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
sext around sext(shorter IV + constant), using a
longer IV instead, when it can figure out the
add can't overflow. This comes up a lot in
subscripting; mainly affects 64 bit.
llvm-svn: 69123
llvm.dbg.region.end instrinsic. This nested llvm.dbg.func.start/llvm.dbg.region.end pair now enables DW_TAG_inlined_subroutine support in code generator.
llvm-svn: 69118
strncat :(
strncat(foo, "bar", 99)
would be optimized to
memcpy(foo+strlen(foo), "bar", 100, 1)
instead of
memcpy(foo+strlen(foo), "bar", 4, 1)"
Patch by Benjamin Kramer!
llvm-svn: 68905
integer types, unless they are already strange. This prevents it from
turning the code produced by SROA into crazy libcalls and stuff that
the code generator can't handle. In the attached example, the result
was an i96 multiply that caused the x86 backend to assert.
Note that if TargetData had an idea of what the legal types are for
a target that this could be used to stop instcombine from introducing
i64 muls, as Scott wanted.
llvm-svn: 68598
to/from integer types that are not intptr_t to convert to intptr_t
then do an integer conversion to the dest type. This exposes the
cast to the optimizer.
llvm-svn: 67638
1. Make instcombine always canonicalize trunc x to i1 into an icmp(x&1). This
exposes the AND to other instcombine xforms and is more of what the code
generator expects.
2. Rewrite the remaining trunc pattern match to use 'match', which
simplifies it a lot.
llvm-svn: 67635
linkage: the value may be replaced with something
different at link time. (Frontends that want to
allow values to be loaded out of weak constants can
give their constants weak_odr linkage).
llvm-svn: 67407
the inliner; prevents nondeterministic behavior
when the same address is reallocated.
Don't build call graph nodes for debug intrinsic calls;
they're useless, and there were typically a lot of them.
llvm-svn: 67311
and was deleting Instructions without clearing the
corresponding map entry. This led to nondeterministic
behavior if the same address got allocated to another
Instruction within a short time.
llvm-svn: 67306
it is not APInt clean, but even when it is it needs to be evaluated carefully
to determine whether it is actually profitable.
This fixes a crash on PR3806
llvm-svn: 67134
changes.
For InvokeInst now all arguments begin at op_begin().
The Callee, Cont and Fail are now faster to get by
access relative to op_end().
This patch introduces some temporary uglyness in CallSite.
Next I'll bring CallInst up to a similar scheme and then
the uglyness will magically vanish.
This patch also exposes all the reliance of the libraries
on InvokeInst's operand ordering. I am thinking of taking
care of that too.
llvm-svn: 66920
right; did the wrong thing when there are exactly 11
non-debug instructions, followed by debug info.
Remove a FIXME since it's apparently been fixed along the way.
llvm-svn: 66840
in the Ada testcase. Reverting this only covers up
the real problem, which is a nasty conceptual difficulty
in the phi elimination pass: when eliminating phi nodes
in landing pads, the register copies need to come before
the invoke, not at the end of the basic block which is
too late... See PR3784.
llvm-svn: 66826
from a switch table. Multiple table entries that
branch to the same place were being sorted by the
pointer value of the ConstantInt*; changed to sort
by the actual value of the ConstantInt.
llvm-svn: 66749
allocations. Apparently the assumption is there is an
instruction (terminator?) following the allocation so I
am allowing the same assumption.
llvm-svn: 66716
and extern_weak_odr. These are the same as the non-odr versions,
except that they indicate that the global will only be overridden
by an *equivalent* global. In C, a function with weak linkage can
be overridden by a function which behaves completely differently.
This means that IP passes have to skip weak functions, since any
deductions made from the function definition might be wrong, since
the definition could be replaced by something completely different
at link time. This is not allowed in C++, thanks to the ODR
(One-Definition-Rule): if a function is replaced by another at
link-time, then the new function must be the same as the original
function. If a language knows that a function or other global can
only be overridden by an equivalent global, it can give it the
weak_odr linkage type, and the optimizers will understand that it
is alright to make deductions based on the function body. The
code generators on the other hand map weak and weak_odr linkage
to the same thing.
llvm-svn: 66339
to find a tiny mouse hole to squeeze through, it struck
me that globals without a name can be considered internal
since they can't be referenced from outside the current
module. This patch makes GlobalOpt give them internal
linkage. Also done for aliases even though they always
have names, since in my opinion anonymous aliases should
be allowed for consistency with global variables and
functions. So if that happens one day, this code is ready!
llvm-svn: 66267
If non constant local GV named A is used by a constant local GV named B (e.g. llvm.dbg.variable) and B is not used by anyone else then eliminate A as well as B.
