This is a minor extension of SROA to handle a special case that is
important for some ARM NEON operations. Some of the NEON intrinsics
return multiple values, which are handled as struct types containing
multiple elements of the same vector type. The corresponding return
types declared in the arm_neon.h header have equivalent arrays. We
need SROA to recognize that it can split up those arrays and structs
into separate vectors, even though they are not always accessed with
the same type. SROA already handles loads and stores of an entire
alloca by using insertvalue/extractvalue to access the individual
pieces, and that code works the same regardless of whether the type
is a struct or an array. So, all that needs to be done is to check
for compatible arrays and homogeneous structs.
llvm-svn: 123381
SROA only split up structs and arrays one level at a time, so padding can
only cause trouble if it is located in between the struct or array elements.
llvm-svn: 123380
phi nodes. It is called from MergeBlockIntoPredecessor which is
called from GVN, which claims to preserve these.
I'm skeptical that this is the actual problem behind PR8954, but
this is a stab in the right direction.
llvm-svn: 123222
without informing memdep. This could cause nondeterminstic weirdness
based on where instructions happen to get allocated, and will hopefully
breath some life into some broken testers.
llvm-svn: 123124
larger memsets. Among other things, this fixes rdar://8760394 and
allows us to handle "Example 2" from http://blog.regehr.org/archives/320,
compiling it into a single 4096-byte memset:
_mad_synth_mute: ## @mad_synth_mute
## BB#0: ## %entry
pushq %rax
movl $4096, %esi ## imm = 0x1000
callq ___bzero
popq %rax
ret
llvm-svn: 123089
that it was leaving in loops after rotation (between the original latch
block and the original header.
With this change, it is possible for rotated loops to have just a single
basic block, which is useful.
llvm-svn: 123075
1. Rip out LoopRotate's domfrontier updating code. It isn't
needed now that LICM doesn't use DF and it is super complex
and gross.
2. Make DomTree updating code a lot simpler and faster. The
old loop over all the blocks was just to find a block??
3. Change the code that inserts the new preheader to just use
SplitCriticalEdge instead of doing an overcomplex
reimplementation of it.
No behavior change, except for the name of the inserted preheader.
llvm-svn: 123072
they all ready do). This removes two dominator recomputations prior to isel,
which is a 1% improvement in total llc time for 403.gcc.
The only potentially suspect thing is making GCStrategy recompute dominators if
it used a custom lowering strategy.
llvm-svn: 123064
them into the loop preheader, eliminating silly instructions like
"icmp i32 0, 100" in fixed tripcount loops. This also better exposes the
bigger problem with loop rotate that I'd like to fix: once this has been
folded, the duplicated conditional branch *often* turns into an uncond branch.
Not aggressively handling this is pessimizing later loop optimizations
somethin' fierce by making "dominates all exit blocks" checks fail.
llvm-svn: 123060
1. Take a flags argument instead of a bool. This makes
it more clear to the reader what it is used for.
2. Add a flag that says that "remapping a value not in the
map is ok".
3. Reimplement MapValue to share a bunch of code and be a lot
more efficient. For lookup failures, don't drop null values
into the map.
4. Using the new flag a bunch of code can vaporize in LinkModules
and LoopUnswitch, kill it.
No functionality change.
llvm-svn: 123058
map from ValueMapper.h (giving us access to its utilities)
and add a fastpath in the loop rotation code, avoiding expensive
ssa updator manipulation for values with nothing to update.
llvm-svn: 123057
skipping them, but it should probably use a worklist and only revisit those
instructions in subloops that have actually changed. It should probably also
use a worklist after the first iteration like instsimplify now does. Regardless,
it's only 0.3% of opt -O2 time on 403.gcc if it replaces the instcombine placed
in the middle of the loop passes.
llvm-svn: 122868
when safe.
The testcase is basically this nested loop:
void foo(char *X) {
for (int i = 0; i != 100; ++i)
for (int j = 0; j != 100; ++j)
X[j+i*100] = 0;
}
which gets turned into a single memset now. clang -O3 doesn't optimize
this yet though due to a phase ordering issue I haven't analyzed yet.
llvm-svn: 122806
instruction *after* the store. The store will always be deleted
if the transformation kicks in, so we'd do an N^2 scan of every
loop block. Whoops.
llvm-svn: 122805
FunctionPass. It probably doesn't have a reason to be a LoopPass, as it will
probably drop the simple fixed point and either use RPO iteration or Duncan's
approach in instsimplify of only revisiting instructions that have changed.
The next step is to preserve LoopSimplify. This looks like it won't be too hard,
although the pass manager doesn't actually seem to respect when non-loop passes
claim to preserve LCSSA or LoopSimplify. This will have to be fixed.
llvm-svn: 122791
that are allowed to have metadata operands are intrinsic calls,
and the only ones that take metadata currently return void.
Just reject all void instructions, which should not be value
numbered anyway. To future proof things, add an assert to the
getHashValue impl for calls to check that metadata operands
aren't present.
llvm-svn: 122759
nested values, so they can change and drop to null, which can
change the hash and cause havok.
It turns out that it isn't a good idea to value number stuff
with metadata operands anyway, so... don't.
llvm-svn: 122758
capacity on the Visited SmallPtrSet. On 403.gcc, this is about a 4.5% speedup of
CodeGenPrepare time (which itself is 10% of time spent in the backend).
This is progress towards PR8889.
llvm-svn: 122741
of instcombine that is currently in the middle of the loop pass pipeline. This
commit only checks in the pass; it will hopefully be enabled by default later.
llvm-svn: 122719
sure that the loop we're promoting into a memcpy doesn't mutate the input
of the memcpy. Before we were just checking that the dest of the memcpy
wasn't mod/ref'd by the loop.
llvm-svn: 122712
isExitBlockDominatedByBlockInLoop is a relic of the days when domtree was
*just* a tree and didn't have DFS numbers. Checking DFS numbers is faster
and easier than "limiting the search of the tree".
llvm-svn: 122702
header for now for memset/memcpy opportunities. It turns out that loop-rotate
is successfully rotating loops, but *DOESN'T MERGE THE BLOCKS*, turning "for
loops" into 2 basic block loops that loop-idiom was ignoring.
