If the original FPU specification involved a restricted VFP unit (d16), ensure
that we reset the functionality when we encounter a new FPU type. In
particular, if the user specified vfpv3-d16, but switched to a VFPv3 (which has
32 double precision registers), we would fail to reset the D16 feature, and
treat it as being equivalent to vfpv3-d16.
------------------------------------------------------------------------
The FPU directive permits the user to switch the target FPU, enabling
instructions that would be otherwise unavailable. However, when configuring the
new subtarget features, we would not enable the implied functions for newer
FPUs. This would result in invalid rejection of valid input. Ensure that we
inherit the implied FPU functionality when enabling newer versions of the FPU.
Fortunately, these are mostly hierarchical, unlike the CPUs.
[GVN] don't propagate equality comparisons of FP zero (PR22376)
In http://reviews.llvm.org/D6911, we allowed GVN to propagate FP equalities
to allow some simple value range optimizations. But that introduced a bug
when comparing to -0.0 or 0.0: these compare equal even though they are not
bitwise identical.
This patch disallows propagating zero constants in equality comparisons. Fixes: http://llvm.org/bugs/show_bug.cgi?id=22376
Differential Revision: http://reviews.llvm.org/D7257
Hans Wennborg [Thu, 29 Jan 2015 20:57:17 +0000 (20:57 +0000)]
Merging r227290:
------------------------------------------------------------------------
r227290 | dblaikie | 2015-01-27 18:34:53 -0800 (Tue, 27 Jan 2015) | 1 line
PR22356: DebugInfo: Handle the size of a member where the type of that member is a typedef (or other sugar) of a declaration.
------------------------------------------------------------------------
Fix LLVMSetMetadata and LLVMAddNamedMetadataOperand for single value MDNodes
Summary:
MetadataAsValue uses a canonical format that strips the MDNode if it
contains only a single constant value. This triggers an assertion when
trying to cast the value to a MDNode.
R600/SI: Move i64 -> v2i32 load promotion into AMDGPUDAGToDAGISel::Select()
We used to do this promotion during DAG legalization, but this
caused an infinite loop in ExpandUnalignedLoad() because it assumed
that i64 loads were legal if i64 was a legal type.
It also seems better to report i64 loads as legal, since they actually
are and we were just promoting them to simplify our tablegen files.
Hans Wennborg [Wed, 28 Jan 2015 19:15:34 +0000 (19:15 +0000)]
Merging r227250:
------------------------------------------------------------------------
r227250 | ab | 2015-01-27 13:52:16 -0800 (Tue, 27 Jan 2015) | 31 lines
[SimplifyLibCalls] Don't confuse strcpy_chk for stpcpy_chk.
This was introduced in a faulty refactoring (r225640, mea culpa):
the tests weren't testing the return values, so, for both
__strcpy_chk and __stpcpy_chk, we would return the end of the
buffer (matching stpcpy) instead of the beginning (for strcpy).
The root cause was the prefix "__" being ignored when comparing,
which made us always pick LibFunc::stpcpy_chk.
Pass the LibFunc::Func directly to avoid this kind of error.
Also, make the testcases as explicit as possible to prevent this.
The now-useful testcases expose another, entangled, stpcpy problem,
with the further simplification. This was introduced in a
refactoring (r225640) to match the original behavior.
However, this leads to problems when successive simplifications
generate several similar instructions, none of which are removed
by the custom replaceAllUsesWith.
For instance, InstCombine (the main user) doesn't erase the
instruction in its custom RAUW. When trying to simplify say
__stpcpy_chk:
- first, an stpcpy is created (fortified simplifier),
- second, a memcpy is created (normal simplifier), but the
stpcpy call isn't removed.
- third, InstCombine later revisits the instructions,
and simplifies the first stpcpy to a memcpy. We now have
two memcpys.
[LPM] Rip all of ManagedStatic and ThreadLocal out of the pretty stack
tracing code.
Managed static was just insane overhead for this. We took memory fences
and external function calls in every path that pushed a pretty stack
frame. This includes a multitude of layers setting up and tearing down
passes, the parser in Clang, everywhere. For the regression test suite
or low-overhead JITs, this was contributing to really significant
overhead.
Even the LLVM ThreadLocal is really overkill here because it uses
pthread_{set,get}_specific logic, and has careful code to both allocate
and delete the thread local data. We don't actually want any of that,
and this code in particular has problems coping with deallocation. What
we want is a single TLS pointer that is valid to use during global
construction and during global destruction, any time we want. That is
exactly what every host compiler and OS we use has implemented for
a long time, and what was standardized in C++11. Even though not all of
our host compilers support the thread_local keyword, we can directly use
the platform-specific keywords to get the minimal functionality needed.