In other words, debug info should not interfere in removal of unused GV.
--This life, and those below, will be ignored--
M test/Transforms/GlobalOpt/2009-03-03-dbg.ll
M lib/Transforms/IPO/GlobalOpt.cpp
llvm-svn: 66167
use, check also for the case where it has two uses,
the other being a llvm.dbg.declare. This is needed so
debug info doesn't affect codegen.
llvm-svn: 65970
testsuite:
Running /Volumes/Sandbox/Buildbot/llvm/full-llvm/build/llvmCore/test/CodeGen/X86/dg.exp ...
FAIL: /Volumes/Sandbox/Buildbot/llvm/full-llvm/build/llvmCore/test/CodeGen/X86/nancvt.ll
Failed with exit(1) at line 2
while running: grep 2147027116 nancvt.ll.tmp | count 3
count: expected 3 lines and got 0.
child process exited abnormally
FAIL: /Volumes/Sandbox/Buildbot/llvm/full-llvm/build/llvmCore/test/CodeGen/X86/vec_ins_extract.ll
Failed with exit(1) at line 1
while running: llvm-as < /Volumes/Sandbox/Buildbot/llvm/full-llvm/build/llvmCore/test/CodeGen/X86/vec_ins_extract.ll | opt -scalarrepl -instcombine | llc -march=x86 -mcpu=yonah | not /usr/bin/grep sub.*esp
subl $28, %esp
subl $28, %esp
child process exited abnormally
And more.
llvm-svn: 65758
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 counts that use signed comparisons. It's not obviously the best
approach for preserving trip count information, and at any rate there
isn't anything in the tree right now that makes use of that, so for
now always using zero-extensions is preferable.
llvm-svn: 65347
so that ScalarEvolution doesn't hang onto a dangling Loop*, which
could be a problem if another Loop happens to get allocated at the
same address.
llvm-svn: 65323
memcpy to match the alignment of the destination. It isn't necessary
for making loads and stores handled like the SSE loadu/storeu
intrinsics, and it was causing a performance regression in
MultiSource/Applications/JM/lencod.
The problem appears to have been a memcpy that copies from some
highly aligned array into an alloca; the alloca was then being
assigned a large alignment, which required codegen to perform
dynamic stack-pointer re-alignment, which forced the enclosing
function to have a frame pointer, which led to increased spilling.
llvm-svn: 65289
as legality. Make load sinking and gep sinking more careful: we only
do it when it won't pessimize loads from the stack. This has the added
benefit of not producing code that is unanalyzable to SROA.
llvm-svn: 65209
addresses, part 1. This fixes an obvious logic bug. Previously if the only
in-loop use is a PHI, it would return AllUsesAreAddresses as true.
llvm-svn: 65178
Currently this pass will delete the variable declaration info,
and keep the line number info. But the kept line number info is not updated,
and some is redundant or not correct, this patch just updates those info.
llvm-svn: 65123
reduction of address calculations down to basic pointer arithmetic.
This is currently off by default, as it needs a few other features
before it becomes generally useful. And even when enabled, full
strength reduction is only performed when it doesn't increase
register pressure, and when several other conditions are true.
This also factors out a bunch of exisiting LSR code out of
StrengthReduceStridedIVUsers into separate functions, and tidies
up IV insertion. This actually decreases register pressure even
in non-superhero mode. The change in iv-users-in-other-loops.ll
is an example of this; there are two more adds because there are
two fewer leas, and there is less spilling.
llvm-svn: 65108
here. Since we only do the transform if there is
one use, strip off any such users in the hope of
making the transform fire more often.
llvm-svn: 64926
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
are multiple IV's in a loop, some of them may under go signed
or unsigned wrapping even if the IV that's used in the loop
exit condition doesn't. Restrict sign-extension-elimination
and zero-extension-elimination to only those that operate on
the original loop-controlling IV.
llvm-svn: 64866
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
Enhance instcombine to use the preferred field of
GetOrEnforceKnownAlignment in more cases, so that regular IR operations are
optimized in the same way that the intrinsics currently are.
llvm-svn: 64623
when I was looking at functions used by python.
