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
where the induction variable has a non-unit stride, such as {0,+,2}, and
there are expressions such as {1,+,2} inside the loop formed with
or or add nsw operators.
llvm-svn: 82151
SCEVUnknowns, as the non-SCEVUnknown cases in the getSCEVAtScope code
can also end up repeatedly climing through the same expression trees,
which can be unusably slow when the trees are very tall.
Also, add a quick check for SCEV pointer equality to the main
SCEV comparison routine, as the full comparison code can be expensive
in the case of large expression trees.
These fix compile-time problems in some pathlogical cases.
llvm-svn: 80623
will always return the same value. This isn't currently necessary,
since this code doesn't currently ever get called under circumstances
where it would matter, but it may some day.
llvm-svn: 80017
TargetData is not present. It still uses TargetData when available.
This generalization also fixed some limitations in the TargetData
case; the attached testcase covers this.
llvm-svn: 79344
affected after a PHI node has been analyzed, just remove affected
SCEVs from the Scalars map, so that they'll be (lazily) recreated as
needed. This avoids creating SCEV objects that aren't actually needed.
Also, rewrite the associated def-use walking code to be non-recursive
and to continue traversing past Instructions that don't have an
entry in the Scalars map.
llvm-svn: 77032
(x pred y) with more thorough code that does more complete canonicalization
before resorting to range checks. This helps it find more cases where
the canonicalized expressions match.
llvm-svn: 76671
Getelementptrs that are defined to wrap are virtually useless to
optimization, and getelementptrs that are undefined on any kind
of overflow are too restrictive -- it's difficult to ensure that
all intermediate addresses are within bounds. I'm going to take
a different approach.
Remove a few optimizations that depended on this flag.
llvm-svn: 76437
in a convenient manner, factoring out some common code from
InstructionCombining and ValueTracking. Move the contents of
BinaryOperators.h into Operator.h and use Operator to generalize them
to support ConstantExprs as well as Instructions.
llvm-svn: 76232
This adds location info for all llvm_unreachable calls (which is a macro now) in
!NDEBUG builds.
In NDEBUG builds location info and the message is off (it only prints
"UREACHABLE executed").
llvm-svn: 75640
SCEVZeroExtendExpr ahead of the most expensive analysis. This
speeds up analysis and helps avoid pathologically bad behavior
on the testcase in PR4534.
llvm-svn: 75496
This involves temporarily hard wiring some parts to use the global context. This isn't ideal, but it's
the only way I could figure out to make this process vaguely incremental.
llvm-svn: 75445
Make llvm_unreachable take an optional string, thus moving the cerr<< out of
line.
LLVM_UNREACHABLE is now a simple wrapper that makes the message go away for
NDEBUG builds.
llvm-svn: 75379
to a loop deletion more thorough. Don't prune the def-use tree search at
instructions that don't have SCEVs computed, because an instruction with
a user that has a computed SCEV may itself lack a computed SCEV. Also,
remove loop-related values from the ValuesAtScopes and
ConstantEvolutionLoopExitValues maps as well.
This fixes a regression in 483.xalancbmk.
llvm-svn: 75030
Constant. This lets ConstantInts be handled as SCEVConstant instead
of SCEVUnknown, as getUnknown no longer has special-case code for
ConstantInt and friends. This usually doesn't affect the final
output, since the constants end up getting folded later, but it
does make intermediate expressions more obvious in many cases.
llvm-svn: 74459
an individual exhaustive evaluation reflects only the exit value
implied by an individual exit, which may differ from the actual
exit value of the loop if there are other exits. This fixes PR4477.
llvm-svn: 74447
nesting order of nested AddRec expressions to skip the transformation
if it would introduce an AddRec with operands not loop-invariant
with respect to its loop.
llvm-svn: 74343
trip counts in more cases.
Generalize ScalarEvolution's isLoopGuardedByCond code to recognize
And and Or conditions, splitting the code out into an
isNecessaryCond helper function so that it can evaluate Ands and Ors
recursively, and make SCEVExpander be much more aggressive about
hoisting instructions out of loops.
test/CodeGen/X86/pr3495.ll has an additional instruction now, but
it appears to be due to an arbitrary register allocation difference.
llvm-svn: 74048
createSCEV. Also, recognize UndefValue in createSCEV.
