that can have a big effect :). The first is to enable the
iterative SCC passmanager juice that kicks in when the
scc passmgr detects that a function pass has devirtualized
a call. In this case, it will rerun all the passes it
manages on the SCC, up to the iteration count limit (4). This
is useful because a function pass may devirualize a call, and
we want the inliner to inline it, or pruneeh to infer stuff
about it, etc.
The second patch is to add *all* call sites to the
DevirtualizedCalls list the inliner uses. This list is
about to get renamed, but the jist of this is that the
inliner now reconsiders *all* inlined call sites as candidates
for further inlining. The intuition is this that in cases
like this:
f() { g(1); } g(int x) { h(x); }
We analyze this bottom up, and may decide that it isn't
profitable to inline H into G. Next step, we decide that it is
profitable to inline G into F, and do so, which means that F
now calls H. Even though the call from G -> H may not have been
profitable to inline, the call from F -> H may be (in this case
because a constant allows folding etc).
In my spot checks, this doesn't have a big impact on code. For
example, the LLC output for 252.eon grew from 0.02% (from
317252 to 317308) and 176.gcc actually shrunk by .3% (from 1525612
to 1520964 bytes). 252.eon never iterated in the SCC Passmgr,
176.gcc iterated at most 1 time.
llvm-svn: 102823
were still inlining self-recursive functions into other functions.
Inlining a recursive function into itself has the potential to
reduce recursion depth by a factor of 2, inlining a recursive
function into something else reduces recursion depth by exactly
1. Since inlining a recursive function into something else is a
weird form of loop peeling, turn this off.
The deleted testcase was added by Dale in r62107, since then
we're leaning towards not inlining recursive stuff ever. In any
case, if we like inlining recursive stuff, it should be done
within the recursive function itself to get the algorithm
recursion depth win.
llvm-svn: 102798
doesn't dominate the header is needed, don't check whether the increment
expression has computable loop evolution. While the operands of an
addrec are required to be loop-invariant, they're not required to
dominate any part of the loop. This fixes PR6914.
llvm-svn: 102389
Also, generalize ScalarEvolutions's min and max recognition to handle
some new forms of min and max that this change makes more common.
llvm-svn: 102234
Add the instruction pointer value for debuggability.
We now get dump output that looks like this:
Call graph node for function: 'f1'<<0x1017086b0>> #uses=1
CS<0x1017046f8> calls external node
Call graph node for function: '_ZNSt6vectorIdSaIdEEC1EmRKdRKS0_'<<0x1017086f0>> #uses=1
CS<0x0> calls external node
Call graph node for function: 'f4'<<0x1017087a0>> #uses=1
CS<0x101708c88> calls function 'f3'
llvm-svn: 102194
Fix RefreshCallGraph to use CGN->replaceCallEdge instead of hand
rolling its own loop. replaceCallEdge properly maintains the
reference counts of the nodes, fixing a crash exposed by the
iterative callgraph stuff.
llvm-svn: 102120
we have RefreshCallGraph detect when a function pass devirtualizes
a call, and have CGSCCPassMgr iterate (up to a count) when this
happens. This allows (in the example) GVN to devirtualize the
call in foo, then the inliner to inline it away.
This is not currently enabled because I haven't done any analysis
on the (potentially substantial) code size or performance impact of
doing this, and guess what, it exposes callgraph updating bugs in
various passes. This is progress though, and you can play with it
by passing -max-cg-scc-iterations=5 to opt.
llvm-svn: 101973
recursive callsites, inlining can reduce the number of calls by
exponential factors, as it does in
MultiSource/Benchmarks/Olden/treeadd. More involved heuristics
will be needed.
llvm-svn: 101969
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
by switching CachedFunctionInfo from a std::map to a
ValueMap (which is implemented in terms of a DenseMap).
DenseMap has different iterator invalidation semantics
than std::map.
This should hopefully fix the dragonegg builder.
llvm-svn: 101658
CGSCC can delete nodes in regions of the callgraph that
have already been visited. If new CG nodes are allocated
to the same pointer, we shouldn't abort, just handle it
correctly by assigning a new number. This should restore
stability by removing invalidated pointers that *will* be
reused from the densemap in the iterator.
llvm-svn: 101628
to keep the node entries in scc_iterator up to date instead of dangling as
the SCC mutates.
