Skip to content

Commit a7758a8

Browse files
jvanburenjvanburen
committed
Refactored shift optimizations
1 parent 0822f30 commit a7758a8

File tree

1 file changed

+119
-48
lines changed

1 file changed

+119
-48
lines changed

‎backend/cmm_helpers.ml

+119-48
Original file line numberDiff line numberDiff line change
@@ -385,46 +385,110 @@ let ignore_low_bit_int = function
385385
| Cop (Cor, [c; Cconst_int (1, _)], _) -> c
386386
| c -> c
387387

388+
let is_defined_shift n = 0 <= n && n < arch_bits
389+
390+
let[@inline] get_const = function
391+
| Cconst_int (i, _) -> Some (Nativeint.of_int i)
392+
| Cconst_natint (i, _) -> Some i
393+
| _ -> None
394+
395+
let or_int c1 c2 dbg =
396+
match get_const c1, get_const c2 with
397+
| Some c1, Some c2 -> natint_const_untagged dbg (Nativeint.logor c1 c2)
398+
| _, Some 0n | Some -1n, _ -> c1
399+
| Some 0n, _ | _, Some -1n -> c2
400+
| Some _, None ->
401+
(* prefer putting constants on the right *)
402+
Cop (Cor, [c2; c1], dbg)
403+
| _, _ -> Cop (Cor, [c1; c2], dbg)
404+
405+
let and_int c1 c2 dbg =
406+
match get_const c1, get_const c2 with
407+
| Some c1, Some c2 -> natint_const_untagged dbg (Nativeint.logand c1 c2)
408+
| _, Some 0n | Some -1n, _ -> c2
409+
| Some 0n, _ | _, Some -1n -> c1
410+
| Some _, None ->
411+
(* prefer putting constants on the right *)
412+
Cop (Cand, [c2; c1], dbg)
413+
| _, _ -> Cop (Cand, [c1; c2], dbg)
414+
415+
let xor_int c1 c2 dbg =
416+
match get_const c1, get_const c2 with
417+
| Some c1, Some c2 -> natint_const_untagged dbg (Nativeint.logxor c1 c2)
418+
| _, Some 0n -> c1
419+
| Some 0n, _ -> c2
420+
| Some _, None ->
421+
(* prefer putting constants on the right *)
422+
Cop (Cxor, [c2; c1], dbg)
423+
| _, _ -> Cop (Cxor, [c1; c2], dbg)
424+
425+
let replace x ~with_ =
426+
match x with
427+
| Cconst_int _ | Cconst_natint _ | Cconst_symbol _ | Cvar _ | Ctuple [] ->
428+
with_
429+
| inner -> Csequence (inner, with_)
430+
388431
let rec lsr_int c1 c2 dbg =
389432
match c1, c2 with
390433
| c1, Cconst_int (0, _) -> c1
391-
| Cop (Clsr, [c; Cconst_int (n1, _)], _), Cconst_int (n2, _)
392-
when n1 > 0 && n2 > 0 && n1 + n2 < size_int * 8 ->
393-
Cop (Clsr, [c; Cconst_int (n1 + n2, dbg)], dbg)
394-
| c1, Cconst_int (n, _) when n > 0 ->
395-
Cop (Clsr, [ignore_low_bit_int c1; c2], dbg)
396-
| _ -> Cop (Clsr, [c1; c2], dbg)
434+
| c1, Cconst_int (n, _) when is_defined_shift n -> (
435+
let c1 = ignore_low_bit_int c1 in
436+
match get_const c1 with
437+
| Some x -> natint_const_untagged dbg (Nativeint.shift_right_logical x n)
438+
| None -> (
439+
match c1 with
440+
| Cop (Clsr, [inner; Cconst_int (n', _)], _) when is_defined_shift n' ->
441+
if is_defined_shift (n + n')
442+
then lsr_const inner (n + n') dbg
443+
else replace inner ~with_:(Cconst_int (0, dbg))
444+
| Cop (Cor, [x; ((Cconst_int _ | Cconst_natint _) as y)], _) ->
445+
or_int (lsr_int x c2 dbg) (lsr_int y c2 dbg) dbg
446+
| Cop (Cand, [x; ((Cconst_int _ | Cconst_natint _) as y)], _) ->
447+
and_int (lsr_int x c2 dbg) (lsr_int y c2 dbg) dbg
448+
| Cop (Cxor, [x; ((Cconst_int _ | Cconst_natint _) as y)], _) ->
449+
xor_int (lsr_int x c2 dbg) (lsr_int y c2 dbg) dbg
450+
| c1 -> Cop (Clsr, [c1; c2], dbg)))
451+
| Cop (Clsr, [x; (Cconst_int (n', _) as y)], z), c2 when is_defined_shift n'
452+
->
453+
(* prefer putting the constant shift on the outside to help enable further
454+
peephole optimizations *)
455+
Cop (Clsr, [Cop (Clsr, [x; c2], dbg); y], z)
456+
| c1, c2 -> Cop (Clsr, [c1; c2], dbg)
397457

