We don't have tests for the effect of if-conversion loops because it requires a big test (that includes if-converted loops) and it is difficult to find and balance a loop to do the right thing.
llvm-svn: 186845
helper function. This leaves both trivial cases handled entirely in
helper functions and merely manages the list of allocas to process in
the run method.
The next step will be to handle all of the trivial promotion work prior
to even creating the core class and the subsequent simplifications that
enables.
llvm-svn: 186784
a single block into the helper routine. This takes advantage of the fact
that we can directly replace uses prior to any store with undef to
simplify matters and unconditionally promote allocas only used within
one block.
I've removed the special handling for the case of no stores existing.
This has no semantic effect but might slow things down. I'll fix that in
a later patch when I refactor this entire thing to be easier to manage
the different cases.
llvm-svn: 186783
handles the general cases.
The hope is to refactor this so that we don't end up building the entire
class for the trivial cases. I also want to lift a lot of the early
pre-processing in the initial segment of run() into a separate routine,
and really none of it needs to happen inside the primary promotion
class.
These routines in particular used none of the actual state in the
promotion class, so they don't really make sense as members.
llvm-svn: 186781
This struct is nicely independent of everything else, and we already
needed a foward declaration here. It's simpler to just define it
immediately.
llvm-svn: 186780
GlobalOpt simplifies llvm.compiler.used by removing any members that are also
in the more strict llvm.used. Handle the special case where llvm.compiler.used
becomes empty.
llvm-svn: 186778
We were incorrectly using compiler_used instead of compiler.used. Unfortunately
the passes using the broken name had tests also using the broken name.
llvm-svn: 186705
implementation of the SROA algorithm. We were using the term 'partition'
in many places that no longer ever represented an actual partition, but
rather just an arbitrary slice of an alloca.
No functionality change intended here. Mostly just renaming of types,
functions, variables, and rewording of comments. Several comments were
rewritten to make a lot more sense in the new structure of things.
The stats are still weird and not reflective of how this really works.
I'll fix those up in a separate patch as it is a touch more semantic of
a change...
llvm-svn: 186659
SROA.
The crux of the issue is that now we track uses of a partition of the
alloca in two places: the iterators over the partitioning uses and the
previously collected split uses vector. We weren't accounting for the
fact that the split uses might invalidate integer widening in ways other
than due to their width (in this case due to being volatile).
Further reduced testcase added to the tests.
llvm-svn: 186655
end of a vector. This was found with ASan. I've had one other report of
a crasher, but thus far been unable to reproduce the crash. It may well
be fixed with this version, and if not I'd like to get more information
from the build bots about what is happening.
See r186316 for the full commit log for the new implementation of the
SROA algorithm.
llvm-svn: 186565
Duncan pointed out a mistake in my fix in r186425 when only one of the allocas
being compared had the target-default alignment. This is essentially his
suggested solution. Thanks!
llvm-svn: 186510
This check does not always work because not all of the GEPs use a constant offset, but it happens often enough to reduce the number of times we use SCEV.
llvm-svn: 186465
For safety, the inliner cannot decrease the allignment on an alloca when
merging it with another.
I've included two variants of the test case for this: one with DataLayout
available, and one without. When DataLayout is not available, if only one of
the allocas uses the default alignment (getAlignment() == 0), then they cannot
be safely merged.
llvm-svn: 186425
a bot.
This reverts the commit which introduced a new implementation of the
fancy SROA pass designed to reduce its overhead. I'll skip the huge
commit log here, refer to r186316 if you're looking for how this all
works and why it works that way.
llvm-svn: 186332
different core implementation strategy.
Previously, SROA would build a relatively elaborate partitioning of an
alloca, associate uses with each partition, and then rewrite the uses of
each partition in an attempt to break apart the alloca into chunks that
could be promoted. This was very wasteful in terms of memory and compile
time because regardless of how complex the alloca or how much we're able
to do in breaking it up, all of the datastructure work to analyze the
partitioning was done up front.
The new implementation attempts to form partitions of the alloca lazily
and on the fly, rewriting the uses that make up that partition as it
goes. This has a few significant effects:
1) Much simpler data structures are used throughout.
2) No more double walk of the recursive use graph of the alloca, only
walk it once.
3) No more complex algorithms for associating a particular use with
a particular partition.
4) PHI and Select speculation is simplified and happens lazily.
5) More precise information is available about a specific use of the
alloca, removing the need for some side datastructures.
Ultimately, I think this is a much better implementation. It removes
about 300 lines of code, but arguably removes more like 500 considering
that some code grew in the process of being factored apart and cleaned
up for this all to work.
I've re-used as much of the old implementation as possible, which
includes the lion's share of code in the form of the rewriting logic.
The interesting new logic centers around how the uses of a partition are
sorted, and split into actual partitions.
Each instruction using a pointer derived from the alloca gets
a 'Partition' entry. This name is totally wrong, but I'll do a rename in
a follow-up commit as there is already enough churn here. The entry
describes the offset range accessed and the nature of the access. Once
we have all of these entries we sort them in a very specific way:
increasing order of begin offset, followed by whether they are
splittable uses (memcpy, etc), followed by the end offset or whatever.
Sorting by splittability is important as it simplifies the collection of
uses into a partition.
Once we have these uses sorted, we walk from the beginning to the end
building up a range of uses that form a partition of the alloca.
Overlapping unsplittable uses are merged into a single partition while
splittable uses are broken apart and carried from one partition to the
next. A partition is also introduced to bridge splittable uses between
the unsplittable regions when necessary.
I've looked at the performance PRs fairly closely. PR15471 no longer
will even load (the module is invalid). Not sure what is up there.
PR15412 improves by between 5% and 10%, however it is nearly impossible
to know what is holding it up as SROA (the entire pass) takes less time
than reading the IR for that test case. The analysis takes the same time
as running mem2reg on the final allocas. I suspect (without much
evidence) that the new implementation will scale much better however,
and it is just the small nature of the test cases that makes the changes
small and noisy. Either way, it is still simpler and cleaner I think.
llvm-svn: 186316
If an outside loop user of the reduction value uses the header phi node we
cannot just reduce the vectorized phi value in the vector code epilog because
we would loose VF-1 reductions.
lp:
p = phi (0, lv)
lv = lv + 1
...
brcond , lp, outside
outside:
usr = add 0, p
(Say the loop iterates two times, the value of p coming out of the loop is one).
We cannot just transform this to:
vlp:
p = phi (<0,0>, lv)
lv = lv + <1,1>
..
brcond , lp, outside
outside:
p_reduced = p[0] + [1];
usr = add 0, p_reduced
(Because the original loop iterated two times the vectorized loop would iterate
one time, but p_reduced ends up being zero instead of one).
We would have to execute VF-1 iterations in the scalar remainder loop in such
cases. For now, just disable vectorization.
PR16522
llvm-svn: 186256
In general, one should always complete CFG modifications first, update
CFG-based analyses, like Dominatores and LoopInfo, then generate
instruction sequences.
LoopVectorizer was creating a new loop, calling SCEVExpander to
generate checks, then updating LoopInfo. I just changed the order.
llvm-svn: 186241
Address calculation for gather/scather in vectorized code can incur a
significant cost making vectorization unbeneficial. Add infrastructure to add
cost.
Tests and cost model for targets will be in follow-up commits.
radar://14351991
llvm-svn: 186187
against a constant."
This reverts commit r186107. It didn't handle wrapping arithmetic in the
loop correctly and thus caused the following C program to count from
0 to UINT64_MAX instead of from 0 to 255 as intended:
#include <stdio.h>
int main() {
unsigned char first = 0, last = 255;
do { printf("%d\n", first); } while (first++ != last);
}
Full test case and instructions to reproduce with just the -indvars pass
sent to the original review thread rather than to r186107's commit.
llvm-svn: 186152
Before we could vectorize PHINodes scanning successors was a good way of finding candidates. Now we can vectorize the phinodes which is simpler.
llvm-svn: 186139
Patch by Michele Scandale!
Adds a special handling of the case where, during the loop exit
condition rewriting, the exit value is a constant of bitwidth lower
than the type of the induction variable: instead of introducing a
trunc operation in order to match correctly the operand types, it
allows to convert the constant value to an equivalent constant,
depending on the initial value of the induction variable and the trip
count, in order have an equivalent comparison between the induction
variable and the new constant.
llvm-svn: 186107
We can vectorize them because in the case where we wrap in the address space the
unvectorized code would have had to access a pointer value of zero which is
undefined behavior in address space zero according to the LLVM IR semantics.
(Thank you Duncan, for pointing this out to me).
Fixes PR16592.
llvm-svn: 186088
predecessors of the two blocks it is attempting to merge supply the
same incoming values to any phi in the successor block. This change
allows merging in the case where there is one or more incoming values
that are undef. The undef values are rewritten to match the non-undef
value that flows from the other edge. Patch by Mark Lacey.
llvm-svn: 186069
Without the changes introduced into this patch, if TRE saw any allocas at all,
TRE would not perform TRE *or* mark callsites with the tail marker.
Because TRE runs after mem2reg, this inadequacy is not a death sentence. But
given a callsite A without escaping alloca argument, A may not be able to have
the tail marker placed on it due to a separate callsite B having a write-back
parameter passed in via an argument with the nocapture attribute.