This is a really terrible problem which was causing -g to affect codegen
because it would permute the memory image of the compiler process.
Thanks to Dale for expertly hunting it down.
llvm-svn: 101565
to CallGraphSCCPass's instead of passing around a
std::vector<CallGraphNode*>. No functionality change,
but now we have a much tidier interface.
llvm-svn: 101558
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
The information is already available with "opt -analyze". The DominatorTree
does also not have this in its runOnFunction. So they behave now
more consistent.
llvm-svn: 101038
AddRecs so that it checks for overflow in the computation that it is
performing, rather than just checking hasNo{Signed,Unsigned}Wrap, since
those flags are for a different computation. This fixes a bug that
impacts an upcoming change.
llvm-svn: 101028
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
source addition. Apparently the buildbots were wrong about failures.
---
Add some switches helpful for debugging:
-print-before=<Pass Name>
Dump IR before running pass <Pass Name>.
-print-before-all
Dump IR before running each pass.
-print-after-all
Dump IR after running each pass.
These are helpful when tracking down a miscompilation. It is easy to
get IR dumps and do diffs on them, etc.
To make this work well, add a new getPrinterPass API to Pass so that
each kind of pass (ModulePass, FunctionPass, etc.) can create a Pass
suitable for dumping out the kind of object the Pass works on.
llvm-svn: 100249
materializing an MDNode for every debugloc. don't do that! :)
"clang -g -S t.c" really no longer makes mdnodes for location
tuples now.
llvm-svn: 100224
representation. This eliminates the 'DILocation' MDNodes for
file/line/col tuples from -O0 -g codegen.
This remove the old DebugLoc class, making it a typedef for DebugLoc,
I'll rename NewDebugLoc next.
I didn't update the JIT to use the new apis, so it will continue to
work, but be as slow as before. Someone should eventually do this
or, better yet, rip out the JIT debug info stuff and build the JIT
on top of MC.
llvm-svn: 100209
<string> include. For some reason the buildbot choked on this while my
builds did not. It's probably due to a difference in system headers.
---
Add some switches helpful for debugging:
-print-before=<Pass Name>
Dump IR before running pass <Pass Name>.
-print-before-all
Dump IR before running each pass.
-print-after-all
Dump IR after running each pass.
These are helpful when tracking down a miscompilation. It is easy to
get IR dumps and do diffs on them, etc.
To make this work well, add a new getPrinterPass API to Pass so that
each kind of pass (ModulePass, FunctionPass, etc.) can create a Pass
suitable for dumping out the kind of object the Pass works on.
llvm-svn: 100204
-print-before=<Pass Name>
Dump IR before running pass <Pass Name>.
-print-before-all
Dump IR before running each pass.
-print-after-all
Dump IR after running each pass.
These are helpful when tracking down a miscompilation. It is easy to
get IR dumps and do diffs on them, etc.
To make this work well, add a new getPrinterPass API to Pass so that
each kind of pass (ModulePass, FunctionPass, etc.) can create a Pass
suitable for dumping out the kind of object the Pass works on.
llvm-svn: 100143
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
use-before-def errors in SCEVExpander-produced code in sqlite3 when debug
info with optimization is enabled, though the testcases for this are
dependent on use-list order.
llvm-svn: 99001
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
BumpPtrAllocator-allocated region to allow it to be stored in a more
compact form and to avoid the need for a non-trivial destructor call.
Use this new mechanism in ScalarEvolution instead of
FastFoldingSetNode to avoid leaking memory in the case where a
FoldingSetNodeID uses heap storage, and to reduce overall memory
usage.
llvm-svn: 98829
pointer and length, and allocate the arrays in ScalarEvolution's
BumpPtrAllocator, so that they get released when their owning
SCEV gets released. SCEVs are immutable, so they don't need to worry
about operand array resizing. This fixes a memory leak reported
in PR6637.
llvm-svn: 98755
The Caller cost info would be reset everytime a callee was inlined. If the
caller has lots of calls and there is some mutual recursion going on, the
caller cost info could be calculated many times.
