ARM: issue with constant islands and large switch instructions by xavierleroy · Pull Request #1022 · ocaml/ocaml

xavierleroy · 2017-01-27T14:44:27Z

This is an alternate fix for the issue reported in #994 .

Additionally, it fixes another potential issue with switch instructions whereas the size of the instruction encoding could have been underestimated.

… a new constant island As reported in GPR#994, the previous heuristic for starting a new constant island could start it too late if the instruction just emitted produces a lot of asm code. This is typically the case for Lswitch instruction with a large jump table. This commit checks for new constant island before emitting the instruction, and takes into account (an upper bound for) the size of the instruction's code. Also: report a more accurate size for the encoding of Lswitch in Thumb2 mode. The one computed before overestimated the size most of the time, but could also underestimate in some cases.

xavierleroy · 2017-02-21T10:00:36Z

We need to decide quicky whether this goes in 4.05 or is kept for future work. @mshinwell : an opinion?

mshinwell · 2017-02-21T10:42:27Z

I'm planning to read this one soon, so we should be able to get this into 4.05.

mshinwell · 2017-03-07T09:43:55Z

asmcomp/arm/emit.mlp

+    emit_all (ninstr' + n) (has_fallthrough i.desc) i.next
  end

+


Unrelated whitespace change

mshinwell · 2017-03-07T09:44:07Z

asmcomp/arm/emit.mlp

  end;
  `{emit_label !tailrec_entry_point}:\n`;
-  emit_all 0 fundecl.fun_body;
+  emit_all 0 true fundecl.fun_body;


I think labelling ~fallthrough would be clearer.

mshinwell · 2017-03-07T09:44:38Z

asmcomp/arm/emit.mlp

-      emit_all 0 i.next
-    end else
-      emit_all ninstr' i.next
+    let limit = limit - !num_literals - max_instruction_size i in


The name num_literals is confusing since it sounds like, well, the number of literals---but actually it's the number of words they occupy. Could we rename this?

mshinwell · 2017-03-07T09:46:37Z

asmcomp/arm/emit.mlp

+      then 1 + (Array.length jumptbl + 1) / 2 + Array.length jumptbl
+      else 2 + Array.length jumptbl
+  | _ ->
+      4   (* conservative upper bound *)


This number doesn't appear to be conservative, e.g. for Iextcall. Could you double-check all the cases?
Maybe there should be an assertion in emit_all that emit_instr doesn't return a number greater than what max_instruction_size returned.

(Aside: problems like this make me increasingly think that the right way to do code emitters is via a DSL, where absolutely everything including layout is described; then it's easier to do things such as measuring.)

…ze_literals Continuation of GPR#1022 and commit 19757e0.

xavierleroy · 2017-03-13T10:03:27Z

I fixed max_instruction_size, erring on the side of caution, and renamed num_literals. This is good to go, in my opinion.

xavierleroy · 2017-03-14T13:27:07Z

The failing Travis check is unrelated (camlp4 issue).

Is this good to go?

mshinwell · 2017-03-15T08:53:27Z

I still reckon there should be an assertion that the max instruction size wasn't an underestimate, but if you don't want to do that, OK.

xavierleroy · 2017-03-20T09:26:11Z

Squashed and merged in trunk, commit 07b28ed .