Assume that B is the only other callsite besides A and B only has nocapture
escaping alloca arguments (*NOTE* B may have other arguments that are not passed
allocas). In this case not marking A with the tail marker is unnecessarily
conservative since:
1. By assumption A has no escaping alloca arguments itself so it can not
access the caller's stack via its arguments.
2. Since all of B's escaping alloca arguments are passed as parameters with
the nocapture attribute, we know that B does not stash said escaping
allocas in a manner that outlives B itself and thus could be accessed
indirectly by A.
With the changes introduced by this patch:
1. If we see any escaping allocas passed as a capturing argument, we do
nothing and bail early.
2. If we do not see any escaping allocas passed as captured arguments but we
do see escaping allocas passed as nocapture arguments:
i. We do not perform TRE to avoid PR962 since the code generator produces
significantly worse code for the dynamic allocas that would be created
by the TRE algorithm.
ii. If we do not return twice, mark call sites without escaping allocas
with the tail marker. *NOTE* This excludes functions with escaping
nocapture allocas.
3. If we do not see any escaping allocas at all (whether captured or not):
i. If we do not have usage of setjmp, mark all callsites with the tail
marker.
ii. If there are no dynamic/variable sized allocas in the function,
attempt to perform TRE on all callsites in the function.
Based off of a patch by Nick Lewycky.
rdar://14324281.
llvm-svn: 186057
A special case list can now specify categories for specific globals,
which can be used to instruct an instrumentation pass to treat certain
functions or global variables in a specific way, such as by omitting
certain aspects of instrumentation while keeping others, or informing
the instrumentation pass that a specific uninstrumentable function
has certain semantics, thus allowing the pass to instrument callers
according to those semantics.
For example, AddressSanitizer now uses the "init" category instead of
global-init prefixes for globals whose initializers should not be
instrumented, but which in all other respects should be instrumented.
The motivating use case is DataFlowSanitizer, which will have a
number of different categories for uninstrumentable functions, such
as "functional" which specifies that a function has pure functional
semantics, or "discard" which indicates that a function's return
value should not be labelled.
Differential Revision: http://llvm-reviews.chandlerc.com/D1092
llvm-svn: 185978
The following transforms are valid if -C is a power of 2:
(icmp ugt (xor X, C), ~C) -> (icmp ult X, C)
(icmp ult (xor X, C), -C) -> (icmp uge X, C)
These are nice, they get rid of the xor.
llvm-svn: 185915
Commit 185883 fixes a bug in the IRBuilder that should fix the ASan bot. AssertingVH can help in exposing some RAUW problems.
Thanks Ben and Alexey!
llvm-svn: 185886
Back in r179493 we determined that two transforms collided with each
other. The fix back then was to reorder the transforms so that the
preferred transform would give it a try and then we would try the
secondary transform. However, it was noted that the best approach would
canonicalize one transform into the other, removing the collision and
allowing us to optimize IR given to us in that form.
llvm-svn: 185808
This is a complete re-write if the bottom-up vectorization class.
Before this commit we scanned the instruction tree 3 times. First in search of merge points for the trees. Second, for estimating the cost. And finally for vectorization.
There was a lot of code duplication and adding the DCE exposed bugs. The new design is simpler and DCE was a part of the design.
In this implementation we build the tree once. After that we estimate the cost by scanning the different entries in the constructed tree (in any order). The vectorization phase also works on the built tree.
llvm-svn: 185774
This is the first patch in a series of 3 patches which clean up how we create
runtime function declarations in the ARC optimizer when they do not exist
already in the IR.
Currently we have a bunch of duplicated code in ObjCARCOpts, ObjCARCContract
that does this. This patch refactors that code into a separate class called
ARCRuntimeEntryPoints which lazily creates the declarations for said
entrypoints.
The next two patches will consist of the work of refactoring
ObjCARCContract/ObjCARCOpts to use this new code.
llvm-svn: 185740
This transform was originally added in r185257 but later removed in
r185415. The original transform would create instructions speculatively
and then discard them if the speculation was proved incorrect. This has
been replaced with a scheme that splits the transform into two parts:
preflight and fold. While we preflight, we build up fold actions that
inform the folding stage on how to act.
llvm-svn: 185667
This allows us to create switches even if instcombine has munged two of the
incombing compares into one and some bit twiddling. This was motivated by enum
compares that are common in clang.
llvm-svn: 185632
This implies annotating it as nounwind and its arguments as nocapture. To be
conservative, we do not annotate the arguments with noalias since some platforms
do not have restrict on the declaration for gettimeofday.
llvm-svn: 185502
I'm reverting this commit because:
1. As discussed during review, it needs to be rewritten (to avoid creating and
then deleting instructions).
2. This is causing optimizer crashes. Specifically, I'm seeing things like
this:
While deleting: i1 %
Use still stuck around after Def is destroyed: <badref> = select i1 <badref>, i32 0, i32 1
opt: /src/llvm-trunk/lib/IR/Value.cpp:79: virtual llvm::Value::~Value(): Assertion `use_empty() && "Uses remain when a value is destroyed!"' failed.
I'd guess that these will go away once we're no longer creating/deleting
instructions here, but just in case, I'm adding a regression test.
Because the code is bring rewritten, I've just XFAIL'd the original regression test. Original commit message:
InstCombine: Be more agressive optimizing 'udiv' instrs with 'select' denoms
Real world code sometimes has the denominator of a 'udiv' be a
'select'. LLVM can handle such cases but only when the 'select'
operands are symmetric in structure (both select operands are a constant
power of two or a left shift, etc.). This falls apart if we are dealt a
'udiv' where the code is not symetric or if the select operands lead us
to more select instructions.
Instead, we should treat the LHS and each select operand as a distinct
divide operation and try to optimize them independently. If we can
to simplify each operation, then we can replace the 'udiv' with, say, a
'lshr' that has a new select with a bunch of new operands for the
select.
llvm-svn: 185415
Math functions are mark as readonly because they read the floating point
rounding mode. Because we don't vectorize loops that would contain function
calls that set the rounding mode it is safe to ignore this memory read.
llvm-svn: 185299
Changing the sign when comparing the base pointer would introduce all
sorts of unexpected things like:
%gep.i = getelementptr inbounds [1 x i8]* %a, i32 0, i32 0
%gep2.i = getelementptr inbounds [1 x i8]* %b, i32 0, i32 0
%cmp.i = icmp ult i8* %gep.i, %gep2.i
%cmp.i1 = icmp ult [1 x i8]* %a, %b
%cmp = icmp ne i1 %cmp.i, %cmp.i1
ret i1 %cmp
into:
%cmp.i = icmp slt [1 x i8]* %a, %b
%cmp.i1 = icmp ult [1 x i8]* %a, %b
%cmp = xor i1 %cmp.i, %cmp.i1
ret i1 %cmp
By preserving the original sign, we now get:
ret i1 false
This fixes PR16483.
llvm-svn: 185259
Real world code sometimes has the denominator of a 'udiv' be a
'select'. LLVM can handle such cases but only when the 'select'
operands are symmetric in structure (both select operands are a constant
power of two or a left shift, etc.). This falls apart if we are dealt a
'udiv' where the code is not symetric or if the select operands lead us
to more select instructions.
Instead, we should treat the LHS and each select operand as a distinct
divide operation and try to optimize them independently. If we can
to simplify each operation, then we can replace the 'udiv' with, say, a
'lshr' that has a new select with a bunch of new operands for the
select.
llvm-svn: 185257
We may, after other optimizations, find ourselves with IR that looks
like:
%shl = shl i32 1, %y
%cmp = icmp ult i32 %shl, 32
Instead, we should just compare the shift count:
%cmp = icmp ult i32 %y, 5
llvm-svn: 185242
To support this we have to insert 'extractelement' instructions to pick the right lane.
We had this functionality before but I removed it when we moved to the multi-block design because it was too complicated.
llvm-svn: 185230
In this code we keep track of pointers that we are allowed to read from, if they are accessed by non-predicated blocks.
We use this list to allow vectorization of conditional loads in predicated blocks because we know that these addresses don't segfault.
llvm-svn: 185214
- Build debug metadata for 'bare' Modules using DIBuilder
- DebugIR can be constructed to generate an IR file (to be seen by a debugger)
or not in cases where the user already has an IR file on disk.
llvm-svn: 185193
I used the class to safely reset the state of the builder's debug location. I
think I have caught all places where we need to set the debug location to a new
one. Therefore, we can replace the class by a function that just sets the debug
location.
llvm-svn: 185165
No functionality change.
It should suffice to check the type of a debug info metadata, instead of
calling Verify. For cases where we know the type of a DI metadata, use
assert.
Also update testing cases to make them conform to the format of DI classes.
llvm-svn: 185135
This reverts commit r185099.
Looks like both the ppc-64 and mips bots are still failing after I reverted this
change.
Since:
1. The mips bot always performs a clean build,
2. The ppc64-bot failed again after a clean build (I asked the ppc-64
maintainers to clean the bot which they did... Thanks Will!),
I think it is safe to assume that this change was not the cause of the failures
that said builders were seeing. Thus I am recomitting.
llvm-svn: 185111
This reverts commit r185095. This is causing a FileCheck failure on
the 3dnow intrinsics on at least the mips/ppc bots but not on the x86
bots.
Reverting while I figure out what is going on.
llvm-svn: 185099
The category which an APFloat belongs to should be dependent on the
actual value that the APFloat has, not be arbitrarily passed in by the
user. This will prevent inconsistency bugs where the category and the
actual value in APFloat differ.
