The ARM/ARM64 AESE and AESD instructions have a builtin XOR as the first step in
the instruction. Therefore, if the AES key is zero and the AES data was
previously XORed, it can be combined into a single instruction.
Differential Revision: https://reviews.llvm.org/D47239
Patch by Michael Brase!
llvm-svn: 333193
Summary:
Finally fixes [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]].
Now that the backend is all done, we can finally fold it!
The canonical unfolded masked merge pattern is
```(x & m) | (y & ~m)```
There is a second, equivalent variant:
```(x | ~m) & (y | m)```
Only one of them (the or-of-and's i think) is canonical.
And if the mask is not a constant, we should fold it to:
```((x ^ y) & M) ^ y```
https://rise4fun.com/Alive/ndQw
Reviewers: spatel, craig.topper
Reviewed By: spatel
Subscribers: nicholas, RKSimon, llvm-commits
Differential Revision: https://reviews.llvm.org/D46814
llvm-svn: 333106
We can eliminate old value if bound_ctrl = 1 and row_mask = bank_mask = 0xf.
This is alternative implementation working with the intrinsic in InstCombine.
Original review for past-ISel optimization: D46570.
Differential Revision: https://reviews.llvm.org/D46596
llvm-svn: 332956
Summary:
This patch fixes PR37526 by simplifying the newly generated LoadInst
instructions. If the pointer address is a bitcast from the pointer to
the NewType, we can just remove this extra bitcast instead of creating
the new one. This fixes the PR37526 + may speed up the whole compilation
process.
Reviewers: spatel, RKSimon, hfinkel
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D47144
llvm-svn: 332855
We already do this for min/max (see the blob above the diff),
so we should do the same for abs/nabs.
A sign-bit check (<s 0) is used as a predicate for other IR
transforms and it's likely the best for codegen.
This might solve the motivating cases for D47037 and D47041,
but I think those patches still make sense. We can't guarantee
this canonicalization if the icmp has more than one use.
Differential Revision: https://reviews.llvm.org/D47076
llvm-svn: 332819
Summary:
- Add wasm personality function
- Re-categorize the existing `isFuncletEHPersonality()` function into
two different functions: `isFuncletEHPersonality()` and
`isScopedEHPersonality(). This becomes necessary as wasm EH uses scoped
EH instructions (catchswitch, catchpad/ret, and cleanuppad/ret) but not
outlined funclets.
- Changed some callsites of `isFuncletEHPersonality()` to
`isScopedEHPersonality()` if they are related to scoped EH IR-level
stuff.
Reviewers: majnemer, dschuff, rnk
Subscribers: jfb, sbc100, jgravelle-google, eraman, JDevlieghere, sunfish, llvm-commits
Differential Revision: https://reviews.llvm.org/D45559
llvm-svn: 332667
According to alive this is valid. I'm hoping to use this to make an assumption that the sign bit is zero after this sequence. The only way it wouldn't be is if the input was INT__MIN, but by preserving the flags we can make doing this to INT_MIN UB.
The nuw flags is weird because it creates such a contradiction that the original number would have to be positive meaning we could remove the select entirely, but we don't get that far.
Differential Revision: https://reviews.llvm.org/D46988
llvm-svn: 332623
The canonicalization was restricted to shuffle masks with
a 1-to-1 mapping to the constant vector, but that disqualifies
the common splat pattern. This is part of solving PR37463:
https://bugs.llvm.org/show_bug.cgi?id=37463
llvm-svn: 332479
Summary:
Part of the InstCombine code for simplifying GEPs looks through
addrspacecasts. However, this was done by updating a variable
also used by the next transformation, for marking GEPs as
inbounds. This led to replacing a GEP with a similar instruction
in a different addrspace, which caused an assertion failure in RAUW.
This caused julia issue https://github.com/JuliaLang/julia/issues/27055
Patch by Jeff Bezanson <jeff@juliacomputing.com>
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D46722
llvm-svn: 332302
The DEBUG() macro is very generic so it might clash with other projects.
