Summary:
It can hurt performance to prefetch ahead too much. Be conservative for
now and don't prefetch ahead more than 3 iterations on Cyclone.
Reviewers: hfinkel
Subscribers: llvm-commits, mzolotukhin
Differential Revision: http://reviews.llvm.org/D17949
llvm-svn: 263772
Summary:
And use this TTI for Cyclone. As it was explained in the original RFC
(http://thread.gmane.org/gmane.comp.compilers.llvm.devel/92758), the HW
prefetcher work up to 2KB strides.
I am also adding tests for this and the previous change (D17943):
* Cyclone prefetching accesses with a large stride
* Cyclone not prefetching accesses with a small stride
* Generic Aarch64 subtarget not prefetching either
Reviewers: hfinkel
Subscribers: aemerson, rengolin, llvm-commits, mzolotukhin
Differential Revision: http://reviews.llvm.org/D17945
llvm-svn: 263771
Summary:
Use the new LoopVersioning facility (D16712) to add noalias metadata in
the vector loop if we versioned with memchecks. This can enable some
optimization opportunities further down the pipeline (see the included
test or the benchmark improvement quoted in D16712).
The test also covers the bug I had in the initial version in D16712.
The vectorizer did not previously use LoopVersioning. The reason is
that the vectorizer performs its transformations in single shot. It
creates an empty single-block vector loop that it then populates with
the widened, if-converted instructions. Thus creating an intermediate
versioned scalar loop seems wasteful.
So this patch (rather than bringing in LoopVersioning fully) adds a
special interface to LoopVersioning to allow the vectorizer to add
no-alias annotation while still performing its own versioning.
As the vectorizer propagates metadata from the instructions in the
original loop to the vector instructions we also check the pointer in
the original instruction and see if LoopVersioning can add no-alias
metadata based on the issued memchecks.
Reviewers: hfinkel, nadav, mzolotukhin
Subscribers: mzolotukhin, llvm-commits
Differential Revision: http://reviews.llvm.org/D17191
llvm-svn: 263744
Summary:
If we decide to version a loop to benefit a transformation, it makes
sense to record the now non-aliasing accesses in the newly versioned
loop. This allows non-aliasing information to be used by subsequent
passes.
One example is 456.hmmer in SPECint2006 where after loop distribution,
we vectorize one of the newly distributed loops. To vectorize we
version this loop to fully disambiguate may-aliasing accesses. If we
add the noalias markers, we can use the same information in a later DSE
pass to eliminate some dead stores which amounts to ~25% of the
instructions of this hot memory-pipeline-bound loop. The overall
performance improves by 18% on our ARM64.
The scoped noalias annotation is added in LoopVersioning. The patch
then enables this for loop distribution. A follow-on patch will enable
it for the vectorizer. Eventually this should be run by default when
versioning the loop but first I'd like to get some feedback whether my
understanding and application of scoped noalias metadata is correct.
Essentially my approach was to have a separate alias domain for each
versioning of the loop. For example, if we first version in loop
distribution and then in vectorization of the distributed loops, we have
a different set of memchecks for each versioning. By keeping the scopes
in different domains they can conveniently be defined independently
since different alias domains don't affect each other.
As written, I also have a separate domain for each loop. This is not
necessary and we could save some metadata here by using the same domain
across the different loops. I don't think it's a big deal either way.
Probably the best is to review the tests first to see if I mapped this
problem correctly to scoped noalias markers. I have plenty of comments
in the tests.
Note that the interface is prepared for the vectorizer which needs the
annotateInstWithNoAlias API. The vectorizer does not use LoopVersioning
so we need a way to pass in the versioned instructions. This is also
why the maps have to become part of the object state.
Also currently, we only have an AA-aware DSE after the vectorizer if we
also run the LTO pipeline. Depending how widely this triggers we may
want to schedule a DSE toward the end of the regular pass pipeline.
Reviewers: hfinkel, nadav, ashutosh.nema
Subscribers: mssimpso, aemerson, llvm-commits, mcrosier
Differential Revision: http://reviews.llvm.org/D16712
llvm-svn: 263743
This is similar to D18133 where we allowed profile weights on select instructions.
This extends that change to also allow the 'unpredictable' attribute of branches to apply to selects.
A test to check that 'unpredictable' metadata is preserved when cloning instructions was checked in at:
http://reviews.llvm.org/rL263648
Differential Revision: http://reviews.llvm.org/D18220
llvm-svn: 263716
I'm testing out a script that auto-generates the check lines.
It's 98% copied from utils/update_llc_test_checks.py.
If others think this is useful, please let me know.
llvm-svn: 263668
I'm testing out a script that auto-generates the check lines.