I also fixed up all of the references to this constructor (which were
only in LLVM).
llvm-svn: 185095
When we store values for reversed induction stores we must not store the
reversed value in the vectorized value map. Another instruction might use this
value.
This fixes 3 test cases of PR16455.
llvm-svn: 185051
The Builtin attribute is an attribute that can be placed on function call site that signal that even though a function is declared as being a builtin,
rdar://problem/13727199
llvm-svn: 185049
debug statements to add a missing newline. Also canonicalize to '\n' instead of
"\n"; the latter calls a function with a loop the former does not.
llvm-svn: 184897
When a 1-element vector alloca is promoted, a store instruction can often be
rewritten without converting the value to a scalar and using an insertelement
instruction to stuff it into the new alloca. This patch just adds a check
to skip that conversion when it is unnecessary. This turns out to be really
important for some ARM Neon operations where <1 x i64> is used to get around
the fact that i64 is not a legal type.
llvm-svn: 184870
This should hopefully have fixed the stage2/stage3 miscompare on the dragonegg
testers.
"LoopVectorize: Use the dependence test utility class
We now no longer need alias analysis - the cases that alias analysis would
handle are now handled as accesses with a large dependence distance.
We can now vectorize loops with simple constant dependence distances.
for (i = 8; i < 256; ++i) {
a[i] = a[i+4] * a[i+8];
}
for (i = 8; i < 256; ++i) {
a[i] = a[i-4] * a[i-8];
}
We would be able to vectorize about 200 more loops (in many cases the cost model
instructs us no to) in the test suite now. Results on x86-64 are a wash.
I have seen one degradation in ammp. Interestingly, the function in which we
now vectorize a loop is never executed so we probably see some instruction
cache effects. There is a 2% improvement in h264ref. There is one or the other
TSCV loop kernel that speeds up.
radar://13681598"
llvm-svn: 184724
CGSCC pass manager. This should insulate the inlining decisions from the
vectorization decisions, however it may have both compile time and code
size problems so it is just an experimental option right now.
Adding this based on a discussion with Arnold and it seems at least
worth having this flag for us to both run some experiments to see if
this strategy is workable. It may solve some of the regressions seen
with the loop vectorizer.
llvm-svn: 184698
We now no longer need alias analysis - the cases that alias analysis would
handle are now handled as accesses with a large dependence distance.
We can now vectorize loops with simple constant dependence distances.
for (i = 8; i < 256; ++i) {
a[i] = a[i+4] * a[i+8];
}
for (i = 8; i < 256; ++i) {
a[i] = a[i-4] * a[i-8];
}
We would be able to vectorize about 200 more loops (in many cases the cost model
instructs us no to) in the test suite now. Results on x86-64 are a wash.
I have seen one degradation in ammp. Interestingly, the function in which we
now vectorize a loop is never executed so we probably see some instruction
cache effects. There is a 2% improvement in h264ref. There is one or the other
TSCV loop kernel that speeds up.
radar://13681598
llvm-svn: 184685
This class checks dependences by subtracting two Scalar Evolution access
functions allowing us to catch very simple linear dependences.
The checker assumes source order in determining whether vectorization is safe.
We currently don't reorder accesses.
Positive true dependencies need to be a multiple of VF otherwise we impede
store-load forwarding.
llvm-svn: 184684
Sets of dependent accesses are built by unioning sets based on underlying
objects. This class will be used by the upcoming dependence checker.
llvm-svn: 184683
Untill now we detected the vectorizable tree and evaluated the cost of the
entire tree. With this patch we can decide to trim-out branches of the tree
that are not profitable to vectorizer.
Also, increase the max depth from 6 to 12. In the worse possible case where all
of the code is made of diamond-shaped graph this can bring the cost to 2**10,
but diamonds are not very common.
llvm-svn: 184681
Rewrote the SLP-vectorization as a whole-function vectorization pass. It is now able to vectorize chains across multiple basic blocks.
It still does not vectorize PHIs, but this should be easy to do now that we scan the entire function.
I removed the support for extracting values from trees.
We are now able to vectorize more programs, but there are some serious regressions in many workloads (such as flops-6 and mandel-2).
llvm-svn: 184647
This is apart of a series of patches to encapsulate PtrState.RRI and
make PtrState.RRI a private field of PtrState.
*NOTE* This is actually the second commit in the patch stream. I should
have put this note on the first such commit r184528.
llvm-svn: 184532
This commit completely removes what is left of the simplify-libcalls
pass. All of the functionality has now been migrated to the instcombine
and functionattrs passes. The following C API functions are now NOPs:
1. LLVMAddSimplifyLibCallsPass
2. LLVMPassManagerBuilderSetDisableSimplifyLibCalls
llvm-svn: 184459
We collect gather sequences when we vectorize basic blocks. Gather sequences are excellent
hints for vectorization of other basic blocks.
llvm-svn: 184444
Prior to this change, the considered addressing modes may be invalid since the
maximum and minimum offsets were not taking into account.
This was causing an assertion failure.
The added test case exercices that behavior.
<rdar://problem/14199725> Assertion failed: (CurScaleCost >= 0 && "Legal
addressing mode has an illegal cost!")
llvm-svn: 184341
The type <3 x i8> is a common in graphics and we want to be able to vectorize it.
This changes accelerates bullet by 12% and 471_omnetpp by 5%.
llvm-svn: 184317
This pass was assuming that if hasAddressTaken() returns false for a
function, the function's only uses are call sites. That's not true
because there can be references by BlockAddresses too.
Fix the pass to handle this case. Fix
BlockAddress::replaceUsesOfWithOnConstant() to allow a function's type
to be changed by RAUW'ing the function with a bitcast of the recreated
function.
Patch by Mark Seaborn.
llvm-svn: 183933
Instead of a custom implementation of replaceAllUsesWith, we just call
replaceAllUsesWith and recreate llvm.used and llvm.compiler-used.
This change is particularity interesting because it makes llvm see
through what clang is doing with static used functions in extern "C"
contexts. With this change, running clang -O2 in
extern "C" {
__attribute__((used)) static void foo() {}
}
produces
@llvm.used = appending global [1 x i8*] [i8* bitcast (void ()* @foo to
i8*)], section "llvm.metadata"
define internal void @foo() #0 {
entry:
ret void
}
llvm-svn: 183756
r183584 tries to derive some info from the code *AFTER* a call and apply
these derived info to the code *BEFORE* the call, which is not always safe
as the call in question may never return, and in this case, the derived
info is invalid.
Thank Duncan for pointing out this potential bug.
rdar://14073661
llvm-svn: 183606
The MemCpyOpt pass is capable of optimizing:
callee(&S); copy N bytes from S to D.
into:
callee(&D);
subject to some legality constraints.
Assertion is triggered when the compiler tries to evalute "sizeof(typeof(D))",
while D is an opaque-typed, 'sret' formal argument of function being compiled.
i.e. the signature of the func being compiled is something like this:
T caller(...,%opaque* noalias nocapture sret %D, ...)
The fix is that when come across such situation, instead of calling some
utility functions to get the size of D's type (which will crash), we simply
assume D has at least N bytes as implified by the copy-instruction.
rdar://14073661
llvm-svn: 183584
IndVarSimplify is willing to move divide instructions outside of their
loop bodies if they are invariant of the loop. However, it may not be
safe to expand them if we do not know if they can trap.
Instead, check to see if it is not safe to expand the instruction and
skip the expansion.
This fixes PR16041.
Testcase by Rafael Ávila de Espíndola.
llvm-svn: 183239
The problem this time seems to be a thinko. We were assuming that in the CFG
A
| \
| B
| /
C
speculating the basic block B would cause only the phi value for the B->C edge
to be speculated. That is not true, the phi's are semantically in the edges, so
if the A->B->C path is taken, any code needed for A->C is not executed and we
have to consider it too when deciding to speculate B.
llvm-svn: 183226
PR16069 is an interesting case where an incoming value to a PHI is a
trap value while also being a 'ConstantExpr'.
We do not consider this case when performing the 'HoistThenElseCodeToIf'
optimization.
Instead, make our modifications more conservative if we detect that we
cannot transform the PHI to a select.
llvm-svn: 183152
index greater than the size of the vector is invalid. The shuffle may be
shrinking the size of the vector. Fixes a crash!
Also drop the maximum recursion depth of the safety check for this
optimization to five.
llvm-svn: 183080
Use ScalarEvolution's getBackedgeTakenCount API instead of getExitCount since
that is really what we want to know. Using the more specific getExitCount was
safe because we made sure that there is only one exiting block.
No functionality change.
llvm-svn: 183047
Account for the cost of scaling factor in Loop Strength Reduce when rating the
formulae. This uses a target hook.
The default implementation of the hook is: if the addressing mode is legal, the
scaling factor is free.
<rdar://problem/13806271>
llvm-svn: 183045
We check that instructions in the loop don't have outside users (except if
they are reduction values). Unfortunately, we skipped this check for
if-convertable PHIs.
Fixes PR16184.
llvm-svn: 183035
Namely, check if the target allows to fold more that one register in the
addressing mode and if yes, adjust the cost accordingly.
Prior to this commit, reg1 + scale * reg2 accesses were artificially preferred
to reg1 + reg2 accesses. Indeed, the cost model wrongly assumed that reg1 + reg2
needs a temporary register for the computation, whereas it was correctly
estimated for reg1 + scale * reg2.