Provided this limited trial survives the build bots, I will move this to
Compiler.h so it is more widely available as a light weight if limited
alternative to the ThreadLocal class. Many thanks to David Majnemer for
helping me think through the implications across platforms and craft the
MSVC-compatible syntax.
The end result is *substantially* faster. When running llc in a tight
loop over a small IR file targeting the aarch64 backend, this improves
its performance by over 10% for me. It also seems likely to fix the
remaining regressions seen by JIT users with threading enabled.
This may actually have more impact on real-world compile times due to
the use of the pretty stack tracing utility throughout the rest of Clang
or LLVM, but I've not collected any detailed measurements.
------------------------------------------------------------------------
[LPM] A targeted but somewhat horrible fix to the legacy pass manager's
querying of the pass registry.
The pass manager relies on the static registry of PassInfo objects to
perform all manner of its functionality. I don't understand why it does
much of this. My very vague understanding is that this registry is
touched both during static initialization *and* while each pass is being
constructed. As a consequence it is hard to make accessing it not
require a acquiring some lock. This lock ends up in the hot path of
setting up, tearing down, and invaliditing analyses in the legacy pass
manager.
On most systems you can observe this as a non-trivial % of the time
spent in 'ninja check-llvm'. However, I haven't really seen it be more
than 1% in extreme cases of compiling more real-world software,
including LTO.
Unfortunately, some of the GPU JITs are seeing this taking essentially
all of their time because they have very small IR running through
a small pass pipeline very many times (at least, this is the vague
understanding I have of it).
This patch tries to minimize the cost of looking up PassInfo objects by
leveraging the fact that the objects themselves are immutable and they
are allocated separately on the heap and so don't have their address
change. It also requires a change I made the last time I tried to debug
this problem which removed the ability to de-register a pass from the
registry. This patch creates a single access path to these objects
inside the PMTopLevelManager which memoizes the result of querying the
registry. This is somewhat gross as I don't really know if
PMTopLevelManager is the *right* place to put it, and I dislike using
a mutable member to memoize things, but it seems to work.
For long-lived pass managers this should completely eliminate
the cost of acquiring locks to look into the pass registry once the
memoized cache is warm. For 'ninja check' I measured about 1.5%
reduction in CPU time and in total time on a machine with 32 hardware
threads. For normal compilation, I don't know how much this will help,
sadly. We will still pay the cost while we populate the memoized cache.
I don't think it will hurt though, and for LTO or compiles with many
small functions it should still be a win. However, for tight loops
around a pass manager with many passes and small modules, this will help
tremendously. On the AArch64 backend I saw nearly 50% reductions in time
to complete 2000 cycles of spinning up and tearing down the pipeline.
Measurements from Owen of an actual long-lived pass manager show more
along the lines of 10% improvements.
[LPM] Stop using the string based preservation API. It is an
abomination.
For starters, this API is incredibly slow. In order to lookup the name
of a pass it must take a memory fence to acquire a pointer to the
managed static pass registry, and then potentially acquire locks while
it consults this registry for information about what passes exist by
that name. This stops the world of LLVMs in your process no matter
how little they cared about the result.
To make this more joyful, you'll note that we are preserving many passes
which *do not exist* any more, or are not even analyses which one might
wish to have be preserved. This means we do all the work only to say
"nope" with no error to the user.
String-based APIs are a *bad idea*. String-based APIs that cannot
produce any meaningful error are an even worse idea. =/
I have a patch that simply removes this API completely, but I'm hesitant
to commit it as I don't really want to perniciously break out-of-tree
users of the old pass manager. I'd rather they just have to migrate to
the new one at some point. If others disagree and would like me to kill
it with fire, just say the word. =]
------------------------------------------------------------------------
COMDATs must be identically named to the symbol. When support for COMDATs was
introduced, the symbol rewriter was not updated, resulting in rewriting failing
for symbols which were placed into COMDATs. This corrects the behaviour and
adds test cases for this.
------------------------------------------------------------------------
The rewrite for the pattern based rewrite is unnecessary if the existing name
matches the pattern.
------------------------------------------------------------------------
Summary:
At the moment, address calculation is taking the debug line info from the
address node (e.g. TargetGlobalAddress). When a function is called multiple
times, this results in output of the form:
This patch makes address calculations for function calls take the debug line
info for the call node and results in output of the form:
.loc $first_call_location
.. address calculation ..
.. function call ..
.loc $second_call_location
.. address calculation ..
.. function call ..
.loc $third_call_location
.. address calculation ..
.. function call ..
All other address calculations continue to use the address node.
Test Plan: Fixes test/DebugInfo/multiline.ll on a mips host.
InstCombine: Don't strip bitcasts off of callsites marked 'thunk'
The return type of a thunk is meaningless, we just want the arguments
and return value to be forwarded.
------------------------------------------------------------------------
Make ScalarEvolution less aggressive with respect to no-wrap flags.
ScalarEvolution currently lowers a subtraction recurrence to an add
recurrence with the same no-wrap flags as the subtraction. This is
incorrect because `sub nsw X, Y` is not the same as `add nsw X, -Y`
and `sub nuw X, Y` is not the same as `add nuw X, -Y`. This patch
fixes the issue, and adds two test cases demonstrating the bug.
We were passing the scratch buffer address to the shaders via user sgprs,
but now we use external symbols and have the driver patch the shader
using reloc information.
R600/SI: Don't store scratch buffer frame index in MUBUF offset field
We don't have a good way of legalizing this if the frame index offset
is more than the 12-bits, which is size of MUBUF's offset field, so
now we store the frame index in the vaddr field.
v2: modify hasVALU32BitEncoding instead
v3: - add pseudoToMCOpcode helper to AMDGPUInstInfo, which is used by both
hasVALU32BitEncoding and AMDGPUMCInstLower::lower
- report an error if a pseudo can't be lowered
Fix the C-API MCJIT test for 32-bit big endian machines.
Avoid using unions for storing the return value from
LLVMGetGlobalValueAddress() and LLVMGetFunctionAddress() and accessing it as
a pointer through another pointer member. This causes problems on 32-bit big
endian machines since the pointer gets the higher part of the return value of
the aforementioned functions.
------------------------------------------------------------------------
The bug was introduced in r225282. r225282 assumed that sub X, Y is
the same as add X, -Y. This is not correct if we are going to upgrade
the sub to sub nuw. This change fixes the issue by making the
optimization ignore sub instructions.
This commit moves `MDLocation`, finishing off PR21433. There's an
accompanying clang commit for frontend testcases. I'll attach the
testcase upgrade script I used to PR21433 to help out-of-tree
frontends/backends.
This changes the schema for `DebugLoc` and `DILocation` from:
Note that empty fields (line/column: 0 and inlinedAt: null) don't get
printed by the assembly writer.
------------------------------------------------------------------------
IR: Drop metadata references more aggressively during teardown
Sometimes teardown happens before the debug info graph is complete
(e.g., when clang throws an error). In that case, `MDNode`s will still
have RAUW, so deleting constants that the `MDNode`s point at will be
relatively expensive -- it'll cause re-uniquing all up the chain (what
I've been referring to as "teardown madness").
So, drop references *before* deleting constants. We need to drop a few
more references now: the metadata side of the metadata/value bridges
needs to be dropped off the cliff along with the rest of it (previously,
the bridges were cleaned before we did anything with the `MDNode`s).
There's no real functionality change here -- state before and after
`LLVMContextImpl::~LLVMContextImpl()` is unchanged -- so no testcase.
------------------------------------------------------------------------
Happened pretty commonly during `LLVMContext` teardown when `clang -g`
hit an error. This fixes the use-after-free. Next I'll clean up
teardown so that it's not RAUW'ing when metadata-tracked values are
deleted (only really causes a problem if the graph is mid-construction
when teardown starts, but it's still unnecessary work).
------------------------------------------------------------------------
Override the TLI callback enableAggressiveFMAFusion and return true. Indeed, fmul, fmadd and fadd nodes cost the same number of cycles, so we can enable more combining heuristics to produce more fmadd nodes.
Erik Eckstein [Wed, 14 Jan 2015 11:24:47 +0000 (11:24 +0000)]
reapply: SLPVectorizer: Cache results from memory alias checking.
This speeds up the dependency calculations for blocks with many load/store/call instructions.
Beside the improved runtime, there is no functional change.
Compared to the original commit, this re-applied commit contains a bug fix which ensures that there are
no incorrect collisions in the alias cache.
Chandler Carruth [Wed, 14 Jan 2015 11:23:27 +0000 (11:23 +0000)]
[cleanup] Re-sort all the #include lines in LLVM using
utils/sort_includes.py.
I clearly haven't done this in a while, so more changed than usual. This
even uncovered a missing include from the InstrProf library that I've
added. No functionality changed here, just mechanical cleanup of the
include order.
