Currently, this pass only focuses on *trivial* loop unswitching. At that
reduced problem it remains significantly better than the current loop
unswitch:
- Old pass is worse than cubic complexity. New pass is (I think) linear.
- New pass is much simpler in its design by focusing on full unswitching. (See
below for details on this).
- New pass doesn't carry state for thresholds between pass iterations.
- New pass doesn't carry state for correctness (both miscompile and
infloop) between pass iterations.
- New pass produces substantially better code after unswitching.
- New pass can handle more trivial unswitch cases.
- New pass doesn't recompute the dominator tree for the entire function
and instead incrementally updates it.
I've ported all of the trivial unswitching test cases from the old pass
to the new one to make sure that major functionality isn't lost in the
process. For several of the test cases I've worked to improve the
precision and rigor of the CHECKs, but for many I've just updated them
to handle the new IR produced.
My initial motivation was the fact that the old pass carried state in
very unreliable ways between pass iterations, and these mechansims were
incompatible with the new pass manager. However, I discovered many more
improvements to make along the way.
This pass makes two very significant assumptions that enable most of these
improvements:
1) Focus on *full* unswitching -- that is, completely removing whatever
control flow construct is being unswitched from the loop. In the case
of trivial unswitching, this means removing the trivial (exiting)
edge. In non-trivial unswitching, this means removing the branch or
switch itself. This is in opposition to *partial* unswitching where
some part of the unswitched control flow remains in the loop. Partial
unswitching only really applies to switches and to folded branches.
These are very similar to full unrolling and partial unrolling. The
full form is an effective canonicalization, the partial form needs
a complex cost model, cannot be iterated, isn't canonicalizing, and
should be a separate pass that runs very late (much like unrolling).
2) Leverage LLVM's Loop machinery to the fullest. The original unswitch
dates from a time when a great deal of LLVM's loop infrastructure was
missing, ineffective, and/or unreliable. As a consequence, a lot of
complexity was added which we no longer need.
With these two overarching principles, I think we can build a fast and
effective unswitcher that fits in well in the new PM and in the
canonicalization pipeline. Some of the remaining functionality around
partial unswitching may not be relevant today (not many test cases or
benchmarks I can find) but if they are I'd like to add support for them
as a separate layer that runs very late in the pipeline.
Purely to make reviewing and introducing this code more manageable, I've
split this into first a trivial-unswitch-only pass and in the next patch
I'll add support for full non-trivial unswitching against a *fixed*
threshold, exactly like full unrolling. I even plan to re-use the
unrolling thresholds, as these are incredibly similar cost tradeoffs:
we're cloning a loop body in order to end up with simplified control
flow. We should only do that when the total growth is reasonably small.
One of the biggest changes with this pass compared to the previous one
is that previously, each individual trivial exiting edge from a switch
was unswitched separately as a branch. Now, we unswitch the entire
switch at once, with cases going to the various destinations. This lets
us unswitch multiple exiting edges in a single operation and also avoids
numerous extremely bad behaviors, where we would introduce 1000s of
branches to test for thousands of possible values, all of which would
take the exact same exit path bypassing the loop. Now we will use
a switch with 1000s of cases that can be efficiently lowered into
a jumptable. This avoids relying on somehow forming a switch out of the
branches or getting horrible code if that fails for any reason.
Another significant change is that this pass actively updates the CFG
based on unswitching. For trivial unswitching, this is actually very
easy because of the definition of loop simplified form. Doing this makes
the code coming out of loop unswitch dramatically more friendly. We
still should run loop-simplifycfg (at the least) after this to clean up,
but it will have to do a lot less work.
Finally, this pass makes much fewer attempts to simplify instructions
based on the unswitch. Something like loop-instsimplify, instcombine, or
GVN can be used to do increasingly powerful simplifications based on the
now dominating predicate. The old simplifications are things that
something like loop-instsimplify should get today or a very, very basic
loop-instcombine could get. Keeping that logic separate is a big
simplifying technique.
Most of the code in this pass that isn't in the old one has to do with
achieving specific goals:
- Updating the dominator tree as we go
- Unswitching all cases in a switch in a single step.
I think it is still shorter than just the trivial unswitching code in
the old pass despite having this functionality.
Differential Revision: https://reviews.llvm.org/D32409
llvm-svn: 301576
Just calling dropAllReferences leaves pointers to the ConstantExpr
behind, so we would eventually crash with a null pointer dereference.