@poll

b11eea1 flambda-backend: Introduce Import_info (ocaml#1036) bc5b135 flambda-backend: Fix `ocamlobjinfo` on flambda2 .cmx files (ocaml#1029) c8babbd flambda-backend: Compilation_unit optimisations (ocaml#1035) e8d3e22 flambda-backend: Use 4.14.0 opam switch for building (includes upgrading ocamlformat to 0.24.1) (ocaml#1030) eb14a86 flambda-backend: Port PR81 from ocaml-jst (ocaml#1024) 131bc12 flambda-backend: Merge ocaml-jst 2022-12-13 (ocaml#1022) 06c189a flambda-backend: Make stack allocation the default (ocaml#1013) 98debd5 flambda-backend: Initial support for value slots not of value kind (ocaml#946) deb1714 flambda-backend: Add is_last flag to closinfo words (ocaml#938) d07fce1 flambda-backend: Disable poll insertion in Configure (ocaml#967) 0f1ce0e flambda-backend: Regenerate ocaml/configure autoconf 2.69 (instead of 2.71) (ocaml#1012) 27132d8 flambda-backend: Fix for spurious typing error related to expanding through functor arguments (ocaml#997) 724fb68 flambda-backend: Use `Compilation_unit.t` instead of `Ident.t` for globals (ocaml#871) 396d5b8 flambda-backend: Add a test for frametable setup in natdynlinked libraries (ocaml#983) b73ab12 flambda-backend: Fix invocation of `caml_shared_startup` in native dynlink (ocaml#980) 7c7d75a flambda-backend: Fix split_default_wrapper which did not trigger anymore with flambda2 (ocaml#970) 8fb75bd flambda-backend: Port ocaml#11727 and ocaml#11732 (ocaml#965) fdb7987 flambda-backend: Fix include functor issue after 4.14 merge. (ocaml#948) 9745cdb flambda-backend: Print -dprofile/-dtimings output to stdout like 4.12 (ocaml#943) 5f51f21 flambda-backend: Merge pull request ocaml#932 from mshinwell/4.14-upgrade 841687d flambda-backend: Run make alldepend in ocaml/ (ocaml#936) 72a7658 flambda-backend: Remove reformatting changes only in dynlink/dune (preserving PR889 and adjusting to minimise diff) 6d758cd flambda-backend: Revert whitespace changes in dune files, to match upstream c86bf6e flambda-backend: Remove duplicate tests for polling 971dbeb flambda-backend: Testsuite fixes 32f8356 flambda-backend: Topeval fix for symbols patch befea01 flambda-backend: Compilation fixes / rectify merge faults a84543f flambda-backend: Merge ocaml-jst 8e65056 flambda-backend: Merge ocaml-jst 4d70045 flambda-backend: Remove filename from system frametable (amd64) (ocaml#920) 5e57b7d flambda-backend: Bugfix for runtime frame_descr logic for C frames (ocaml#918) 6423d5e flambda-backend: Merge pull request ocaml#914 from mshinwell/merge-ocaml-jst-2022-10-24 ead605c flambda-backend: Add a missing Extract_exception (ocaml#916) c8f1481 flambda-backend: Resolve conflicts and add specialise/specialised attributes to Builtin_attributes cf4d0d3 flambda-backend: Merge fixes (ocaml#21) c2f742f flambda-backend: Re-enable some tests for Flambda2 (ocaml#881) 3d38d13 flambda-backend: Long frames in frametable (ocaml#797) 85aec7b flambda-backend: Add loop attribute to Builtin_attributes c0f16e3 flambda-backend: Compilation fixes 90dea23 flambda-backend: Merge flambda-backend/main 5acc6ea flambda-backend: Fixes after merge e501946 flambda-backend: Merge ocaml-jst 115083b flambda-backend: Merge ocaml-jst 9943b2e flambda-backend: Revert "Revert "Transform tail-recursive functions into recursive continuations (ocaml#893)"" (ocaml#909) ce339f1 flambda-backend: Fix alloc modes and call kinds for overapplications (ocaml#902) e6a317c flambda-backend: Revert "Transform tail-recursive functions into recursive continuations (ocaml#893)" 853c488 flambda-backend: Transform tail-recursive functions into recursive continuations (ocaml#893) 5a977e4 flambda-backend: Fix missing End_region primitives on switch arms (ocaml#898) 7fa7f9d flambda-backend: Add missing dependencies to Dune files (ocaml#889) 3cd36f0 flambda-backend: Have Lambda `Pgetglobal` and `Psetglobal` take `Compilation_unit.t` (ocaml#896) 7565915 flambda-backend: [@poll error] attribute (ocaml#745) 9eb9448 flambda-backend: Backport the main safepoints PRs (ocaml#740) 689bdda flambda-backend: Add strict mode for ocamldep (ocaml#892) git-subtree-dir: ocaml git-subtree-split: b11eea1

xavierleroy mentioned this pull request Jan 27, 2017

ARM: fix excessive immediate values for pc-relative ldr #994

Closed

mshinwell requested changes Mar 7, 2017

View reviewed changes

whitequark mentioned this pull request Mar 7, 2017

Add iOS support #1084

Closed

ARM emitter: fix max_instruction_size and rename num_literals into si…

c3b532d

…ze_literals Continuation of GPR#1022 and commit 19757e0.

mshinwell approved these changes Mar 15, 2017

View reviewed changes

mshinwell added the approved label Mar 15, 2017

xavierleroy closed this Mar 20, 2017

xavierleroy deleted the arm-literals branch October 3, 2017 13:44

sadiqj pushed a commit to sadiqj/ocaml that referenced this pull request Feb 21, 2023

flambda-backend: Merge ocaml-jst 2022-12-13 (ocaml#1022)

131bc12

stedolan pushed a commit to stedolan/ocaml that referenced this pull request Mar 21, 2023

Merge ocaml-jst 2022-12-13 (ocaml#1022)

004de82

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

ARM: issue with constant islands and large switch instructions#1022

ARM: issue with constant islands and large switch instructions#1022
xavierleroy wants to merge 2 commits intotrunkfrom
arm-literals

xavierleroy commented Jan 27, 2017

Uh oh!

xavierleroy commented Feb 21, 2017

Uh oh!

mshinwell commented Feb 21, 2017

Uh oh!

mshinwell Mar 7, 2017

Uh oh!

mshinwell Mar 7, 2017

Uh oh!

mshinwell Mar 7, 2017

Uh oh!

mshinwell Mar 7, 2017

Uh oh!

xavierleroy commented Mar 13, 2017

Uh oh!

xavierleroy commented Mar 14, 2017

Uh oh!

mshinwell commented Mar 15, 2017

Uh oh!

xavierleroy commented Mar 20, 2017

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

xavierleroy commented Jan 27, 2017

Uh oh!

xavierleroy commented Feb 21, 2017

Uh oh!

mshinwell commented Feb 21, 2017

Uh oh!

mshinwell Mar 7, 2017

Choose a reason for hiding this comment

Uh oh!

mshinwell Mar 7, 2017

Choose a reason for hiding this comment

Uh oh!

mshinwell Mar 7, 2017

Choose a reason for hiding this comment

Uh oh!

mshinwell Mar 7, 2017

Choose a reason for hiding this comment

Uh oh!

xavierleroy commented Mar 13, 2017

Uh oh!

xavierleroy commented Mar 14, 2017

Uh oh!

mshinwell commented Mar 15, 2017

Uh oh!

xavierleroy commented Mar 20, 2017

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants