Summary:
Make Scalar Evolution able to propagate NSW and NUW flags from instructions to SCEVs in some cases. This is based on reasoning about when poison from instructions with these flags would trigger undefined behavior. This gives a 13% speed-up on some Eigen3-based Google-internal microbenchmarks for NVPTX.
There does not seem to be clear agreement about when poison should be considered to propagate through instructions. In this analysis, poison propagates only in cases where that should be uncontroversial.
This change makes LSR able to create induction variables for expressions like &ptr[i + offset] for loops like this:
for (int i = 0; i < limit; ++i) {
sum += ptr[i + offset];
}
Here ptr is a 64 bit pointer and offset is a 32 bit integer. For NVPTX, LSR currently creates an induction variable for i + offset instead, which is not as fast. Improving this situation is what brings the 13% speed-up on some Eigen3-based Google-internal microbenchmarks for NVPTX.
There are more details in this discussion on llvmdev.
June: http://lists.cs.uiuc.edu/pipermail/llvmdev/2015-June/thread.html#87234
July: http://lists.cs.uiuc.edu/pipermail/llvmdev/2015-July/thread.html#87392
Patch by Bjarke Roune
Reviewers: eliben, atrick, sanjoy
Subscribers: majnemer, hfinkel, jingyue, meheff, llvm-commits
Differential Revision: http://reviews.llvm.org/D11212
llvm-svn: 243460
The 'common' section TLS is not implemented.
Current C/C++ TLS variables are not placed in common section.
DWARF debug info to get the address of TLS variables is not generated yet.
clang and driver changes in http://reviews.llvm.org/D10524
Added -femulated-tls flag to select the emulated TLS model,
which will be used for old targets like Android that do not
support ELF TLS models.
Added TargetLowering::LowerToTLSEmulatedModel as a target-independent
function to convert a SDNode of TLS variable address to a function call
to __emutls_get_address.
Added into lib/Target/*/*ISelLowering.cpp to call LowerToTLSEmulatedModel
for TLSModel::Emulated. Although all targets supporting ELF TLS models are
enhanced, emulated TLS model has been tested only for Android ELF targets.
Modified AsmPrinter.cpp to print the emutls_v.* and emutls_t.* variables for
emulated TLS variables.
Modified DwarfCompileUnit.cpp to skip some DIE for emulated TLS variabls.
TODO: Add proper DIE for emulated TLS variables.
Added new unit tests with emulated TLS.
Differential Revision: http://reviews.llvm.org/D10522
llvm-svn: 243438
Summary:
If a scale or a base register can be rewritten as "Zext({A,+,1})" then
LSR will now consider a formula of that form in its normal cost
computation.
Depends on D9180
Reviewers: qcolombet, atrick
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D9181
llvm-svn: 243348
Summary:
Was D9784: "Remove loop variant range check when induction variable is
strictly increasing"
This change re-implements D9784 with the two differences:
1. It does not use SCEVExpander and does not generate new
instructions. Instead, it does a quick local search for existing
`llvm::Value`s that it needs when modifying the `icmp`
instruction.
2. It is more general -- it deals with both increasing and decreasing
induction variables.
I've added all of the tests included with D9784, and two more.
As an example on what this change does (copied from D9784):
Given C code:
```
for (int i = M; i < N; i++) // i is known not to overflow
if (i < 0) break;
a[i] = 0;
}
```
This transformation produces:
```
for (int i = M; i < N; i++)
if (M < 0) break;
a[i] = 0;
}
```
Which can be unswitched into:
```
if (!(M < 0))
for (int i = M; i < N; i++)
a[i] = 0;
}
```
I went back and forth on whether the top level logic should live in
`SimplifyIndvar::eliminateIVComparison` or be put into its own
routine. Right now I've put it under `eliminateIVComparison` because
even though the `icmp` is not *eliminated*, it no longer is an IV
comparison. I'm open to putting it in its own helper routine if you
think that is better.
Reviewers: reames, nicholas, atrick
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D11278
llvm-svn: 243331
Now that we are generating sane codegen for vector sext/zext nodes on SSE targets, this patch uses instcombine to replace the SSE41/AVX2 pmovsx and pmovzx intrinsics with the equivalent native IR code.
Differential Revision: http://reviews.llvm.org/D11503
llvm-svn: 243303
The pointer size of the addrspacecasted pointer might not have matched,
so this would have hit an assert in accumulateConstantOffset.
I think this was here to allow constant folding of a load of an
addrspacecasted constant. Accumulating the offset through the
addrspacecast doesn't make much sense, so something else is necessary
to allow folding the load through this cast.
llvm-svn: 243300
Summary:
Fix the cost of interleaved accesses for ARM/AArch64.
We were calling getTypeAllocSize and using it to check
the number of bits, when we should have called
getTypeAllocSizeInBits instead.
This would pottentially cause the vectorizer to
generate loads/stores and shuffles which cannot
be matched with an interleaved access instruction.
No performance changes are expected for now since
matching/generating interleaved accesses is still
disabled by default.
Reviewers: rengolin
Subscribers: aemerson, llvm-commits, rengolin
Differential Revision: http://reviews.llvm.org/D11524
llvm-svn: 243270
r243250 appeared to break clang/test/Analysis/dead-store.c on one of the build
slaves, but I couldn't reproduce this failure locally. Probably a false
positive as I saw this test was broken by r243246 or r243247 too but passed
later without people fixing anything.
llvm-svn: 243253
Summary:
This patch updates TargetTransformInfoImplCRTPBase::getGEPCost to consider
addressing modes. It now returns TCC_Free when the GEP can be completely folded
to an addresing mode.
I started this patch as I refactored SLSR. Function isGEPFoldable looks common
and is indeed used by some WIP of mine. So I extracted that logic to getGEPCost.
Furthermore, I noticed getGEPCost wasn't directly tested anywhere. The best
testing bed seems CostModel, but its getInstructionCost method invokes
getAddressComputationCost for GEPs which provides very coarse estimation. So
this patch also makes getInstructionCost call the updated getGEPCost for GEPs.
This change inevitably breaks some tests because the cost model changes, but
nothing looks seriously wrong -- if we believe the new cost model is the right
way to go, these tests should be updated.
This patch is not perfect yet -- the comments in some tests need to be updated.
I want to know whether this is a right approach before fixing those details.
Reviewers: chandlerc, hfinkel
Subscribers: aschwaighofer, llvm-commits, aemerson
Differential Revision: http://reviews.llvm.org/D9819
llvm-svn: 243250
Summary:
This patch improves trivial loop unswitch.
The current trivial loop unswitch only checks if loop header's terminator contains a trivial unswitch condition. But if the loop header only has one reachable successor (due to intentionally or unintentionally missed code simplification), we should consider the successor as part of the loop header. Therefore, instead of stopping at loop header's terminator, we should keep traversing its successors within loop until reach a *real* conditional branch or switch (whose condition can not be constant folded). This change will enable a single -loop-unswitch pass to unswitch multiple trivial conditions (unswitch one trivial condition could open opportunity to unswitch another one in the same loop), while the old implementation can unswitch only one per pass.
Reviewers: reames, broune
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D11481
llvm-svn: 243203
Add a verifier check that `DILocalVariable`s of tag
`DW_TAG_arg_variable` always have a non-zero 'arg:' field, and those of
tag `DW_TAG_auto_variable` always have a zero 'arg:' field. These are
the only configurations that are properly understood by the backend.
(Also, fix the bad examples in LangRef and test/Assembler, and fix the
bug in Kaleidoscope Ch8.)
A large number of testcases seem to have bitrotted their way forward
from some ancient version of the debug info hierarchy that didn't have
`arg:` parameters. If you have out-of-tree testcases that start failing
in the verifier and you don't care enough to get the `arg:` right, you
may have some luck just calling:
sed -e 's/, arg: 0/, arg: 1/'
or some such, but I hand-updated the ones in tree.
llvm-svn: 243183
This patch extend LoopReroll pass to hand the loops which
is similar to the following:
while (len > 1) {
sum4 += buf[len];
sum4 += buf[len-1];
len -= 2;
}
llvm-svn: 243171
The names for instructions inserted were previous dependent on iteration order. By deriving the names from the original instructions, we can avoid instability in tests without resorting to ordered traversals. It also makes the IR mildly easier to read at large scale.
llvm-svn: 243140
Summary:
Resolving a branch allows us to ignore blocks that won't be executed, and thus make our estimate more accurate.
This patch is intended to be applied after D10205 (though it could be applied independently).
Reviewers: chandlerc
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D10206
llvm-svn: 243084
Summary:
Scalarizer has two data structures that hold information about changes
to the function, Gathered and Scattered. These are cleared in finish()
at the end of runOnFunction() if finish() detects any changes to the
function.
However, finish() was checking for changes by only checking if
Gathered was non-empty. The function visitStore() only modifies
Scattered without touching Gathered. As a result, Scattered could have
ended up having stale data if Scalarizer only scalarized store
instructions. Since the data in Scattered is used during the execution
of the pass, this introduced dangling pointer errors.
The fix is to check whether both Scattered and Gathered are empty
before deciding what to do in finish(). This also fixes a problem
where the Function can be modified although the pass returns false.
Reviewers: rnk
Subscribers: rnk, srhines, llvm-commits
Differential Revision: http://reviews.llvm.org/D10459
llvm-svn: 243040
The MSVC ABI requires that we generate an alias for the vtable which
means looking through a GlobalAlias which cannot be overridden improves
our ability to devirtualize.
Found while investigating PR20801.
Patch by Andrew Zhogin!
Differential Revision: http://reviews.llvm.org/D11306
llvm-svn: 242955
Currently, a load from an alloca that is used in as single block and is not preceded
by a store is replaced by undef. This is not always correct if the single block is
inside a loop.
Fix the logic so that:
1) If there are no stores in the block, replace the load with an undef, as before.
2) If there is a store (regardless of where it is in the block w.r.t the load), bail
out, and let the rest of mem2reg handle this alloca.
Patch by: gil.rapaport@intel.com
Differential Revision: http://reviews.llvm.org/D11355
llvm-svn: 242884
Summary: The current code in LoopUnswtich::processCurrentLoop() mixes trivial loop unswitch and non-trivial loop unswitch together. It goes over all basic blocks in the loop and checks if a condition is trivial or non-trivial unswitch condition. However, trivial unswitch condition can only occur in the loop header basic block (where it controls whether or not the loop does something at all). This refactoring separate trivial loop unswitch and non-trivial loop unswitch. Before going over all basic blocks in the loop, it checks if the loop header contains a trivial unswitch condition. If so, unswitch it. Otherwise, go over all blocks like before but don't check trivial condition any more since they are not possible to be in the other blocks. This code has no functionality change.
Reviewers: meheff, reames, broune
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D11276
llvm-svn: 242873
types and loads, loads or stores widened past the size of an alloca,
etc.
This started off with a bug report about big-endian behavior with
bitfields and loads and stores to a { i32, i24 } struct. An initial
attempt to fix this was sent for review in D10357, but that didn't
really get to the root of the problem.
The core issue was that canConvertValue and convertValue in SROA were
handling different bitwidth integers by doing a zext of the integer. It
wouldn't do a trunc though, only a zext! This would in turn lead SROA to
form an i24 load from an i24 alloca, zext it to i32, and then use it.
This would at least produce the wrong value for big-endian systems.
One of my many false starts here was to correct the computation for
big-endian systems by shifting. But this doesn't actually work because
the original code has a 64-bit store to the entire 8 bytes, and a 32-bit
load of the last 4 bytes, and because the alloc size is 8 bytes, we
can't lose that last (least significant if bigendian) byte! The real
problem here is that we're forming an i24 load in SROA which is actually
not sufficiently wide to load all of the necessary bits here. The source
has an i32 load, and SROA needs to form that as well.
The straightforward way to do this is to disable the zext logic in
canConvertValue and convertValue, forcing us to actually load all
32-bits. This seems like a really good change, but it in turn breaks
several other parts of SROA.
First in the chain of knock-on failures, we had places where we were
doing integer-widening promotion even though some of the integer loads
or stores extended *past the end* of the alloca's memory! There was even
a comment about preventing this, but it only prevented the case where
the type had a different bit size from its store size. So I added checks
to handle the cases where we actually have a widened load or store and
to avoid trying to special integer widening promotion in those cases.
Second, we actually rely on the ability to promote in the face of loads
past the end of an alloca! This is important so that we can (for
example) speculate loads around PHI nodes to do more promotion. The bits
loaded are garbage, but as long as they aren't used and the alignment is
suitable high (which it wasn't in the test case!) this is "fine". And we
can't stop promoting here, lots of things stop working well if we do. So
we need to add specific logic to handle the extension (and truncation)
case, but *only* where that extension or truncation are over bytes that
*are outside the alloca's allocated storage* and thus totally bogus to
load or store.
And of course, once we add back this correct handling of extension or
truncation, we need to correctly handle bigendian systems to avoid
re-introducing the exact bug that started us off on this chain of misery
in the first place, but this time even more subtle as it only happens
along speculated loads atop a PHI node.
I've ported an existing test for PHI speculation to the big-endian test
file and checked that we get that part correct, and I've added several
more interesting big-endian test cases that should help check that we're
getting this correct.
Fun times.
llvm-svn: 242869
We insert a bitcast which obfuscates the getCalledFunction for the utility
function which looks up attributes from the called function. Loosing ABI
changing parameter attributes is a bad thing.
rdar://21516488
llvm-svn: 242807
A patch by Chakshu Grover!
This patch allows constfolding of trunc,rint,nearbyint,ceil and floor intrinsics using APFloat class.
Differential Revision: http://reviews.llvm.org/D11144
llvm-svn: 242763
Summary:
SpeculativeExecution enables a series straight line optimizations (such
as SLSR and NaryReassociate) on conditional code. For example,
if (...)
... b * s ...
if (...)
... (b + 1) * s ...
speculative execution can hoist b * s and (b + 1) * s from then-blocks,
so that we have
... b * s ...
if (...)
...
... (b + 1) * s ...
if (...)
...
Then, SLSR can rewrite (b + 1) * s to (b * s + s) because after
speculative execution b * s dominates (b + 1) * s.
The performance impact of this change is significant. It speeds up the
benchmarks running EigenFloatContractionKernelInternal16x16
(ba68f42fa6/unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h (cl-526))
by roughly 2%. Some internal benchmarks that have the above code pattern
are improved by up to 40%. No significant slowdowns are observed on
Eigen CUDA microbenchmarks.
Reviewers: jholewinski, broune, eliben
Subscribers: llvm-commits, jholewinski
Differential Revision: http://reviews.llvm.org/D11201
llvm-svn: 242437
Internalizing an individual comdat group member without also internalizing
the other members of the comdat can break comdat semantics. For example,
if a module contains a reference to an internalized comdat member, and the
linker chooses a comdat group from a different object file, this will break
the reference to the internalized member.
This change causes the internalizer to only internalize comdat members if all
other members of the comdat are not externally visible. Once a comdat group
has been fully internalized, there is no need to apply comdat rules to its
members; later optimization passes (e.g. globaldce) can legally drop individual
members of the comdat. So we drop the comdat attribute from all comdat members.
Differential Revision: http://reviews.llvm.org/D10679
llvm-svn: 242423
Self-referential constants containing references to a merged function
no longer cause the MergeFunctions pass to infinite loop. Also adds a
reproduction IR which would otherwise fail, which was isolated from a similar
issue in Chromium.
Author: jrkoenig
Reviewers: nlewycky, jfb
Subscribers: llvm-commits, nlewycky, jfb
Differential Revision: http://reviews.llvm.org/D11208
llvm-svn: 242337
During estimation of unrolling effect we should be able to propagate
constants through casts.
Differential Revision: http://reviews.llvm.org/D10207
llvm-svn: 242257
Sometimes an incidentally created instruction can duplicate a Value used
elsewhere. It then often doesn't end up in the leader table. If it's later
removed, we attempt to remove it from the leader table and segfault.
Instead we should just ignore the removal request, which won't cause any
problems. The reverse situation, where the original instruction is replaced by
the new one (which you might think could leave the leader table empty) cannot
occur, because the incidental instruction will never be found in the first
place.
llvm-svn: 242199
Volatile loads and stores are made visible in global state regardless of
what memory is involved. It is not correct to disregard the ordering
and synchronization scope because it is possible to synchronize with
memory operations performed by hardware.
This partially addresses PR23737.
llvm-svn: 242126
Previously we would refrain from attempting to increase the linkage of
available_externally globals because they were considered weak for the
linker. Now they are treated more like a declaration instead of a weak
definition.
This was causing SSE alignment faults in Chromuim, when some code
assumed it could increase the alignment of a dllimported global that it
didn't control. http://crbug.com/509256
llvm-svn: 242091
This test case was breaking the hexagon elf bot. The failing lines
were actually unnecessary as checking that the store still reads the
correct value demonstrates that everything is working fine now.
llvm-svn: 242073
When spotting that a loop can use ctpop, we were incorrectly replacing all uses of a value with a value derived from ctpop.
The bug here was exposed because we were replacing a use prior to the ctpop with the ctpop value and so we have a use before def, i.e., we changed
%tobool.5 = icmp ne i32 %num, 0
store i1 %tobool.5, i1* %ptr
br i1 %tobool.5, label %for.body.lr.ph, label %for.end
to
store i1 %1, i1* %ptr
%0 = call i32 @llvm.ctpop.i32(i32 %num)
%1 = icmp ne i32 %0, 0
br i1 %1, label %for.body.lr.ph, label %for.end
Even if we inserted the ctpop so that it dominates the store here, that would still be incorrect. The store doesn’t want the result of ctpop.
The fix is very simple, and involves replacing only the branch condition with the ctpop instead of all uses.
Reviewed by Hal Finkel.
llvm-svn: 242068
Enable runtime unrolling for loops with unroll count metadata ("#pragma unroll N")
and a runtime trip count. Also, do not unroll loops with unroll full metadata if the
loop has a runtime loop count. Previously, such loops would be unrolled with a
very large threshold (pragma-unroll-threshold) if runtime unrolled happened to be
enabled resulting in a very large (and likely unwise) unroll factor.
llvm-svn: 242047
Summary:
This at least saves compile time. I also encountered a case where
ephemeral values affect whether other variables are promoted, causing
performance issues. It may be a bug in LSR, but I didn't manage to
reduce it yet. Anyhow, I believe it's in general not worth considering
ephemeral values in LSR.
Reviewers: atrick, hfinkel
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D11115
llvm-svn: 242011
There is no suitable basic block to sink instructions in loops without
exits. The only way an instruction in a loop without exits can be used
is as an incoming value to a PHI. In such cases, the incoming block for
the corresponding value is unreachable.
This fixes PR24013.
Differential Revision: http://reviews.llvm.org/D10903
llvm-svn: 241987
After changes in rL231820 loop re-rotation is performed even in -Oz mode. Since loop rotation is disabled for -Oz, it seems loop re-rotation should be disabled too.
Differential Revision: http://reviews.llvm.org/D10961
llvm-svn: 241897