With this fix, we form many *many* more memcpy and memsets than before, including
on the "history" loops in the viterbi benchmark, which look like this:
for (j=0; j<MAX_history; ++j) {
history_new[i][j+1] = history[2*i][j];
}
Transforming these loops into memcpy's speeds up the viterbi benchmark from
11.98s to 3.55s on my machine. Woo.
llvm-svn: 122685
pipeline to be caught by instcombine, and it's not feasible to catch them in SimplifyCFG because the
use-lists are in an inconsistent state at the point where it could know that it need to simplify them.
Instead, have CodeGenPrepare look for trivially redundant PHIs as part of its general cleanup effort.
llvm-svn: 122516
I still think that LVI should be handling this, but that capability is some ways off in the future,
and this matters for some significant benchmarks.
llvm-svn: 122378
which have trapping constant exprs in them due to PHI nodes.
Eliminating them can cause the constant expr to be evalutated
on new paths if the input edges are critical.
llvm-svn: 122164
a null endptr argument, because they may write to errno.
This fixes a seflhost miscompile observed on Linux targets when TBAA
was enabled.
llvm-svn: 122014
When it sees a promising select it now tries to figure out whether the condition of the select is known in any of the predecessors and if so it maps the operands appropriately.
llvm-svn: 121859
zextOrTrunc(), and APSInt methods extend(), extOrTrunc() and new method
trunc(), to be const and to return a new value instead of modifying the
object in place.
llvm-svn: 121120
memcpy's like:
memcpy(A, B)
memcpy(A, C)
we cannot delete the first memcpy as dead if A and C might be aliases.
If so, we actually get:
memcpy(A, B)
memcpy(A, A)
which is not correct to transform into:
memcpy(A, A)
This patch was heavily influenced by Jakub Staszak's patch in PR8728, thanks
Jakub!
llvm-svn: 120974
Should have no functional change other than the order of two transformations that are mutually-exclusive and the exact formatting of debug output.
Internally, it now stores the ConstantInt*s as Constant*s, and actual undef values instead of nulls.
llvm-svn: 120946
1. if the underlying pointer passed in can be resolved
to any argument or alloca, then we don't need to scan.
Previously we would only avoid the scan if the alloca
or byval was actually considered dead.
2. The dead store processing code is itself completely
dead and didn't handle volatile stores right anyway,
so delete it. This allows simplifying the interface
to RemoveAccessedObjects.
llvm-svn: 120467
made sense to me. We now have a set of dead stack objects, and
they become live when loaded. Fix a theoretical problem where
we'd pass in the wrong pointer to the alias query.
llvm-svn: 120465
If the call might read all the allocas, stop scanning early.
Convert a vector to smallvector, shrink SmallPtrSet to 16 instead
of 64 to avoid crazy linear scans.
llvm-svn: 120463
about pairs of AA::Location's instead of looking for MemDep's
"Def" predicate. This is more powerful and general, handling
memset/memcpy/store all uniformly, and implementing PR8701 and
probably obsoleting parts of memcpyoptimizer.
This also fixes an obscure bug with init.trampoline and i8
stores, but I'm not surprised it hasn't been hit yet. Enhancing
init.trampoline to carry the size that it stores would allow
DSE to be much more aggressive about optimizing them.
llvm-svn: 120406
contains "ref".
Enhance DSE to use a modref query instead of a store-specific hack
to generalize the "ignore may-alias stores" optimization to handle
memset and memcpy.
llvm-svn: 120368
1. Don't bother trying to optimize:
lifetime.end(ptr)
store(ptr)
as it is undefined, and therefore shouldn't exist.
2. Move the 'storing a loaded pointer' xform up, simplifying
the may-aliased store code.
llvm-svn: 120359
by my recent GVN improvement. Looking through a single layer of
PHI nodes when attempting to sink GEPs, we need to iteratively
look through arbitrary PHI nests.
llvm-svn: 120202
allowing the memcpy to be eliminated.
Unfortunately, the requirements on byval's without explicit
alignment are really weak and impossible to predict in the
mid-level optimizer, so this doesn't kick in much with current
frontends. The fix is to change clang to set alignment on all
byval arguments.
llvm-svn: 119916
preserves LCSSA form out of ScalarEvolution and into the LoopInfo
class. Use it to check that SimplifyInstruction simplifications
are not breaking LCSSA form. Fixes PR8622.
llvm-svn: 119727
this was a tree of hashtables, and a query recursed into the table for the immediate dominator ad infinitum
if the initial lookup failed. This led to really bad performance on tall, narrow CFGs.
We can instead replace it with what is conceptually a multimap of value numbers to leaders (actually
represented by a hashtable with a list of Value*'s as the value type), and then
determine which leader from that set to use very cheaply thanks to the DFS numberings maintained by
DominatorTree. Because there are typically few duplicates of a given value, this scan tends to be
quite fast. Additionally, we use a custom linked list and BumpPtr allocation to avoid any unnecessary
allocation in representing the value-side of the multimap.
This change brings with it a 15% (!) improvement in the total running time of GVN on 403.gcc, which I
think is pretty good considering that includes all the "real work" being done by MemDep as well.