Differential Revision: https://reviews.llvm.org/D32551
llvm-svn: 301575
also a discussion about exactly what we should do prior to re-enabling
it.
The current bug is http://llvm.org/PR32821 and the discussion about this
is in the review thread for r300200.
llvm-svn: 301505
This patch is part of D28975's breakdown.
induction.ll encodes the specific (and rather arbitrary) numbers given to
predicated basic blocks by the unique naming mechanism, which makes it
sensitive to changes in LV's instruction generation order. This patch replaces
those specific numbers with a numeric pattern.
Differential Revision: https://reviews.llvm.org/D32404
llvm-svn: 301345
The code Sanjay Patel moved over from InstCombine doesn't work properly if the 'and' has both inputs as nots because we used a commuted op matcher on the 'and' first. But this will bind to the first 'not' on 'and' when there could be two 'not's. InstCombine could rely on DeMorgan to ensure the 'and' wouldn't have two 'not's eventually, but InstSimplify can't rely on that.
This patch matches the xor first then checks for the ands and allows a not of either operand of the xor.
Differential Revision: https://reviews.llvm.org/D32458
llvm-svn: 301329
The matching here wasn't able to handle all the possible commutes. It always assumed the not would be on the left of the xor, but that's not guaranteed.
Differential Revision: https://reviews.llvm.org/D32474
llvm-svn: 301316
One of the fast-math optimizations is to replace calls to standard double
functions with their float equivalents, e.g. exp -> expf. However, this can
cause infinite loops for the following:
float expf(float val) { return (float) exp((double) val); }
A similar inline declaration exists in the MinGW-w64 math.h header file which
when compiled with -O2/3 and fast-math generates infinite loops.
So this fix checks that the calling function to the standard double function
that is being replaced does not match the float equivalent.
Differential Revision: https://reviews.llvm.org/D31806
llvm-svn: 301304
We need to do this to prevent a miscompile which sinks an objc_retain
past an objc_release that releases the object objc_retain retains. This
happens because the top-down and bottom-up traversals each determines
the insert point for retain or release individually without knowing
where the other instruction is moved.
For example, when the following IR is fed to the ARC optimizer, the
top-down traversal decides to insert objc_retain right before
objc_release and the bottom-up traversal decides to insert objc_release
right after clang.arc.use.
(IR before ARC optimizer)
%11 = call i8* @objc_retain(i8* %10)
call void (...) @clang.arc.use(%0* %5)
call void @llvm.dbg.value(...)
call void @objc_release(i8* %6)
This reverses the order of objc_release and objc_retain, which causes
the object to be destructed prematurely.
(IR after ARC optimizer)
call void (...) @clang.arc.use(%0* %5)
call void @objc_release(i8* %6)
call void @llvm.dbg.value(...)
%11 = call i8* @objc_retain(i8* %10)
rdar://problem/30530580
llvm-svn: 301289
We can simplify (and (icmp X, C1), (icmp X, C2)) to one of the icmps in many cases.
I had to check some of these with Alive to prove to myself it's right, but everything
seems to check out. Eg, the code in instcombine was completely ignoring predicates with
mismatched signedness.
Handling or-of-icmps would be a follow-up step.
Differential Revision: https://reviews.llvm.org/D32143
llvm-svn: 301260
Summary:
Ensure that the new merge BB (which contains the rest of the original BB
after the mem op being optimized) gets a profile frequency, in case
there are additional mem ops later in the BB. Otherwise they get skipped
as the merge BB looks cold.
Reviewers: davidxl, xur
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D32447
llvm-svn: 301244
This reverts commit r300732. This breaks a few tests.
I think the problem is related to adding more uses of
the condition that don't yet exist at this point.
llvm-svn: 301242
The current Loop Unroll implementation works with loops having a
single latch that contains a conditional branch to a block outside
the loop (the other successor is, by defition of latch, the header).
If this precondition doesn't hold, avoid unrolling the loop as
the code is not ready to handle such circumstances.
Differential Revision: https://reviews.llvm.org/D32261
llvm-svn: 301239
Summary:
llvm.invariant.group.barrier returns pointer that mustalias
pointer it takes. It can't be marked with `returned` attribute,
because it would be remove easily. The other reason is that
only Alias Analysis can know about this, because if any other
pass would know it, then the result would be replaced with it's
argument, which would be invalid.