<rdar://problem/13973908>
llvm-svn: 183021
Before this change, each module defined a weak_odr global __msan_track_origins
with a value of 1 if origin tracking is enabled, 0 if disabled. If there are
modules with different values, any of them may win. If 0 wins, and there is at
least one module with 1, the program will most likely crash.
With this change, __msan_track_origins is only emitted if origin tracking is
on. Then runtime library detects if there is at least one module with origin
tracking, and enables runtime support for it.
llvm-svn: 182997
- llvm.loop.parallel metadata has been renamed to llvm.loop to be more generic
by making the root of additional loop metadata.
- Loop::isAnnotatedParallel now looks for llvm.loop and associated
llvm.mem.parallel_loop_access
- document llvm.loop and update llvm.mem.parallel_loop_access
- add support for llvm.vectorizer.width and llvm.vectorizer.unroll
- document llvm.vectorizer.* metadata
- add utility class LoopVectorizerHints for getting/setting loop metadata
- use llvm.vectorizer.width=1 to indicate already vectorized instead of
already_vectorized
- update existing tests that used llvm.loop.parallel and
llvm.vectorizer.already_vectorized
Reviewed by: Nadav Rotem
llvm-svn: 182802
Extend LinkModules to pass a ValueMaterializer to RemapInstruction and friends to lazily create Functions for lazily linked globals. This is a big win when linking small modules with large (mostly unused) library modules.
llvm-svn: 182776
as the BinaryOperator, *not* in the block where the IRBuilder is currently
inserting into. Fixes a bug where scalarizePHI would create instructions
that would not dominate all uses.
llvm-svn: 182639
- move AsmWriter.h from public headers into lib
- marked all AssemblyWriter functions as non-virtual; no need to override them
- DebugIR now "plugs into" AssemblyWriter with an AssemblyAnnotationWriter helper
- exposed flags to control hiding of a) debug metadata b) debug intrinsic calls
C/R: Paul Redmond
llvm-svn: 182617
We are not working on a DAG and I ran into a number of problems when I enabled the vectorizations of 'diamond-trees' (trees that share leafs).
* Imroved the numbering API.
* Changed the placement of new instructions to the last root.
* Fixed a bug with external tree users with non-zero lane.
* Fixed a bug in the placement of in-tree users.
llvm-svn: 182508
The earlier change list introduced the following inst combines:
B * (uitofp i1 C) —> select C, B, 0
A * (1 - uitofp i1 C) —> select C, 0, A
select C, 0, B + select C, A, 0 —> select C, A, B
Together these 3 changes would simplify :
A * (1 - uitofp i1 C) + B * uitofp i1 C
down to :
select C, B, A
In practice we found that the first two substitutions can have a
negative effect on performance, because they reduce opportunities to
use FMA contractions; between the two options FMAs are often the
better choice. This change list amends the previous one to enable
just these inst combines:
select C, B, 0 + select C, 0, A —> select C, B, A
A * (1 - uitofp i1 C) + B * uitofp i1 C —> select C, B, A
llvm-svn: 182499
The Value pointers we store in the induction variable list can be RAUW'ed by a
call to SCEVExpander::expandCodeFor, use a TrackingVH instead. Do the same thing
in some other places where we store pointers that could potentially be RAUW'ed.
Fixes PR16073.
llvm-svn: 182485
Other passes, PPC counter-loop formation for example, also need to add loop
preheaders outside of the regular loop simplification pass. This makes
InsertPreheaderForLoop a global function so that it can be used by other
passes.
No functionality change intended.
llvm-svn: 182299
We only want to check this once, not for every conditional block in the loop.
No functionality change (except that we don't perform a check redudantly
anymore).
llvm-svn: 181942
InstCombine can be uncooperative to vectorization and sink loads into
conditional blocks. This prevents vectorization.
Undo this optimization if there are unconditional memory accesses to the same
addresses in the loop.
radar://13815763
llvm-svn: 181860
CXAAtExitFn was set outside a loop and before optimizations where functions
can be deleted. This patch will set CXAAtExitFn inside the loop and after
optimizations.
Seg fault when running LTO because of accesses to a deleted function.
rdar://problem/13838828
llvm-svn: 181838
We used to give up if we saw two integer inductions. After this patch, we base
further induction variables on the chosen one like we do in the reverse
induction and pointer induction case.
Fixes PR15720.
radar://13851975
llvm-svn: 181746
In the presense of a block being initialized, the frontend will emit the
objc_retain on the original pointer and the release on the pointer loaded from
the alloca. The optimizer will through the provenance analysis realize that the
two are related (albiet different), but since we only require KnownSafe in one
direction, will match the inner retain on the original pointer with the guard
release on the original pointer. This is fixed by ensuring that in the presense
of allocas we only unconditionally remove pointers if both our retain and our
release are KnownSafe (i.e. we are KnownSafe in both directions) since we must
deal with the possibility that the frontend will emit what (to the optimizer)
appears to be unbalanced retain/releases.
An example of the miscompile is:
%A = alloca
retain(%x)
retain(%x) <--- Inner Retain
store %x, %A
%y = load %A
... DO STUFF ...
release(%y)
call void @use(%x)
release(%x) <--- Guarding Release
getting optimized to:
%A = alloca
retain(%x)
store %x, %A
%y = load %A
... DO STUFF ...
release(%y)
call void @use(%x)
rdar://13750319
llvm-svn: 181743
This makes the statistics gathering completely independent of the actual
optimization occuring, preventing any sort of bleeding over from occuring.
Additionally, it simplifies a switch statement in the non-statistic gathering case.
llvm-svn: 181719
The external user does not have to be in lane #0. We have to save the lane for each scalar so that we know which vector lane to extract.
llvm-svn: 181674
There are two transforms in visitUrem that conflict with each other.
*) One, if a divisor is a power of two, subtracts one from the divisor
and turns it into a bitwise-and.
*) The other unwraps both operands if they are surrounded by zext
instructions.
Flipping the order allows the subtraction to go beneath the sign
extension.
llvm-svn: 181668
Use the widest induction type encountered for the cannonical induction variable.
We used to turn the following loop into an empty loop because we used i8 as
induction variable type and truncated 1024 to 0 as trip count.
int a[1024];
void fail() {
int reverse_induction = 1023;
unsigned char forward_induction = 0;
while ((reverse_induction) >= 0) {
forward_induction++;
a[reverse_induction] = forward_induction;
--reverse_induction;
}
}
radar://13862901
llvm-svn: 181667
The shift amount may be larger than the type leading to undefined behavior.
Limit the transform to constant shift amounts. While there update the bits to
clear in the result which may enable additional optimizations.
PR15959.
llvm-svn: 181604
iteration.
This on step toward non-iterative GVN. My local hack suggests that getting rid
of iteration will speedup GVN by 30%+ on a medium sized input (2k LOC, C++).
I cannot explain why not 2x or more at this moment.
llvm-svn: 181532
That's obviously wrong. Conservatively restrict it to the sign bit, which
matches the original intention of this analysis. Fixes PR15940.
llvm-svn: 181518
A computable loop exit count does not imply the presence of an induction
variable. Scalar evolution can return a value for an infinite loop.
Fixes PR15926.
llvm-svn: 181495
- requires existing debug information to be present
- fixes up file name and line number information in metadata
- emits a "<orig_filename>-debug.ll" succinct IR file (without !dbg metadata
or debug intrinsics) that can be read by a debugger
- initialize pass in opt tool to enable the "-debug-ir" flag
- lit tests to follow
llvm-svn: 181467
The two nested loops were confusing and also conservative in identifying
reduction variables. This patch replaces them by a worklist based approach.
llvm-svn: 181369
We were passing an i32 to ConstantInt::get where an i64 was needed and we must
also pass the sign if we pass negatives numbers. The start index passed to
getConsecutiveVector must also be signed.
Should fix PR15882.
llvm-svn: 181286
Test case by Michele Scandale!
Fixes PR10293: Load not hoisted out of loop with multiple exits.
There are few regressions with this patch, now tracked by
rdar:13817079, and a roughly equal number of improvements. The
regressions are almost certainly back luck because LoopRotate has very
little idea of whether rotation is profitable. Doing better requires a
more comprehensive solution.
This checkin is a quick fix that lacks generality (PR10293 has
a counter-example). But it trivially fixes the case in PR10293 without
interfering with other cases, and it does satify the criteria that
LoopRotate is a loop canonicalization pass that should avoid
heuristics and special cases.
I can think of two approaches that would probably be better in
the long run. Ultimately they may both make sense.
(1) LoopRotate should check that the current header would make a good
loop guard, and that the loop does not already has a sufficient
guard. The artifical SimplifiedLoopLatch check would be unnecessary,
and the design would be more general and canonical. Two difficulties:
- We need a strong guarantee that we won't endlessly rotate, so the
analysis would need to be precise in order to avoid the
SimplifiedLoopLatch precondition.
- Analysis like this are usually based on SCEV, which we don't want to
rely on.
(2) Rotate on-demand in late loop passes. This could even be done by
shoving the loop back on the queue after the optimization that needs
it. This could work well when we find LICM opportunities in
multi-branch loops. This requires some work, and it doesn't really
solve the problem of SCEV wanting a loop guard before the analysis.
llvm-svn: 181230
A * (1 - (uitofp i1 C)) -> select C, 0, A
B * (uitofp i1 C) -> select C, B, 0
select C, 0, A + select C, B, 0 -> select C, B, A
These come up in code that has been hand-optimized from a select to a linear blend,
on platforms where that may have mattered. We want to undo such changes
with the following transform:
A*(1 - uitofp i1 C) + B*(uitofp i1 C) -> select C, A, B
llvm-svn: 181216
We used to disable constant merging not only if a constant is llvm.used, but
also if an alias of a constant is llvm.used. This change fixes that.
llvm-svn: 181175
Add support for min/max reductions when "no-nans-float-math" is enabled. This
allows us to assume we have ordered floating point math and treat ordered and
unordered predicates equally.
radar://13723044
llvm-svn: 181144
This function consists of following steps:
1. Collect dependent memory accesses.
2. Analyze availability.
3. Perform fully redundancy elimination, or
4. Perform PRE, depending on the availability
Step 2, 3 and 4 are now moved to three helper routines.
llvm-svn: 181047
By supporting the vectorization of PHINodes with more than two incoming values we can increase the complexity of nested if statements.
We can now vectorize this loop:
int foo(int *A, int *B, int n) {
for (int i=0; i < n; i++) {
int x = 9;
if (A[i] > B[i]) {
if (A[i] > 19) {
x = 3;
} else if (B[i] < 4 ) {
x = 4;
} else {
x = 5;
}
}
A[i] = x;
}
}
llvm-svn: 181037
Actually it took me couple of hours trying to make sense of them and
only to find they are dead code. I guess the original author used
"allSingleSucc" to indicate if there are any critial edge emanating
from some blocks, and tried to perform code motion (actually speculation)
in the presence of these critical edges; but later on he/she changed mind
and decided to perform edge-splitting first.
llvm-svn: 180951
the things, and renames it to CBindingWrapping.h. I also moved
CBindingWrapping.h into Support/.
This new file just contains the macros for defining different wrap/unwrap
methods.
The calls to those macros, as well as any custom wrap/unwrap definitions
(like for array of Values for example), are put into corresponding C++
headers.
Doing this required some #include surgery, since some .cpp files relied
on the fact that including Wrap.h implicitly caused the inclusion of a
bunch of other things.
This also now means that the C++ headers will include their corresponding
C API headers; for example Value.h must include llvm-c/Core.h. I think
this is harmless, since the C API headers contain just external function
declarations and some C types, so I don't believe there should be any
nasty dependency issues here.
llvm-svn: 180881
This reverts commit r180802
There's ongoing discussion about whether this is the right place to make
this transformation. Reverting for now while we figure it out.
llvm-svn: 180834
Always fold a shuffle-of-shuffle into a single shuffle when there's only one
input vector in the first place. Continue to be more conservative when there's
multiple inputs.
rdar://13402653
PR15866
llvm-svn: 180802
This fixes the optimization introduced in r179748 and reverted in r179750.
While the optimization was sound, it did not properly respect differences in
bit-width.
llvm-svn: 180777
This resurrects r179957, but adds code that makes sure we don't touch
atomic/volatile stores:
This transformation will transform a conditional store with a preceeding
uncondtional store to the same location:
a[i] =
may-alias with a[i] load
if (cond)
a[i] = Y
into an unconditional store.
a[i] = X
may-alias with a[i] load
tmp = cond ? Y : X;
a[i] = tmp
We assume that on average the cost of a mispredicted branch is going to be
higher than the cost of a second store to the same location, and that the
secondary benefits of creating a bigger basic block for other optimizations to
work on outway the potential case where the branch would be correctly predicted
and the cost of the executing the second store would be noticably reflected in
performance.
hmmer's execution time improves by 30% on an imac12,2 on ref data sets. With
this change we are on par with gcc's performance (gcc also performs this
transformation). There was a 1.2 % performance improvement on a ARM swift chip.
Other tests in the test-suite+external seem to be mostly uninfluenced in my
experiments:
This optimization was triggered on 41 tests such that the executable was
different before/after the patch. Only 1 out of the 40 tests (dealII) was
reproducable below 100% (by about .4%). Given that hmmer benefits so much I
believe this to be a fair trade off.
llvm-svn: 180731
Turning retains into retainRV calls disrupts the data flow analysis in
ObjCARCOpts. Thus we move it as late as we can by moving it into
ObjCARCContract.
We leave in the conversion from retainRV -> retain in ObjCARCOpt since
it enables the dataflow analysis.
rdar://10813093
llvm-svn: 180698
When Reassociator optimize "(x | C1)" ^ "(X & C2)", it may swap the two
subexpressions, however, it forgot to swap cached constants (of C1 and C2)
accordingly.
rdar://13739160
llvm-svn: 180676
Since we can't guarantee that the original dbg.declare instrinsic
is removed by LowerDbgDeclare(), we need to make sure that we are
not inserting the same dbg.value intrinsic over and over.
This removes tons of redundant DIEs when compiling optimized code.
rdar://problem/13056109
llvm-svn: 180615
This reverts commit r180222.
I think this might tie in with a different problem which will require a
different approach potentially. I am reverting this in the case I need to go
down that second path.
My apologies for the noise. = /.
llvm-svn: 180590
Due to the semantics of ARC, we must be extremely conservative with autorelease
calls inserted by the frontend since ARC gaurantees that said object will be in
the autorelease pool after that point, an optimization invariant that the
optimizer must respect.
On the other hand, we are allowed significantly more flexibility with
autoreleaseRV instructions.
Often times though this flexibility is disrupted by early transformations which
transform objc_autoreleaseRV => objc_autorelease if said instruction is no
longer being used as part of an RV pair (generally due to inlining). Since we
can not tell the difference in between an autorelease put into place by the
frontend and one created through said ``strength reduction'' we can not perform
these optimizations.
The addition of this set gets around said issues by allowing us to differentiate
in between said two cases.
rdar://problem/13697741.
llvm-svn: 180222
This patch disables memory-instruction vectorization for types that need padding
bytes, e.g., x86_fp80 has 10 bytes store size with 6 bytes padding in darwin on
x86_64. Because the load/store vectorization is performed by the bit casting to
a packed vector, which has incompatible memory layout due to the lack of padding
bytes, the present vectorizer produces inconsistent result for memory
instructions of those types.
This patch checks an equality of the AllocSize of a scalar type and allocated
size for each vector element, to ensure that there is no padding bytes and the
array can be read/written using vector operations.
Patch by Daisuke Takahashi!
Fixes PR15758.
llvm-svn: 180196
debug location. This solves a problem where range of an inlined
subroutine is emitted wrongly.
Patch by Manman Ren.
Fixes rdar://problem/12415623
llvm-svn: 180140
even if erroneously annotated with the parallel loop metadata.
Fixes Bug 15794:
"Loop Vectorizer: Crashes with the use of llvm.loop.parallel metadata"
llvm-svn: 180081
This is an edge case that can happen if we modify a chain of multiple selects.
Update all operands in that case and remove the assert. PR15805.
llvm-svn: 179982
There is the temptation to make this tranform dependent on target information as
it is not going to be beneficial on all (sub)targets. Therefore, we should
probably do this in MI Early-Ifconversion.
This reverts commit r179957. Original commit message:
"SimplifyCFG: If convert single conditional stores
This transformation will transform a conditional store with a preceeding
uncondtional store to the same location:
a[i] =
may-alias with a[i] load
if (cond)
a[i] = Y
into an unconditional store.
a[i] = X
may-alias with a[i] load
tmp = cond ? Y : X;
a[i] = tmp
We assume that on average the cost of a mispredicted branch is going to be
higher than the cost of a second store to the same location, and that the
secondary benefits of creating a bigger basic block for other optimizations to
work on outway the potential case were the branch would be correctly predicted
and the cost of the executing the second store would be noticably reflected in
performance.
hmmer's execution time improves by 30% on an imac12,2 on ref data sets. With
this change we are on par with gcc's performance (gcc also performs this
transformation). There was a 1.2 % performance improvement on a ARM swift chip.
Other tests in the test-suite+external seem to be mostly uninfluenced in my
experiments:
This optimization was triggered on 41 tests such that the executable was
different before/after the patch. Only 1 out of the 40 tests (dealII) was
reproducable below 100% (by about .4%). Given that hmmer benefits so much I
believe this to be a fair trade off.
I am going to watch performance numbers across the builtbots and will revert
this if anything unexpected comes up."
llvm-svn: 179980
This will make it clearer when we are actually resetting a sequence's progress
vs just changing state. This is an important distinction because the former case
clears any pointers that we are tracking while the later does not.
llvm-svn: 179963
This transformation will transform a conditional store with a preceeding
uncondtional store to the same location:
a[i] =
may-alias with a[i] load
if (cond)
a[i] = Y
into an unconditional store.
a[i] = X
may-alias with a[i] load
tmp = cond ? Y : X;
a[i] = tmp
We assume that on average the cost of a mispredicted branch is going to be
higher than the cost of a second store to the same location, and that the
secondary benefits of creating a bigger basic block for other optimizations to
work on outway the potential case were the branch would be correctly predicted
and the cost of the executing the second store would be noticably reflected in
performance.
hmmer's execution time improves by 30% on an imac12,2 on ref data sets. With
this change we are on par with gcc's performance (gcc also performs this
transformation). There was a 1.2 % performance improvement on a ARM swift chip.
Other tests in the test-suite+external seem to be mostly uninfluenced in my
experiments:
This optimization was triggered on 41 tests such that the executable was
different before/after the patch. Only 1 out of the 40 tests (dealII) was
reproducable below 100% (by about .4%). Given that hmmer benefits so much I
believe this to be a fair trade off.
I am going to watch performance numbers across the builtbots and will revert
this if anything unexpected comes up.
llvm-svn: 179957
The logic that actually compares the types considers pointers and integers the
same if they are of the same size. This created a strange mismatch between hash
and reality and made the test case for this fail on some platforms (yay,
test cases).
llvm-svn: 179905
Also make some static function class functions to avoid having to mention the
class namespace for enums all the time.
No functionality change intended.
llvm-svn: 179886
A min/max operation is represented by a select(cmp(lt/le/gt/ge, X, Y), X, Y)
sequence in LLVM. If we see such a sequence we can treat it just as any other
commutative binary instruction and reduce it.
This appears to help bzip2 by about 1.5% on an imac12,2.
radar://12960601
llvm-svn: 179773
This occurs due to an alloca representing a separate ownership from the
original pointer. Thus consider the following pseudo-IR:
objc_retain(%a)
for (...) {
objc_retain(%a)
%block <- %a
F(%block)
objc_release(%block)
}
objc_release(%a)
From the perspective of the optimizer, the %block is a separate
provenance from the original %a. Thus the optimizer pairs up the inner
retain for %a and the outer release from %a, resulting in segfaults.
This is fixed by noting that the signature of a mismatch of
retain/releases inside the for loop is a Use/CanRelease top down with an
None bottom up (since bottom up the Retain-CanRelease-Use-Release
sequence is completed by the inner objc_retain, but top down due to the
differing provenance from the objc_release said sequence is not
completed). In said case in CheckForCFGHazards, we now clear the state
of %a implying that no pairing will occur.
Additionally a test case is included.
rdar://12969722
llvm-svn: 179747
If a switch instruction has a case for every possible value of its type,
with the same successor, SimplifyCFG would replace it with an icmp ult,
but the computation of the bound overflows in that case, which inverts
the test.
Patch by Jed Davis!
llvm-svn: 179587
Two return types are not equivalent if one is a pointer and the other is an
integral. This is because we cannot bitcast a pointer to an integral value.
PR15185
llvm-svn: 179569
One performs: (X == 13 | X == 14) -> X-13 <u 2
The other: (A == C1 || A == C2) -> (A & ~(C1 ^ C2)) == C1
The problem is that there are certain values of C1 and C2 that
trigger both transforms but the first one blocks out the second,
this generates suboptimal code.
Reordering the transforms should be better in every case and
allows us to do interesting stuff like turn:
%shr = lshr i32 %X, 4
%and = and i32 %shr, 15
%add = add i32 %and, -14
%tobool = icmp ne i32 %add, 0
into:
%and = and i32 %X, 240
%tobool = icmp ne i32 %and, 224
llvm-svn: 179493
This is basically the same fix in three different places. We use a set to avoid
walking the whole tree of a big ConstantExprs multiple times.
For example: (select cmp, (add big_expr 1), (add big_expr 2))
We don't want to visit big_expr twice here, it may consist of thousands of
nodes.
The testcase exercises this by creating an insanely large ConstantExprs out of
a loop. It's questionable if the optimizer should ever create those, but this
can be triggered with real C code. Fixes PR15714.
llvm-svn: 179458
When trying to collapse sequences of insertelement/extractelement
instructions into single shuffle instructions, there is one specific
case where the Instruction Combiner wrongly updates the resulting
Mask of shuffle indexes.
The problem is in function CollectShuffleElments.
If we have a sequence of insert/extract element instructions
like the one below:
%tmp1 = extractelement <4 x float> %LHS, i32 0
%tmp2 = insertelement <4 x float> %RHS, float %tmp1, i32 1
%tmp3 = extractelement <4 x float> %RHS, i32 2
%tmp4 = insertelement <4 x float> %tmp2, float %tmp3, i32 3
Where:
. %RHS will have a mask of [4,5,6,7]
. %LHS will have a mask of [0,1,2,3]
The Mask of shuffle indexes is wrongly computed to [4,1,6,7]
instead of [4,0,6,7].
When analyzing %tmp2 in order to compute the Mask for the
resulting shuffle instruction, the algorithm forgets to update
the mask index at position 1 with the index associated to the
element extracted from %LHS by instruction %tmp1.
Patch by Andrea DiBiagio!
llvm-svn: 179291
This commit adds the infrastructure for performing bottom-up SLP vectorization (and other optimizations) on parallel computations.
The infrastructure has three potential users:
1. The loop vectorizer needs to be able to vectorize AOS data structures such as (sum += A[i] + A[i+1]).
2. The BB-vectorizer needs this infrastructure for bottom-up SLP vectorization, because bottom-up vectorization is faster to compute.
3. A loop-roller needs to be able to analyze consecutive chains and roll them into a loop, in order to reduce code size. A loop roller does not need to create vector instructions, and this infrastructure separates the chain analysis from the vectorization.
This patch also includes a simple (100 LOC) bottom up SLP vectorizer that uses the infrastructure, and can vectorize this code:
void SAXPY(int *x, int *y, int a, int i) {
x[i] = a * x[i] + y[i];
x[i+1] = a * x[i+1] + y[i+1];
x[i+2] = a * x[i+2] + y[i+2];
x[i+3] = a * x[i+3] + y[i+3];
}
llvm-svn: 179117
I brazenly think this change is slightly simpler than r178793 because:
- no "state" in functor
- "OpndPtrs[i]" looks simpler than "&Opnds[OpndIndices[i]]"
While I can reproduce the probelm in Valgrind, it is rather difficult to come up
a standalone testing case. The reason is that when an iterator is invalidated,
the stale invalidated elements are not yet clobbered by nonsense data, so the
optimizer can still proceed successfully.
Thank Benjamin for fixing this bug and generously providing the test case.
llvm-svn: 179062
The fix for PR14972 in r177055 introduced a real think-o in the *store*
side, likely because I was much more focused on the load side. While we
can arbitrarily widen (or narrow) a loaded value, we can't arbitrarily
widen a value to be stored, as that changes the width of memory access!
Lock down the code path in the store rewriting which would do this to
only handle the intended circumstance.
All of the existing tests continue to pass, and I've added a test from
the PR.
llvm-svn: 178974
The normal dataflow sequence in the ARC optimizer consists of the following
states:
Retain -> CanRelease -> Use -> Release
The optimizer before this patch stored the uses that determine the lifetime of
the retainable object pointer when it bottom up hits a retain or when top down
it hits a release. This is correct for an imprecise lifetime scenario since what
we are trying to do is remove retains/releases while making sure that no
``CanRelease'' (which is usually a call) deallocates the given pointer before we
get to the ``Use'' (since that would cause a segfault).
If we are considering the precise lifetime scenario though, this is not
correct. In such a situation, we *DO* care about the previous sequence, but
additionally, we wish to track the uses resulting from the following incomplete
sequences:
Retain -> CanRelease -> Release (TopDown)
Retain <- Use <- Release (BottomUp)
*NOTE* This patch looks large but the most of it consists of updating
test cases. Additionally this fix exposed an additional bug. I removed
the test case that expressed said bug and will recommit it with the fix
in a little bit.
llvm-svn: 178921
This optimization is unstable at this moment; it
1) block us on a very important application
2) PR15200
3) test6 and test7 in test/Transforms/ScalarRepl/dynamic-vector-gep.ll
(the CHECK command compare the output against wrong result)
I personally believe this optimization should not have any impact on the
autovectorized code, as auto-vectorizer is supposed to put gather/scatter
in a "right" way. Although in theory downstream optimizaters might reveal
some gather/scatter optimization opportunities, the chance is quite slim.
For the hand-crafted vectorizing code, in term of redundancy elimination,
load-CSE, copy-propagation and DSE can collectively achieve the same result,
but in much simpler way. On the other hand, these optimizers are able to
improve the code in a incremental way; in contrast, SROA is sort of all-or-none
approach. However, SROA might slighly win in stack size, as it tries to figure
out a stretch of memory tightenly cover the area accessed by the dynamic index.
rdar://13174884
PR15200
llvm-svn: 178912
Pass down the fact that an operand is going to be a vector of constants.
This should bring the performance of MultiSource/Benchmarks/PAQ8p/paq8p on x86
back. It had degraded to scalar performance due to my pervious shift cost change
that made all shifts expensive on x86.
radar://13576547
llvm-svn: 178809
OpndPtrs stored pointers into the Opnd vector that became invalid when the
vector grows. Store indices instead. Sadly I only have a large testcase that
only triggers under valgrind, so I didn't include it.
llvm-svn: 178793
Cleaned up trailing whitespace and added extra slashes in front of a
function level comment so that it follow the convention of having 3
slashes.
llvm-svn: 178712
The semantics of ARC implies that a pointer passed into an objc_autorelease
must live until some point (potentially down the stack) where an
autorelease pool is popped. On the other hand, an
objc_autoreleaseReturnValue just signifies that the object must live
until the end of the given function at least.
Thus objc_autorelease is stronger than objc_autoreleaseReturnValue in
terms of the semantics of ARC* implying that performing the given
strength reduction without any knowledge of how this relates to
the autorelease pool pop that is further up the stack violates the
semantics of ARC.
*Even though objc_autoreleaseReturnValue if you know that no RV
optimization will occur is more computationally expensive.
llvm-svn: 178612
The iterator could be invalidated when it's recursively deleting a whole bunch
of constant expressions in a constant initializer.
Note: This was only reproducible if `opt' was run on a `.bc' file. If `opt' was
run on a `.ll' file, it wouldn't crash. This is why the test first pushes the
`.ll' file through `llvm-as' before feeding it to `opt'.
PR15440
llvm-svn: 178531
clang.arc.used is an interesting call for ARC since ObjCARCContract
needs to run to remove said intrinsic to avoid a linker error (since the
call does not exist).
llvm-svn: 178369
Since we handle optimizable objc_retainBlocks through strength reduction
in OptimizableIndividualCalls, we know that all code after that point
will only see non-optimizable objc_retainBlock calls. IsForwarding is
only called by functions after that point, so it is ok to just classify
objc_retainBlock as non-forwarding.
<rdar://problem/13249661>.
llvm-svn: 178285
If an objc_retainBlock has the copy_on_escape metadata attached to it
AND if the block pointer argument only escapes down the stack, we are
allowed to strength reduce the objc_retainBlock to to an objc_retain and
thus optimize it.
Current there is logic in the ARC data flow analysis to handle
this case which is complicated and involved making distinctions in
between objc_retainBlock and objc_retain in certain places and
considering them the same in others.
This patch simplifies said code by:
1. Performing the strength reduction in the initial ARC peephole
analysis (ObjCARCOpts::OptimizeIndividualCalls).
2. Changes the ARC dataflow analysis (which runs after the peephole
analysis) to consider all objc_retainBlock calls to not be optimizable
(since if the call was optimizable, we would have strength reduced it
already).
This patch leaves in the infrastructure in the ARC dataflow analysis to
handle this case, which due to 2 will just be dead code. I am doing this
on purpose to separate the removal of the old code from the testing of
the new code.
<rdar://problem/13249661>.
llvm-svn: 178284
If we compile a single source program, the `.gcda' file will be generated where
the program was executed. This isn't desirable, because that place may be at an
unpredictable place (the program could call `chdir' for instance).
Instead, we will output the `.gcda' file in the same place we output the `.gcno'
file. I.e., the directory where the executable was generated. This matches GCC's
behavior.
<rdar://problem/13061072> & PR11809
llvm-svn: 178084
The OptimizeIntToFloatBitCast converts shift-truncate sequences
into extractelement operations. The computation of the element
index to be used in the resulting operation is currently only
correct for little-endian targets.
This commit fixes the element index computation to be correct
for big-endian targets as well. If the target byte order is
unknown, the optimization cannot be performed at all.
llvm-svn: 178031
This will allow for verification and analysis of the merge function of
the data flow analyses in the ARC optimizer.
The actual implementation of this feature is by introducing calls to
the functions llvm.arc.annotation.{bottomup,topdown}.{bbstart,bbend}
which are only declared. Each such call takes in a pointer to a global
with the same name as the pointer whose provenance is being tracked and
a pointer whose name is one of our Sequence states and points to a
string that contains the same name.
To ensure that the optimizer does not consider these annotations in any
way, I made it so that the annotations are considered to be of IC_None
type.
A test case is included for this commit and the previous
ObjCARCAnnotation commit.
llvm-svn: 177952
Previously the inner works of the data flow analysis in ObjCARCOpts was hard to
get out of the optimizer for analysis of bugs or testing. All of the current ARC
unit tests are based off of testing the effect of the data flow
analysis (i.e. what statements are removed or moved, etc.). This creates
weakness in the current unit testing regimem since we are not actually testing
what effects various instructions have on the modeled pointer state.
Additionally in order to analyze a bug in the optimizer, one would need to track
by hand what the optimizer was actually doing either through use of DEBUG
statements or through the usage of a debugger, both yielding large loses in
developer productivity.
This patch deals with these two issues by providing ARC annotation
metadata that annotates instructions with the state changes that they cause in
various pointers as well as provides metadata to annotate provenance sources.
Specifically, we introduce the following metadata types:
1. llvm.arc.annotation.bottomup.
2. llvm.arc.annotation.topdown.
3. llvm.arc.annotation.provenancesource.
llvm.arc.annotation.{bottomup,topdown}: These annotations describes a state
change in a pointer when we are visiting instructions bottomup/topdown
respectively. The output format for both is the same:
!1 = metadata !{metadata !"(test,%x)", metadata !"S_Release", metadata !"S_Use"}
The first element is a string tuple with the following format:
(function,variable name)
The second two elements of the metadata show the previous state of the
pointer (in this case S_Release) and the new state of the pointer (S_Use). We
write the metadata in such a manner to ensure that it is easy for outside tools
to parse. This is important since I am currently working on a tool for taking
this information and pretty printing it besides the IR and that can be used for
LIT style testing via the generation of an index.
llvm.arc.annotation.provenancesource: This metadata is used to annotate
instructions which act as provenance sources, i.e. ones that introduce a
new (from the optimizer's perspective) non-argument pointer to track. This
enables cross-referencing in between provenance sources and the state changes
that occur to them.
This is still a work in progress. Additionally I plan on committing
later today additions to the annotations that annotate at the top/bottom
of basic blocks the state of the various pointers being tracked.
*NOTE* The metadata support is conditionally compiled into libObjCARCOpts only
when we are producing a debug build of llvm/clang and even so are
disabled by default. To enable the annotation metadata, pass in
-enable-objc-arc-annotations to opt.
llvm-svn: 177951
The problem is that the code mistakenly took for granted that following constructor
is able to create an APFloat from a *SIGNED* integer:
APFloat::APFloat(const fltSemantics &ourSemantics, integerPart value)
rdar://13486998
llvm-svn: 177906
This simplification happens at 2 places :
- using the nsw attribute when the shl / mul is used by a sign test
- when the shl / mul is compared for (in)equality to zero
llvm-svn: 177856
Before: the function name was stored by the compiler as a constant string
and the run-time was printing it.
Now: the PC is stored instead and the run-time prints the full symbolized frame.
This adds a couple of instructions into every function with non-empty stack frame,
but also reduces the binary size because we store less strings (I saw 2% size reduction).
This change bumps the asan ABI version to v3.
llvm part.
Example of report (now):
==31711==ERROR: AddressSanitizer: stack-buffer-overflow on address 0x7fffa77cf1c5 at pc 0x41feb0 bp 0x7fffa77cefb0 sp 0x7fffa77cefa8
READ of size 1 at 0x7fffa77cf1c5 thread T0
#0 0x41feaf in Frame0(int, char*, char*, char*) stack-oob-frames.cc:20
#1 0x41f7ff in Frame1(int, char*, char*) stack-oob-frames.cc:24
#2 0x41f477 in Frame2(int, char*) stack-oob-frames.cc:28
#3 0x41f194 in Frame3(int) stack-oob-frames.cc:32
#4 0x41eee0 in main stack-oob-frames.cc:38
#5 0x7f0c5566f76c (/lib/x86_64-linux-gnu/libc.so.6+0x2176c)
#6 0x41eb1c (/usr/local/google/kcc/llvm_cmake/a.out+0x41eb1c)
Address 0x7fffa77cf1c5 is located in stack of thread T0 at offset 293 in frame
#0 0x41f87f in Frame0(int, char*, char*, char*) stack-oob-frames.cc:12 <<<<<<<<<<<<<< this is new
This frame has 6 object(s):
[32, 36) 'frame.addr'
[96, 104) 'a.addr'
[160, 168) 'b.addr'
[224, 232) 'c.addr'
[288, 292) 's'
[352, 360) 'd'
llvm-svn: 177724
The original code used i32, and i64 if legal. This introduced unneeded
casts when they aren't legal, or when the index variable i has another
type. In order of preference: try to use i's type; use the smallest
fitting legal type (using an added DataLayout method); default to i32.
A testcase checks that this works when the index gep operand is i16.
Patch by : Ahmed Bougacha <ahmed.bougacha@gmail.com>
Reviewed by : Duncan
llvm-svn: 177712
How did this ever work?
Basically, if you have a function that's inlined into the caller, it may not
have any 'call' instructions, but any 'resume' instructions it may have should
still be forwarded to the outer (caller's) landing pad. This requires that all
of the 'landingpad' instructions in the callee have their clauses merged with
the caller's outer 'landingpad' instruction (hence the bit of ugly code in the
`forwardResume' method).
Testcase in a follow commit to the test-suite repository.
<rdar://problem/13360379> & PR15555
llvm-svn: 177680
The key part of this is ensuring that name prefixes remain in a Twine
form until we get to a point where we can nuke them under NDEBUG. This
is tricky using the old APIs as they played fast and loose with Twine,
which is prone to serious error. The inserter is much cleaner as it is
actually in the call stack leading to the setName call, and so has
a good opportunity to prepend the prefix.
This matters more than you might imagine because most runs over an
alloca find a single partition, and rewrite 3 or 4 instructions
referring to it. As a consequence doing this lazily and exclusively with
Twine allows the optimizer to delete more of it and shaves another 2% to
3% off of the release build's SROA run time for PR15412. I also think
the APIs are cleaner, and the use of Twine is more reliable, so
I consider it a win-win despite the churn required to reach this state.
llvm-svn: 177631
The simplify-libcalls pass implemented a doInitialization hook to infer
function prototype attributes for well-known functions. Given that the
simplify-libcalls pass is going away *and* that the functionattrs pass
is already in place to deduce function attributes, I am moving this logic
to the functionattrs pass. This approach was discussed during patch
review:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20121126/157465.html.
llvm-svn: 177619
Use the new `llvm_gcov_init' function to register the writeout and flush
functions. The initialization function will also call `atexit' for some cleanups
and final writout calls. But it does this only once. This is better than
checking for the `main' function, because in a library that function may not
exist.
<rdar://problem/12439551>
llvm-svn: 177579
This is espcially important because the new SROA pass goes to great
lengths to provide helpful names for debugging, and as a consequence
they can become very slow to render.
Good for between 5% and 15% of the SROA runtime on some slow test cases
such as the one in PR15412.
llvm-svn: 177495
We don't want to write out >1000 files at the same time. That could make things
prohibitively expensive. Instead, register the "writeout" function so that it's
emitted serially.
<rdar://problem/12439551>
llvm-svn: 177437
- it is trivially known to be used inside the loop in a way that can not be optimized away
- there is no use outside of the loop which can take advantage of the computation hoisting
llvm-svn: 177432
For each compile unit, we want to register a function that will flush that
compile unit. Otherwise, __gcov_flush() would only flush the counters within the
current compile unit, and not any outside of it.
PR15191 & <rdar://problem/13167507>
llvm-svn: 177340
Rules include:
1)1 x*y +/- x*z => x*(y +/- z)
(the order of operands dosen't matter)
2) y/x +/- z/x => (y +/- z)/x
The transformation is disabled if the new add/sub expr "y +/- z" is a
denormal/naz/inifinity.
rdar://12911472
llvm-svn: 177088
The fundamental problem is that SROA didn't allow for overly wide loads
where the bits past the end of the alloca were masked away and the load
was sufficiently aligned to ensure there is no risk of page fault, or
other trapping behavior. With such widened loads, SROA would delete the
load entirely rather than clamping it to the size of the alloca in order
to allow mem2reg to fire. This was exposed by a test case that neatly
arranged for GVN to run first, widening certain loads, followed by an
inline step, and then SROA which miscompiles the code. However, I see no
reason why this hasn't been plaguing us in other contexts. It seems
deeply broken.
Diagnosing all of the above took all of 10 minutes of debugging. The
really annoying aspect is that fixing this completely breaks the pass.
;] There was an implicit reliance on the fact that no loads or stores
extended past the alloca once we decided to rewrite them in the final
stage of SROA. This was used to encode information about whether the
loads and stores had been split across multiple partitions of the
original alloca. That required threading explicit tracking of whether
a *use* of a partition is split across multiple partitions.
Once that was done, another problem arose: we allowed splitting of
integer loads and stores iff they were loads and stores to the entire
alloca. This is a really arbitrary limitation, and splitting at least
some integer loads and stores is crucial to maximize promotion
opportunities. My first attempt was to start removing the restriction
entirely, but currently that does Very Bad Things by causing *many*
common alloca patterns to be fully decomposed into i8 operations and
lots of or-ing together to produce larger integers on demand. The code
bloat is terrifying. That is still the right end-goal, but substantial
work must be done to either merge partitions or ensure that small i8
values are eagerly merged in some other pass. Sadly, figuring all this
out took essentially all the time and effort here.
So the end result is that we allow splitting only when the load or store
at least covers the alloca. That ensures widened loads and stores don't
hurt SROA, and that we don't rampantly decompose operations more than we
have previously.
All of this was already fairly well tested, and so I've just updated the
tests to cover the wide load behavior. I can add a test that crafts the
pass ordering magic which caused the original PR, but that seems really
brittle and to provide little benefit. The fundamental problem is that
widened loads should Just Work.
llvm-svn: 177055
constructs default arguments. It can now take default arguments from
cl::opt'ions. Add a new -default-gcov-version=... option, and actually test it!
Sink the reverse-order of the version into GCOVProfiling, hiding it from our
users.
llvm-svn: 177002
emitProfileNotes(), similar to emitProfileArcs(). Also update its comment.
Also add a comment on Version[4] (there will be another comment in clang later),
and compress lines that exceeded 80 columns.
llvm-svn: 176994
Nadav reported a performance regression due to the work I did to
merge the library call simplifier into instcombine [1]. The issue
is that a new LibCallSimplifier object is being created whenever
InstCombiner::runOnFunction is called. Every time a LibCallSimplifier
object is used to optimize a call it creates a hash table to map from
a function name to an object that optimizes functions of that name.
For short-lived LibCallSimplifier instances this is quite inefficient.
Especially for cases where no calls are actually simplified.
This patch fixes the issue by dropping the hash table and implementing
an explicit lookup function to correlate the function name to the object
that optimizes functions of that name. This avoids the cost of always
building and destroying the hash table in cases where the LibCallSimplifier
object is short-lived and avoids the cost of building the table when no
simplifications are actually preformed.
On a benchmark containing 100,000 calls where none of them are simplified
I noticed a 30% speedup. On a benchmark containing 100,000 calls where
all of them are simplified I noticed an 8% speedup.
[1] http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20130304/167639.html
llvm-svn: 176840
We want vectorization to happen at -g. Ignore calls to the dbg.value intrinsic
and don't transfer them to the vectorized code.
radar://13378964
llvm-svn: 176768
it. Fortunately, versions of gcov that predate the extra checksum also ignore
any extra data, so this isn't a problem. There will be a matching commit in
compiler-rt.
llvm-svn: 176745
into the actual gcov file.
Instead of using the bottom 4 bytes as the function identifier, use a counter.
This makes the identifier numbers stable across multiple runs.
llvm-svn: 176616
Fixes rdar:13349374.
Volatile loads and stores need to be preserved even if the language
standard says they are undefined. "volatile" in this context means "get
out of the way compiler, let my platform handle it".
Additionally, this is the only way I know of with llvm to write to the
first page (when hardware allows) without dropping to assembly.
llvm-svn: 176599
When considering folding a bitcast of an alloca into the alloca itself,
make sure we don't shrink the amount of memory being allocated, or
things rapidly go sideways.
rdar://13324424
llvm-svn: 176547
* Only apply divide bypass optimization when not optimizing for size.
* Fixed bug caused by constant for 0 value of type Int32,
used dividend type to generate the constant instead.
* For atom x86-64 apply the divide bypass to use 16-bit divides instead of
64-bit divides when operand values are small enough.
* Added lit tests for 64-bit divide bypass.
Patch by Tyler Nowicki!
llvm-svn: 176442
The LoopVectorizer often runs multiple times on the same function due to inlining.
When this happens the loop vectorizer often vectorizes the same loops multiple times, increasing code size and adding unneeded branches.
With this patch, the vectorizer during vectorization puts metadata on scalar loops and marks them as 'already vectorized' so that it knows to ignore them when it sees them a second time.
PR14448.
llvm-svn: 176399
The instcombine recognized pattern looks like:
a = b * c
d = a +/- Cst
or
a = b * c
d = Cst +/- a
When creating the new operands for fadd or fsub instruction following the related fmul, the first operand was created with the second original operand (M0 was created with C1) and the second with the first (M1 with Opnd0).
The fix consists in creating the new operands with the appropriate original operand, i.e., M0 with Opnd0 and M1 with C1.
llvm-svn: 176300
Shadow checks are disabled and memory loads always produce fully initialized
values in functions that don't have a sanitize_memory attribute. Value and
argument shadow is propagated as usual.
This change also updates blacklist behaviour to match the above.
llvm-svn: 176247
This properly asks TargetLibraryInfo if a call is available and if it is, it
can be translated into the corresponding LLVM builtin. We don't vectorize sqrt()
yet because I'm not sure about the semantics for negative numbers. The other
intrinsic should be exact equivalents to the libm functions.
Differential Revision: http://llvm-reviews.chandlerc.com/D465
llvm-svn: 176188
enhancement done the trivial way; by extending inputs and truncating outputs
which is addequate for targets with little or no support for integer arithmetic
on integer types less than 32 bits.
llvm-svn: 176139
These are two related changes (one in llvm, one in clang).
LLVM:
- rename address_safety => sanitize_address (the enum value is the same, so we preserve binary compatibility with old bitcode)
- rename thread_safety => sanitize_thread
- rename no_uninitialized_checks -> sanitize_memory
CLANG:
- add __attribute__((no_sanitize_address)) as a synonym for __attribute__((no_address_safety_analysis))
- add __attribute__((no_sanitize_thread))
- add __attribute__((no_sanitize_memory))
for S in address thread memory
If -fsanitize=S is present and __attribute__((no_sanitize_S)) is not
set llvm attribute sanitize_S
llvm-svn: 176075
This is a common pattern with dyn_cast and similar constructs, when the
PHI no longer depends on the select it can often be turned into a simpler
construct or even get hoisted out of the loop.
PR15340.
llvm-svn: 175995
The 'nobuiltin' attribute is applied to call sites to indicate that LLVM should
not treat the callee function as a built-in function. I.e., it shouldn't try to
replace that function with different code.
llvm-svn: 175835
Storing the load/store instructions with the values
and inspect them using Alias Analysis to make sure
they don't alias, since the GEP pointer operand doesn't
take the offset into account.
Trying hard to not add any extra cost to loads and stores
that don't overlap on global values, AA is *only* calculated
if all of the previous attempts failed.
Using biggest vector register size as the stride for the
vectorization access, as we're being conservative and
the cost model (which calculates the real vectorization
factor) is only run after the legalization phase.
We might re-think this relationship in the future, but
for now, I'd rather be safe than sorry.
llvm-svn: 175818
Nick Lewycky