The one downside to this approach is that we can no longer use GVN to perform simple conditional progation,
but that seems like an acceptable loss since we now have LVI and CorrelatedValuePropagation to pick up
the slack. If you see conditional propagation that's not happening, please file bugs against LVI or CVP.
llvm-svn: 119714
refusing to optimize two memcpy's like this:
copy A <- B
copy C <- A
if it couldn't prove that noalias(B,C). We can eliminate
the copy by producing a memmove instead of memcpy.
llvm-svn: 119694
if it is passed as a byval argument. The byval argument will just be a
read, so it is safe to read from the original global instead. This allows
us to promote away the %agg.tmp alloca in PR8582
llvm-svn: 119686
systematically, CollapsePhi will always return null here. Note
that CollapsePhi did an extra check, isSafeReplacement, which
the SimplifyInstruction logic does not do. I think that check
was bogus - I guess we will soon find out! (It was originally
added in commit 41998 without a testcase).
llvm-svn: 119456
"%z = %x and %y". If GVN can prove that %y equals %x, then it turns
this into "%z = %x and %x". With the new code, %z will be replaced
with %x everywhere (and then deleted). Previously %z would be value
numbered too, which is a waste of time. Also, while a clever value
numbering algorithm would give %z the same value number as %x, our
current one doesn't do so (at least I don't think it does). The new
logic has an essentially equivalent effect to what you would get if
%z was given the same value number as %x, i.e. it should make value
numbering smarter. While there, get hold of target data once at the
start rather than a gazillion times all over the place.
llvm-svn: 118923
references. For example, this allows gvn to eliminate the load in
this example:
void foo(int n, int* p, int *q) {
p[0] = 0;
p[1] = 1;
if (n) {
*q = p[0];
}
}
llvm-svn: 118714
needs to be guaranteed never to be run on an unreachable block. However, earlier block simplifications may have
changed the CFG to make block that were reachable when we began our iteration unreachable by the time we try to
simplify them. (Note that this also means that our depth-first iterators were potentially being invalidated).
This should not have a large impact on code quality, since later runs of instcombine should pick up these simplifications.
Fixes PR8506.
llvm-svn: 117709
must be called in the pass's constructor. This function uses static dependency declarations to recursively initialize
the pass's dependencies.
Clients that only create passes through the createFooPass() APIs will require no changes. Clients that want to use the
CommandLine options for passes will need to manually call the appropriate initialization functions in PassInitialization.h
before parsing commandline arguments.
I have tested this with all standard configurations of clang and llvm-gcc on Darwin. It is possible that there are problems
with the static dependencies that will only be visible with non-standard options. If you encounter any crash in pass
registration/creation, please send the testcase to me directly.
llvm-svn: 116820
perform initialization without static constructors AND without explicit initialization
by the client. For the moment, passes are required to initialize both their
(potential) dependencies and any passes they preserve. I hope to be able to relax
the latter requirement in the future.
llvm-svn: 116334
formulae which become illegal as a result of the offset updating don't
escape.
This is for rdar://8529692. No testcase yet, because the given cases
hit use-list ordering differences.
llvm-svn: 116093
This doesn't usually matter, because the other heuristics usually
succeed regardless, but it's good to keep the register use
bookkeeping consistent.
llvm-svn: 116005
Anyone interested in more general PRE would be better served by implementing it separately, to get real
anticipation calculation, etc.
llvm-svn: 115337
The x86_mmx type is used for MMX intrinsics, parameters and
return values where these use MMX registers, and is also
supported in load, store, and bitcast.
Only the above operations generate MMX instructions, and optimizations
do not operate on or produce MMX intrinsics.
MMX-sized vectors <2 x i32> etc. are lowered to XMM or split into
smaller pieces. Optimizations may occur on these forms and the
result casted back to x86_mmx, provided the result feeds into a
previous existing x86_mmx operation.
The point of all this is prevent optimizations from introducing
MMX operations, which is unsafe due to the EMMS problem.
llvm-svn: 115243
Because of this, we cannot use the Simplify* APIs, as they can assert-fail on unreachable code. Since it's not easy to determine
if a given threading will cause a block to become unreachable, simply defer simplifying simplification to later InstCombine and/or
DCE passes.
llvm-svn: 115082
register pressure and thus excess spills, which we don't currently recover from well. This should
be re-evaluated in the future if our ability to generate good spills/splits improves.
Partial fix for <rdar://problem/7635585>.
llvm-svn: 114919
This reverts revision 114633. It was breaking llvm-gcc-i386-linux-selfhost.
It seems there is a downstream bug that is exposed by
-cgp-critical-edge-splitting=0. When that bug is fixed, this patch can go back
in.
Note that the changes to tailcallfp2.ll are not reverted. They were good are
required.
llvm-svn: 114859
Splitting critical edges at the merge point only addressed part of the issue; it is also possible for non-post-domination
to occur when the path from the load to the merge has branches in it. Unfortunately, full anticipation analysis is
time-consuming, so for now approximate it. This is strictly more conservative than real anticipation, so we will miss
some cases that real PRE would allow, but we also no longer insert loads into paths where they didn't exist before. :-)
This is a very slight net positive on SPEC for me (0.5% on average). Most of the benchmarks are largely unaffected, but
when it pays off it pays off decently: 181.mcf improves by 4.5% on my machine.
llvm-svn: 114785
truncates are free only in the case where the extended type is legal but the
load type is not. If both types are illegal, such as when they are too big,
the load may not be legalized into an extended load.
llvm-svn: 114568
load when the type of the load is not legal, even if truncates are not free.
The load is going to be legalized to an extending load anyway.
llvm-svn: 114488
walking the asm arguments once and stashing their Values. This is
wrong because the same memory location can be in the list twice, and
if the first one has a sunkaddr substituted, the stashed value for the
second one will be wrong (use-after-free). PR 8154.
llvm-svn: 114104
not unrolling loops that contain calls that would be better off getting inlined. This mostly
comes up when an interleaved devirtualization pass has devirtualized a call which the inliner
will inline on a future pass. Thus, rather than blocking all loops containing calls, add
a metric for "inline candidate calls" and block loops containing those instead.
llvm-svn: 113535
unrolling threshold to the optimize-for-size threshold. Basically, for loops containing calls, unrolling
can still be profitable as long as the loop is REALLY small.
llvm-svn: 113439
The threshold value of 50 is arbitrary, and I chose it simply by analogy to the inlining thresholds, where
the baseline unrolling threshold is slightly smaller than the baseline inlining threshold. This could
undoubtedly use some tuning.
llvm-svn: 113306
in the duplicated block instead of duplicating them.
