***** FILE 'db.lua'*****
14:21:22 testing debug library and debug information
14:21:22 ..............................................................................................................................................
14:21:22 thread '<unnamed>' panicked at ykrt/src/compile/jitc_yk/opt/analyse.rs:73:29:
14:21:22 not yet implemented
14:21:22 note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace
14:21:22 
14:21:22 thread '<unnamed>' panicked at ykrt/src/job_queue.rs:122:44:
14:21:22 called `Result::unwrap()` on an `Err` value: Any { .. }
14:21:22 fatal runtime error: failed to initiate panic, error 5, aborting
14:21:22 Aborted (core dumped)

I will look into the slowdowns.

[ltratt]

You can comment that todo out and turn it into a FIXME. A "proper" fix would improve optimisations, but I don't want to go down that route right now.

[vext01]

Why are the slower ones slower?

Looking to bounce which is the one showing the largest slowdown -- when using multitime instead of rebench, it's difficult to claim that there's a significant performance difference.

One thing that is evident is that there is more variation at the process execution level after my change:

 1: /home/vext01/research/yklua/src/lua harness.lua bounce 10 1500
             Mean        Std.Dev.    Min         Median      Max
-real        11.460      0.152       11.269      11.409      11.692
-user        11.787      0.157       11.578      11.753      12.023
-sys         0.202       0.021       0.160       0.208       0.220
+real        11.471      0.516       10.978      11.151      12.330
+user        11.787      0.512       11.286      11.475      12.628
+sys         0.197       0.012       0.180       0.192       0.216

I'll now fix the trace optimiser todo!

[vext01]

16:57:15 ***** FILE 'math.lua'*****
16:57:15 .testing numbers and math lib
16:57:15 64-bit integers, 53-bit (mantissa) floats
16:57:16 
16:57:16 thread '<unnamed>' panicked at ykrt/src/compile/jitc_yk/opt/mod.rs:758:17:
16:57:16 assertion `left == right` failed
16:57:16   left: 1
16:57:16  right: 64
16:57:16 note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

Will investigate tomorrow.