was centered around the premise of laying out a loop in a chain, and
then rotating that chain. This is good for preserving contiguous layout,
but bad for actually making sane rotations. In order to keep it safe,
I had to essentially make it impossible to rotate deeply nested loops.
The information needed to correctly reason about a deeply nested loop is
actually available -- *before* we layout the loop. We know the inner
loops are already fused into chains, etc. We lose information the moment
we actually lay out the loop.
The solution was the other alternative for this algorithm I discussed
with Benjamin and some others: rather than rotating the loop
after-the-fact, try to pick a profitable starting block for the loop's
layout, and then use our existing layout logic. I was worried about the
complexity of this "pick" step, but it turns out such complexity is
needed to handle all the important cases I keep teasing out of benchmarks.
This is, I'm afraid, a bit of a work-in-progress. It is still
misbehaving on some likely important cases I'm investigating in Olden.
It also isn't really tested. I'm going to try to craft some interesting
nested-loop test cases, but it's likely to be extremely time consuming
and I don't want to go there until I'm sure I'm testing the correct
behavior. Sadly I can't come up with a way of getting simple, fine
grained test cases for this logic. We need complex loop structures to
even trigger much of it.
llvm-svn: 145183
heavily on AnalyzeBranch. That routine doesn't behave as we want given
that rotation occurs mid-way through re-ordering the function. Instead
merely check that there are not unanalyzable branching constructs
present, and then reason about the CFG via successor lists. This
actually simplifies my mental model for all of this as well.
The concrete result is that we now will rotate more loop chains. I've
added a test case from Olden highlighting the effect. There is still
a bit more to do here though in order to regain all of the performance
in Olden.
llvm-svn: 145179
pass. This is designed to achieve one of the important optimizations
that the old code placement pass did, but more simply.
This is a somewhat rough and *very* conservative version of the
transform. We could get a lot fancier here if there are profitable cases
to do so. In particular, this only looks for a single pattern, it
insists that the loop backedge being rotated away is the last backedge
in the chain, and it doesn't provide any means of doing better in-loop
placement due to the rotation. However, it appears that it will handle
the important loops I am finding in the LLVM test suite.
llvm-svn: 145158
was returning incorrect values in rare cases, and incorrectly marking
exact conversions as inexact in some more common cases. Fixes PR11406, and a
missed optimization in test/CodeGen/X86/fp-stack-O0.ll.
llvm-svn: 145141
need lots of fanciness around retaining a reference to a Chain's slot in
the BlockToChain map, but that's all gone now. We can just go directly
to allocating the new chain (which will update the mapping for us) and
using it.
Somewhat gross mechanically generated test case replicates the issue
Duncan spotted when actually testing this out.
llvm-svn: 145120
conflicts, we should only be adding the first block of the chain to the
list, lest we try to merge into the middle of that chain. Most of the
places we were doing this we already happened to be looking at the first
block, but there is no reason to assume that, and in some cases it was
clearly wrong.
I've added a couple of tests here. One already worked, but I like having
an explicit test for it. The other is reduced from a test case Duncan
reduced for me and used to crash. Now it is handled correctly.
llvm-svn: 145119
and positive: positive, because it could be directly computed to be positive;
negative, because the nsw flags means it is either negative or undefined (the
multiplication always overflowed).
llvm-svn: 145104
Before:
movabsq $4294967296, %rax ## encoding: [0x48,0xb8,0x00,0x00,0x00,0x00,0x01,0x00,0x00,0x00]
testq %rax, %rdi ## encoding: [0x48,0x85,0xf8]
jne LBB0_2 ## encoding: [0x75,A]
After:
btq $32, %rdi ## encoding: [0x48,0x0f,0xba,0xe7,0x20]
jb LBB0_2 ## encoding: [0x72,A]
btq is usually slower than testq because it doesn't fuse with the jump, but here we're better off
saving one register and a giant movabsq.
llvm-svn: 145103
further. This invariant just wasn't going to work in the face of
unanalyzable branches; we need to be resillient to the phenomenon of
chains poking into a loop and poking out of a loop. In fact, we already
were, we just needed to not assert on it.