Duplicating them into the end of the loop and the preheader
means that we got a phi node in the header of the loop,
which prevented LICM from hoisting them. GVN would
usually come around later and merge the duplicated
instructions so we'd get reasonable output... except that
anything dependent on the shoulda-been-hoisted value can't
be hoisted. In PR5319 (which this fixes), a memory value
didn't get promoted.
llvm-svn: 113134
Loop::hasLoopInvariantOperands method. Remove
a useless and confusing Loop::isLoopInvariant(Instruction)
method, which didn't do what you thought it did.
No functionality change.
llvm-svn: 113133
location is being re-stored to the memory location. We would get
a dangling pointer from the SSAUpdate data structure and miss a
use. This fixes PR8068
llvm-svn: 113042
I'm sure it is harmless. Original commit message:
If PrototypeValue is erased in the middle of using the SSAUpdator
then the SSAUpdator may access freed memory. Instead, simply pass
in the type and name explicitly, which is all that was used anyway.
llvm-svn: 112810
on llvmdev: SRoA is introducing MMX datatypes like <1 x i64>,
which then cause random problems because the X86 backend is
producing mmx stuff without inserting proper emms calls.
In the short term, force off MMX datatypes. In the long term,
the X86 backend should not select generic vector types to MMX
registers. This is being worked on, but won't be done in time
for 2.8. rdar://8380055
llvm-svn: 112696
This actually exposed an infinite recursion bug in ComputeValueKnownInPredecessors which theoretically already existed (in JumpThreading's
handling of and/or of i1's), but never manifested before. This patch adds a tracking set to prevent this case.
llvm-svn: 112589
instead of PromoteMemToReg. This allows it to stop using DF and DT,
eliminating a computation of DT and DF from clang -O3. Clang is now
down to 2 runs of DomFrontier.
llvm-svn: 112457
assertingvh so we get a violent explosion if the pointer dangles.
2) Fix AliasSetTracker::deleteValue to remove call sites with
by-pointer comparisons instead of by-alias queries. Using
findAliasSetForCallSite can cause alias sets to get merged
when they shouldn't, and can also miss alias sets when the
call is readonly.
#2 fixes PR6889, which only repros with a .c file :(
llvm-svn: 112452
LICM correctly. When sinking an instruction, it should not add
entries for the sunk instruction to the AST, it should remove
the entry for the sunk instruction. The blocks being sunk to
are not in the loop, so their instructions shouldn't be in the
AST (yet)!
llvm-svn: 112447
keeping them around until the pass is destroyed, keep them
around a) just when useful (not for outer loops) and b) destroy
them right after we use them. This should reduce memory use
and fixes potential bugs where a loop is deleted and another
loop gets allocated to the same address.
llvm-svn: 112446
LSRInstance data structures up to date. This fixes some
pessimizations caused by stale data which will be exposed
in an upcoming change.
llvm-svn: 112440
from the LHS should disable reconsidering that pred on the
RHS. However, knowing something about the pred on the RHS
shouldn't disable subsequent additions on the RHS from
happening.
llvm-svn: 111349
uninteresting, just put all the operands on one list and make
GenerateReassociations make the decision about what's interesting.
This is simpler, and it avoids an extra ScalarEvolution::getAddExpr call.
llvm-svn: 111133
- Eliminate redundant successors.
- Convert an indirectbr with one successor into a direct branch.
Also, generalize SimplifyCFG to be able to be run on a function entry block.
It knows quite a few simplifications which are applicable to the entry
block, and it only needs a few checks to avoid trouble with the entry block.
llvm-svn: 111060
dependence on DominanceFrontier. Instead, add an explicit DominanceFrontier
pass in StandardPasses.h to ensure that it gets scheduled at the right
time.
Declare that loop unrolling preserves ScalarEvolution, and shuffle some
getAnalysisUsages.
This eliminates one LoopSimplify and one LCCSA run in the standard
compile opts sequence.
llvm-svn: 109413
different widths. In a use with a narrower fixup, formulae
may be wider than the fixup, in which case the high bits
aren't necessarily meaningful, so it isn't safe to reuse
them for uses with wider fixups.
This fixes PR7618, though the testcase is too large for a
reasonable regression test, since it heavily dependes on
hitting LSR's heuristics in a certain way.
llvm-svn: 108455
by a return that returns a constant, while elsewhere in the function
another return instruction returns a different constant. This is a
special case of accumulator recursion, so just generalize the existing
logic a bit.
llvm-svn: 108241
operation, but the way it's implemented requires the operation to also be
commutative. So add a check for commutativity (and tweak the corresponding
comments). This makes no difference in practice since every associative
LLVM instruction is also commutative! Here's an example to show the need
for commutativity: the accum_recursion.ll testcase calculates the factorial
function. Before the transformation the result of a call is
((((1*1)*2)*3)...)*x
while afterwards it is
(((1*x)*(x-1))...*2)*1
which clearly requires both associativity and commutativity of * to be equal
to the original.
llvm-svn: 108056
have any effect, and second, deleting stores can potentially invalidate
an AliasAnalysis, and there's currently no notification for this.
llvm-svn: 107496
large integers, the first inserted value would always create
an 'or X, 0'. Even though this is trivially zapped by
instcombine, don't bother creating this pointless instruction.
llvm-svn: 106979
the returned value after the tail call if it differs from other return
values. The optimal thing to do would be to introduce a phi node for
the return value, but for the moment just fix the miscompile.
llvm-svn: 106947
SCEVUnknown values which are loop-variant, as LSR can't do anything
interesting with these values in any case. This fixes very slow compile
times on loops which have large numbers of such values.
llvm-svn: 106897
for an "i" constraint should get lowered; PR 6309. While
this argument was passed around a lot, this is the only
place it was used, so it goes away from a lot of other
places.
llvm-svn: 106893
use sharing map. The reconcileNewOffset logic already forces a
separate use if the kinds differ, so incorporating the kind in the
key means we can track more sharing opportunities.
More sharing means fewer total uses to track, which means smaller
problem sizes, which means the conservative throttles don't kick
in as often.
llvm-svn: 106396
Changed directly instead of using a return value.