Chandler Carruth [Wed, 14 Jan 2015 10:33:21 +0000 (10:33 +0000)]
[dom] Make the DominatorTreeBase not a dynamic class!
Now that the passes are wrappers around this, we no longer need
a vtable, virtual destructor, and other associated mess. This is
particularly nice to me as this is a class template. =]
Chandler Carruth [Wed, 14 Jan 2015 10:19:28 +0000 (10:19 +0000)]
[PM] Port domtree to the new pass manager (at last).
This adds the domtree analysis to the new pass manager. The analysis
returns the same DominatorTree result entity used by the old pass
manager and essentially all of the code is shared. We just have
different boilerplate for running and printing the analysis.
I've converted one test to run in both modes just to make sure this is
exercised while both are live in the tree.
This commit refines the pattern for the octeon seq/seqi/sne/snei instructions.
The target register is set to 0 or 1 according to the result of the comparison.
In C, this is something like
rd = (unsigned long)(rs == rt)
This commit adds a zext to bring the result to i64. With this change the
instruction is selected for this type of code. (gcc produces the same code for
the above C code.)
Chandler Carruth [Wed, 14 Jan 2015 10:07:19 +0000 (10:07 +0000)]
[PM] Make DominatorTrees (corectly) movable so that we can move them
into the new pass manager's analysis cache which stores results
by-value.
Technically speaking, the dom trees were originally not movable but
copyable! This, unsurprisingly, didn't work at all -- the copy was
shallow and just resulted in rampant memory corruption. This change
explicitly forbids copying (as it would need to be a deep copy) and
makes them explicitly movable with the unsurprising boiler plate to
member-wise move them because we can't rely on MSVC to generate this
code for us. =/
If there is no associated immediate (MS style inline asm), do not try to access
the operand, assume that it is valid. This should fix the buildbots after SVN
r225941.
Mehdi Amini [Wed, 14 Jan 2015 05:33:01 +0000 (05:33 +0000)]
Fold a loop for array processing in ComputeLinearIndex
When processing an array, every Elt has the same layout, it is
useless to recursively call each ComputeLinearIndex on each element.
Just do it once and multiply by the number of elements.
NVPTX: Use MapMetadata() instead of custom/stale/untested logic
Copy the `GVMap` over to a standard `ValueToValueMapTy` so that we can
reuse the `MapMetadata()` logic. Unfortunately the `GVMap` can't just
be replaced, since `MapMetadata()` likes to modify the map, but at least
this will prevent NVPTX from bitrotting.
Chandler Carruth [Wed, 14 Jan 2015 03:34:55 +0000 (03:34 +0000)]
[dom] Add a basic dominator tree test.
Correct, we have *zero* basic testing of the dominator tree in the
regression test suite. There is a single test that even prints it out,
and that test only checks a single line of the output. There are
a handful of tests that check post dominators, but all of those are
looking for bugs rather than just exercising the basic machinery.
This test is super boring and unexciting. But hey, it's something.
I needed there to be something so I could switch the basic test to run
with both the old and new pass manager.
Hao Liu [Wed, 14 Jan 2015 03:02:16 +0000 (03:02 +0000)]
Fix a wrong comment in LoopVectorize.
I.E. more than two -> exactly two
Fix a typo function name in LoopVectorize.
I.E. collectStrideAcccess() -> collectStrideAccess()
Hal Finkel [Wed, 14 Jan 2015 01:37:21 +0000 (01:37 +0000)]
[PowerPC] Fix the noop-insert test
The form of nops used is CPU-specific (some CPUs, such as the POWER7, have
special group-terminating nops). We probably want a different callback for this
kind of nop insertion (something more like MCAsmBackend::writeNopData), or for
PPC to use a different mechanism for scheduling nops, but this will stop the
test from failing for now.
Matt Arsenault [Wed, 14 Jan 2015 01:35:22 +0000 (01:35 +0000)]
R600/SI: Fix bad code with unaligned byte vector loads
Don't do the v4i8 -> v4f32 combine if the load will need to
be expanded due to alignment. This stops adding instructions
to repack into a single register that the v_cvt_ubyteN_f32
instructions read.
Matt Arsenault [Wed, 14 Jan 2015 01:35:17 +0000 (01:35 +0000)]
Implement new way of expanding extloads.
Now that the source and destination types can be specified,
allow doing an expansion that doesn't use an EXTLOAD of the
result type. Try to do a legal extload to an intermediate type
and extend that if possible.
This generalizes the special case custom lowering of extloads
R600 has been using to work around this problem.
This also happens to fix a bug that would incorrectly use more
aligned loads than should be used.
projects / llvm / log