The renaming was done as follows:
- git grep -l 'DEBUG' | xargs sed -i 's/\bDEBUG\s\?(/LLVM_DEBUG(/g'
- git diff -U0 master | ../clang/tools/clang-format/clang-format-diff.py -i -p1 -style LLVM
- Manual change to APInt
- Manually chage DOCS as regex doesn't match it.
In the transition period the DEBUG() macro is still present and aliased
to the LLVM_DEBUG() one.
Differential Revision: https://reviews.llvm.org/D43624
llvm-svn: 332240
Summary:
This change adds handling of the atomic memset intrinsic to the
code path that simplifies the regular memset. In practice this means
that we will now also expand a small constant-length atomic memset
into a single unordered atomic store.
Reviewers: apilipenko, skatkov, mkazantsev, anna, reames
Reviewed By: reames
Subscribers: reames, llvm-commits
Differential Revision: https://reviews.llvm.org/D46660
llvm-svn: 332132
Summary:
This change reworks the handling of atomic memcpy within the instcombine pass.
Previously, a constant length atomic memcpy would be lowered into loads & stores
as long as no more than 16 load/store pairs are created. This is quite different
from the lowering done for a non-atomic memcpy; which only ever lowers into a single
load/store pair of no more than 8 bytes. Larger constant-sized memcpy calls are
expanded to load/stores in later passes, such as SelectionDAG lowering.
In this change the behaviour for atomic memcpy is unified with non-atomic memcpy;
atomic memcpy is now treated in the same was as non-atomic memcpy has always been.
We leave it to later passes to lower longer-length atomic memcpy calls.
Due to the structure of the pass's handling of memtransfer intrinsics, this change
also gives us handling of atomic memmove that we did not previously have.
Reviewers: apilipenko, skatkov, mkazantsev, anna, reames
Reviewed By: reames
Subscribers: reames, llvm-commits
Differential Revision: https://reviews.llvm.org/D46658
llvm-svn: 332093
The bitwidth of the operation should always be wider than the result width of the truncate since we don't recurse through any width changing operations.
llvm-svn: 332055
Put in a conservatively correct estimate for now. Avoids miscompiling
clang in FDO mode. This is really tricky to trigger in reality as
basically all interesting cases will be folded away by computeKnownBits
earlier, I was unable to find a reasonably small test case.
llvm-svn: 331975
The previous handling for guard widening in InstCombine was extremely restrictive. In particular, it didn't handle the common case where we had two guards separated by a single icmp. Handle this by scanning through a small fixed window of instructions to find the next guard if needed.
Differential Revision: https://reviews.llvm.org/D46203
llvm-svn: 331935
This is safe as long as the udiv is not exact. The pattern is not common in
C++ code, but comes up all the time in code generated by XLA's GPU backend.
Differential Revision: https://reviews.llvm.org/D46647
llvm-svn: 331933
Inspired by r331508, I did a grep and found these.
Mostly just change from dyn_cast to cast. Some cases also showed a dyn_cast result being converted to bool, so those I changed to isa.
llvm-svn: 331577
Add logic for the special case when a cmp+select can clearly be
reduced to just a bitwise logic instruction, and remove an
over-reaching chunk of general purpose bit magic. The primary goal
is to remove cases where we are not improving the IR instruction
count when doing these select transforms, and in all cases here that
is true.
In the motivating 3-way compare tests, there are further improvements
because we can combine/propagate select values (not sure if that
belongs in instcombine, but it's there for now).
DAGCombiner has folds to turn some of these selects into bit magic,
so there should be no difference in the end result in those cases.
Not all constant combinations are handled there yet, however, so it
is possible that some targets will see more cmov/csel codegen with
this change in IR canonicalization.
Ideally, we'll go further to *not* turn selects into multiple
logic/math ops in instcombine, and we'll canonicalize to selects.
But we should make sure that this step does not result in regressions
first (and if it does, we should fix those in the backend).
The general direction for this change was discussed here:
http://lists.llvm.org/pipermail/llvm-dev/2016-September/105373.htmlhttp://lists.llvm.org/pipermail/llvm-dev/2017-July/114885.html
Alive proofs for the new bit magic:
https://rise4fun.com/Alive/XG7
Differential Revision: https://reviews.llvm.org/D46086
llvm-svn: 331486
This is a follow-up to r331272.
We've been running doxygen with the autobrief option for a couple of
years now. This makes the \brief markers into our comments
redundant. Since they are a visual distraction and we don't want to
encourage more \brief markers in new code either, this patch removes
them all.
Patch produced by
for i in $(git grep -l '\@brief'); do perl -pi -e 's/\@brief //g' $i & done
https://reviews.llvm.org/D46290
llvm-svn: 331275
We've been running doxygen with the autobrief option for a couple of
years now. This makes the \brief markers into our comments
redundant. Since they are a visual distraction and we don't want to
encourage more \brief markers in new code either, this patch removes
them all.
Patch produced by
for i in $(git grep -l '\\brief'); do perl -pi -e 's/\\brief //g' $i & done
Differential Revision: https://reviews.llvm.org/D46290
llvm-svn: 331272
Summary:
As discussed in D45733, we want to do this in InstCombine.
https://rise4fun.com/Alive/LGk
Reviewers: spatel, craig.topper
Reviewed By: spatel
Subscribers: chandlerc, xbolva00, llvm-commits
Differential Revision: https://reviews.llvm.org/D45867
llvm-svn: 331205
Summary:
Masked merge has a pattern of: `((x ^ y) & M) ^ y`.
But, there is no difference between `((x ^ y) & M) ^ y` and `((x ^ y) & ~M) ^ x`,
We should canonicalize the pattern to non-inverted mask.
https://rise4fun.com/Alive/Yol
Reviewers: spatel, craig.topper
Reviewed By: spatel
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D45664
llvm-svn: 331112
Summary:
Currently, we
1. match `LHS` matcher to the `first` operand of binary operator,
2. and then match `RHS` matcher to the `second` operand of binary operator.
If that does not match, we swap the `LHS` and `RHS` matchers:
1. match `RHS` matcher to the `first` operand of binary operator,
2. and then match `LHS` matcher to the `second` operand of binary operator.
This works ok.
But it complicates writing of commutative matchers, where one would like to match
(`m_Value()`) the value on one side, and use (`m_Specific()`) it on the other side.
This is additionally complicated by the fact that `m_Specific()` stores the `Value *`,
not `Value **`, so it won't work at all out of the box.
The last problem is trivially solved by adding a new `m_c_Specific()` that stores the
`Value **`, not `Value *`. I'm choosing to add a new matcher, not change the existing
one because i guess all the current users are ok with existing behavior,
and this additional pointer indirection may have performance drawbacks.
Also, i'm storing pointer, not reference, because for some mysterious-to-me reason
it did not work with the reference.
The first one appears trivial, too.
Currently, we
1. match `LHS` matcher to the `first` operand of binary operator,
2. and then match `RHS` matcher to the `second` operand of binary operator.
If that does not match, we swap the ~~`LHS` and `RHS` matchers~~ **operands**:
1. match ~~`RHS`~~ **`LHS`** matcher to the ~~`first`~~ **`second`** operand of binary operator,
2. and then match ~~`LHS`~~ **`RHS`** matcher to the ~~`second`~ **`first`** operand of binary operator.
Surprisingly, `$ ninja check-llvm` still passes with this.
But i expect the bots will disagree..
The motivational unittest is included.
I'd like to use this in D45664.
Reviewers: spatel, craig.topper, arsenm, RKSimon
Reviewed By: craig.topper
Subscribers: xbolva00, wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D45828
llvm-svn: 331085
Summary:
Simplify integer add expression X % C0 + (( X / C0 ) % C1) * C0 to
X % (C0 * C1). This is a common pattern seen in code generated by the XLA
GPU backend.
Add test cases for this new optimization.
Patch by Bixia Zheng!
Reviewers: sanjoy
Reviewed By: sanjoy
Subscribers: efriedma, craig.topper, lebedev.ri, llvm-commits, jlebar
Differential Revision: https://reviews.llvm.org/D45976
llvm-svn: 330992
As discussed in D45862, we want to delete parts of
this code because it can create more instructions
than it removes. But we also want to preserve some
folds that are winners, so tidy up what's here to
make splitting the good from bad a bit easier.
llvm-svn: 330841
(notionally Scalar.h is part of libLLVMScalarOpts, so it shouldn't be
included by InstCombine which doesn't/shouldn't need to depend on
ScalarOpts)
llvm-svn: 330669
This is the last step in getting constant pattern matchers to allow
undef elements in constant vectors.
