cases where the alloca type, the load types, and the store types used
all disagree.
Previously, the only way that vector-based promotion occured was if the
alloca type was a vector type. This was one of the *very* few remaining
uses of the alloca's type to guide SROA/mem2reg left in LLVM. It turns
out it was a bad idea.
The alloca type can change very easily based on the mixture of types
loaded and stored to that alloca. We shouldn't be relying on it as
a signal for very much. Instead, the source of truth should be loads and
stores. We should canonicalize the loads and stores as much as possible
and then rely on them exclusively in SROA.
When looking and loads and stores, we may find many different candidate
vector types. This change will let SROA try all of them to find a vector
type which is a viable way to promote the entire alloca to a vector
register.
With this change, it becomes possible to do better canonicalization and
optimization of loads and stores without breaking SROA in random ways,
and that should allow fixing a core source of performance loss in hot
numerical loops such as those in Eigen.
llvm-svn: 220116
This reverts commit r219899.
This also updates byval-tail-call.ll to make it clear what was breaking.
Adding r219899 again will cause the load/store to disappear.
llvm-svn: 220093
DSE's overlap checking contained special logic, used only when no DataLayout
was available, which inferred a complete overwrite when the pointee types were
equal. This logic seems fine for regular loads/stores, but does not work for
memcpy and friends. Instead of fixing this, I'm just removing it.
Philosophically, transformations should not contain enhanced behavior used only
when data layout is lacking (data layout should be strictly additive), and
maintaining these rarely-tested code paths seems not worthwhile at this stage.
Credit to Aliaksei Zasenka for the bug report and the diagnosis. The test case
(slightly reduced from that provided by Aliaksei) replaces the original
contents of test/Transforms/DeadStoreElimination/no-targetdata.ll -- a few
other tests have been updated to have a data layout.
llvm-svn: 220035
'AS'.
Using 'S' as this was a terrible idea. Arguably, 'AS' is not much
better, but it at least follows the idea of using initialisms and
removes active confusion about the AllocaSlices variable and a Slice
variable.
llvm-svn: 219963
clang-modernize.
I did have to clean up the variable types and whitespace a bit because
the use of auto made the code much less readable here.
llvm-svn: 219962
iterators.
There are a ton of places where it essentially wants ranges
rather than just iterators. This is just the first step that adds the
core slice range typedefs and uses them in a couple of places. I still
have to explicitly construct them because they've not been punched
throughout the entire set of code. More range-based cleanups incoming.
llvm-svn: 219955
Summary:
Currently, call slot optimization requires that if the destination is an
argument, the argument has the sret attribute. This is to ensure that
the memory access won't trap. In addition to sret, we can also allow the
optimization to happen for arguments that have the new dereferenceable
attribute, which gives the same guarantee.
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D5832
llvm-svn: 219950
Make tail recursion elimination a bit more aggressive. This allows us to get
tail recursion on functions that are just branches to a different function. The
fact that the function takes a byval argument does not restrict it from being
optimised into just a tail call.
llvm-svn: 219899
routines and fix all of the bugs they expose.
I hit a test case that crashed even without these asserts due to passing
a non-exiting latch to the ExitingBlock parameter of the trip count
computation machinery. However, when I add the nice asserts, it turns
out we have plenty of coverage of these bugs, they just didn't manifest
in crashers.
The core problem seems to stem from an assumption that the latch *is*
the exiting block. While this is often true, and somewhat the "normal"
way to think about loops, it isn't necessarily true. The correct way to
call the trip count routines in a *generic* fashion (that is, without
a particular exit in mind) is to just use the loop's single exiting
block if it has one. The trip count can't be computed generically unless
it does. This works great for the loop vectorizer. The loop unroller
actually *wants* to select the latch when it has to chose between
multiple exits because for unrolling it is the latch trips that matter.
But if this is the desire, it needs to explicitly guard for non-exiting
latches and check for the generic trip count in that case.
I've added the asserts, and added convenience APIs for querying the trip
count generically that check for a single exit block. I've kept the APIs
consistent between computing trip count and trip multiples.
Thansk to Mark for the help debugging and tracking down the *right* fix
here!
llvm-svn: 219550
Particularly, it addresses cases where Reassociate breaks Subtracts but then fails to optimize combinations like I1 + -I2 where I1 and I2 have the same rank and are identical.
Patch by Dmitri Shtilman.
llvm-svn: 219092
My commit rL216160 introduced a bug PR21014: IndVars widens code 'for (i = ; i < ...; i++) arr[ CONST - i]' into 'for (i = ; i < ...; i++) arr[ i - CONST]'
thus inverting index expression. This patch fixes it.
