Prevent more unnecessary float boxing, especially in if and match#107
Prevent more unnecessary float boxing, especially in if and match#107vbrankov wants to merge 5 commits intoocaml:trunkfrom
if and match#107Conversation
|
This seams correct to me |
|
I'm a bit worried by the compilation of pattern-matching issued from GADTs, that may return distinct types in each branch. The code never decides to unbox when two branches return an integer and a float, but it seems to me that it may unbox when one or several branches return a float, and the other are staticraises. I'm not sure which input source patterns lead to static raises in these positions, but I don't see a clear reason why those couldn't be another GADT clause returning a non-float type. |
|
It is not a problem to unbox if a branch of a match returns something unboxable and another one never returns. The whole expression is not unboxable, because 2 branch returns different kind of values, but the if could be unboxed. |
asmcomp/cmmgen.ml
Outdated
There was a problem hiding this comment.
Any non returning primitive could return Undetermined. Praise for instance. (I am not certain there are no other one)
|
I've added handling of |
|
I'm currently integrating the patch and will merge it shortly. I've renamed Undetermined to No_result, to make it clear that it really means that the sub-expression can never return a value (and so it is fine to apply any kind of unboing to it). And I've also changed the code in the Ulet case to avoid emitting code for the let-body when one knows that the let-bound expression cannot return. |
|
Applied to trunk, commit 16271. |
|
This ticket can be closed. |
Backtrace improvements
c703f5f Incorporate upstream comments into type-variable refactor (ocaml#121) 362ba23 Constrain curry modes to increase along applications (ocaml#108) b1f0cf9 Simplify the extension handling (ocaml#114) 4fd53a1 Remove pat_mode from typedtree (ocaml#105) cf6fcbc Handle attributes on lambdas with locally abstract types (ocaml#120) 5fa80fe Don't track attributes inside attributes for warning 53 (ocaml#115) 8a69777 Handle unclosed `[: ... :]` patterns (via `Generic_array` machinery) (ocaml#117) b0737f4 Add promote-one Makefile target (ocaml#118) c6ad684 Refactoring and fixes around module lookup (ocaml#107) b0a6495 Add documentation for global constructor arguments (ocaml#69) dd79aec Print `nlocal` in the `-d(raw)lambda` output (ocaml#112) 8035026 Fix `nlocal` in the generated Lambda for list comprehensions (ocaml#113) afbcdf0 Immutable arrays (ocaml#47) bfe1490 fix several issues when removing exp_mode (ocaml#110) 8f46060 Better error message for under-applied functions (ocaml#74) 27331d8 Consistently use Lmutvar or Lvar in comprehensions (ocaml#111) 01e965b Skip failing test for now 0131357 Fix test case to use comprehensions_experimental 22a7368 Temporarily disable list comprehensions tests due to locals bug e08377d Make `comprehensions` into `comprehensions_experimental` for now (ocaml#109) 947cf89 List and array comprehensions (ocaml#46) bd9e051 remove exp_mode from typedtree (ocaml#100) a9268d2 Fix misplaced attribute warning when using external parser (and some cleanup) (ocaml#101) 2b33f24 Refactor toplevel local escape check (ocaml#104) ed2aec6 Comment functions exported from TyVarEnv. 87838ba Move new variable creation into TyVarEnv. a3f60ab Encapsulate functions that work with tyvars 43d83a6 Prevent possibility of forgetting to re-widen 2f3dd34 Encapsulate context when narrowing type env't d78ff6d Make immediate64 things mode cross (ocaml#97) aa25ab9 Fix version number (ocaml#94) d01ffa0 Fix .depend file (ocaml#93) 942f2ab Bootstrap (ocaml#92) 05f7e38 Check Menhir version (ocaml#91) 1569b58 Move the CI jobs from 4.12 to 4.14. (ocaml#90) git-subtree-dir: ocaml git-subtree-split: c703f5f
Co-authored-by: Patrick Ferris <[email protected]>
4 participants
This patch prevents unnecessary boxing in more language constructs. Especially important are
ifandmatchblocks, for example:All OCaml tests passed. The test code that shows that the extra allocations are gone is here: