Previously, despite the names 'enqueue' and 'dequeue', it behaved as
a stack and visited blocks in a LIFO fashion. This interacts badly with
extremely broad CFGs *inside* of a loop (such as a large switch inside
a state machine) where every block updates a different variable.
When encountering such a CFG, the checker visited blocks in essentially
a "depth first" order due to the stack-like behavior of the work list.
Combined with each block updating a different variable, the saturation
logic of the checker caused it to re-traverse blocks [1,N-1] of the
broad CFG inside the loop after traversing block N. These re-traversals
were to propagate the variable values derived from block N. Assuming
approximately the same number of variables as inner blocks exist, the
end result is O(N^2) updates. By making this a queue, we also make the
traversal essentially "breadth-first" across each of the N inner blocks
of the loop. Then all of this state is propagated around to all N inner
blocks of the loop. The result is O(N) updates.
The truth is in the numbers:
Before, gcc.c: 96409 block visits (max: 61546, avg: 591)
After, gcc.c: 69958 block visits (max: 33090, avg: 429)
Before, PR10183: 2540494 block vists (max: 2536495, avg: 37360)
After, PR10183: 137803 block visits (max: 134406, avg: 2026)
The nearly 20x reduction in work for PR10183 corresponds to a roughly
100x speedup in compile time.
I've tested it on all the code I can get my hands on, and I've seen no
slowdowns due to this change. Where I've collected stats, the ammount of
work done is on average less. I'll also commit shortly some synthetic
test cases useful in analyzing the performance of CFG-based warnings.
Submitting this based on Doug's feedback that post-commit review should
be good. Ted, please review! Hopefully this helps compile times until
then.
llvm-svn: 134697
change.
Previously clang was passing the following feature strings to the ARM backend
when CPU is cortex-a8: +neon,-vfp2,-vfp3
This used to work because -vfp2,-vfp3 had no effect after +neon. Now that the
features are controlled by individual bits (with implied hierarchy), the net
effect is all three features will be turned off.
llvm-svn: 134691
argument expansion to use the macro argument source locations as well.
Add a few tests to exercise this. There is still a bit more work needed
here though.
llvm-svn: 134674
instantiation and improve diagnostics which are stem from macro
arguments to trace the argument itself back through the layers of macro
expansion.
This requires some tricky handling of the source locations, as the
argument appears to be expanded in the opposite direction from the
surrounding macro. This patch provides helper routines that encapsulate
the logic and explain the reasoning behind how we step through macros
during diagnostic printing.
This fixes the rest of the test cases originially in PR9279, and later
split out into PR10214 and PR10215.
There is still some more work we can do here to improve the macro
backtrace, but those will follow as separate patches.
llvm-svn: 134660
the normal case.
Before, for this:
$ cat t.c
int test(int x) { return x * 2; }
We would get this:
addl %edi, %edi
jno LBB0_2
## BB#1: ## %overflow
ud2
LBB0_2: ## %nooverflow
movl %edi, %eax
popq %rbp
ret
Now we get this:
addl %edi, %edi
jo LBB0_2
## BB#1: ## %nooverflow
movl %edi, %eax
popq %rbp
ret
LBB0_2: ## %overflow
ud2
<rdar://problem/8283919>
llvm-svn: 134642
clang_codeCompleteGetContexts(), that provides the client with
information about the context in which code completion has occurred
and what kinds of entities make sense as completions at that
point. Patch by Connor Wakamo!
llvm-svn: 134615
where we have an immediate need of a retained value.
As an exception, don't do this when the call is made as the immediate
operand of a __bridge retain. This is more in the way of a workaround
than an actual guarantee, so it's acceptable to be brittle here.
rdar://problem/9504800
llvm-svn: 134605
When a macro instantiation occurs, reserve a SLocEntry chunk with length the
full length of the macro definition source. Set the spelling location of this chunk
to point to the start of the macro definition and any tokens that are lexed directly
from the macro definition will get a location from this chunk with the appropriate offset.
For any tokens that come from argument expansion, '##' paste operator, etc. have their
instantiation location point at the appropriate place in the instantiated macro definition
(the argument identifier and the '##' token respectively).
This improves macro instantiation diagnostics:
Before:
t.c:5:9: error: invalid operands to binary expression ('struct S' and 'int')
int y = M(/);
^~~~
t.c:5:11: note: instantiated from:
int y = M(/);
^
After:
t.c:5:9: error: invalid operands to binary expression ('struct S' and 'int')
int y = M(/);
^~~~
t.c:3:20: note: instantiated from:
\#define M(op) (foo op 3);
~~~ ^ ~
t.c:5:11: note: instantiated from:
int y = M(/);
^
The memory savings for a candidate boost library that abuses the preprocessor are:
- 32% less SLocEntries (37M -> 25M)
- 30% reduction in PCH file size (900M -> 635M)
- 50% reduction in memory usage for the SLocEntry table (1.6G -> 800M)
llvm-svn: 134587
It would add up relative (decomposed) offsets like in getDecomposedSpellingLocSlowCase, but while
it makes sense to preserve the offset among lexed spelling locations, it doesn't make
sense to add anything to the offset of the instantiation location. The instantiation
location will be the same regardless of the relative offset in the tokens that were
instantiated.
This bug didn't actually affect anything because, currently, in practice we never create macro
locations with relative offset greater than 0.
llvm-svn: 134586
rather than a computed std::distance(). At some point I had convinced
myself that these two were different; but as far as I can tell on
re-exampination they aren't, and the number of block IDs is actually
just a count of the blocks in the CFG.
While this removes the primary motivation for guarding all of this with
CollectStats, I have a patch coming up that will almost certainly make
it important again.
llvm-svn: 134552
throw-expressions, such that we don't consider the NRVO when the
non-volatile automatic object comes from outside the innermost try
scope (C++0x [class.copymove]p13). In C++98/03, our ASTs were
incorrect but it didn't matter because IR generation doesn't actually
apply the NRVO here. In C++0x, however, we were moving from an object
when in fact we should have copied from it. Fixes PR10142 /
<rdar://problem/9714312>.
llvm-svn: 134548
The small number of elements was determined by taking the median
file length in clang+llvm and /usr/include on OS X with xcode installed.
llvm-svn: 134496
Special detail is added for uninitialized variable analysis as this has
serious performance problems than need to be tracked.
Computing some of this data is expensive, for example walking the CFG to
determine its size. To avoid doing that unless the stats data is going
to be used, we thread a bit into the Sema object to track whether
detailed stats should be collected or not. This bit is used to avoid
computations whereever the computations are likely to be more expensive
than checking the state of the flag. Thus, counters are in some cases
unconditionally updated, but the more expensive (and less frequent)
aggregation steps are skipped.
With this patch, we're able to see that for 'gcc.c':
*** Analysis Based Warnings Stats:
232 functions analyzed (0 w/o CFGs).
7151 CFG blocks built.
30 average CFG blocks per function.
1167 max CFG blocks per function.
163 functions analyzed for uninitialiazed variables
640 variables analyzed.
3 average variables per function.
94 max variables per function.
96409 block visits.
591 average block visits per function.
61546 max block visits per function.
And for the reduced testcase in PR10183:
*** Analysis Based Warnings Stats:
98 functions analyzed (0 w/o CFGs).
8526 CFG blocks built.
87 average CFG blocks per function.
7277 max CFG blocks per function.
68 functions analyzed for uninitialiazed variables
1359 variables analyzed.
19 average variables per function.
1196 max variables per function.
2540494 block visits.
37360 average block visits per function.
2536495 max block visits per function.
That last number is the somewhat scary one that indicates the problem in
PR10183.
llvm-svn: 134494