398458
and asr_int c1 c2 dbg =
399459
match c1, c2 with
400460
| c1, Cconst_int (0, _) -> c1
401-
| c1, Cconst_int (n, _) when 0 < n && n < arch_bits -> (
402-
match ignore_low_bit_int c1 with
403-
| Cop (Casr, [inner; Cconst_int (n', _)], _) when 0 <= n' && n' < arch_bits
404-
->
405-
asr_const inner (Int.min (n + n') (arch_bits - 1)) dbg
406-
| Cop (Clsr, [inner; Cconst_int (n', _)], _) when 0 < n' && n' < arch_bits
407-
->
408-
(* If the inner unsigned shift is guaranteed positive, then we know the
409-
sign bit is 0 and we can weaken this operation to a logical shift *)
410-
if n + n' < arch_bits
411-
then lsr_const inner (n + n') dbg
412-
else Csequence (inner, Cconst_int (0, dbg))
413-
| Cop (Cor, [inner; Cconst_int (x, _)], _) when n <= Sys.int_size ->
414-
let inner = asr_const inner n dbg in
415-
let x = x asr n in
416-
if x = 0 then inner else Cop (Cor, [inner; Cconst_int (x, dbg)], dbg)
417-
| Cop (Clsl, [c; Cconst_int (1, _)], _)
418-
(* some operations always return small enough integers that it is safe and
419-
correct to optimise [asr (lsl x 1) 1] into [x]. *)
420-
when n = 1 && guaranteed_to_be_small_int c ->
421-
c
422-
| Cop (Cor, [inner; Cconst_int (x, _)], _) when n < Sys.int_size ->
423-
let inner = asr_const inner n dbg in
424-
let x = x asr n in
425-
if x = 0 then inner else Cop (Cor, [inner; Cconst_int (x, dbg)], dbg)
426-
| c1' -> Cop (Casr, [c1'; c2], dbg))
427-
| Cop (Casr, [x; (Cconst_int _ as y)], z), c2 ->
461+
| c1, Cconst_int (n, _) when is_defined_shift n -> (
462+
let c1 = ignore_low_bit_int c1 in
463+
match get_const c1 with
464+
| Some x -> natint_const_untagged dbg (Nativeint.shift_right x n)
465+
| None -> (
466+
match c1 with
467+
| Cconst_int (x, _) -> Cconst_int (x asr n, dbg)
468+
| Cconst_natint (x, _) ->
469+
natint_const_untagged dbg (Nativeint.shift_right x n)
470+
| Cop (Casr, [inner; Cconst_int (n', _)], _) when is_defined_shift n' ->
471+
(* saturating add, since the sign bit extends to the left *)
472+
asr_const inner (Int.min (n + n') (arch_bits - 1)) dbg
473+
| Cop (Clsr, [_; Cconst_int (n', _)], _)
474+
when n' > 0 && is_defined_shift n' ->
475+
(* If the argument is guaranteed non-negative, then we know the sign bit
476+
is 0 and we can weaken this operation to a logical shift *)
477+
lsr_const c1 n dbg
478+
| Cop (Clsl, [c; Cconst_int (1, _)], _)
479+
when n = 1 && guaranteed_to_be_small_int c ->
480+
(* some operations always return small enough integers that it is safe
481+
and correct to optimise [asr (lsl x 1) 1] into [x]. *)
482+
c
483+
| Cop (Cor, [x; ((Cconst_int _ | Cconst_natint _) as y)], _) ->
484+
or_int (asr_int x c2 dbg) (asr_int y c2 dbg) dbg
485+
| Cop (Cand, [x; ((Cconst_int _ | Cconst_natint _) as y)], _) ->
486+
and_int (asr_int x c2 dbg) (asr_int y c2 dbg) dbg
487+