Thanks to Jörg Sonnenberger for pointing.
Differential Revision: http://reviews.llvm.org/D5576
llvm-svn: 218867
argument of the llvm.dbg.declare/llvm.dbg.value intrinsics.
Previously, DIVariable was a variable-length field that has an optional
reference to a Metadata array consisting of a variable number of
complex address expressions. In the case of OpPiece expressions this is
wasting a lot of storage in IR, because when an aggregate type is, e.g.,
SROA'd into all of its n individual members, the IR will contain n copies
of the DIVariable, all alike, only differing in the complex address
reference at the end.
By making the complex address into an extra argument of the
dbg.value/dbg.declare intrinsics, all of the pieces can reference the
same variable and the complex address expressions can be uniqued across
the CU, too.
Down the road, this will allow us to move other flags, such as
"indirection" out of the DIVariable, too.
The new intrinsics look like this:
declare void @llvm.dbg.declare(metadata %storage, metadata %var, metadata %expr)
declare void @llvm.dbg.value(metadata %storage, i64 %offset, metadata %var, metadata %expr)
This patch adds a new LLVM-local tag to DIExpressions, so we can detect
and pretty-print DIExpression metadata nodes.
What this patch doesn't do:
This patch does not touch the "Indirect" field in DIVariable; but moving
that into the expression would be a natural next step.
http://reviews.llvm.org/D4919
rdar://problem/17994491
Thanks to dblaikie and dexonsmith for reviewing this patch!
Note: I accidentally committed a bogus older version of this patch previously.
llvm-svn: 218787
argument of the llvm.dbg.declare/llvm.dbg.value intrinsics.
Previously, DIVariable was a variable-length field that has an optional
reference to a Metadata array consisting of a variable number of
complex address expressions. In the case of OpPiece expressions this is
wasting a lot of storage in IR, because when an aggregate type is, e.g.,
SROA'd into all of its n individual members, the IR will contain n copies
of the DIVariable, all alike, only differing in the complex address
reference at the end.
By making the complex address into an extra argument of the
dbg.value/dbg.declare intrinsics, all of the pieces can reference the
same variable and the complex address expressions can be uniqued across
the CU, too.
Down the road, this will allow us to move other flags, such as
"indirection" out of the DIVariable, too.
The new intrinsics look like this:
declare void @llvm.dbg.declare(metadata %storage, metadata %var, metadata %expr)
declare void @llvm.dbg.value(metadata %storage, i64 %offset, metadata %var, metadata %expr)
This patch adds a new LLVM-local tag to DIExpressions, so we can detect
and pretty-print DIExpression metadata nodes.
What this patch doesn't do:
This patch does not touch the "Indirect" field in DIVariable; but moving
that into the expression would be a natural next step.
http://reviews.llvm.org/D4919
rdar://problem/17994491
Thanks to dblaikie and dexonsmith for reviewing this patch!
llvm-svn: 218778
Summary:
This patch adds a threshold that controls the number of bonus instructions
allowed for folding branches with common destination. The original code allows
at most one bonus instruction. With this patch, users can customize the
threshold to allow multiple bonus instructions. The default threshold is still
1, so that the code behaves the same as before when users do not specify this
threshold.
The motivation of this change is that tuning this threshold significantly (up
to 25%) improves the performance of some CUDA programs in our internal code
base. In general, branch instructions are very expensive for GPU programs.
Therefore, it is sometimes worth trading more arithmetic computation for a more
straightened control flow. Here's a reduced example:
__global__ void foo(int a, int b, int c, int d, int e, int n,
const int *input, int *output) {
int sum = 0;
for (int i = 0; i < n; ++i)
sum += (((i ^ a) > b) && (((i | c ) ^ d) > e)) ? 0 : input[i];
*output = sum;
}
The select statement in the loop body translates to two branch instructions "if
((i ^ a) > b)" and "if (((i | c) ^ d) > e)" which share a common destination.
With the default threshold, SimplifyCFG is unable to fold them, because
computing the condition of the second branch "(i | c) ^ d > e" requires two
bonus instructions. With the threshold increased, SimplifyCFG can fold the two
branches so that the loop body contains only one branch, making the code
conceptually look like:
sum += (((i ^ a) > b) & (((i | c ) ^ d) > e)) ? 0 : input[i];
Increasing the threshold significantly improves the performance of this
particular example. In the configuration where both conditions are guaranteed
to be true, increasing the threshold from 1 to 2 improves the performance by
18.24%. Even in the configuration where the first condition is false and the
second condition is true, which favors shortcuts, increasing the threshold from
1 to 2 still improves the performance by 4.35%.
We are still looking for a good threshold and maybe a better cost model than
just counting the number of bonus instructions. However, according to the above
numbers, we think it is at least worth adding a threshold to enable more
experiments and tuning. Let me know what you think. Thanks!
Test Plan: Added one test case to check the threshold is in effect
Reviewers: nadav, eliben, meheff, resistor, hfinkel
Reviewed By: hfinkel
Subscribers: hfinkel, llvm-commits
Differential Revision: http://reviews.llvm.org/D5529
llvm-svn: 218711
The doFinalization method checks that the LoopToAliasSetMap is
empty. LICM populates that map as it runs through the loop nest,
deleting the entries for child loops as it goes. However, if a child
loop is deleted by another pass (e.g. unrolling) then the loop will
never be deleted from the map because LICM walks the loop nest to
find entries it can delete.
The fix is to delete the loop from the map and free the alias set
when the loop is deleted from the loop nest.
Differential Revision: http://reviews.llvm.org/D5305
llvm-svn: 218387
shim between the TargetTransformInfo immutable pass and the Subtarget
via the TargetMachine and Function. Migrate a single call from
BasicTargetTransformInfo as an example and provide shims where TargetMachine
begins taking a Function to determine the subtarget.
No functional change.
llvm-svn: 218004
This improves other optimizations such as LSR. A sext may be added to the
compare's other operand, but this can often be hoisted outside of the loop.
llvm-svn: 217953
We used to crash processing any relevant @llvm.assume on a 32-bit target
(because we'd ask SE to subtract expressions of differing types). I've copied
our 'simple.ll' test, but with the data layout from arm-linux-gnueabihf to get
some meaningful test coverage here.
llvm-svn: 217574
The routine that determines an alignment given some SCEV returns zero if the
answer is unknown. In a case where we could determine the increment of an
AddRec but not the starting alignment, we would compute the integer modulus by
zero (which is illegal and traps). Prevent this by returning early if either
the start or increment alignment is unknown (zero).
llvm-svn: 217544
Summary:
This patch moves the profile reading logic out of the Sample Profile
transformation into a generic profile reader facility in
lib/ProfileData.
The intent is to use this new reader to implement a sample profile
reader/writer that can be used to convert sample profiles from external
sources into LLVM.
This first patch introduces no functional changes. It moves the profile
reading code from lib/Transforms/SampleProfile.cpp into
lib/ProfileData/SampleProfReader.cpp.
In subsequent patches I will:
- Add a bitcode format for sample profiles to allow for more efficient
encoding of the profile.
- Add a writer for both text and bitcode format profiles.
- Add a 'convert' command to llvm-profdata to be able to convert between
the two (and serve as entry point for other sample profile formats).
Reviewers: bogner, echristo
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D5250
llvm-svn: 217437
This change teaches LazyValueInfo to use the @llvm.assume intrinsic. Like with
the known-bits change (r217342), this requires feeding a "context" instruction
pointer through many functions. Aside from a little refactoring to reuse the
logic that turns predicates into constant ranges in LVI, the only new code is
that which can 'merge' the range from an assumption into that otherwise
computed. There is also a small addition to JumpThreading so that it can have
LVI use assumptions in the same block as the comparison feeding a conditional
branch.
With this patch, we can now simplify this as expected:
int foo(int a) {
__builtin_assume(a > 5);
if (a > 3) {
bar();
return 1;
}
return 0;
}
llvm-svn: 217345
This adds a ScalarEvolution-powered transformation that updates load, store and
memory intrinsic pointer alignments based on invariant((a+q) & b == 0)
expressions. Many of the simple cases we can get with ValueTracking, but we
still need something like this for the more complicated cases (such as those
with an offset) that require some algebra. Note that gcc's
__builtin_assume_aligned's optional third argument provides exactly for this
kind of 'misalignment' offset for which this kind of logic is necessary.
The primary motivation is to fixup alignments for vector loads/stores after
vectorization (and unrolling). This pass is added to the optimization pipeline
just after the SLP vectorizer runs (which, admittedly, does not preserve SE,
although I imagine it could). Regardless, I actually don't think that the
preservation matters too much in this case: SE computes lazily, and this pass
won't issue any SE queries unless there are any assume intrinsics, so there
should be no real additional cost in the common case (SLP does preserve DT and
LoopInfo).
llvm-svn: 217344
This change, which allows @llvm.assume to be used from within computeKnownBits
(and other associated functions in ValueTracking), adds some (optional)
parameters to computeKnownBits and friends. These functions now (optionally)
take a "context" instruction pointer, an AssumptionTracker pointer, and also a
DomTree pointer, and most of the changes are just to pass this new information
when it is easily available from InstSimplify, InstCombine, etc.
As explained below, the significant conceptual change is that known properties
of a value might depend on the control-flow location of the use (because we
care that the @llvm.assume dominates the use because assumptions have
control-flow dependencies). This means that, when we ask if bits are known in a
value, we might get different answers for different uses.
The significant changes are all in ValueTracking. Two main changes: First, as
with the rest of the code, new parameters need to be passed around. To make
this easier, I grouped them into a structure, and I made internal static
versions of the relevant functions that take this structure as a parameter. The
new code does as you might expect, it looks for @llvm.assume calls that make
use of the value we're trying to learn something about (often indirectly),
attempts to pattern match that expression, and uses the result if successful.
By making use of the AssumptionTracker, the process of finding @llvm.assume
calls is not expensive.
Part of the structure being passed around inside ValueTracking is a set of
already-considered @llvm.assume calls. This is to prevent a query using, for
example, the assume(a == b), to recurse on itself. The context and DT params
are used to find applicable assumptions. An assumption needs to dominate the
context instruction, or come after it deterministically. In this latter case we
only handle the specific case where both the assumption and the context
instruction are in the same block, and we need to exclude assumptions from
being used to simplify their own ephemeral values (those which contribute only
to the assumption) because otherwise the assumption would prove its feeding
comparison trivial and would be removed.
This commit adds the plumbing and the logic for a simple masked-bit propagation
(just enough to write a regression test). Future commits add more patterns
(and, correspondingly, more regression tests).
llvm-svn: 217342
This adds a set of utility functions for collecting 'ephemeral' values. These
are LLVM IR values that are used only by @llvm.assume intrinsics (directly or
indirectly), and thus will be removed prior to code generation, implying that
they should be considered free for certain purposes (like inlining). The
inliner's cost analysis, and a few other passes, have been updated to account
for ephemeral values using the provided functionality.
This functionality is important for the usability of @llvm.assume, because it
limits the "non-local" side-effects of adding llvm.assume on inlining, loop
unrolling, etc. (these are hints, and do not generate code, so they should not
directly contribute to estimates of execution cost).
llvm-svn: 217335
This adds an immutable pass, AssumptionTracker, which keeps a cache of
@llvm.assume call instructions within a module. It uses callback value handles
to keep stale functions and intrinsics out of the map, and it relies on any
code that creates new @llvm.assume calls to notify it of the new instructions.
The benefit is that code needing to find @llvm.assume intrinsics can do so
directly, without scanning the function, thus allowing the cost of @llvm.assume
handling to be negligible when none are present.
The current design is intended to be lightweight. We don't keep track of
anything until we need a list of assumptions in some function. The first time
this happens, we scan the function. After that, we add/remove @llvm.assume
calls from the cache in response to registration calls and ValueHandle
callbacks.
There are no new direct test cases for this pass, but because it calls it
validation function upon module finalization, we'll pick up detectable
inconsistencies from the other tests that touch @llvm.assume calls.
This pass will be used by follow-up commits that make use of @llvm.assume.
llvm-svn: 217334
LinearFunctionTestReplace tries to use the *next* indvar to compare
against when possible. However, it may be the case that the calculation
for the next indvar has NUW/NSW flags and that it may only be safely
used inside the loop. Using it in a comparison to calculate the exit
condition could result in observing poison.
This fixes PR20680.
Differential Revision: http://reviews.llvm.org/D5174
llvm-svn: 217102
Summary:
BBs might contain non-LCSSA'd values after the LCSSA pass is run if they
are unreachable from the entry block.
Normally, the users of the instruction would be PHIs but the unreachable
BBs have normal users; rewrite their uses to be undef values.
An alternative fix could involve fixing this at LCSSA but that would
require this invariant to hold after subsequent transforms. If a BB
created an unreachable block, they would be in violation of this.
This fixes PR19798.
Differential Revision: http://reviews.llvm.org/D5146
llvm-svn: 216911
SROA may decide that it needs to insert a bitcast and would set it's
insertion point before a PHI. This will create an invalid module
right quick.
Instead, choose the first insertion point in the basic block that holds
our PHI.
This fixes PR20822.
Differential Revision: http://reviews.llvm.org/D5141
llvm-svn: 216891
chain became completely broken here as *all* intrinsic users ended up
being skipped, and the ones that seemed to be singled out were actually
the exact wrong set.
This is a great example of why long else-if chains can be easily
confusing. Switch the entire code to use early exits and early continues
to have simpler (and more importantly, correct) logic here, as well as
fixing the reversed logic for detecting and continuing on lifetime
intrinsics.
I've also significantly cleaned up the test case and added another test
case demonstrating an example where the optimization is not (trivially)
safe to perform.
llvm-svn: 216871