Rename FilterOutUndesirableDedicatedRegisters's Changed variable to
distinguish it from LSRInstance's Changed member.
llvm-svn: 104269
operand on the left, the interesting operand is on the right. This
fixes a bug where LSR was failing to recognize ICmpZero uses,
which led it to be unable to reverse the induction variable in the
attached testcase.
Delete test/CodeGen/X86/stack-color-with-reg-2.ll, because its test
is extremely fragile and hard to meaningfully update.
llvm-svn: 104262
when it detects undefined behavior. llvm.trap generally codegens into some
thing really small (e.g. a 2 byte ud2 instruction on x86) and debugging this
sort of thing is "nontrivial". For example, we now compile:
void foo() { *(int*)0 = 42; }
into:
_foo:
pushl %ebp
movl %esp, %ebp
ud2
Some may even claim that this is a security hole, though that seems dubious
to me. This addresses rdar://7958343 - Optimizing away null dereference
potentially allows arbitrary code execution
llvm-svn: 103356
LSRUse's Regs set after all pruning is done, rather than trying
to do it on the fly, which can produce an incomplete result.
This fixes a case where heuristic pruning was stripping all
formulae from a use, which led the solver to enter an infinite
loop.
Also, add a few asserts to diagnose this kind of situation.
llvm-svn: 103328
indirect branches in all the predecessors. This avoids unnecessarily
splitting edges in cases where load PRE is not possible anyway.
Thanks to Jakub Staszak for pointing this out.
llvm-svn: 103034
arguments are handled with a new InlineFunctionInfo class. This
makes it easier to extend InlineFunction to return more info in the
future.
llvm-svn: 102137
condition we're unswitching on. In this case, don't try to
simplify the second copy of the loop which may be dead or not,
but is probably a constant now. This fixes PR6879
llvm-svn: 101870
just ask ScalarEvolution for it on demand. This helps IVUsers be more robust
in the case of expressions changing underneath it. This fixes PR6862.
llvm-svn: 101819
with a fix for self-hosting
rotate CallInst operands, i.e. move callee to the back
of the operand array
the motivation for this patch are laid out in my mail to llvm-commits:
more efficient access to operands and callee, faster callgraph-construction,
smaller compiler binary
llvm-svn: 101465
with a fix
rotate CallInst operands, i.e. move callee to the back
of the operand array
the motivation for this patch are laid out in my mail to llvm-commits:
more efficient access to operands and callee, faster callgraph-construction,
smaller compiler binary
llvm-svn: 101397
of the operand array
the motivation for this patch are laid out in my mail to llvm-commits:
more efficient access to operands and callee, faster callgraph-construction,
smaller compiler binary
llvm-svn: 101364
numerator is an induction variable. For example, with code like this:
for (i=0;i<n;++i)
x[i%n] = 0;
IndVarSimplify will now recognize that i is always less than n inside
the loop, and eliminate the remainder.
llvm-svn: 101113
expression is a UDiv and it doesn't appear that the UDiv came from
the user's source.
ScalarEvolution has recently figured out how to compute a tripcount
expression for the inner loop in
SingleSource/Benchmarks/Shootout/sieve.c, using a udiv. Emitting a
udiv instruction dramatically slows down the enclosing loop.
llvm-svn: 101068
a ScalarEvolution bug with overflow handling is fixed, the normal analysis
code will automatically decline to operate on the icmp instructions which
are responsible for the loop exit.
llvm-svn: 101032
instead of deleting just the user. This makes it more consistent with
other code in IndVarSimplify, and theoretically can eliminate more users
earlier.
llvm-svn: 101027
the loop exit test. This usually doesn't come up for a variety of
reasons, but it isn't impossible, so make IndVarSimplify handle it
conservatively.
llvm-svn: 101008
variables. For example, with code like this:
for (i=0;i<n;++i)
if (i<n)
x[i] = 0;
IndVarSimplify will now recognize that i is always less than n inside
the loop, and eliminate the if.
llvm-svn: 101000
into adjacent loops. Also, ensure that the insert position is
dominated by the loop latch of any loop in the post-inc set which
has a latch.
llvm-svn: 100906
forced constant is changed to a constant, we would end
up adding the instruction to the wrong worklist,
preventing it from being properly revisited. This fixes
rdar://7832370
llvm-svn: 100837
explicitly split into stride-and-offset pairs. Also, add the
ability to track multiple post-increment loops on the same expression.
This refines the concept of "normalizing" SCEV expressions used for
to post-increment uses, and introduces a dedicated utility routine for
normalizing and denormalizing expressions.
This fixes the expansion of expressions which are post-increment users
of more than one loop at a time. More broadly, this takes LSR another
step closer to being able to reason about more than one loop at a time.
llvm-svn: 100699
undefs in branches/switches, we have two cases: a branch on a literal
undef or a branch on a symbolic value which is undef. If we have a
literal undef, the code was correct: forcing it to a constant is the
right thing to do.
If we have a branch on a symbolic value that is undef, we should force
the symbolic value to a constant, which then makes the successor block
live. Forcing the condition of the branch to being a constant isn't
safe if later paths become live and the value becomes overdefined. This
is the case that 'forcedconstant' is designed to handle, so just use it.
This fixes rdar://7765019 but there is no good testcase for this, the
one I have is too insane to be useful in the future.
llvm-svn: 100478
Added support for address spaces and added a isVolatile field to memcpy, memmove, and memset,
e.g., llvm.memcpy.i32(i8*, i8*, i32, i32) -> llvm.memcpy.p0i8.p0i8.i32(i8*, i8*, i32, i32, i1)
llvm-svn: 100304
exits the loop. With this information we can guarantee
the iteration count of the loop is bounded by the
compare. I think this xforms is finally safe now.
llvm-svn: 100285
checker. Amusingly, we already had tests that we should
have rejects because they would be miscompiled in the
testsuite.
