Hard-coded operand indices were scattered throughout lowering stages
and layers. It was super bug prone.
------------------------------------------------------------------------
[weak vtables] Place class definitions into anonymous namespaces to prevent weak vtables.
This patch places class definitions in implementation files into anonymous
namespaces to prevent weak vtables. This eliminates the need of providing an
out-of-line definition to pin the vtable explicitly to the file.
------------------------------------------------------------------------
Jack Carter [Tue, 19 Nov 2013 01:25:18 +0000 (01:25 +0000)]
[Mips] Support for MicroMips STO refactoring.
No true functional changes.
Change the "hack" name of emitMipsHackSTOCG to emitSymSTO.
Remove demonstration code in AsmParser for emitMipsHackSTOCG and
emitMipsHackELFFlags. The STO field is in an ELF symbol and is not
an explicit directive. That said, we are missing the compliment call
in AsmParser and that will need to be addressed soon.
XFAIL dummy tests for emitMipsHackELFFlags and emitMipsHackELFFlags.
These will built out with following patches.
Juergen Ributzka [Tue, 19 Nov 2013 00:57:56 +0000 (00:57 +0000)]
[weak vtables] Remove a bunch of weak vtables
This patch removes most of the trivial cases of weak vtables by pinning them to
a single object file. The memory leaks in this version have been fixed. Thanks
Alexey for pointing them out.
Reid Kleckner [Mon, 18 Nov 2013 23:08:12 +0000 (23:08 +0000)]
Revert "COFF: Emit all MCSymbols rather than filtering out some of them"
This reverts commit r190888, to fix PR17967. The original change wasn't
the right way to get @feat.00 into the object file. The right fix is to
make @feat.00 be a global symbol.
Adrian Prantl [Mon, 18 Nov 2013 23:04:38 +0000 (23:04 +0000)]
Debug info: Let LowerDbgDeclare perfom the dbg.declare -> dbg.value
lowering only for load/stores to scalar allocas. The resulting values
confuse the backend and don't add anything because we can describe
array-allocas with a dbg.declare intrinsic just fine.
Paul Robinson [Mon, 18 Nov 2013 21:44:03 +0000 (21:44 +0000)]
The 'optnone' attribute means don't inline anything into this function
(except functions marked always_inline).
Functions with 'optnone' must also have 'noinline' so they don't get
inlined into any other function.
Daniel Sanders [Mon, 18 Nov 2013 13:12:43 +0000 (13:12 +0000)]
[mips] Fix 'ran out of registers' in MIPS32 with FP64 when generating code for (ConstantFP 0.0)
Fixed an inappropriate use of BuildPairF64 when compiling for MIPS32 with FP64
which resulted in an impossible constraint on the register allocation. It now
uses BuildPairF64_64.
test-release.sh: tweak RPATH for the binary packages.
libtool sets RPATH to "$ORIGIN/../lib:/the/directory/where/it/was/built/lib" so that a developper can use the built or the installed version seamlessly. Our binary packages should not have this developer friendly tweak, as the users of the binaries will not have the build tree.
Beside, in case the development tree is a possibly on an automounted share, this can create very bad user experience : they will incur an automount timeout penalty and will get a very bad feeling of llvm/clang's speed.
Alexey Samsonov [Mon, 18 Nov 2013 09:31:53 +0000 (09:31 +0000)]
Revert r194865 and r194874.
This change is incorrect. If you delete virtual destructor of both a base class
and a subclass, then the following code:
Base *foo = new Child();
delete foo;
will not cause the destructor for members of Child class. As a result, I observe
plently of memory leaks. Notable examples I investigated are:
ObjectBuffer and ObjectBufferStream, AttributeImpl and StringSAttributeImpl.
Manman Ren [Mon, 18 Nov 2013 03:19:31 +0000 (03:19 +0000)]
Debug Info Verifier: disable it by default.
Debug info verifier is part of the verifier which is a Function Pass.
Tot currently tries to pull all reachable debug info MDNodes in each function,
which is too time-consuming. The correct fix seems to be separating debug info
verification to its own module pass.
I will disable the debug info verifier until a correct fix is found.
For Bill's testing case, enabling debug info verifier increase compile
time from 11s to 11m.
Manman Ren [Sun, 17 Nov 2013 18:48:57 +0000 (18:48 +0000)]
Debug Info Verifier: fix when to find debug info nodes and when to verify them.
We used to collect debug info MDNodes in doInitialization and verify them in
doFinalization. That is incorrect since MDNodes can be modified by passes run
between doInitialization and doFinalization.
To fix the problem, we handle debug info MDNodes that can be reached from a
function in runOnFunction (i.e we collect those nodes by calling processDeclare,
processValue and processLocation, and then verify them in runOnFunction).
We handle debug info MDNodes that can be reached from named metadata in
doFinalization. This is in line with how Verifier handles module-level data
(they are verified in doFinalization).
Manman Ren [Sun, 17 Nov 2013 18:42:37 +0000 (18:42 +0000)]
Debug Info Verifier: enable public functions of Finder to update the type map.
We used to depend on running processModule before the other public functions
such as processDeclare, processValue and processLocation. We are now relaxing
the constraint by adding a module argument to the three functions and
letting the three functions to initialize the type map. This will be used in
a follow-on patch that collects nodes reachable from a Function.
