optimize cmm shifts and tags #3669
Conversation
|
Array.length gets compiled worse now, i need to fix that the latter doesn't turn into a |
jvanburen
force-pushed
the
jvb/optimize-cmm-shifts-and-tags
branch
from
3468a12 to
529a6d8
Compare
This will be used by a future pr to simplify casting without impacting performance as much
jvanburen
force-pushed
the
jvb/optimize-cmm-shifts-and-tags
branch
from
2d05f61 to
e1c12d7
Compare
jvanburen
force-pushed
the
jvb/optimize-cmm-shifts-and-tags
branch
from
cd768b1 to
93513c0
Compare
|
array length looks strictly better now, the only problems are in cmm-scalar-type, which i'll fix in yet another PR |
| | Cop (Csubi, [c; Cconst_int (x, _)], _) when Misc.no_overflow_sub n x -> | ||
| add_const c (n - x) dbg | ||
| | c -> Cop (Caddi, [c; Cconst_int (n, dbg)], dbg) | ||
| map_tail1 c ~f:(fun c -> |
There was a problem hiding this comment.
did you encounter cases where the map_tail1 was important ? if so it would be nice to document them (at least some of the general shape/conditions when it can happen)
| | Cop (Caddi, [c1; Cconst_int (n1, _)], _), c2 -> | ||
| add_const (sub_int c1 c2 dbg) n1 dbg | ||
| | c1, c2 -> Cop (Csubi, [c1; c2], dbg) | ||
| map_tail2 c1 c2 ~f:(fun c1 c2 -> |
There was a problem hiding this comment.
same question as above: if only for curiosity/better understanding, what kind of situations lead to map_tail2 being needed/useful ?
| add_const (add_int c1 c2 dbg) n2 dbg | ||
| | c1, Cop (Cor, [c2; Cconst_int (n2, _)], _) | ||
| when can_interchange_add_with_or c2 n2 -> | ||
| add_const (add_int c1 c2 dbg) n2 dbg |
There was a problem hiding this comment.
wouldn't it be simpler in this case to call prefer_add on c1 and c2 before the match ?
| | Cop (Cxor, [x; ((Cconst_int _ | Cconst_natint _) as y)], _) -> | ||
| xor_int (lsl_int x c2 dbg) (lsl_int y c2 dbg) dbg | ||
| | c1 -> Cop (Clsl, [c1; c2], dbg))) | ||
| | Cop (Clsl, [x; (Cconst_int (n', _) as y)], z), c2 |
There was a problem hiding this comment.
| | Cop (Clsl, [x; (Cconst_int (n', _) as y)], z), c2 | |
| | Cop (Clsl, [x; (Cconst_int (n', _) as y)], dbg'), c2 |
this code is already tricky as is without having to remember that z is a debuginfo, contrary to x and y, ^^
| Cop | ||
| (Cor, [asr_int c (Cconst_int (n - 1, dbg)) dbg; Cconst_int (1, dbg)], dbg) | ||
| | c -> incr_int (lsl_int c (Cconst_int (1, dbg)) dbg) dbg | ||
| or_const (asr_const c (n - 1) dbg) 1n dbg |
There was a problem hiding this comment.
Should we check whether is_defined_shift n here ? This code is correct regardless , so it might deserve a comment to explain/explicit that it doesn't matter here.
| or_const (asr_const c (n - 1) dbg) 1n dbg | ||
| | Cop (Clsr, [c; Cconst_int (n, _)], _) when n > 0 -> | ||
| or_const (lsr_const c (n - 1) dbg) 1n dbg | ||
| | c -> incr_int (lsl_const c 1 dbg) dbg | ||
|
|
||
| let untag_int i dbg = | ||
| match i with | ||
| | Cconst_int (n, _) -> Cconst_int (n asr 1, dbg) | ||
| | Cop (Cor, [Cop (Casr, [c; Cconst_int (n, _)], _); Cconst_int (1, _)], _) | ||
| when n > 0 && n < (size_int * 8) - 1 -> |
There was a problem hiding this comment.
| when n > 0 && n < (size_int * 8) - 1 -> | |
| when n > 0 && is_defined_shift (n + 1) -> |
might be simpler and more consistent ?
| or_int (sign_extend ~bits x ~dbg) (sign_extend ~bits y ~dbg) dbg | ||
| | Cop (Cxor, [x; y], _) when is_constant y -> | ||
| xor_int (sign_extend ~bits x ~dbg) (sign_extend ~bits y ~dbg) dbg | ||
| | Cop (((Casr | Clsr) as op), [inner; Cconst_int (n, _)], _) as e | ||
| when 0 <= n && n < arch_bits -> |
There was a problem hiding this comment.
| when 0 <= n && n < arch_bits -> | |
| when (is_defined_shift n) -> |
add some optimizations for shifts and tags in preparation for cmm-scalar-type, which needs these to keep decent performance