The remaining issue with this is that we don't check that
the branch causes us to exit the loop if it fails, so we
don't actually know if we remain in bounds.
llvm-svn: 100284
to a signed vs unsigned value depending on the sign of the
constant fp means that we can't distinguish between a
truly negative number and a positive number so large the
32nd bit is set. So, do don't this!
llvm-svn: 100283
this cleans up a bunch of code and also fixes several crashes and
miscompiles. More to come unfortunately, this optimization
is quite broken.
llvm-svn: 100270
Added support for address spaces and added a isVolatile field to memcpy, memmove, and memset,
e.g., llvm.memcpy.i32(i8*, i8*, i32, i32) -> llvm.memcpy.p0i8.p0i8.i32(i8*, i8*, i32, i32, i1)
llvm-svn: 100191
is necessary. Inherits from new templated baseclass CallSiteBase<>
which is highly customizable. Base CallSite on it too, in a configuration
that allows full mutation.
Adapt some call sites in analyses to employ ImmutableCallSite.
llvm-svn: 100100
generate wrong code pretty much anywhere AFAICT.
A case that hits the bug reproducibly is impossible,
but the situation was like this:
Addr = ...
Store -> Addr
Addr2 = GEP , 0, 0
Store -> Addr2
Handling the first store, the code changed replaced Addr
with a sunkaddr and deleted Addr, but not its table
entry. Code in OptimizedBlock replaced Addr2 with a
bitcast; if that happened to reuse the memory of Addr,
the old table entry was erroneously found when handling
the second store.
llvm-svn: 100044
e.g., llvm.memcpy.i32(i8*, i8*, i32, i32) -> llvm.memcpy.p0i8.p0i8.i32(i8*, i8*, i32, i32, i1)
A update of langref will occur in a subsequent checkin.
llvm-svn: 99928
I have audited all getOperandNo calls now, fixing
hidden assumptions. CallSite related uglyness will
be eliminated successively.
Note this patch has a long and griveous history,
for all the back-and-forths have a look at
CallSite.h's log.
llvm-svn: 99399
so that the SCEVExpander doesn't retain a dangling pointer as its
insert position. The dangling pointer in this case wasn't ever used
to insert new instructions, but it was causing trouble with
SCEVExpander's code for automatically advancing its insert position
past debug intrinsics.
This fixes use-after-free errors that valgrind noticed in
test/Transforms/IndVarSimplify/2007-06-06-DeleteDanglesPtr.ll and
test/Transforms/IndVarSimplify/exit_value_tests.ll.
llvm-svn: 99036
This time I did a self-hosted bootstrap on Linux x86-64,
with no problems. Let's see how darwin 64-bit self-hosting
goes. At the first sign of failure I'll back this out.
Maybe the valgrind bots give me a hint of what may be wrong
(it at all).
llvm-svn: 98957
out the remainder of the calls that we should lower in some way and
move the tests to the new correct directory. Fix up tests that are now
optimized more than they were before by -instcombine.
llvm-svn: 97875
can be used in more places. Add an argument for the TargetData that
most of them need. Update for the getInt8PtrTy() change. Should be
no functionality change.
llvm-svn: 97844
predecessors before returning. Otherwise, if multiple predecessor edges need
splitting, we only get one of them per iteration. This makes a small but
measurable compile time improvement with -enable-full-load-pre.
llvm-svn: 97521
which branch on undef to branch on a boolean constant for the edge
exiting the loop. This helps ScalarEvolution compute trip counts for
loops.
Teach ScalarEvolution to recognize single-value PHIs, when safe, and
ForgetSymbolicName to forget such single-value PHI nodes as apprpriate
in ForgetSymbolicName.
llvm-svn: 97126
argument is non-null, pass it along to PHITranslateSubExpr so that it can
prefer using existing values that dominate the PredBB, instead of just
blindly picking the first equivalent value that it finds on a uselist.
Also when the DominatorTree is specified, have PHITranslateValue filter
out any result that does not dominate the PredBB. This is basically just
refactoring the check that used to be in GetAvailablePHITranslatedSubExpr
and also in GVN.
Despite my initial expectations, this change does not affect the results
of GVN for any testcases that I could find, but it should help compile time.
Before this change, if PHITranslateSubExpr picked a value that does not
dominate, PHITranslateWithInsertion would then insert a new value, which GVN
would later determine to be redundant and would replace. By picking a good
value to begin with, we save GVN the extra work of inserting and then
replacing a new value.
llvm-svn: 97010
induction variable value and a loop-variant value, don't force the
insert position to be at the post-increment position, because it may
not be dominated by the loop-variant value. This fixes a
use-before-def problem noticed on PPC.
llvm-svn: 96774
strides in foreign loops. This helps locate reuse opportunities
with existing induction variables in foreign loops and reduces
the need for inserting new ones. This fixes rdar://7657764.
llvm-svn: 96629
a loop exit value, so that if a loop gets deleted, ScalarEvolution
isn't stick holding on to dangling SCEVAddRecExprs for that loop. This
fixes PR6339.
llvm-svn: 96626
with multiplication by constants distributed through, occasionally
those subexpressions can include both x and -x. For now, if this
condition is discovered within LSR, just prune such cases away,
as they won't be profitable. This fixes a "zero allocated in a
base register" assertion failure.
llvm-svn: 96177
and add a doxygen comment.
Cache the phi entry to avoid doing tons of
PHINode::getBasicBlockIndex calls in the common case.
On my insane testcase from re2c, this speeds up CGP from
617.4s to 7.9s (78x).
llvm-svn: 96083
bug fixes, and with improved heuristics for analyzing foreign-loop
addrecs.
This change also flattens IVUsers, eliminating the stride-oriented
groupings, which makes it easier to work with.
llvm-svn: 95975
block. Other blocks may have pointer cycles that will crash
basicaa and other alias analyses. In any case, there is no
point wasting cycles optimizing dead blocks. This fixes
rdar://7635088
llvm-svn: 95852
Initial skeleton and SCEVUnknown lowering implemented,
the rest should come relatively quickly. Move testcase
to new directory.