Highlights include, better largefile support (64-bit file sizes on 32-bit
systems), fputs string is nocapture, popen/pclose added (popen being noalias
return), modf and frexp and friends. Also added some missing 'break' statements
and combined identical sections.
llvm-svn: 64615
- Test for signed and unsigned wrapping conditions, instead of just
testing for non-negative induction ranges.
- Handle loops with GT comparisons, in addition to LT comparisons.
- Support more cases of induction variables that don't start at 0.
llvm-svn: 64532
addrec in a different loop to check the value being added to
the accumulated Start value, not the Start value before it has
the new value added to it. This prevents LSR from going crazy
on the included testcase. Dale, please review.
llvm-svn: 64440
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
accessed at least once as a vector. This prevents it from
compiling the example in not-a-vector into:
define double @test(double %A, double %B) {
%tmp4 = insertelement <7 x double> undef, double %A, i32 0
%tmp = insertelement <7 x double> %tmp4, double %B, i32 4
%tmp2 = extractelement <7 x double> %tmp, i32 4
ret double %tmp2
}
instead, producing the integer code. Producing vectors when they
aren't otherwise in the program is dangerous because a lot of other
code treats them carefully and doesn't want to break them down.
OTOH, many things want to break down tasty i448's.
llvm-svn: 63638
With the new world order, it can handle cases where the first
store into the alloca is an element of the vector, instead of
requiring the first analyzed store to have the vector type
itself. This allows us to un-xfail
test/CodeGen/X86/vec_ins_extract.ll.
llvm-svn: 63590
turn icmp eq a+x, b+x into icmp eq a, b if a+x or b+x has other uses. This
may have been increasing register pressure leading to the bzip2 slowdown.
llvm-svn: 63487
improvements to the EvaluateInDifferentType code. This code works
by just inserted a bunch of new code and then seeing if it is
useful. Instcombine is not allowed to do this: it can only insert
new code if it is useful, and only when it is converging to a more
canonical fixed point. Now that we iterate when DCE makes progress,
this causes an infinite loop when the code ends up not being used.
llvm-svn: 63483
simplifydemandedbits to simplify instructions with *multiple
uses* in contexts where it can get away with it. This allows
it to simplify the code in multi-use-or.ll into a single 'add
double'.
This change is particularly interesting because it will cover
up for some common codegen bugs with large integers created due
to the recent SROA patch. When working on fixing those bugs,
this should be disabled.
llvm-svn: 63481
Now, if it detects that "V" is the same as some other value,
SimplifyDemandedBits returns the new value instead of RAUW'ing it immediately.
This has two benefits:
1) simpler code in the recursive SimplifyDemandedBits routine.
2) it allows future fun stuff in instcombine where an operation has multiple
uses and can be simplified in one context, but not all.
#2 isn't implemented yet, this patch should have no functionality change.
llvm-svn: 63479
be able to handle *ANY* alloca that is poked by loads and stores of
bitcasts and GEPs with constant offsets. Before the code had a number
of annoying limitations and caused it to miss cases such as storing into
holes in structs and complex casts (as in bitfield-sroa) where we had
unions of bitfields etc. This also handles a number of important cases
that are exposed due to the ABI lowering stuff we do to pass stuff by
value.
One case that is pretty great is that we compile
2006-11-07-InvalidArrayPromote.ll into:
define i32 @func(<4 x float> %v0, <4 x float> %v1) nounwind {
%tmp10 = call <4 x i32> @llvm.x86.sse2.cvttps2dq(<4 x float> %v1)
%tmp105 = bitcast <4 x i32> %tmp10 to i128
%tmp1056 = zext i128 %tmp105 to i256
%tmp.upgrd.43 = lshr i256 %tmp1056, 96
%tmp.upgrd.44 = trunc i256 %tmp.upgrd.43 to i32
ret i32 %tmp.upgrd.44
}
which turns into:
_func:
subl $28, %esp
cvttps2dq %xmm1, %xmm0
movaps %xmm0, (%esp)
movl 12(%esp), %eax
addl $28, %esp
ret
Which is pretty good code all things considering :).