This patch reduces inliner runtime from 240s to 0.5s for a function with 20000
small function calls.
This is a more conservative version of r98089 that doesn't break the clang
test CodeGenCXX/temp-order.cpp. That test relies on rather extreme inlining
for constant folding.
llvm-svn: 98099
The Caller cost info would be reset everytime a callee was inlined. If the
caller has lots of calls and there is some mutual recursion going on, the
caller cost info could be calculated many times.
This patch reduces inliner runtime from 240s to 0.5s for a function with 20000
small function calls.
llvm-svn: 98089
This patch updates LLVMDebugVersion to 8.
Debug info descriptors encoded using LLVMDebugVersion 7 is supported.
Corresponding llvmgcc and clang FE commits are required.
llvm-svn: 98020
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
getelementptr. Despite only doing so in the case where x is a known
array object and c can be converted to an index within range, this
could still be invalid if c is actually the address of an object
allocated outside of LLVM. Also, SCEVExpander, the original motivation
for this code, has since been improved to avoid inttoptr+ptroint in
more cases.
llvm-svn: 96950
operators.
The test difference is just due to the multiplication operands
being commuted (and thus requiring a more elaborate match). In
optimized code, that expression would be folded.
llvm-svn: 96816
true or false as its exit condition. These are usually eliminated by
SimplifyCFG, but the may be left around during a pass which wishes
to preserve the CFG.
llvm-svn: 96683
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
case where there are loop-invariant instructions somehow left
inside the loop, and in a position where they won't dominate
the IV increment position.
llvm-svn: 96448
name, test whether the SCEV itself is that temporary symbolic name,
in addition to checking whether the symbolic name appears as a
possibly-indirect operand.
llvm-svn: 96216
insert location has become an "inserted" instruction since the time
it was saved. If so, advance to the first non-"inserted" instruction.
llvm-svn: 96203
current insertion point, advance the current insertion point.
This avoids a use-before-def situation in a testcase extracted
from clang which is difficult to reduce to a reasonable-sized
regression test.
llvm-svn: 96151
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
cases, and implement target-independent folding rules for alignof and
offsetof. Also, reassociate reassociative operators when it leads to
more folding.
Generalize ScalarEvolution's isOffsetOf to recognize offsetof on
arrays. Rename getAllocSizeExpr to getSizeOfExpr, and getFieldOffsetExpr
to getOffsetOfExpr, for consistency with analagous ConstantExpr routines.
Make the target-dependent folder promote GEP array indices to
pointer-sized integers, to make implicit casting explicit and exposed
to subsequent folding.
And add a bunch of testcases for this new functionality, and a bunch
of related existing functionality.
llvm-svn: 94987
use plain SCEVUnknowns with ConstantExpr::getSizeOf and
ConstantExpr::getOffsetOf constants. This eliminates a bunch of
special-case code.
Also add code for pattern-matching these expressions, for clients that
want to recognize them.
Move ScalarEvolution's logic for expanding array and vector sizeof
expressions into an element count times the element size, to expose
the multiplication to subsequent folding, into the regular constant
folder.
llvm-svn: 94737
After running a batch of measurements, it is clear that the inliner metrics
need some adjustments:
Own argument bonus: 20 -> 5
Outgoing argument penalty: 0 -> 5
Alloca bonus: 10 -> 5
Constant instr bonus: 7 -> 5
Dead successor bonus: 40 -> 5*(avg instrs/block)
The new cost metrics are generaly 25 points higher than before, so we may need
to move thresholds.
With this change, InlineConstants::CallPenalty becomes a political correction:
if (!isa<IntrinsicInst>(II) && !callIsSmall(CS.getCalledFunction()))
NumInsts += InlineConstants::CallPenalty + CS.arg_size();
The code size is accurately modelled by CS.arg_size(). CallPenalty is added
because calls tend to take a long time, so it may not be worth it to inline a
function with lots of calls.
All of the political corrections are in the InlineConstants namespace:
IndirectCallBonus, CallPenalty, LastCallToStaticBonus, ColdccPenalty,
NoreturnPenalty.
llvm-svn: 94615
A GEP with all constant indices is already considered free by
analyzeBasicBlock(), so don't give it an extra bonus in
CountCodeReductionForAlloca().