Move pass to right before SimplifyLibCalls - which is
moved down a bit so we can take advantage of a few opts.
llvm-svn: 95628
short-circuited conditions to AND/OR expressions, and those expressions
are often converted back to a short-circuited form in code gen. The
original source order may have been optimized to take advantage of the
expected values, and if we reassociate them, we change the order and
subvert that optimization. Radar 7497329.
llvm-svn: 95333
The SRThreshold value makes perfect sense for checking if an entire aggregate
should be promoted to a scalar integer, but it is not so good for splitting
an aggregate into its separate elements. A struct may contain a large embedded
array along with some scalar fields that would benefit from being split apart
by SROA. Even if the total aggregate size is large, it may still be good to
perform SROA. Thus, the most important piece of this patch is simply moving
the aggregate size comparison vs. SRThreshold so that it guards only the
aggregate promotion.
We have also been checking the number of elements to decide if an aggregate
should be split up. The limit of "SRThreshold/4" seemed rather arbitrary,
and I don't think it's very useful to derive this limit from SRThreshold
anyway. I've collected some data showing that the current default limit of
32 (since SRThreshold defaults to 128) is a reasonable cutoff for struct
types. One thing suggested by the data is that distinguishing between structs
and arrays might be useful. There are (obviously) a lot more large arrays
than large structs (as measured by the number of elements and not the total
size -- a large array inside a struct still counts as a single element given
the way we do SROA right now). Out of 8377 arrays where we successfully
performed SROA while compiling a large set of benchmarks, only 16 of them had
more than 8 elements. And, for those 16 arrays, it's not at all clear that
SROA was actually beneficial. So, to offset the compile time cost of
investigating more large structs for SROA, the patch lowers the limit on array
elements to 8.
This fixes Apple Radar 7563690.
llvm-svn: 95224
disabled by default. This divides the existing load PRE code into 2 phases:
first it checks that it is safe to move the load to each of the predecessors
where it is unavailable, and then if it is safe, the code is changed to move
the load. Radar 7571861.
llvm-svn: 95007
parameter with a default value, instead of just hardcoding it in the
implementation. The limit of MaxLookup = 6 was introduced in r69151 to fix
a performance problem with O(n^2) behavior in instcombine, but the scalarrepl
pass is relying on getUnderlyingObject to go all the way back to an AllocaInst.
Making the limit part of the method signature makes it clear that by default
the result is limited and should help avoid similar problems in the future.
This fixes pr6126.
llvm-svn: 94433
for arbitrary terminators in predecessors, don't assume
it is a conditional or uncond branch. The testcase shows
an example where they can happen with switches.
llvm-svn: 94323
handle the case when we can infer an input to the xor
from all inputs that agree, instead of going into an
infinite loop. Another part of PR6199
llvm-svn: 94321
missing ones are libsupport, libsystem and libvmcore. libvmcore is
currently blocked on bugpoint, which uses EH. Once it stops using
EH, we can switch it off.
This #if 0's out 3 unit tests, because gtest requires RTTI information.
Suggestions welcome on how to fix this.
llvm-svn: 94164
loop-variant components, adds must be inserted after the increment.
Keep track of the increment position for this case, and insert
these adds in the correct location.
llvm-svn: 94110
This new version is much more aggressive about doing "full" reduction in
cases where it reduces register pressure, and also more aggressive about
rewriting induction variables to count down (or up) to zero when doing so
reduces register pressure.
It currently uses fairly simplistic algorithms for finding reuse
opportunities, but it introduces a new framework allows it to combine
multiple strategies at once to form hybrid solutions, instead of doing
all full-reduction or all base+index.
llvm-svn: 94061
than the scaled register. This makes it more likely that subsequent
AddrModeMatcher queries will match the new address the same way as the
old, instead of accidentally matching what had been the base register
as the new scaled register, and then failing to match the scaled register.
This fixes some problems with address-mode sinking multiple muls into a
block, which will be a lot more common with some upcoming
LoopStrengthReduction changes.
llvm-svn: 93935
are the same. I had already fixed a similar problem where the source and
destination were different bitcasts derived from the same alloca, but the
previous fix still did not handle the case where both operands are exactly
the same value. Radar 7552893.
llvm-svn: 93848
in JT.
2) When cloning blocks for PHI or xor conditions, use
instsimplify to simplify the code as we go. This allows us to
squish common cases early in JT which opens up opportunities for
subsequent iterations, and allows it to completely simplify the
testcase.
llvm-svn: 93253
condition is a xor with a phi node. This eliminates nonsense
like this from 176.gcc in several places:
LBB166_84:
testl %eax, %eax
- setne %al
- xorb %cl, %al
- notb %al
- testb $1, %al
- je LBB166_85
+ je LBB166_69
+ jmp LBB166_85
This is rdar://7391699
llvm-svn: 93221
on the example in PR4216. This doesn't trigger in the testsuite,
so I'd really appreciate someone scrutinizing the logic for
correctness.
llvm-svn: 92458
when a consequtive sequence of elements all satisfies the
predicate. Like the double compare case, this generates better
code than the magic constant case and generalizes to more than
32/64 element array lookups.
Here are some examples where it triggers. From 403.gcc, most
accesses to the rtx_class array are handled, e.g.:
@rtx_class = constant [153 x i8] c"xxxxxmmmmmmmmxxxxxxxxxxxxmxxxxxxiiixxxxxxxxxxxxxxxxxxxooxooooooxxoooooox3x2c21c2222ccc122222ccccaaaaaa<<<<<<<<<<<<<<<<<<111111111111bbooxxxxxxxxxxcc2211x", align 32 ; <[153 x i8]*> [#uses=547]
%142 = icmp eq i8 %141, 105
@rtx_class = constant [153 x i8] c"xxxxxmmmmmmmmxxxxxxxxxxxxmxxxxxxiiixxxxxxxxxxxxxxxxxxxooxooooooxxoooooox3x2c21c2222ccc122222ccccaaaaaa<<<<<<<<<<<<<<<<<<111111111111bbooxxxxxxxxxxcc2211x", align 32 ; <[153 x i8]*> [#uses=543]
%165 = icmp eq i8 %164, 60
Also, most of the 59-element arrays (mode_class/rid_to_yy, etc)
optimized before are actually range compares. This lets 32-bit
machines optimize them.