One effect of this is that SROA will start generating arbitrary bitwidth
integers that are a multiple of 8 bits. In the case above, we got a
256 bit integer, but the codegen guys assure me that it can handle the
simple and/or/shift/zext stuff that we're doing on these operations.
This addresses rdar://6532315
llvm-svn: 63469
There is now a direct way from value-use-iterator to incoming block in PHINode's API.
This way we avoid the iterator->index->iterator trip, and especially the costly
getOperandNo() invocation. Additionally there is now an assertion that the iterator
really refers to one of the PHI's Uses.
llvm-svn: 62869
we assumed a CFG structure that would be valid when all code in
the function is reachable, but not all code is necessarily
reachable. Do a simple, but horrible, CFG walk to check for this
case.
llvm-svn: 62487
because of dead code, a phi could use the speculated instruction
that was not in "BB2". Make this check explicit and tighten up
some other corners. This fixes PR3292. No testcase becauase this
depends entirely on visitation order of blocks and requires a
sequence of 8 passes to repro.
llvm-svn: 62476
doing very similar pointer capture analysis.
Factor out the common logic. The new version
is from FunctionAttrs since it does a better
job than the version in BasicAliasAnalysis
llvm-svn: 62461
putc, puts, perror, vscanf and vsscanf from getting annotations.
Add annotations for eight printf functions, memalign, pread and pwrite.
On Linux, llvm-gcc sometimes renames strdup, getc, putc, strtok_r, scanf and
sscanf. Match the alternate function names.
Fix a crash annotating opendir.
Don't mark fsetpos's second parameter as nocapture. It's supposed to be
captured.
Do mark fopen's path and mode strings as nocapture. Mark ferror as readonly,
but not fileno which may set errno.
llvm-svn: 62456
- Looking at the number of sign bits of the a sext instruction to determine whether new trunc + sext pair should be added when its source is being evaluated in a different type.
llvm-svn: 62263
my earlier patch to this file.
The issue there was that all uses of an IV inside a loop
are actually references to Base[IV*2], and there was one
use outside that was the same but LSR didn't see the base
or the scaling because it didn't recurse into uses outside
the loop; thus, it used base+IV*scale mode inside the loop
instead of pulling base out of the loop. This was extra bad
because register pressure later forced both base and IV into
memory. Doing that recursion, at least enough
to figure out addressing modes, is a good idea in general;
the change in AddUsersIfInteresting does this. However,
there were side effects....
It is also possible for recursing outside the loop to
introduce another IV where there was only 1 before (if
the refs inside are not scaled and the ref outside is).
I don't think this is a common case, but it's in the testsuite.
It is right to be very aggressive about getting rid of
such introduced IVs (CheckForIVReuse and the handling of
nonzero RewriteFactor in StrengthReduceStridedIVUsers).
In the testcase in question the new IV produced this way
has both a nonconstant stride and a nonzero base, neither
of which was handled before. And when inserting
new code that feeds into a PHI, it's right to put such
code at the original location rather than in the PHI's
immediate predecessor(s) when the original location is outside
the loop (a case that couldn't happen before)
(RewriteInstructionToUseNewBase); better to avoid making
multiple copies of it in this case.
Also, the mechanism for keeping SCEV's corresponding to GEP's
no longer works, as the GEP might change after its SCEV
is remembered, invalidating the SCEV, and we might get a bad
SCEV value when looking up the GEP again for a later loop.
This also couldn't happen before, as we weren't recursing
into GEP's outside the loop.
Also, when we build an expression that involves a (possibly
non-affine) IV from a different loop as well as an IV from
the one we're interested in (containsAddRecFromDifferentLoop),
don't recurse into that. We can't do much with it and will
get in trouble if we try to create new non-affine IVs or something.
More testcases are coming.
llvm-svn: 62212
vector and extraneous loop over it, 2) not delete globals used by
phis/selects etc which could actually be useful. This fixes PR3321.