I'm adding a dedicated m_ZeroInt() function and building m_Zero() from
that. In most cases, calling code can be updated to use m_ZeroInt()
directly when there's no need to match pointers, but I'm leaving that
efficiency optimization as a follow-up step because it's not always
clear when that's ok.
There are just enough icmp folds in InstSimplify that can be used for
integer or pointer types, that we probably still want a generic m_Zero()
for those cases. Otherwise, we could eliminate it (and possibly add a
m_NullPtr() as an alias for isa<ConstantPointerNull>()).
We're conservatively returning a full zero vector (zeroinitializer) in
InstSimplify/InstCombine on some of these folds (see diffs in InstSimplify),
but I'm not sure if that's actually necessary in all cases. We may be
able to propagate an undef lane instead. One test where this happens is
marked with 'TODO'.
llvm-svn: 330550
Summary:
When sinking an instruction in InstCombine we now also sink
the DbgInfoIntrinsics that are using the sunken value.
Example)
When sinking the load in this input
bb.X:
%0 = load i64, i64* %start, align 4, !dbg !31
tail call void @llvm.dbg.value(metadata i64 %0, ...)
br i1 %cond, label %for.end, label %for.body.lr.ph
for.body.lr.ph:
br label %for.body
we now also move the dbg.value, like this
bb.X:
br i1 %cond, label %for.end, label %for.body.lr.ph
for.body.lr.ph:
%0 = load i64, i64* %start, align 4, !dbg !31
tail call void @llvm.dbg.value(metadata i64 %0, ...)
br label %for.body
In the past we haven't moved the dbg.value so we got
bb.X:
tail call void @llvm.dbg.value(metadata i64 %0, ...)
br i1 %cond, label %for.end, label %for.body.lr.ph
for.body.lr.ph:
%0 = load i64, i64* %start, align 4, !dbg !31
br label %for.body
So in the past we got a debug-use before the def of %0.
And that dbg.value was also on the path jumping to %for.end, for
which %0 never was defined.
CodeGenPrepare normally comes to rescue later (when not moving
the dbg.value), since it moves dbg.value instrinsics quite
brutally, without really analysing if it is correct to move
the intrinsic (see PR31878).
So at the moment this patch isn't expected to have much impact,
besides that it is moving the dbg.value already in opt, making
the IR look more sane directly.
This can be seen as a preparation to (hopefully) make it possible
to turn off CodeGenPrepare::placeDbgValues later as a solution
to PR31878.
I also adjusted test/DebugInfo/X86/sdagsplit-1.ll to make the
IR in the test case up-to-date with this behavior in InstCombine.
Reviewers: rnk, vsk, aprantl
Reviewed By: vsk, aprantl
Subscribers: mattd, JDevlieghere, llvm-commits
Tags: #debug-info
Differential Revision: https://reviews.llvm.org/D45425
llvm-svn: 330243
The bitcast may be interfering with other combines or vectorization
as shown in PR16739:
https://bugs.llvm.org/show_bug.cgi?id=16739
Most pointer-related optimizations are probably able to look through
this bitcast, but removing the bitcast shrinks the IR, so it's at
least a size savings.
Differential Revision: https://reviews.llvm.org/D44833
llvm-svn: 330237
Two cleanups:
1. As noted in D45453, we had tests that don't need FMF that were misplaced in the 'fast-math.ll' test file.
2. This removes the final uses of dyn_castFNegVal, so that can be deleted. We use 'match' now.
llvm-svn: 330126
Summary:
In order to get the whole fold as specified in [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]],
let's first handle the simple straight-forward things.
Let's start with the `and` -> `or` simplification.
The one obvious thing missing here: the constant mask is not handled.
