ViewVC logotype

Annotation of /branches/eth/eve/Src/Eiffel/API/evaluated_type/real_a.e

Parent Directory Parent Directory | Revision Log Revision Log

Revision 72419 - (hide annotations)
Sat Feb 23 00:45:27 2008 UTC (11 years, 11 months ago) by manus
Original Path: trunk/Src/Eiffel/API/evaluated_type/real_32_a.e
File size: 2812 byte(s)
Changed SUPPLIER_AS to only create the supplier on demands.

Merged TYPE_I and TYPE_A hierarchies in just TYPE_A one.
1-The major change is that now to compute the associated class type (CLASS_TYPE instance) of a TYPE_A
  one has to provide a context in which the TYPE_A instance appear. If no context is provided, then
  formal generics are not replaced. For example: LIST [G] alone will simply give you the LIST [G]
  CLASS_TYPE, but if you do it in the context of TEST [INTEGER], then it will return the
2-Updated all callers accordingly.
3-Simplified a few calls where before we did `type.associated_class_type.associated_class' to just
4-Added new preconditions to the context provided, it really helped finding out when we were not doing
  things properly especially in code generation when regenerating code from parent into descendant.
5-Eventhough TYPE_I disappeared, I kept TYPE_C and its descendants since they are useful to have
  them for C code generation and quick discimination for TYPE_A instances.
6-Made validity of types better controled by checking more things in `{TYPE_A}.is_valid'. For example
  that if type used to represent a class which was expanded and that it is not expanded anymore, then
  the type is not valid anymore.
7-The most problematic change with TYPE_I/TYPE_A was that they had different notion of `is_external'
  we have added `is_true_external' to simplify the code in certain areas where the old TYPE_I semantic
  was needed.

Code generation change:
1-We now assume that the default context is always `{BYTE_CONTEXT}.context_class_type'. Then in
  `real_type' and `creation_type' we use `{TYPE_A}.evaluated_type_in_descendant' to properly evaluate
  a type from the ancestor code in the current context. Very useful for regeneration of inherited
  assertions and replicated features. Before it worked, but it was not clear how to do things properly,
  now it is clear.
2-Changed the way we generate the type information for like arguments: we simply nest calls to resolve
  the type of the arguments. This affects the interpreter and C code generation which includes a level
  ID for computing the type of generics.

C Code generation change:
1-Because debugging was too difficult at some point during that work and also because the name mapping
  we were using limited us to 32767 types and 131071 routines, I've changed the mechanism to not use
  the `encode.c' modules in C/bench and did everything in Eiffel. So now we simply write for a feature
  name FXXX_YYY where XXX is the type ID and YYY the body index. The same kind of changes have been
  done for other names (See ENCODER class for details.).
2-We also use `type_id' instead of `static_type_id' for names. The reason it is safe to do so is because
  `type_id' never changes in workbench mode and that in finalized mode, even if you have different
  incremental recompilation, 2 projects should generate more or less the same code at the ID levels
  making it easier to compare them.
3-Changed the way type are created. Instead of having a CREATE_INFO instance in the TYPE_I objects
  used to properly generate the type, we simply use the original TYPE_A instance. That way when we
  have an anchor, we either generate its `actual_type' or its anchor spec depending on the value of
  `use_info. We still use CREATE_INFO to generate the type information but instead of being an object
  in each TYPE_I, it is merely a once that we reinitialize each time we need it.
4-ONE CRITICAL ASPECT OF THE CHANGE #3 was that storable depends on the way type descriptions are generated
  and thanks to eweasel test#storable013 I was able to catch this much earlier in the process.
5-Generates less polymorphic tables for attributes and types specification. What happened before was
  that when we needed a polymorphic type specification for example to create `like x' where x is covariantly
  redefined, we were generating at the same time the polymorphic table for `x' even though if `x'
  was not actually used in the system. This reduces by a tiny amount only the size of the executable.
6-Made the DESCRIPTOR entries much smaller than they used to be by computing the number of ancestors
  instead of allocating for the numbers of routines which was really silly.

IL code generation change:
1-Provided a implementation for creation of BIT constants in .NET however it does not completely work yet,
  it has to do with the manner we generate the BIT_REF class.
2-Changed the way we compare signatures, we store a CLASS_TYPE and a routine ID instead. And then when
  needed we refetch the FEATURE_I object to perform the signature comparions. The issue is that the previous
  solution would not work if NATIVE_ARRAY and TYPED_POINTER do not record all the possible genereric
  derivations in which they are present which is now the case for simplicity.
3-Fixed a bug in code generation for a TYPE_FEATURE_I that is instantiated as a basic type. We would
  generate its associated reference type instead of the basic type. It was not caught before because those
  routines where never called at runtime unless you had a formal generic parameter creation constraint.
4-Fixed an inconsistency shown by eweasel test#incr168 where a precondition check was generated eventhough
  none was needed because one inherited routine had a True precondition. The C code was doing it ok, but
  not .NET
5-Found out see eweasel test#final041 that we first generate inherited assertion and then precondition,
  which is the contrary of what is done in melted or C code generation.
6-Simplified creation of manifest ARRAYS and TUPLE by avoiding a local variable.