Many thanks to Duncan for narrowing this down.
llvm-svn: 62201
compensation for turning off gcc's inliner. This gets
us closer to the amount of inlining we were getting before.
It is not a win on everything, of course, but seems to
gain overall.
llvm-svn: 62058
canonicalization transform based on duncan's comments:
1) improve the comment about %.
2) within our index loop make sure the offset stays
within the *type size*, instead of within the *abi size*.
This allows us to reason explicitly about landing in tail
padding and means that issues like non-zero offsets into
[0 x foo] types don't occur anymore.
llvm-svn: 62045
functions that don't already have a (dynamic) alloca.
Dynamic allocas cause inefficient codegen and we shouldn't
propagate this (behavior follows gcc). Two existing tests
assumed such inlining would be done; they are hacked by
adding an alloca in the caller, preserving the point of
the tests.
llvm-svn: 61946
loads from allocas that cover the entire aggregate. This handles
some memcpy/byval cases that are produced by llvm-gcc. This triggers
a few times in kc++ (with std::pair<std::_Rb_tree_const_iterator
<kc::impl_abstract_phylum*>,bool>) and once in 176.gcc (with %struct..0anon).
llvm-svn: 61915
was it not very helpful, it was also wrong! The problem
is shown in the testcase: the alloca might be passed to
a nocapture callee which dereferences it and returns the
original pointer. But because it was a nocapture call we
think we don't need to track its uses, but we do.
llvm-svn: 61876
integer to a (transitive) bitcast the alloca and if that integer
has the full size of the alloca, then it clobbers the whole thing.
Handle this by extracting pieces out of the stored integer and
filing them away in the SROA'd elements.
This triggers fairly frequently because the CFE uses integers to
pass small structs by value and the inliner exposes these. For
example, in kimwitu++, I see a bunch of these with i64 stores to
"%struct.std::pair<std::_Rb_tree_const_iterator<kc::impl_abstract_phylum*>,bool>"
In 176.gcc I see a few i32 stores to "%struct..0anon".
In the testcase, this is a difference between compiling test1 to:
_test1:
subl $12, %esp
movl 20(%esp), %eax
movl %eax, 4(%esp)
movl 16(%esp), %eax
movl %eax, (%esp)
movl (%esp), %eax
addl 4(%esp), %eax
addl $12, %esp
ret
vs:
_test1:
movl 8(%esp), %eax
addl 4(%esp), %eax
ret
The second half of this will be to handle loads of the same form.
llvm-svn: 61853
In fact this also deletes those with linkonce linkage,
however this is currently dead because for the moment
aliases aren't allowed to have this linkage type.
llvm-svn: 61742
Finalization occurs after all the FunctionPasses in the group have run, which
is clearly not what we want.
This also means that we have to make sure that we apply the right param
attributes when creating a new function.
Also, add a missed optimization: strdup and strndup. NoCapture and
NoAlias return!
llvm-svn: 61658
not have pointer type. In particular, it may
be the condition argument for a select or a GEP
index. While I was unable to construct a testcase
for which some bits of the original pointer are
captured due to one of these, it's very very close
to being possible - so play safe and exclude these
possibilities.
llvm-svn: 61580
the argument to be stored to an alloca by tracking uses
of the alloca. This occurs 4 times (out of 7121, 0.05%)
in MultiSource/Applications, so may not be worth it. On
the other hand, it is easy to do and fairly cheap. The
functions it helps are: W_addcom and W_addlit in spiff;
process_args (argv) in d (make_dparser); ercPixConcealIMB
in JM/ldecod.
llvm-svn: 61570
functions that don't write can't leak a pointer except through
the return value, so a void readonly function is implicitly nocapture.
Test these, and add a test that verifies that f1 calling f2 with an
otherwise dead pointer gets both of them marked nocapture.
llvm-svn: 61552
to work out (in a very simplistic way) which function
arguments (pointer arguments only) are only dereferenced
and so do not escape. Mark such arguments 'nocapture'.
llvm-svn: 61525
and select instructions doesn't buy anything here
except extra complexity: the only difference in
the entire testsuite was that a readonly function
became readnone in MiBench/consumer-typeset. Add
a comment about this.
llvm-svn: 61478
constants, since doing so is irrelevant for aliasing
purposes. While this doesn't increase the total number
of functions marked readonly or readnone in MultiSource/
Applications (3089), it does result in 12 functions being
marked readnone rather than readonly.
Before:
readnone: 820
readonly: 2269
After:
readnone: 832
readonly: 2257
llvm-svn: 61469
my last patch to this file.
The issue there was that all uses of an IV inside a loop
are actually references to Base[IV*2], and there was one
use outside that was the same but LSR didn't see the base
or the scaling because it didn't recurse into uses outside
the loop; thus, it used base+IV*scale mode inside the loop
instead of pulling base out of the loop. This was extra bad
because register pressure later forced both base and IV into
memory. Doing that recursion, at least enough
to figure out addressing modes, is a good idea in general;
the change in AddUsersIfInteresting does this. However,
there were side effects....
It is also possible for recursing outside the loop to
introduce another IV where there was only 1 before (if
the refs inside are not scaled and the ref outside is).
I don't think this is a common case, but it's in the testsuite.
It is right to be very aggressive about getting rid of
such introduced IVs (CheckForIVReuse and the handling of
nonzero RewriteFactor in StrengthReduceStridedIVUsers).
In the testcase in question the new IV produced this way
has both a nonconstant stride and a nonzero base, neither
of which was handled before. And when inserting
new code that feeds into a PHI, it's right to put such
code at the original location rather than in the PHI's
immediate predecessor(s) when the original location is outside
the loop (a case that couldn't happen before)
(RewriteInstructionToUseNewBase); better to avoid making
multiple copies of it in this case.
Also, the mechanism for keeping SCEV's corresponding to GEP's
no longer works, as the GEP might change after its SCEV
is remembered, invalidating the SCEV, and we might get a bad
SCEV value when looking up the GEP again for a later loop.
This also couldn't happen before, as we weren't recursing
into GEP's outside the loop.
I owe some testcases for this, want to get it in for nightly runs.
llvm-svn: 61362
- Use SplitBlockPredecessors to factor out common predecessors of the critical edge destination. This is disabled for now due to some regressions.
llvm-svn: 61248
my last patch to this file.
The issue there was that all uses of an IV inside a loop
are actually references to Base[IV*2], and there was one
use outside that was the same but LSR didn't see the base
or the scaling because it didn't recurse into uses outside
the loop; thus, it used base+IV*scale mode inside the loop
instead of pulling base out of the loop. This was extra bad
because register pressure later forced both base and IV into
memory. Doing that recursion, at least enough
to figure out addressing modes, is a good idea in general;
the change in AddUsersIfInteresting does this. However,
there were side effects....
It is also possible for recursing outside the loop to
introduce another IV where there was only 1 before (if
the refs inside are not scaled and the ref outside is).
I don't think this is a common case, but it's in the testsuite.
It is right to be very aggressive about getting rid of
such introduced IVs (CheckForIVReuse and the handling of
nonzero RewriteFactor in StrengthReduceStridedIVUsers).
In the testcase in question the new IV produced this way
has both a nonconstant stride and a nonzero base, neither
of which was handled before. (This patch does not handle
all the cases where this can happen.) And when inserting
new code that feeds into a PHI, it's right to put such
code at the original location rather than in the PHI's
immediate predecessor(s) when the original location is outside
the loop (a case that couldn't happen before)
(RewriteInstructionToUseNewBase); better to avoid making
multiple copies of it in this case.
Everything above is exercised in
CodeGen/X86/lsr-negative-stride.ll (and ifcvt4 in ARM which is
the same IR).
llvm-svn: 61178
nodes. This allows it to do fairly general phi insertion if a
load from a pointer global wants to be SRAd but the load is used
by (recursive) phi nodes. This fixes a pessimization on ppc
introduced by Load PRE.
llvm-svn: 61123
consistently for deleting branches. In addition to being slightly
more readable, this makes SimplifyCFG a bit better
about cleaning up after itself when it makes conditions unused.
llvm-svn: 61100
CFG when there is exactly one predecessor where the load is not available.
This is designed to not increase code size but still eliminate partially
redundant loads. This fires 1765 times on 403.gcc even though it doesn't
do critical edge splitting yet (the most common reason for it to fail).
llvm-svn: 61027
cleans up the generated code a bit. This should have the added benefit of
not randomly renaming functions/globals like my previous patch did. :)
llvm-svn: 61023
llvm[2]: Linking Release executable opt (without symbols)
...
Undefined symbols:
"llvm::APFloat::IEEEsingle", referenced from:
__ZN4llvm7APFloat10IEEEsingleE$non_lazy_ptr in libLLVMCore.a(Constants.o)
__ZN4llvm7APFloat10IEEEsingleE$non_lazy_ptr in libLLVMCore.a(AsmWriter.o)
__ZN4llvm7APFloat10IEEEsingleE$non_lazy_ptr in libLLVMCore.a(ConstantFold.o)
"llvm::APFloat::IEEEdouble", referenced from:
__ZN4llvm7APFloat10IEEEdoubleE$non_lazy_ptr in libLLVMCore.a(Constants.o)
__ZN4llvm7APFloat10IEEEdoubleE$non_lazy_ptr in libLLVMCore.a(AsmWriter.o)
__ZN4llvm7APFloat10IEEEdoubleE$non_lazy_ptr in libLLVMCore.a(ConstantFold.o)
ld: symbol(s) not found
This is in release mode. To replicate, compile llvm and llvm-gcc in optimized
mode. Then build llvm, in optimized mode, with the newly created compiler.
llvm-svn: 60977
of a pointer. This allows is to catch more equivalencies. For example,
the type_lists_compatible_p function used to require two iterations of
the gvn pass (!) to delete its 18 redundant loads because the first pass
would CSE all the addressing computation cruft, which would unblock the
second memdep/gvn passes from recognizing them. This change allows
memdep/gvn to catch all 18 when run just once on the function (as is
typical :) instead of just 3.
On all of 403.gcc, this bumps up the # reundandancies found from:
63 gvn - Number of instructions PRE'd
153991 gvn - Number of instructions deleted
50069 gvn - Number of loads deleted
to:
63 gvn - Number of instructions PRE'd
154137 gvn - Number of instructions deleted
50185 gvn - Number of loads deleted
+120 loads deleted isn't bad.
llvm-svn: 60799
MemDep::getNonLocalPointerDependency method. There are
some open issues with this (missed optimizations) and
plenty of future work, but this does allow GVN to eliminate
*slightly* more loads (49246 vs 49033).
Switching over now allows simplification of the other code
path in memdep.
llvm-svn: 60780
doesn't do its own local caching, and is slightly more aggressive about
free/store dse (see testcase). This eliminates the last external client
of MemDep::getDependenceFrom().
llvm-svn: 60619
loops when they can be subsumed into addressing modes.
Change X86 addressing mode check to realize that
some PIC references need an extra register.
(I believe this is correct for Linux, if not, I'm sure
someone will tell me.)
llvm-svn: 60608
1. Merge the 'None' result into 'Normal', making loads
and stores return their dependencies on allocations as Normal.
2. Split the 'Normal' result into 'Clobber' and 'Def' to
distinguish between the cases when memdep knows the value is
produced from when we just know if may be changed.
3. Move some of the logic for determining whether readonly calls
are CSEs into memdep instead of it being in GVN. This still
leaves verification that the arguments are hte same to GVN to
let it know about value equivalences in different contexts.
4. Change memdep's call/call dependency analysis to use
getModRefInfo(CallSite,CallSite) instead of doing something
very weak. This only really matters for things like DSA, but
someday maybe we'll have some other decent context sensitive
analyses :)
5. This reimplements the guts of memdep to handle the new results.
6. This simplifies GVN significantly:
a) readonly call CSE is slightly simpler
b) I eliminated the "getDependencyFrom" chaining for load
elimination and load CSE doesn't have to worry about
volatile (they are always clobbers) anymore.
c) GVN no longer does any 'lastLoad' caching, leaving it to
memdep.
7. The logic in DSE is simplified a bit and sped up. A potentially
unsafe case was eliminated.
llvm-svn: 60607
This fixes many bugs. I will add more test cases in a separate check-in.
Some day, the code that manipulates CFG and updates dom. info could use refactoring help.
llvm-svn: 60554
1) have it fold "br undef", which does occur with
surprising frequency as jump threading iterates.