This was found during a bootstrap with block placement turned on.
llvm-svn: 145100
successors, they just are all landing pad successors. We handle this the
same way as no successors. Comments attached for the next person to wade
through here and another lovely test case courtesy of Benjamin Kramer's
bugpoint reduction.
llvm-svn: 145098
This was a bug in keeping track of the available domains when merging
domain values.
The wrong domain mask caused ExecutionDepsFix to try to move VANDPSYrr
to the integer domain which is only available in AVX2.
Also add an assertion to catch future attempts at emitting AVX2
instructions.
llvm-svn: 145096
reversed in the function's original ordering, and we happened to
encounter it while handling an outer unnatural CFG structure.
Thanks to the test case reduced from GCC's source by Benjamin Kramer.
This may also fix a crasher in gzip that Duncan reduced for me, but
I haven't yet gotten to testing that one.
llvm-svn: 145094
updateTerminator code didn't correctly handle EH terminators in one very
specific case. AnalyzeBranch would find no terminator instruction, and
so the fallback in updateTerminator is to assume fallthrough. This is
correct, but the destination of the fallthrough was assumed to be the
first successor.
This is *almost always* true, but in certain cases the loop
transformations will cause the landing pad to be the first successor!
Instead of this brittle logic, actually look through the successors for
a non-landing-pad accessor, and to assert if more than one is found.
This will hopefully fix some (if not all) of the self host miscompiles
with block placement. Thanks to Benjamin Kramer for reporting, Nick
Lewycky for an initial stab at a reduction, and Duncan for endless
advice on EH (which I know nothing about) as well as reviewing the
actual fix.
llvm-svn: 145062
properly account for the *global* probability of the edge being taken.
This manifested as a very large number of unconditional branches to
blocks being merged against the CFG even though they weren't
particularly hot within the CFG.
The fix is to check whether the edge being merged is both locally hot
relative to other successors for the source block, and globally hot
compared to other (unmerged) predecessors of the destination block.
This introduces a new crasher on GCC single-source, but it's currently
behind a flag, and Ben has offered to work on the reduction. =]
llvm-svn: 145010
is actually being tested. Also add some FileCheck goodness to much more
carefully ensure that the result is the desired result. Before this test
would only have failed through an assert failure if the underlying fix
were reverted.
Also, add some weight metadata and a comment explaining exactly what is
going on to a trick section of the test case. Originally, we were
getting very unlucky and trying to form a block chain that isn't
actually profitable. I'm working on a fix to avoid forming these
unprofitable chains, and that would also have masked any failure from
this test case. The easy solution is to add some metadata that makes it
*really* profitable to form the bad chain here.
llvm-svn: 145006
formation phase and into the initial walk of the basic blocks. We
essentially pre-merge all blocks where unanalyzable fallthrough exists,
as we won't be able to update the terminators effectively after any
reorderings. This is quite a bit more principled as there may be CFGs
where the second half of the unanalyzable pair has some analyzable
predecessor that gets placed first. Then it may get placed next,
implicitly breaking the unanalyzable branch even though we never even
looked at the part that isn't analyzable. I've included a test case that
triggers this (thanks Benjamin yet again!), and I'm hoping to synthesize
some more general ones as I dig into related issues.
Also, to make this new scheme work we have to be able to handle branches
into the middle of a chain, so add this check. We always fallback on the
incoming ordering.
Finally, this starts to really underscore a known limitation of the
current implementation -- we don't consider broken predecessors when
merging successors. This can caused major missed opportunities, and is
something I'm planning on looking at next (modulo more bug reports).
llvm-svn: 144994
The loop tree's inclusive block lists are painful and expensive to
update. (I have no idea why they're inclusive). The design was
supposed to handle this case but the implementation missed it and my
unit tests weren't thorough enough.
Fixes PR11335: loop unroll update.
llvm-svn: 144970
The right way to check for a binary operation is
cast<BinaryOperator>. The original check: cast<Instruction> &&
numOperands() == 2 would match phi "instructions", leading to an
infinite loop in extreme corner case: a useless phi with operands
[self, constant] that prior optimization passes failed to remove,
being used in the loop by another useless phi, in turn being used by an
lshr or udiv.