It's 98% copied from utils/update_llc_test_checks.py.
If others think this is useful, please let me know.
llvm-svn: 263667
Summary:
Fix LSRInstance::HoistInsertPosition() to check the original insert
position block first for a canonical insertion point that is dominated
by all inputs. This leads to SCEV being able to reuse more instructions
since it currently tracks the instructions it creates for reuse by
keeping a table of <Value, insert point> pairs.
Reviewers: atrick
Subscribers: mcrosier, mzolotukhin, llvm-commits
Differential Revision: http://reviews.llvm.org/D18001
llvm-svn: 263644
This was a latent bug that got exposed by the change to add LoopSimplify
as a dependence to LoopLoadElimination. Since LoopInfo was corrupted
after LV, LoopSimplify mis-compiled nbench in the test-suite (more
details in the PR).
The problem was that when we create the blocks for predicated stores we
didn't add those to any loops.
The original testcase for store predication provides coverage for this
assuming we verify LI on the way out of LV.
Fixes PR26952.
llvm-svn: 263565
(Resubmitting after fixing missing file issue)
With the changes in r263275, there are now more than just functions in
the summary. Completed the renaming of data structures (started in
r263275) to reflect the wider scope. In particular, changed the
FunctionIndex* data structures to ModuleIndex*, and renamed related
variables and comments. Also renamed the files to reflect the changes.
A companion clang patch will immediately succeed this patch to reflect
this renaming.
llvm-svn: 263513
Summary:
Specifically, when we perform runtime loop unrolling of a loop that
contains a convergent op, we can only unroll k times, where k divides
the loop trip multiple.
Without this change, we'll happily unroll e.g. the following loop
for (int i = 0; i < N; ++i) {
if (i == 0) convergent_op();
foo();
}
into
int i = 0;
if (N % 2 == 1) {
convergent_op();
foo();
++i;
}
for (; i < N - 1; i += 2) {
if (i == 0) convergent_op();
foo();
foo();
}.
This is unsafe, because we've just added a control-flow dependency to
the convergent op in the prelude.
In general, runtime unrolling loops that contain convergent ops is safe
only if we don't have emit a prelude, which occurs when the unroll count
divides the trip multiple.
Reviewers: resistor
Subscribers: llvm-commits, mzolotukhin
Differential Revision: http://reviews.llvm.org/D17526
llvm-svn: 263509
Summary: This now try to reorder instructions in order to help create the optimizable pattern.
Reviewers: craig.topper, spatel, dexonsmith, Prazek, chandlerc, joker.eph, majnemer
Differential Revision: http://reviews.llvm.org/D16523
llvm-svn: 263503
With the changes in r263275, there are now more than just functions in
the summary. Completed the renaming of data structures (started in
r263275) to reflect the wider scope. In particular, changed the
FunctionIndex* data structures to ModuleIndex*, and renamed related
variables and comments. Also renamed the files to reflect the changes.
A companion clang patch will immediately succeed this patch to reflect
this renaming.
llvm-svn: 263490
As noted in:
https://llvm.org/bugs/show_bug.cgi?id=26636
This doesn't accomplish anything on its own. It's the first step towards preserving
and using branch weights with selects.
The next step would be to make sure we're propagating the info in all of the other
places where we create selects (SimplifyCFG, InstCombine, etc). I don't think there's
an easy fix to make this happen; we have to look at each transform individually to
determine how to correctly propagate the weights.
Along with that step, we need to then use the weights when making subsequent transform
decisions such as discussed in http://reviews.llvm.org/D16836.
The inliner test is independent but closely related. It verifies that metadata is
preserved when both branches and selects are cloned.
Differential Revision: http://reviews.llvm.org/D18133
llvm-svn: 263482
Summary:
Previously we had a notion of convergent functions but not of convergent
calls. This is insufficient to correctly analyze calls where the target
is unknown, e.g. indirect calls.
Now a call is convergent if it targets a known-convergent function, or
if it's explicitly marked as convergent. As usual, we can remove
convergent where we can prove that no convergent operations are
performed in the call.
Originally landed as r261544, then reverted in r261544 for (incidental)
build breakage. Re-landed here with no changes.
Reviewers: chandlerc, jingyue
Subscribers: llvm-commits, tra, jhen, hfinkel
Differential Revision: http://reviews.llvm.org/D17739
llvm-svn: 263481
The motivating example is this
for (j = n; j > 1; j = i) {
i = j / 2;
}
The signed division is safely to be changed to an unsigned division (j is known
to be larger than 1 from the loop guard) and later turned into a single shift
without considering the sign bit.
llvm-svn: 263406
This follows up on the related AVX instruction transforms, but this
one is too strange to do anything more with. Intel's behavioral
description of this instruction in its Software Developer's Manual
is tragi-comic.
llvm-svn: 263340
Summary: As we now have unit-tests for UnrollAnalyzer, we can convert some existing tests to this format. It should make the tests more robust.