We can think about returned pointer as something that mustalias, but
it doesn't have to be bitwise the same as the argument.
Reviewers: dberlin, chandlerc, hfinkel, sanjoy
Subscribers: reames, nlewycky, rsmith, anna, amharc
Differential Revision: https://reviews.llvm.org/D31585
llvm-svn: 301227
This is a straight cut and paste, but there's a bigger problem: if this
fold exists for simplifyOr, there should be a DeMorganized version for
simplifyAnd. But more than that, we have a patchwork of ad hoc logic
optimizations in InstCombine. There should be some structure to ensure
that we're not missing sibling folds across and/or/xor.
llvm-svn: 301213
When the location description of a source variable involves arithmetic
on the value itself, it needs to be marked with DW_OP_stack_value since it
is not describing the variable's location, but rather its value.
This is a follow-up to r297971 and fixes the source testcase quoted in
the comment in debuginfo-dce.ll.
rdar://problem/30725338
This reapplies r301093 without modifications.
llvm-svn: 301210
There is logic to track the expected number of instructions
produced. It thought in this case an instruction would
be necessary to negate the result, but here it folded
into a ConstantExpr fneg when the non-undef value operand
was cancelled out by the second fsub.
I'm not sure why we don't fold constant FP ops with undef currently,
but I think that would also avoid this problem.
llvm-svn: 301199
Summary:
Instead of keeping a variable indicating whether there are early exits
in the loop. We keep all the early exits. This improves LICM's ability to
move instructions out of the loop based on is-guaranteed-to-execute.
I am going to update compilation time as well soon.
Reviewers: hfinkel, sanjoy, efriedma, mkuper
Reviewed By: hfinkel
Subscribers: llvm-commits, mzolotukhin
Differential Revision: https://reviews.llvm.org/D32433
llvm-svn: 301196
Summary:
The return value of these intrinsics should always have 0 bits for
inactive threads. This means that when all arguments are constant
and the comparison evaluates to true, the intrinsic should return
the current exec mask.
Fixes some GL_ARB_shader_ballot tests.
Reviewers: arsenm
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D32344
llvm-svn: 301195
We handled all of the commuted variants for plain xor already,
although they were scattered around and sometimes folded less
efficiently using distributive laws. We had no folds for not-xor.
Handling all of these patterns consistently is part of trying to
reinstate:
https://reviews.llvm.org/rL300977
llvm-svn: 301144
Summary:
In case all predecessor go to a single successor of current BB. We want to fold (not thread).
I failed to update the phi nodes properly in the last patch https://reviews.llvm.org/rL300657.
Phi nodes values are per predecessor in LLVM.
Reviewers: sanjoy
Reviewed By: sanjoy
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D32400
llvm-svn: 301139
There's probably some better way to write this that eliminates the
code duplication without hurting readability, but at least this
eliminates the logic holes and is hopefully slightly more efficient
than creating new instructions.
llvm-svn: 301129
When the location description of a source variable involves arithmetic
on the value itself, it needs to be marked with DW_OP_stack_value since it
is not describing the variable's location, but rather its value.
This is a follow-up to r297971 and fixes the source testcase quoted in
the comment in debuginfo-dce.ll.
rdar://problem/30725338
llvm-svn: 301093
The bug was introduced by r301018 "[InstCombine] fadd double (sitofp x), y check that the promotion is valid". The patch didn't expect that fadd can be on vectors not necessarily scalars. Add vector support along with the test.
llvm-svn: 301070
Fixes leaving intermediate flat addressing computations
where a GEP instruction's source is a constant expression.
Still leaves behind a trivial addrspacecast + gep pair that
instcombine is able to handle, which ideally could be folded
here directly.
llvm-svn: 301044
Doing these transformations check that the result of integer addition is representable in the FP type.
(fadd double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
(fadd double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
This is a fix for https://bugs.llvm.org//show_bug.cgi?id=27036
Reviewed By: andrew.w.kaylor, scanon, spatel
Differential Revision: https://reviews.llvm.org/D31182
llvm-svn: 301018
Currently we choose PostBB as the single successor of QFB, but its possible that QTB's single successor is QFB which would make QFB the correct choice.
Differential Revision: https://reviews.llvm.org/D32323
llvm-svn: 300992
Sanjay Patel