I have an idea how to handle it, but it will require some thinking,
and is not strictly required here, so i've left that for later.
https://rise4fun.com/Alive/Pkmg
Reviewers: spatel, craig.topper, eli.friedman, jingyue
Reviewed By: spatel
Subscribers: llvm-commits
Was reviewed as part of https://reviews.llvm.org/D45631
llvm-svn: 330103
These simplifications were previously enabled only with isFast(), but that
is more restrictive than required. Since r317488, FMF has 'reassoc' to
control these cases at a finer level.
llvm-svn: 330089
Summary:
The fold added in D45108 did not account for the fact that
the and instruction is commutative, and if the mask is a variable,
the mask variable and the fold variable may be swapped.
I have noticed this by accident when looking into [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]]
This extends/generalizes that fold, so it is handled too.
Reviewers: spatel, craig.topper
Reviewed By: spatel
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D45539
llvm-svn: 330001
This completes the work started in r329604 and r329605 when we changed clang to no longer use the intrinsics.
We lost some InstCombine SimplifyDemandedBit optimizations through this change as we aren't able to fold 'and', bitcast, shuffle very well.
llvm-svn: 329990
This restores what was lost with rL73243 but without
re-introducing the bug that was present in the old code.
Note that we already have these transforms if the ops are
marked 'fast' (and I assume that's happening somewhere in
the code added with rL170471), but we clearly don't need
all of 'fast' for these transforms.
llvm-svn: 329362
Summary:
This is a fix to PR37005.
Essentially, rL328539 ([InstCombine] reassociate loop invariant GEP chains to enable LICM) contains a bug
whereby it will convert:
%src = getelementptr inbounds i8, i8* %base, <2 x i64> %val
%res = getelementptr inbounds i8, <2 x i8*> %src, i64 %val2
into:
%src = getelementptr inbounds i8, i8* %base, i64 %val2
%res = getelementptr inbounds i8, <2 x i8*> %src, <2 x i64> %val
By swapping the index operands if the GEPs are in a loop, and %val is loop variant while %val2
is loop invariant.
This fix recreates new GEP instructions if the index operand swap would result in the type
of %src changing from vector to scalar, or vice versa.
Reviewers: sebpop, spatel
Reviewed By: sebpop
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D45287
llvm-svn: 329331
The tests marked with 'FIXME' require loosening the check
in SimplifyAssociativeOrCommutative() to optimize completely;
that's still checking isFast() in Instruction::isAssociative().
llvm-svn: 329121
Summary:
Folding patterns like:
%vec = shufflevector <4 x i8> %insvec, <4 x i8> undef, <4 x i32> zeroinitializer
%cast = bitcast <4 x i8> %vec to i32
%cond = icmp eq i32 %cast, 0
into:
%ext = extractelement <4 x i8> %insvec, i32 0
%cond = icmp eq i32 %ext, 0
Combined with existing rules, this allows us to fold patterns like:
%insvec = insertelement <4 x i8> undef, i8 %val, i32 0
%vec = shufflevector <4 x i8> %insvec, <4 x i8> undef, <4 x i32> zeroinitializer
%cast = bitcast <4 x i8> %vec to i32
%cond = icmp eq i32 %cast, 0
into:
%cond = icmp eq i8 %val, 0
When we construct a splat vector via a shuffle, and bitcast the vector into an integer type for comparison against an integer constant. Then we can simplify the the comparison to compare the splatted value against the integer constant.
Reviewers: spatel, anna, mkazantsev
Reviewed By: spatel
Subscribers: efriedma, rengolin, llvm-commits
Differential Revision: https://reviews.llvm.org/D44997
llvm-svn: 329087
Summary:
The cast simplifications that instcombine does here do not make any
attempt to obey the verifier rules for musttail calls. Therefore we have
to disable them.