Reviewers: chandlerc, sanjoy
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D17904
llvm-svn: 263318
Summary:
This intrinsic, together with deoptimization operand bundles, allow
frontends to express transfer of control and frame-local state from
one (typically more specialized, hence faster) version of a function
into another (typically more generic, hence slower) version.
In languages with a fully integrated managed runtime this intrinsic can
be used to implement "uncommon trap" like functionality. In unmanaged
languages like C and C++, this intrinsic can be used to represent the
slow paths of specialized functions.
Note: this change does not address how `@llvm.experimental_deoptimize`
is lowered. That will be done in a later change.
Reviewers: chandlerc, rnk, atrick, reames
Subscribers: llvm-commits, kmod, mjacob, maksfb, mcrosier, JosephTremoulet
Differential Revision: http://reviews.llvm.org/D17732
llvm-svn: 263281
Value profile instrumentation treats inline asm calls like they are
indirect calls. This causes problems when the 'Callee' is passed to a
ptrtoint cast -- the verifier rightly claims that this is bogus and
crashes opt.
llvm-svn: 263278
tests to run GVN in both modes.
This is mostly the boring refactoring just like SROA and other complex
transformation passes. There is some trickiness in that GVN's
ValueNumber class requires hand holding to get to compile cleanly. I'm
open to suggestions about a better pattern there, but I tried several
before settling on this. I was trying to balance my desire to sink as
much implementation detail into the source file as possible without
introducing overly many layers of abstraction.
Much like with SROA, the design of this system is made somewhat more
cumbersome by the need to support both pass managers without duplicating
the significant state and logic of the pass. The same compromise is
struck here.
I've also left a FIXME in a doxygen comment as the GVN pass seems to
have pretty woeful documentation within it. I'd like to submit this with
the FIXME and let those more deeply familiar backfill the information
here now that we have a nice place in an interface to put that kind of
documentaiton.
Differential Revision: http://reviews.llvm.org/D18019
llvm-svn: 263208
llvm::getDISubprogram walks the instructions in a function, looking for one in the scope of the current function, so that it can find the !dbg entry for the subprogram itself.
Now that !dbg is attached to functions, this should not be necessary. This patch changes all uses to just query the subprogram directly on the function.
Ideally this should be NFC, but in reality its possible that a function:
has no !dbg (in which case there's likely a bug somewhere in an opt pass), or
that none of the instructions had a scope referencing the function, so we used to not find the !dbg on the function but now we will
Reviewed by Duncan Exon Smith.
Differential Revision: http://reviews.llvm.org/D18074
llvm-svn: 263184
The code assumed that we always had a preheader without making the pass
dependent on LoopSimplify.
Thanks to Mattias Eriksson V for reporting this.
llvm-svn: 263173
This patch fixes the problem which occurs when loop-vectorize tries to use @llvm.masked.load/store intrinsic for a non-default addrspace pointer. It fails with "Calling a function with a bad signature!" assertion in CallInst constructor because it tries to pass a non-default addrspace pointer to the pointer argument which has default addrspace.
The fix is to add pointer type as another overloaded type to @llvm.masked.load/store intrinsics.
Reviewed By: reames
Differential Revision: http://reviews.llvm.org/D17270
llvm-svn: 263158
of, and I misdiagnosed for months and months.
Andrea has had a patch for this forever, but I just couldn't see how
it was fixing the root cause of the problem. It didn't make sense to me,
even though the patch was perfectly good and the analysis of the actual
failure event was *fantastic*.
Well, I came back to it today because the patch has sat for *far* too
long and needs attention and decided I wouldn't let it go until I really
understood what was going on. After quite some time in the debugger,
I finally realized that in fact I had just missed an important case with
my previous attempt to fix PR22093 in r225149. Not only do we need to
handle loads that won't be split, but stores-of-loads that we won't
split. We *do* actually have enough logic in the presplitting to form
new slices for split stores.... *unless* we decided not to split them!
I'm so sorry that it took me this long to come to the realization that
this is the issue. It seems so obvious in hind sight (of course).
Anyways, the fix becomes *much* smaller and more focused. The fact that
we're left doing integer smashing is related to the FIXME in my original
commit: fundamentally, we're not aggressive about pre-splitting for
loads and stores to the same alloca. If we want to get aggressive about
this, it'll need both what Andrea had put into the proposed fix, but
also a *lot* more logic to essentially iteratively pre-split the alloca
until we can't do any more. As I said in that commit log, its really
unclear that this is the right call. Instead, the integer blending and
letting targets lower this to narrower stores seems slightly better. But
we definitely shouldn't really go down that path just to fix this bug.