400.perlbmk has stuff like this:
400.perlbmk: PL_regkind, even for 32-bit:
@PL_regkind = constant [62 x i8] c"\00\00\02\02\02\06\06\06\06\09\09\0B\0B\0D\0E\0E\0E\11\12\12\14\14\16\16\18\18\1A\1A\1C\1C\1E\1F !!!$$&'((((,-.///88886789:;8$", align 32 ; <[62 x i8]*> [#uses=4]
%811 = icmp ne i8 %810, 33
@PL_utf8skip = constant [256 x i8] c"\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\01\02\02\02\02\02\02\02\02\02\02\02\02\02\02\02\02\02\02\02\02\02\02\02\02\02\02\02\02\02\02\02\02\03\03\03\03\03\03\03\03\03\03\03\03\03\03\03\03\04\04\04\04\04\04\04\04\05\05\05\05\06\06\07\0D", align 32 ; <[256 x i8]*> [#uses=94]
%12 = icmp ult i8 %10, 2
etc.
llvm-svn: 92426
two elements match or don't match with two comparisons. For
example, the testcase compiles into:
define i1 @test5(i32 %X) {
%1 = icmp eq i32 %X, 2 ; <i1> [#uses=1]
%2 = icmp eq i32 %X, 7 ; <i1> [#uses=1]
%R = or i1 %1, %2 ; <i1> [#uses=1]
ret i1 %R
}
This generalizes the previous xforms when the array is larger than
64 elements (and this case matches) and generates better code for
cases where it overlaps with the magic bitshift case.
This generalizes more cases than you might expect. For example,
400.perlbmk has:
@PL_utf8skip = constant [256 x i8] c"\01\01\01\...
%15 = icmp ult i8 %7, 7
403.gcc has:
@rid_to_yy = internal constant [114 x i16] [i16 259, i16 260, ...
%18 = icmp eq i16 %16, 295
and xalancbmk has a bunch of examples, such as
_ZN11xercesc_2_5L15gCombiningCharsE and _ZN11xercesc_2_5L10gBaseCharsE.
llvm-svn: 92417
arrays with variable indices into a comparison of the index
with a constant. The most common occurrence of this that
I see by far is stuff like:
if ("foobar"[i] == '\0') ...
which we compile into: if (i == 6), saving a load and
materialization of the global address. This also exposes
loop trip count information to later passes in many cases.
This triggers hundreds of times in xalancbmk, which is where I first
noticed it, but it also triggers in many other apps. Here are a few
interesting ones from various apps:
@must_be_connected_without = internal constant [8 x i8*] [i8* getelementptr inbounds ([3 x i8]* @.str64320, i64 0, i64 0), i8* getelementptr inbounds ([3 x i8]* @.str27283, i64 0, i64 0), i8* getelementptr inbounds ([4 x i8]* @.str71327, i64 0, i64 0), i8* getelementptr inbounds ([4 x i8]* @.str72328, i64 0, i64 0), i8* getelementptr inbounds ([3 x i8]* @.str18274, i64 0, i64 0), i8* getelementptr inbounds ([6 x i8]* @.str11267, i64 0, i64 0), i8* getelementptr inbounds ([3 x i8]* @.str32288, i64 0, i64 0), i8* null], align 32 ; <[8 x i8*]*> [#uses=2]
%scevgep.i = getelementptr [8 x i8*]* @must_be_connected_without, i64 0, i64 %indvar.i ; <i8**> [#uses=1]
%17 = load ...
%18 = icmp eq i8* %17, null ; <i1> [#uses=1]
-> icmp eq i64 %indvar.i, 7
@yytable1095 = internal constant [84 x i8] c"\12\01(\05\06\07\08\09\0A\0B\0C\0D\0E1\0F\10\11266\1D: \10\11,-,0\03'\10\11B6\04\17&\18\1945\05\06\07\08\09\0A\0B\0C\0D\0E\1E\0F\10\11*\1A\1B\1C$3+>#%;<IJ=ADFEGH9KL\00\00\00C", align 32 ; <[84 x i8]*> [#uses=2]
%57 = getelementptr inbounds [84 x i8]* @yytable1095, i64 0, i64 %56 ; <i8*> [#uses=1]
%mode.0.in = getelementptr inbounds [9 x i32]* @mb_mode_table, i64 0, i64 %.pn ; <i32*> [#uses=1]
load ...
%64 = icmp eq i8 %58, 4 ; <i1> [#uses=1]
-> icmp eq i64 %.pn, 35 ; <i1> [#uses=0]
@gsm_DLB = internal constant [4 x i16] [i16 6554, i16 16384, i16 26214, i16 32767]
%scevgep.i = getelementptr [4 x i16]* @gsm_DLB, i64 0, i64 %indvar.i ; <i16*> [#uses=1]
%425 = load %scevgep.i
%426 = icmp eq i16 %425, -32768 ; <i1> [#uses=0]
-> false
llvm-svn: 92411
pointer to int casts that confuse later optimizations. See PR3351
for details.
This improves but doesn't complete fix 483.xalancbmk because llvm-gcc
does this xform in GCC's "fold" routine as well. Clang++ will do
better I guess.
llvm-svn: 92408
positive and negative forms of constants together. This
allows us to compile:
int foo(int x, int y) {
return (x-y) + (x-y) + (x-y);
}
into:
_foo: ## @foo
subl %esi, %edi
leal (%rdi,%rdi,2), %eax
ret
instead of (where the 3 and -3 were not factored):
_foo:
imull $-3, 8(%esp), %ecx
imull $3, 4(%esp), %eax
addl %ecx, %eax
ret
this started out as:
movl 12(%ebp), %ecx
imull $3, 8(%ebp), %eax
subl %ecx, %eax
subl %ecx, %eax
subl %ecx, %eax
ret
This comes from PR5359.
llvm-svn: 92381
Bob Wilson