Reviewers: efriedma, majnemer, pcc
Subscribers: hiraditya, llvm-commits
Differential Revision: https://reviews.llvm.org/D45186
llvm-svn: 329027
This change brings performance of zlib up by 10%. The example below is from a
hot loop in longest_match() from zlib.
do.body:
%cur_match.addr.0 = phi i32 [ %cur_match, %entry ], [ %2, %do.cond ]
%idx.ext = zext i32 %cur_match.addr.0 to i64
%add.ptr = getelementptr inbounds i8, i8* %win, i64 %idx.ext
%add.ptr2 = getelementptr inbounds i8, i8* %add.ptr, i64 %idx.ext1
%add.ptr3 = getelementptr inbounds i8, i8* %add.ptr2, i64 -1
In this example %idx.ext1 is a loop invariant. It will be moved above the use of
loop induction variable %idx.ext such that it can be hoisted out of the loop by
LICM. The operands that have dependences carried by the loop will be sinked down
in the GEP chain. This patch will produce the following output:
do.body:
%cur_match.addr.0 = phi i32 [ %cur_match, %entry ], [ %2, %do.cond ]
%idx.ext = zext i32 %cur_match.addr.0 to i64
%add.ptr = getelementptr inbounds i8, i8* %win, i64 %idx.ext1
%add.ptr2 = getelementptr inbounds i8, i8* %add.ptr, i64 -1
%add.ptr3 = getelementptr inbounds i8, i8* %add.ptr2, i64 %idx.ext
llvm-svn: 328539
This replaces a large chunk of code that was looking for compound
patterns that include these sub-patterns. Existing tests ensure that
all of the previous examples are still folded as expected.
We still need to loosen the FMF check.
llvm-svn: 328502
Remove #include of Transforms/Scalar.h from Transform/Utils to fix layering.
Transforms depends on Transforms/Utils, not the other way around. So
remove the header and the "createStripGCRelocatesPass" function
declaration (& definition) that is unused and motivated this dependency.
Move Transforms/Utils/Local.h into Analysis because it's used by
Analysis/MemoryBuiltins.cpp.
llvm-svn: 328165
This is complicated by -0.0 and nan. This is based on the DAG patterns
as shown in D44091. I'm hoping that we can just remove those DAG folds
and always rely on IR canonicalization to handle the matching to fabs.
We would still need to delete the broken code from DAGCombiner to fix
PR36600:
https://bugs.llvm.org/show_bug.cgi?id=36600
Differential Revision: https://reviews.llvm.org/D44550
llvm-svn: 327858
This was supposed to be an NFC refactoring that will eventually allow
eliminating the isFast() predicate, but there's a rare possibility
that we would pessimize the code as shown in the test case because
we failed to check 'hasOneUse()' properly. This version also removes
an inefficiency of the old code; we would look for:
(X * C) * C1 --> X * (C * C1)
...but that pattern is always handled by
SimplifyAssociativeOrCommutative().
llvm-svn: 327404
getNumUses is a linear time operation. It traverses the user linked list to the end and counts as it goes. Since we are only interested in small constant counts, we should use hasNUses or hasNUsesMore more that terminate the traversal as soon as it can provide the answer.
There are still two other locations in InstCombine, but changing those would force a rebase of D44266 which if accepted would remove them.
Differential Revision: https://reviews.llvm.org/D44398
llvm-svn: 327315
Most of the folds based on SelectPatternResult belong in InstSimplify rather than
InstCombine, so the helper code should be available to other passes/analysis.
llvm-svn: 326812
Summary:
Presently, InstCombiner::foldICmpWithCastAndCast() implicitly assumes that it is
only invoked with icmp instructions of integer type. If that assumption is broken,
and it is called with an icmp of vector type, then it fails (asserts/crashes).
This patch addresses the deficiency. It allows it to simplify
icmp (ptrtoint x), (ptrtoint/c) of vector type into a compare of the inputs,
much as is done when the type is integer.
Reviewers: apilipenko, fedor.sergeev, mkazantsev, anna
Reviewed By: anna
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D44063
llvm-svn: 326730
This patch teaches getMinimumFPType to support shrinking a vector of ConstantFPs. This should improve our ability to combine vector fptrunc with fp binops.
Differential Revision: https://reviews.llvm.org/D43774
llvm-svn: 326729
Instead of returning the smaller FP constant we now return the minimal Type the constant can fit into. We also return the Type of the input to any fp extends. The legality checks are then done on just the size of these Types. If we find something profitable we then emit FPTruncs in front of the smaller binop and assume those FPTruncs will be constant folded or combined with any ConstantFPs or fpextends.