Again, tons of thanks are owed to Andrea and others at Sony for working
on this bug. I really should have seen what was going on here and
re-directed them sooner. =////
llvm-svn: 263121
This patch teaches CGP to duplicate addressing mode computations into cold paths (detected via explicit cold attribute on calls) if required to let addressing mode be safely sunk into the basic block containing each load and store.
In general, duplicating code into cold blocks may result in code growth, but should not effect performance. In this case, it's better to duplicate some code than to put extra pressure on the register allocator by making it keep the address through the entirely of the fast path.
This patch only handles addressing computations, but in principal, we could implement a more general cold cold scheduling heuristic which tries to reduce register pressure in the fast path by duplicating code into the cold path. Getting the profitability of the general case right seemed likely to be challenging, so I stuck to the existing case (addressing computation) we already had.
Differential Revision: http://reviews.llvm.org/D17652
llvm-svn: 263074
This patch teaches LICM's implementation of store promotion to exploit the fact that the memory location being accessed might be provable thread local. The fact it's thread local weakens the requirements for where we can insert stores since no other thread can observe the write. This allows us perform store promotion even in cases where the store is not guaranteed to execute in the loop.
Two key assumption worth drawing out is that this assumes a) no-capture is strong enough to imply no-escape, and b) standard allocation functions like malloc, calloc, and operator new return values which can be assumed not to have previously escaped.
In future work, it would be nice to generalize this so that it works without directly seeing the allocation site. I believe that the nocapture return attribute should be suitable for this purpose, but haven't investigated carefully. It's also likely that we could support unescaped allocas with similar reasoning, but since SROA and Mem2Reg should destroy those, they're less interesting than they first might seem.
Differential Revision: http://reviews.llvm.org/D16783
llvm-svn: 263072
The irony of this patch is that one CPU that is affected is AMD Jaguar, and Jaguar
has a completely double-pumped AVX implementation. But getting the cost model to
reflect that is a much bigger problem. The small goal here is simply to improve on
the lie that !AVX2 == SandyBridge.
Differential Revision: http://reviews.llvm.org/D18000
llvm-svn: 263069
When checking whether an smin is positive, we can move the comparison to one of the inputs if the other is known positive. If the known positive one is the min, then the other can't be negative. If the other is the min, then we compute the min.
Differential Revision: http://reviews.llvm.org/D17873
llvm-svn: 263059
I somehow missed this. The case in GCC (global_alloc) was similar to
the new testcase except it had an array of structs rather than a two
dimensional array.
Fixes RP26885.
llvm-svn: 263058
As part of r251146 InstCombine was extended to call computeKnownBits on
every value in the function to determine whether it happens to be
constant. This increases typical compiletime by 1-3% (5% in irgen+opt
time) in my measurements. On the other hand this case did not trigger
once in the whole llvm-testsuite.
This patch introduces the notion of ExpensiveCombines which are only
enabled for OptLevel > 2. I removed the check in InstructionSimplify as
that is called from various places where the OptLevel is not known but
given the rarity of the situation I think a check in InstCombine is
enough.
Differential Revision: http://reviews.llvm.org/D16835
llvm-svn: 263047
Original commit message:
calculate builtin_object_size if argument is a removable pointer
This patch fixes calculating correct value for builtin_object_size function
when pointer is used only in builtin_object_size function call and never
after that.
Patch by Strahinja Petrovic.
Differential Revision: http://reviews.llvm.org/D17337
Reland the original change with a small modification (first do a null check
and then do the cast) to satisfy ubsan.
llvm-svn: 263011
The diff is relatively large since I took a chance to rearrange the code I had to touch in a more obvious way, but the key bit is merely using the !range metadata when we can't analyze the instruction further. The previous !range metadata code was essentially just dead since no binary operator or cast will have !range metadata (per Verifier) and it was otherwise dropped on the floor.
llvm-svn: 262751
This experiment was originally about trying to use facts implied dominating conditions to infer more precise known bits. While the compile time was found to be acceptable on several large code bases, we never found sufficiently profitable examples to justify turning on the code by default. Given this, it's time to abandon the experiment.
Several folks have commented that they've found this useful for experimentation, but nothing has come of those experiments. Given how easy the patch is to apply, there's no reason to leave the code in tree.
For anyone interested in further investigation in this area, I recommend finding the summary email I sent on one of the original review threads. In particular, I now believe the use-list based approach is strictly worse than the dom-tree-walking approach.
llvm-svn: 262646