Hal Finkel [Sun, 17 Nov 2013 16:02:50 +0000 (16:02 +0000)]
Add a loop rerolling flag to the PassManagerBuilder
This adds a boolean member variable to the PassManagerBuilder to control loop
rerolling (just like we have for unrolling and the various vectorization
options). This is necessary for control by the frontend. Loop rerolling remains
disabled by default at all optimization levels.
python: Properly initialize before trying to create disasm
As the "LLVMInitializeAll*" functions are not available as symbols in
the shared library they can't be used, and as a workaround a list of
the targets is kept and the individual symbols tried. As soon as the
"All"-functions are changed to proper symbols (as opposed to static
inlines in the headers) this hack will be replace with simple calls
to the corresponding "LLVMInitializeAll*" functions.
[block-freq] Add BlockFrequency::scale that returns a remainder from the division and make the private scale in BlockFrequency more performant.
This change is the first in a series of changes improving LLVM's Block
Frequency propogation implementation to not lose probability mass in
branchy code when propogating block frequency information from a basic
block to its successors. This patch is a simple infrastructure
improvement that does not actually modify the block frequency
algorithm. The specific changes are:
1. Changes the division algorithm used when scaling block frequencies by
branch probabilities to a short division algorithm. This gives us the
remainder for free as well as provides a nice speed boost. When I
benched the old routine and the new routine on a Sandy Bridge iMac with
disabled turbo mode performing 8192 iterations on an array of length
32768, I saw ~600% increase in speed in mean/median performance.
2. Exposes a scale method that returns a remainder. This is important so
we can ensure that when we scale a block frequency by some branch
probability BP = N/D, the remainder from the division by D can be
retrieved and propagated to other children to ensure no probability mass
is lost (more to come on this).
Chandler Carruth [Sun, 17 Nov 2013 03:18:05 +0000 (03:18 +0000)]
[PM] Completely remove support for explicit 'require' methods on the
AnalysisManager. All this method did was assert something and we have
a perfectly good way to trigger that assert from the query path.
Hal Finkel [Sun, 17 Nov 2013 02:06:35 +0000 (02:06 +0000)]
Add the cold attribute to error-reporting call sites
Generally speaking, control flow paths with error reporting calls are cold.
So far, error reporting calls are calls to perror and calls to fprintf,
fwrite, etc. with stderr as the stream. This can be extended in the future.
The primary motivation is to improve block placement (the cold attribute
affects the static branch prediction heuristics).
Andrew Trick [Sun, 17 Nov 2013 01:36:23 +0000 (01:36 +0000)]
Added a size field to the stack map record to handle subregister spills.
Implementing this on bigendian platforms could get strange. I added a
target hook, getStackSlotRange, per Jakob's recommendation to make
this as explicit as possible.
for (int i = 0; i < 3200; ++i) {
a[i] += alpha * b[i];
}
and loops like this:
for (int i = 0; i < 500; ++i) {
x[3*i] = foo(0);
x[3*i+1] = foo(0);
x[3*i+2] = foo(0);
}
and turn them into this:
for (int i = 0; i < 1500; ++i) {
x[i] = foo(0);
}
There are two motivations for this transformation:
1. Code-size reduction (especially relevant, obviously, when compiling for
code size).
2. Providing greater choice to the loop vectorizer (and generic unroller) to
choose the unrolling factor (and a better ability to vectorize). The loop
vectorizer can take vector lengths and register pressure into account when
choosing an unrolling factor, for example, and a pre-unrolled loop limits that
choice. This is especially problematic if the manual unrolling was optimized
for a machine different from the current target.
The current implementation is limited to single basic-block loops only. The
rerolling recognition should work regardless of how the loop iterations are
intermixed within the loop body (subject to dependency and side-effect
constraints), but the significant restriction is that the order of the
instructions in each iteration must be identical. This seems sufficient to
capture all current use cases.
This pass is not currently enabled by default at any optimization level.
Hal Finkel [Sat, 16 Nov 2013 21:29:08 +0000 (21:29 +0000)]
Apply the InstCombine fptrunc sqrt optimization to llvm.sqrt
InstCombine, in visitFPTrunc, applies the following optimization to sqrt calls:
(fptrunc (sqrt (fpext x))) -> (sqrtf x)
but does not apply the same optimization to llvm.sqrt. This is a problem
because, to enable vectorization, Clang generates llvm.sqrt instead of sqrt in
fast-math mode, and because this optimization is being applied to sqrt and not
applied to llvm.sqrt, sometimes the fast-math code is slower.
This change makes InstCombine apply this optimization to llvm.sqrt as well.
This fixes the specific problem in PR17758, although the same underlying issue
(optimizations applied to libcalls are not applied to intrinsics) exists for
other optimizations in SimplifyLibCalls.
Owen Anderson [Sat, 16 Nov 2013 00:20:01 +0000 (00:20 +0000)]
Small improvement to InstrinsicEmitter::EmitAttributes. This change removes the “pushing” and “clearing” of the SmallVector and instead uses const arrays to pass the attributeKinds to AttributeSet::get .