1-Updated the code accordingly to the TYPE_I/TYPE_A merge
2-Fixed bug in {DBG_EVALUATOR}.prepare_parameters so that we do not use BYTE_CONTEXT to resolve
  the type but `instantiation_in' instead. It seems to solve one case that the previous debugger did
  not handle (See updated comment).
3-Fixed precondition violation in
  {AST_DEBUGGER_EXPRESSION_CHECKER_GENERATOR}.expression_or_instruction_type_check_and_code where
  we were still trying to evaluate in the descendant if the parent evaluation failed. Causing some
  precondition violations since the AST was not completely annotated with IDs needed for resolution.

Bug fixes:
1-Fixed issue with the way we maintained {CLASS_C}.generic_features, because you could rebuild the list
  without recompiling descendants (case of modifying only a contract of the class) and the rebuilding
  would yield new routine IDs. Then during either type checking or code generation the evaluation of 
  `{TYPE_A}.evaluate_type_in_descendant' could fail because it is not find the new routine ID in the
  descendant class. Now we preserve the routine ID as long as we can, same with feature ID.
2-Made `feature_of_rout_id' works to find the invariant feature.
3-Made sure that when removing an ID from a server it is also removed from the TMP server. Added
  a `not_present' postcondition after calling `remove' to ensure that.

1 dinov 3375 indexing
2 manus 47234 description: "Actual type for real 32 bits type."
3 manus 56535 legal: "See notice at end of class."
4     status: "See notice at end of class."
5 manus 9656 date: "$Date$"
6 dinov 3375 revision: "$Revision $"
8 manus 47234 class REAL_32_A
9 grator 18
10     inherit
11     BASIC_A
12     redefine
13 manus 47234 is_real_32, associated_class, same_as, is_numeric,
14 manus 72419 default_create, process, heaviest
15 grator 18 end
17 manus 41732 create
18     default_create
20     feature {NONE} -- Initialization
22     default_create is
23 manus 47234 -- Initialize new instance of REAL_32_A.
24 manus 41732 do
25     make (associated_class.class_id)
26     end
28 manus 57234 feature -- Visitor
30     process (v: TYPE_A_VISITOR) is
31     -- Process current element.
32     do
33     v.process_real_32_a (Current)
34     end
36 dinov 3078 feature -- Property
38 manus 47234 is_real_32: BOOLEAN is True
39     -- Is the current type a real 32 bits type ?
40 dinov 3078
41 manus 72419 feature -- Access
43 manus 11302 associated_class: CLASS_C is
44     -- Class REAL
45 dinov 3078 once
46 manus 47234 Result := System.real_32_class.compiled_class
47 manus 9656 end
48 dinov 3078
49 manus 72419 feature -- IL code generation
51     heaviest (other: TYPE_A): TYPE_A is
52     -- `other' if `other' is heavier than Current,
53     -- Current otherwise.
54     do
55     if other.is_real_64 then
56     Result := other
57     else
58     Result := Current
59     end
60     end
62 dinov 3181 feature {COMPILER_EXPORTER}
63 grator 18
64 manus 9656 is_numeric: BOOLEAN is True
65 grator 18 -- Is the current type a numeric type ?
67 manus 72419 c_type: REAL_32_I is
68 grator 18 -- C type
69 manus 32079 do
70 manus 47234 Result := real32_c_type
71 manus 9656 end
72 grator 18
73     same_as (other: TYPE_A): BOOLEAN is
74     -- Is the current type the same as `other' ?
75 alexk 71645 local
76     o: REAL_32_A
77 grator 18 do
78 alexk 71645 o ?= other
79     Result := o /= Void and then has_same_attachment_marks (o)
80 manus 9656 end
81 grator 18
82 manus 56535 indexing
83 alexk 71150 copyright: "Copyright (c) 1984-2007, Eiffel Software"
84 manus 58027 license: "GPL version 2 (see http://www.eiffel.com/licensing/gpl.txt)"
85 manus 56535 licensing_options: "http://www.eiffel.com/licensing"
86     copying: "[
87     This file is part of Eiffel Software's Eiffel Development Environment.
88 manus 58027
89 manus 56535 Eiffel Software's Eiffel Development Environment is free
90     software; you can redistribute it and/or modify it under
91     the terms of the GNU General Public License as published
92     by the Free Software Foundation, version 2 of the License
93     (available at the URL listed under "license" above).
94 manus 58027
95 manus 56535 Eiffel Software's Eiffel Development Environment is
96     distributed in the hope that it will be useful, but
97     WITHOUT ANY WARRANTY; without even the implied warranty
99     See the GNU General Public License for more details.
100 manus 58027
101 manus 56535 You should have received a copy of the GNU General Public
102     License along with Eiffel Software's Eiffel Development
103     Environment; if not, write to the Free Software Foundation,
104     Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
105     ]"
106     source: "[
107     Eiffel Software
108     356 Storke Road, Goleta, CA 93117 USA
109     Telephone 805-685-1006, Fax 805-685-6869
110     Website http://www.eiffel.com
111     Customer support http://support.eiffel.com
112     ]"
114 manus 47234 end -- class REAL_32_A


Name Value
svn:eol-style native
svn:keywords Author Date Id Revision

  ViewVC Help
Powered by ViewVC 1.1.23