Change getIntegerSCEV's comment to avoid mentioning FP types,
and re-implement it in terms of getConstant instead of getUnknown.
llvm-svn: 74041
This also throws out the SCEV reference counting scheme, as the the SCEVs now have a lifetime controlled by the
ScalarEvolution pass.
Note that SCEVHandle is now a no-op, and will be remove in a future commit.
llvm-svn: 73892
counts for loops with multiple exits, replacing more conservative code
which only handled constants. This is derived from a patch by
Nick Lewycky.
This also fixes llc aborts in ClamAV and others, as
getUMinFromMismatchedTypes takes care of balancing the types before
working with them.
llvm-svn: 73884
blocks, and also exit blocks with multiple conditions (combined
with (bitwise) ands and ors). It's often infeasible to compute an
exact trip count in such cases, but a useful upper bound can often
be found.
llvm-svn: 73866
SCEVUnknowns with identical Instructions to be equal. This allows
it to analze cases such as the attached testcase, where the front-end
has cloned the loop controlling expression. Along with r73805, this
lets IndVarSimplify eliminate all the sign-extend casts in the
loop in the attached testcase.
llvm-svn: 73807
so that it can access the TargetData member (when available) and
use ValueTracking.h information to compute information for
SCEVUnknown Values.
Also add GetMinLeadingZeros and GetMinSignBits functions,
with minimal implementations.
llvm-svn: 73794
obscuring what would otherwise be a low-bits mask. Use ComputeMaskedBits
to compute what ShrinkDemandedConstant knew about to reconstruct a
low-bits mask value.
llvm-svn: 73540
failures.
To support this, add some utility functions to Type to help support
vector/scalar-independent code. Change ConstantInt::get and
ConstantFP::get to support vector types, and add an overload to
ConstantInt::get that uses a static IntegerType type, for
convenience.
Introduce a new getConstant method for ScalarEvolution, to simplify
common use cases.
llvm-svn: 73431
they contain multiplications of constants with add operations.
This helps simplify several kinds of things; in particular it
helps simplify expressions like ((-1 * (%a + %b)) + %a) to %b,
as expressions like this often come up in loop trip count
computations.
llvm-svn: 73361
even though the order doesn't matter at the top level of an expression,
it does matter when the constant is a subexpression of an n-ary
expression, because n-ary expressions are sorted lexicographically.
llvm-svn: 73358
that of the LHS. It doesn't matter for correctness, but the LHS
is more likely than the RHS to be a pointer type in exotic cases,
and it's more tidy to have it return the integer type.
llvm-svn: 72424
in the case where a loop exit value cannot be computed, instead of only in
some cases while using SCEVCouldNotCompute in others. This simplifies
getSCEVAtScope's callers.
llvm-svn: 72375
instructions. It attempts to create high-level multi-operand GEPs,
though in cases where this isn't possible it falls back to casting
the pointer to i8* and emitting a GEP with that. Using GEP instructions
instead of ptrtoint+arithmetic+inttoptr helps pointer analyses that
don't use ScalarEvolution, such as BasicAliasAnalysis.
Also, make the AddrModeMatcher more aggressive in handling GEPs.
Previously it assumed that operand 0 of a GEP would require a register
in almost all cases. It now does extra checking and can do more
matching if operand 0 of the GEP is foldable. This fixes a problem
that was exposed by SCEVExpander using GEPs.
llvm-svn: 72093
getNoopOrSignExtend, and getTruncateOrNoop. These are similar
to getTruncateOrZeroExtend etc., except that they assert that
the conversion is either not widening or narrowing, as
appropriate. These will be used in some upcoming fixes.
llvm-svn: 71632
These values aren't analyzable, so they don't care if more information
about the loop trip count can be had. Also, SCEVUnknown is used for
a PHI while the PHI itself is being analyzed, so it needs to be left
in the Scalars map. This fixes a variety of subtle issues.
llvm-svn: 71533
return the correct value when the cast operand is all zeros. This ought
to be pretty rare, because it would mean that the regular SCEV folding
routines missed a case, though there are cases they might legitimately
miss. Also, it's unlikely anything currently using GetMinTrailingZeros
cares about this case.
llvm-svn: 71532
which are not analyzed with SCEV techniques, which can require
brute-forcing through a large number of instructions. This
fixes a massive compile-time issue on 400.perlbench (in
particular, the loop in MD5Transform).
llvm-svn: 71259
to sorting SCEVs by their kind, sort SCEVs of the same kind according
to their operands. This helps avoid things like (a+b) being a distinct
expression from (b+a).
llvm-svn: 71160
CallbackVH, with fixes. allUsesReplacedWith need to
walk the def-use chains and invalidate all users of a
value that is replaced. SCEVs of users need to be
recalcualted even if the new value is equivalent. Also,
make forgetLoopPHIs walk def-use chains, since any
SCEV that depends on a PHI should be recalculated when
more information about that PHI becomes available.
llvm-svn: 70927
makes ScalarEvolution::deleteValueFromRecords, and it's code that
subtly needed to be called before ReplaceAllUsesWith, unnecessary.
It also makes ValueDeletionListener unnecessary.
llvm-svn: 70645
artificial "ptrtoint", as it tends to clutter up complicated
expressions. The cast operators now print both source and
destination types, which is usually sufficient.
llvm-svn: 70554
compute an upper-bound value for the trip count, in addition to
the actual trip count. Use this to allow getZeroExtendExpr and
getSignExtendExpr to fold casts in more cases.
This may eventually morph into a more general value-range
analysis capability; there are certainly plenty of places where
more complete value-range information would allow more folding.
llvm-svn: 70509
(sext i8 {-128,+,1} to i64) to i64 {-128,+,1}, where the iteration
crosses from negative to positive, but is still safe if the trip
count is within range.
llvm-svn: 70421
information to simplify [sz]ext({a,+,b}) to {zext(a),+,[zs]ext(b)},
as appropriate.
These functions and the trip count code each call into the other, so
this requires careful handling to avoid infinite recursion. During
the initial trip count computation, conservative SCEVs are used,
which are subsequently discarded once the trip count is actually
known.
Among other benefits, this change lets LSR automatically eliminate
some unnecessary zext-inreg and sext-inreg operation where the
operand is an induction variable.
llvm-svn: 70241
type to truncate to should be the number of bits of the value that are
preserved, not the number that are clobbered with sign-extension.
This fixes regressions in ldecod.
llvm-svn: 69704
size from the integer, requiring zero extension or truncation. Don't
create ZExtInsts with pointer types. This fixes a regression in
consumer-jpeg.
llvm-svn: 69307
have pointer types, though in contrast to C pointer types, SCEV
addition is never implicitly scaled. This not only eliminates the
need for special code like IndVars' EliminatePointerRecurrence
and LSR's own GEP expansion code, it also does a better job because
it lets the normal optimizations handle pointer expressions just
like integer expressions.
Also, since LLVM IR GEPs can't directly index into multi-dimensional
VLAs, moving the GEP analysis out of client code and into the SCEV
framework makes it easier for clients to handle multi-dimensional
VLAs the same way as other arrays.
Some existing regression tests show improved optimization.
test/CodeGen/ARM/2007-03-13-InstrSched.ll in particular improved to
the point where if-conversion started kicking in; I turned it off
for this test to preserve the intent of the test.
llvm-svn: 69258
to more accurately describe what it does. Expand its doxygen comment
to describe what the backedge-taken count is and how it differs
from the actual iteration count of the loop. Adjust names and
comments in associated code accordingly.
llvm-svn: 65382
trip count value when the original loop iteration condition is
signed and the canonical induction variable won't undergo signed
overflow. This isn't required for correctness; it just preserves
more information about original loop iteration values.
Add a getTruncateOrSignExtend method to ScalarEvolution,
following getTruncateOrZeroExtend.
llvm-svn: 64918
modified in a way that may effect the trip count calculation. Change
IndVars to use this method when it rewrites pointer or floating-point
induction variables instead of using a doInitialization method to
sneak these changes in before ScalarEvolution has a chance to see
the loop. This eliminates the need for LoopPass to depend on
ScalarEvolution.
llvm-svn: 64810
loop induction on LP64 targets. When the induction variable is
used in addressing, IndVars now is usually able to inserst a
64-bit induction variable and eliminates the sign-extending cast.
This is also useful for code using C "short" types for
induction variables on targets with 32-bit addressing.
Inserting a wider induction variable is easy; the tricky part is
determining when trunc(sext(i)) expressions are no-ops. This
requires range analysis of the loop trip count. A common case is
when the original loop iteration starts at 0 and exits when the
induction variable is signed-less-than a fixed value; this case
is now handled.
This replaces IndVarSimplify's OptimizeCanonicalIVType. It was
doing the same optimization, but it was limited to loops with
constant trip counts, because it was running after the loop
rewrite, and the information about the original induction
variable is lost by that point.
Rename ScalarEvolution's executesAtLeastOnce to
isLoopGuardedByCond, generalize it to be able to test for
ICMP_NE conditions, and move it to be a public function so that
IndVars can use it.
llvm-svn: 64407
Use it to safely handle less-than-or-equals-to exit conditions in loops. These
also occur when the loop exit branch is exit on true because SCEV inverses the
icmp predicate.
Use it again to handle non-zero strides, but only with an unsigned comparison
in the exit condition.
llvm-svn: 59528
If this patch causes a performance regression for anyone, please let me know,
and it can be fixed in a different way with much more effort.
llvm-svn: 59384
- Recognize expressions like "x > -1 ? x : 0" as min/max and turn them
into expressions like "x < 0 ? 0 : x", which is easily recognizable
as a min/max operation.
- Refrain from folding expression like "y/2 < 1" to "y < 2" when the
comparison is being used as part of a min or max idiom, like
"y/2 < 1 ? 1 : y/2". In that case, the division has another use, so
folding doesn't eliminate it, and obfuscates the min/max, making it
harder to recognize as a min/max operation.
These benefit ScalarEvolution, CodeGen, and anything else that wants to
recognize integer min and max.
llvm-svn: 56246
its callers to emit a space character before calling it when a
space is needed.
This fixes several spurious whitespace issues in
ScalarEvolution's debug dumps. See the test changes for
examples.
This also fixes odd space-after-tab indentation in the output
for switch statements, and changes calls from being printed like
this:
call void @foo( i32 %x )
to this:
call void @foo(i32 %x)
llvm-svn: 56196
continue past the first conditional branch when looking for a
relevant test. This helps it avoid using MAX expressions in
loop trip counts in more cases.
llvm-svn: 54697
version uses a new algorithm for evaluating the binomial coefficients
which is significantly more efficient for AddRecs of more than 2 terms
(see the comments in the code for details on how the algorithm works).
It also fixes some bugs: it removes the arbitrary length restriction for
AddRecs, it fixes the silent generation of incorrect code for AddRecs
which require a wide calculation width, and it fixes an issue where we
were incorrectly truncating the iteration count too far when evaluating
an AddRec expression narrower than the induction variable.
There are still a few related issues I know of: I think there's
still an issue with the SCEVExpander expansion of AddRec in terms of
the width of the induction variable used. The hack to avoid generating
too-wide integers shouldn't be necessary; instead, the callers should be
considering the cost of the expansion before expanding it (in addition
to not expanding too-wide integers, we might not want to expand
expressions that are really expensive, especially when optimizing for
size; calculating an length-17 32-bit AddRec currently generates about 250
instructions of straight-line code on X86). Also, for long 32-bit
AddRecs on X86, CodeGen really sucks at scheduling the code. I'm planning on
filing follow-up PRs for these issues.
llvm-svn: 54332
time applying to the implicit comparison in smin expressions. The
correct way to transform an inequality into the opposite
inequality, either signed or unsigned, is with a not expression.
I looked through the SCEV code, and I don't think there are any more
occurrences of this issue.
llvm-svn: 54194
SGT exit condition. Essentially, the correct way to flip an inequality
in 2's complement is the not operator, not the negation operator.
That said, the difference only affects cases involving INT_MIN.
Also, enhance the pre-test search logic to be a bit smarter about
inequalities flipped with a not operator, so it can eliminate the smax
from the iteration count for simple loops.
llvm-svn: 54184
bail after 256-bits to avoid producing code that the backends can't handle.
Previously, we capped it at 64-bits, preferring to miscompile in those cases.
This change also reverts much of r52248 because the invariants the code was
expecting are now being met.
llvm-svn: 53812
with code that was expecting different bit widths for different values.
Make getTruncateOrZeroExtend a method on ScalarEvolution, and use it.
llvm-svn: 52248
manually performing the comparison. This allows the special
case to work correctly even in the case where someone is
experimenting with a different comparison function :-).
llvm-svn: 49670