2) teach j-t to delete dead blocks. This removes the successor
edges, reducing the in-edges of other blocks, allowing
recursive simplification.
3) Fold things like:
br COND, BBX, BBY
BBX:
br COND, BBZ, BBW
which also happens because jump threading iterates.
llvm-svn: 60470
straight-forward implementation. This does not require any extra
alias analysis queries beyond what we already do for non-local loads.
Some programs really really like load PRE. For example, SPASS triggers
this ~1000 times, ~300 times in 255.vortex, and ~1500 times on 403.gcc.
The biggest limitation to the implementation is that it does not split
critical edges. This is a huge killer on many programs and should be
addressed after the initial patch is enabled by default.
The implementation of this should incidentally speed up rejection of
non-local loads because it avoids creating the repl densemap in cases
when it won't be used for fully redundant loads.
This is currently disabled by default.
Before I turn this on, I need to fix a couple of miscompilations in
the testsuite, look at compile time performance numbers, and look at
perf impact. This is pretty close to ready though.
llvm-svn: 60408
constant. If X is a constant, then this is folded elsewhere.
- Added a note to Target/README.txt to indicate that we'd like to implement
this when we're able.
llvm-svn: 60399
figuring out the base of the IV. This produces better
code in the example. (Addresses use (IV) instead of
(BASE,IV) - a significant improvement on low-register
machines like x86).
llvm-svn: 60374
instead of std::sort. This shrinks the release-asserts LSR.o file
by 1100 bytes of code on my system.
We should start using array_pod_sort where possible.
llvm-svn: 60335
buggy rewrite, this notifies ScalarEvolution of a pending instruction
about to be removed and then erases it, instead of erasing it then
notifying.
llvm-svn: 60329
new instructions it simplifies. Because we're threading jumps on edges
with constants coming in from PHI's, we inherently are exposing a lot more
constants to the new block. Folding them and deleting dead conditions
allows the cost model in jump threading to be more accurate as it iterates.
llvm-svn: 60327
elimination: when finding dependent load/stores, realize that
they are the same if aliasing claims must alias instead of relying
on the pointers to be exactly equal. This makes load elimination
more aggressive. For example, on 403.gcc, we had:
< 68 gvn - Number of instructions PRE'd
< 152718 gvn - Number of instructions deleted
< 49699 gvn - Number of loads deleted
< 6153 memdep - Number of dirty cached non-local responses
< 169336 memdep - Number of fully cached non-local responses
< 162428 memdep - Number of uncached non-local responses
now we have:
> 64 gvn - Number of instructions PRE'd
> 153623 gvn - Number of instructions deleted
> 49856 gvn - Number of loads deleted
> 5022 memdep - Number of dirty cached non-local responses
> 159030 memdep - Number of fully cached non-local responses
> 162443 memdep - Number of uncached non-local responses
That's an extra 157 loads deleted and extra 905 other instructions nuked.
This slows down GVN very slightly, from 3.91 to 3.96s.
llvm-svn: 60314
vector instead of a densemap. This shrinks the memory usage of this thing
substantially (the high water mark) as well as making operations like
scanning it faster. This speeds up memdep slightly, gvn goes from
3.9376 to 3.9118s on 403.gcc
This also splits out the statistics for the cached non-local case to
differentiate between the dirty and clean cached case. Here's the stats
for 403.gcc:
6153 memdep - Number of dirty cached non-local responses
169336 memdep - Number of fully cached non-local responses
162428 memdep - Number of uncached non-local responses
yay for caching :)
llvm-svn: 60313
Note that the FoldOpIntoPhi call is dead because it's impossible for the
first operand of a subtraction to be both a ConstantInt and a PHINode.
llvm-svn: 60306
"For signed integers, the determination of overflow of x*y is not so simple. If
x and y have the same sign, then overflow occurs iff xy > 2**31 - 1. If they
have opposite signs, then overflow occurs iff xy < -2**31."
In this case, x == -1.
llvm-svn: 60278
overflowed on negation. This commit checks to make sure that neithe C nor X
overflows. This requires that the RHS of X (a subtract instruction) be a
constant integer.
llvm-svn: 60275
Owen Anderson