Fixes PR11350: runaway iteration assertion.
llvm-svn: 144935
ADDs. MaxOffs is used as a threshold to limit the size of the offset. Tradeoffs
being: (1) If we can't materialize the large constant then we'll cause fast-isel
to bail. (2) Too large of an offset can't be directly encoded in the ADD
resulting in a MOV+ADD. Generally not a bad thing because otherwise we would
have had ADD+ADD, but on Thumb this turns into a MOVS+MOVT+ADD. Working on a fix
for that. (3) Conversely, too low of a threshold we'll miss opportunities to
coalesce ADDs.
rdar://10412592
llvm-svn: 144886
We don't (yet) have the granularity in the fixups to be specific about which
bitranges are affected. That's a future cleanup, but we're not there yet.
llvm-svn: 144852
For example,
vld1.f64 {d2-d5}, [r2,:128]!
Should be equivalent to:
vld1.f64 {d2,d3,d4,d5}, [r2,:128]!
It's not documented syntax in the ARM ARM, but it is consistent with what's
accepted for VLDM/VSTM and is unambiguous in meaning, so it's a good thing to
support.
rdar://10451128
llvm-svn: 144727
When the 3rd operand is not a low-register, and the first two operands are
the same low register, the parser was incorrectly trying to use the 16-bit
instruction encoding.
rdar://10449281
llvm-svn: 144679
has a reference to it. Unfortunately, that doesn't work for codegen passes
since we don't get notified of MBB's being deleted (the original BB stays).
Use that fact to our advantage and after printing a function, check if
any of the IL BBs corresponds to a symbol that was not printed. This fixes
pr11202.
llvm-svn: 144674
block sequence when recovering from unanalyzable control flow
constructs, *always* use the function sequence. I'm not sure why I ever
went down the path of trying to use the loop sequence, it is
fundamentally not the correct sequence to use. We're trying to preserve
the incoming layout in the cases of unreasonable control flow, and that
is only encoded at the function level. We already have a filter to
select *exactly* the sub-set of blocks within the function that we're
trying to form into a chain.
The resulting code layout is also significantly better because of this.
In several places we were ending up with completely unreasonable control
flow constructs due to the ordering chosen by the loop structure for its
internal storage. This change removes a completely wasteful vector of
basic blocks, saving memory allocation in the common case even though it
costs us CPU in the fairly rare case of unnatural loops. Finally, it
fixes the latest crasher reduced out of GCC's single source. Thanks
again to Benjamin Kramer for the reduction, my bugpoint skills failed at
it.
llvm-svn: 144627
Two new TargetInstrInfo hooks lets the target tell ExecutionDepsFix
about instructions with partial register updates causing false unwanted
dependencies.
The ExecutionDepsFix pass will break the false dependencies if the
updated register was written in the previoius N instructions.
The small loop added to sse-domains.ll runs twice as fast with
dependency-breaking instructions inserted.
llvm-svn: 144602
and stores capture) to permit the caller to see each capture point and decide
whether to continue looking.
Use this inside memdep to do an analysis that basicaa won't do. This lets us
solve another devirtualization case, fixing PR8908!
llvm-svn: 144580
instructions of the two-address operands) in order to avoid inserting copies.
This fixes the few regressions introduced when the two-address hack was
disabled (without regressing the improvements).
rdar://10422688
llvm-svn: 144559
the sum of the edge weights not overflowing uint32, and crashed when
they did. This is generally safe as BranchProbabilityInfo tries to
provide this guarantee. However, the CFG can get modified during codegen
in a way that grows the *sum* of the edge weights. This doesn't seem
unreasonable (imagine just adding more blocks all with the default
weight of 16), but it is hard to come up with a case that actually
triggers 32-bit overflow. Fortuately, the single-source GCC build is
good at this. The solution isn't very pretty, but its no worse than the
previous code. We're already summing all of the edge weights on each
query, we can sum them, check for an overflow, compute a scale, and sum
them again.
I've included a *greatly* reduced test case out of the GCC source that
triggers it. It's a pretty lame test, as it clearly is just barely
triggering the overflow. I'd like to have something that is much more
definitive, but I don't understand the fundamental pattern that triggers
an explosion in the edge weight sums.
The buggy code is duplicated within this file. I'll colapse them into
a single implementation in a subsequent commit.
llvm-svn: 144526