| Cop (Cxor, [x; ((Cconst_int _ | Cconst_natint _) as y)], _) ->
488+
xor_int (asr_int x c2 dbg) (asr_int y c2 dbg) dbg
489+
| c1 -> Cop (Casr, [c1; c2], dbg)))
490+
| Cop (Casr, [x; (Cconst_int (n', _) as y)], z), c2 when is_defined_shift n'
491+
->
428492
(* prefer putting the constant shift on the outside to help enable further
429493
peephole optimizations *)
430494
Cop (Casr, [Cop (Casr, [x; c2], dbg); y], z)
@@ -433,16 +497,27 @@ and asr_int c1 c2 dbg =
433497
and lsl_int c1 c2 dbg =
434498
match c1, c2 with
435499
| c1, Cconst_int (0, _) -> c1
436-
| Cop (Clsl, [c; Cconst_int (n1, _)], _), Cconst_int (n2, _)
437-
when n1 > 0 && n2 > 0 && n1 + n2 < size_int * 8 ->
438-
Cop (Clsl, [c; Cconst_int (n1 + n2, dbg)], dbg)
439-
| Cop (Caddi, [c1; Cconst_int (n1, _)], _), Cconst_int (n2, _)
440-
when Misc.no_overflow_lsl n1 n2 ->
441-
add_const (lsl_int c1 c2 dbg) (n1 lsl n2) dbg
442-
| Cop (Cor, [c1; Cconst_int (n1, _)], _), Cconst_int (n2, _)
443-
when Misc.no_overflow_lsl n1 n2 ->
444-
Cop (Cor, [lsl_int c1 c2 dbg; Cconst_int (n1 lsl n2, dbg)], dbg)
445-
| Cop (Clsl, [x; (Cconst_int _ as y)], z), c2 ->
500+
| c1, Cconst_int (n, _) when is_defined_shift n -> (
501+
match get_const c1 with
502+
| Some c1 -> natint_const_untagged dbg (Nativeint.shift_left c1 n)
503+
| None -> (
504+
match c1 with
505+
| Cop (Clsl, [inner; Cconst_int (n', _)], dbg) when is_defined_shift n' ->
506+
if is_defined_shift (n + n')
507+
then lsl_const inner (n + n') dbg
508+
else replace inner ~with_:(Cconst_int (0, dbg))
509+
| Cop (Caddi, [c1; Cconst_int (offset, _)], _)
510+
when Misc.no_overflow_lsl offset n ->
511+
add_const (lsl_int c1 c2 dbg) (offset lsl n) dbg
512+
| Cop (Cor, [x; ((Cconst_int _ | Cconst_natint _) as y)], _) ->
513+
or_int (lsl_int x c2 dbg) (lsl_int y c2 dbg) dbg
514+
| Cop (Cand, [x; ((Cconst_int _ | Cconst_natint _) as y)], _) ->
515+
and_int (lsl_int x c2 dbg) (lsl_int y c2 dbg) dbg
516+
| Cop (Cxor, [x; ((Cconst_int _ | Cconst_natint _) as y)], _) ->
517+
xor_int (lsl_int x c2 dbg) (lsl_int y c2 dbg) dbg
518+
| c1 -> Cop (Clsl, [c1; c2], dbg)))
519+
| Cop (Clsl, [x; (Cconst_int (n', _) as y)], z), c2 when is_defined_shift n'
520+
->
446521
(* prefer putting the constant shift on the outside to help enable further
447522
peephole optimizations *)
448523
Cop (Clsl, [Cop (Clsl, [x; c2], dbg); y], z)
@@ -2020,11 +2095,7 @@ let and_int e1 e2 dbg =
20202095
match e1, e2 with
20212096
| e, m when is_mask32 m -> zero_extend ~bits:32 e ~dbg
20222097
| m, e when is_mask32 m -> zero_extend ~bits:32 e ~dbg
2023-
| e1, e2 -> Cop (Cand, [e1; e2], dbg)
2024-
2025-
let or_int e1 e2 dbg = Cop (Cor, [e1; e2], dbg)
2026-
2027-
let xor_int e1 e2 dbg = Cop (Cxor, [e1; e2], dbg)
2098+
| e1, e2 -> and_int e1 e2 dbg
20282099

20292100
(* Boxed integers *)
20302101

0 commit comments

Comments
 (0)