This motivated implementing a devious clattner inspired solution:-)
This approach uses a small value "Selector" class to point to an IdentifierInfo for the 0/1 case. For multi-keyword selectors, we instantiate a MultiKeywordSelector object (previously known as SelectorInfo). Now, the incremental cost for selectors is only 24,800 for Cocoa.h! This saves 156,592 bytes, or 86%!! The size reduction is also the result of getting rid of the AST slot, which was not strictly necessary (we will associate a selector with it's method using another table...most likely in Sema).
This change was critical to make now, before we have too many clients.
I still need to add some comments to the Selector class...will likely add later today/tomorrow.
llvm-svn: 42452
#1: It is cleaner. I never "liked" storing keyword selectors (i.e. foo:bar:baz) in the IdentifierTable.
#2: It is more space efficient. Since Cocoa keyword selectors can be quite long, this technique is space saving. For Cocoa.h, pulling the keyword selectors out saves ~180k. The cost of the SelectorInfo data is ~100k. Saves ~80k, or 43%.
#3: It results in many API simplifications. Here are some highlights:
- Removed 3 actions (ActOnKeywordMessage, ActOnUnaryMessage, & one flavor of ObjcBuildMethodDeclaration that was specific to unary messages).
- Removed 3 funky structs from DeclSpec.h (ObjcKeywordMessage, ObjcKeywordDecl, and ObjcKeywordInfo).
- Removed 2 ivars and 2 constructors from ObjCMessageExpr (fyi, this space savings has not been measured).
I am happy with the way it turned out (though it took a bit more hacking than I expected). Given the central role of selectors in ObjC, making sure this is "right" will pay dividends later.
Thanks to Chris for talking this through with me and suggesting this approach.
llvm-svn: 42395
integer constant expressions. The only questionable
thing is that we now reject:
void foo() {
switch (1) {
case (int)1.0e10000:
;
}
}
with:
t.c:5:13: error: case label does not reduce to an integer constant
case (int)1.0e10000:
~~~~~^~~~~~~~~
GCC accepts this, emitting the pedwarn:
t.c:5: warning: floating constant exceeds range of 'double'
llvm-svn: 42238
Rationale:
We currently have a separate table to unique ObjC selectors. Since I don't need all the instance data in IdentifierInfo, I thought this would save space (and make more sense conceptually).
It turns out the cost of having duplicate entries for unary selectors (i.e. names without colons) outweighs the cost difference between the IdentifierInfo & SelectorInfo structures. Here is the data:
Two tables:
*** Selector/Identifier Stats:
# Selectors/Identifiers: 51635
Bytes allocated: 1999824
One table:
*** Identifier Table Stats:
# Identifiers: 49500
Bytes allocated: 1990316
llvm-svn: 42139
- Add ObjcMessageExpr AST node and associated constructors.
- Add SourceLocation's to ActOnKeywordMessage/ActOnUnaryMessage API.
- Instantiate message expressions...
- Replace alloca usage with SmallString.
Next step, installing a correct type, among other tweaks...
llvm-svn: 42116
- Added Expr::isConstantExpr().
- Added type checking for InitListExpr elements.
- Added diagnostic for trying to initialize a variable sized object.
llvm-svn: 41674
This fixes the following bug submitted by Neil...
const char *f (void) { return 0; }
...which would incorrectly warn with -pedantic enabled.
llvm-svn: 41559
This fixes the following (recent) regression noticed by Keith Bauer (thanks!).
void func(void *a);
main() {
void *p;
p = 0;
func(0);
}
...which now works as you would expect.
llvm-svn: 41557
eliminates the possibility that the left hand expression is an ImplicitCastExpr.
As a result, I removed the check for ImplicitCastExpr in Expr::isLvalue().
This results in the following AST's...
[dylan:~/llvm/tools/clang] admin% cat fix.c
short x; void test4(char c) {
x += c;
x = x + c;
}
[dylan:~/llvm/tools/clang] admin% ../../Debug/bin/clang fix.c -parse-ast-dump
Read top-level variable decl: 'x'
void test4(char c)
(CompoundStmt 0x2605d30
(CompoundAssignOperator 0x2605c40 'short' '+='
(DeclRefExpr 0x2605c00 'short' Decl='x' 0x2605a80)
(DeclRefExpr 0x2605c20 'char' Decl='c' 0x2605bc0))
(BinaryOperator 0x2605d10 'short' '='
(DeclRefExpr 0x2605c60 'short' Decl='x' 0x2605a80)
(ImplicitCastExpr 0x2605d00 'short'
(BinaryOperator 0x2605ce0 'int' '+'
(ImplicitCastExpr 0x2605cc0 'int'
(DeclRefExpr 0x2605c80 'short' Decl='x' 0x2605a80))
(ImplicitCastExpr 0x2605cd0 'int'
(DeclRefExpr 0x2605ca0 'char' Decl='c' 0x2605bc0))))))
llvm-svn: 41404
This fixes the following bug...
t.c:1:31: error: expression is not assignable
short x; void foo(char c) { x += c; }
This case, among others are now captured in implicit-casts.c.
llvm-svn: 41402
subclass of Stmt will implement child_begin() and child_end(), which will
be used to iterate over all the children (subexpressions/substatements) of
a Stmt object. This will provide for easy traversal over the AST, which
is useful for a variety of purposes.
None of the interfaces to subclasses of Stmt will be changed (other than
adding the child_begin and child_end methods).
The only caveat is that the implementation of subclasses of Stmt will require
colocating all substatements (subexpressions) in an array. This is because
we define child_iterator as Stmt**. All accessor methods to subexpressions
will need to be altered to reflect this new implementation.
This patch includes the typedefs for child_iterator, as well the implementation
for child_begin/child_end for the primary expressions and some postfix
expressions.
llvm-svn: 41363
According to the spec (C++ 5p6[expr]), we need to adjust "T&" to
"T" before further analysis. We do this via the "implicit cast"
thingy.
llvm-svn: 39953
the result type of the expr node.
Implement isIntegerConstantExpr for ImplicitCastExpr nodes the same
was as for CastExpr nodes.
Implement proper sign/zero extension as well as truncation and noop
conversion in the i-c-e evaluator. This allows us to correctly
handle i-c-e's like these:
char array[1024/(sizeof (long))];
int x['\xBb' == (char) 187 ? 1: -1];
this implements test/Sema/i-c-e2.c
llvm-svn: 39888
previous two checkins, which involved lot's of tedious refactoring, this checkin is nice and clean:-)
- Hacked UsualUnaryConversions, UsualArithmeticConversions, and DefaultFunctionArrayConversion
to create the AST node (using a helper function promoteExprToType).
- Added a setType method to Expr.
- Changed Expr::isIntegerConstantExpr to allow for the new node.
llvm-svn: 39866
Chris Lattner