This patch should remove a small positive bias toward inlining functions with
variable-index GEPs, and remove a smaller negative bias from functions with
all-constant index GEPs.
llvm-svn: 94591
Functions containing indirectbr are marked NeverInline by analyzeBasicBlock(),
so there is no point in giving indirectbr special treatment in
CountCodeReductionForConstant. It is never called.
No functional change intended.
llvm-svn: 94590
have trouble with an intermediate add overflowing. Also, be more conservative
about the case where the induction variable in an SLT loop exit can step past
the RHS of the SLT and overflow in a single step.
Make getSignedRange more aggressive, to recover for some common cases which
the above fixes pessimized.
This addresses rdar://7561161.
llvm-svn: 94512
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
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
form of an expression. This is the expression without the
post-increment adjustment made, which is useful in determining
which registers will be used by the expansion.
llvm-svn: 93921
Move the DOTGraphTraits dotty printer/viewer templates, that were developed for
the dominance tree into their own header file. This will allow reuse in future
passes.
llvm-svn: 93632
This patch also cleans up code that expects there to be a bitcast in the first argument and testcases that call llvm.dbg.declare.
It also strips old llvm.dbg.declare intrinsics that did not pass metadata as the first argument.
llvm-svn: 93531
memcpy, memset and other intrinsics that only access their arguments
to be readnone if the intrinsic's arguments all point to local memory.
This improves the testcase in the README to readonly, but it could in
theory be made readnone, however this would involve more sophisticated
analysis that looks through the memcpy.
llvm-svn: 92829
after the MDNode in memory. This eliminates the operands
pointer and saves a new[] per node.
Note that the code in DIDerivedType::replaceAllUsesWith is wrong
and quite scary. A MDNode should not be RAUW'd with something
else: this changes all uses of the mdnode, which may not be debug
info related! Debug info should use something non-mdnode for
declarations.
llvm-svn: 92321
getMDKindID/getMDKindNames methods to LLVMContext (and add
convenience methods to Module), eliminating MetadataContext.
Move the state that it maintains out to LLVMContext.
llvm-svn: 92259
I asked Devang to do back on Sep 27. Instead of going through the
MetadataContext class with methods like getMD() and getMDs(), just
ask the instruction directly for its metadata with getMetadata()
and getAllMetadata().
This includes a variety of other fixes and improvements: previously
all Value*'s were bloated because the HasMetadata bit was thrown into
value, adding a 9th bit to a byte. Now this is properly sunk down to
the Instruction class (the only place where it makes sense) and it
will be folded away somewhere soon.
This also fixes some confusion in getMDs and its clients about
whether the returned list is indexed by the MDID or densely packed.
This is now returned sorted and densely packed and the comments make
this clear.
This introduces a number of fixme's which I'll follow up on.
llvm-svn: 92235
instead of stored. This reduces memdep memory usage, and also eliminates a bunch of
weakvh's. This speeds up gvn on gcc.c-torture/20001226-1.c from 23.9s to 8.45s (2.8x)
on a different machine than earlier.
llvm-svn: 91885
cache a pointer as being unavailable due to phi trans in the
wrong place. This would cause later queries to fail even when
they didn't involve phi trans.
llvm-svn: 91787
contains another loop, or an instruction. The loop form is
substantially more efficient on large loops than the typical
code it replaces.
llvm-svn: 91654
of 91296 that caused trouble -- the Processed list needs to be
preserved for the livetime of the pass, as AddUsersIfInteresting
is called from other passes.
llvm-svn: 91641
isPodLike type trait. This is a generally useful type trait for
more than just DenseMap, and we really care about whether something
acts like a pod, not whether it really is a pod.
llvm-svn: 91421
add, there is no need to scan the world to find the same add again.
This invalidates the previous testcase, which wasn't wonderful anyway,
because it needed a run of instcombine to permute the use-lists in
just the right way to before GVN was run (so it was really fragile).
Not a big loss.
llvm-svn: 90973
David Chisnall