Differential Revision: https://reviews.llvm.org/D44038
llvm-svn: 326617
The code was checking that all of the instructions in the
sequence are 'fast', but that's not necessary. The final
multiply is all that we need to check (tests adjusted).
The fmul doesn't need to be fully 'fast' either, but that
can be another patch.
llvm-svn: 326608
This is a retry of r326502 with updates to the reassociate
test file that I missed the first time.
@test15_reassoc in the supposed -reassociate test file
(except that it tests 2 other passes too...) shows that
there's no clear responsiblity for reassociation transforms.
Instcombine now gets that case, but only because the
constant values are identical. Otherwise, it would still
miss that pattern.
Reassociate doesn't get that case because it hasn't been
updated to use less than 'fast' FMF.
llvm-svn: 326513
I forgot that I added tests for 'reassoc' to -reassociate, but
suprisingly that file calls -instcombine too, so it is affected.
I'll update that file and try again.
llvm-svn: 326510
Currently this code's control flow very much assumes that there are no meaningful checks after determining that it's a ConstantFP. So whenever it wants to stop it just does "return V". But V is also the variable name it uses when it wants to return a new value. So 'return V' appears multiple times with different meanings.
This patch just moves all the code into a helper function and returns nullptr when it wants to stop.
I've split this from D43774 while I try to figure out how to best handle the vector case there. But this change by itself at least seemed like a readability improvement.
Differential Revision: https://reviews.llvm.org/D43833
llvm-svn: 326361
Also, rename 'foldOpWithConstantIntoOperand' because that's annoyingly
vague. The constant check is redundant in some cases, but it allows
removing duplication for most of the calls.
llvm-svn: 326329
Note: gcc appears to allow this fold with -freciprocal-math alone,
but clang/llvm require more than that with this patch. The wording
in the definitions seems fuzzy enough that it could go either way,
but we'll err on the conservative side of FMF interpretation.
This patch also changes the newly created fmul to have FMF propagated
by the last fdiv rather than intersecting the FMF of the fdivs. This
matches the behavior of other folds near here. The new fmul is only
used to produce an intermediate op for the final fdiv result, so it
shouldn't be any stricter than that result. The previous behavior
could result in dropping FMF via other folds in instcombine or CSE.
Differential Revision: https://reviews.llvm.org/D43398
llvm-svn: 326098
The existing code was inefficiently looking for 'nsz' variants.
That's unnecessary because we canonicalize those to the expected
form with -0.0.
We may also want to adjust or remove the fold that sinks negation.
We don't do that for fdiv (or integer ops?). That should be uniform?
It may also lead to missed optimization as in PR21914:
https://bugs.llvm.org/show_bug.cgi?id=21914
...or we just have to fix other passes to avoid that problem.
llvm-svn: 325924
These are fdiv-with-constant-divisor, so they already become
reciprocal multiplies. The last gap for vector ops should be
closed with rL325590.
It's possible that we're missing folds for some edge cases
with denormal intermediate constants after deleting these,
but there are no tests for those patterns, and it would be
better to handle denormals more consistently (and less
conservatively) as noted in TODO comments.
llvm-svn: 325595
It's possible that we could allow this either 'arcp' or 'reassoc' alone, but this
should be conservatively better than what we have right now. GCC allows this with
only -freciprocal-math.
The last test is changed to show a case that is expected to fold, but we need D43398.
llvm-svn: 325533
The last fold that used to be here was not necessary. That's a
combination of 2 folds (and there's a regression test to show that).
The transforms are guarded by isFast(), but that should be loosened.
llvm-svn: 325531
With this patch in place, when a new-format TBAA tag is available
for a memory-transfer intrinsic call, we prefer propagating that
new-format tag. Otherwise, we fallback to the old approach where
we try to construct a proper TBAA access tag from 'tbaa.struct'
metadata.
Differential Revision: https://reviews.llvm.org/D41543
llvm-svn: 325488
...and delete the equivalent local functiona from InstCombine.
These might be useful to other InstCombine files or other passes
and makes FP queries more similar to integer constant queries.
llvm-svn: 325398
The variable name 'AllowReassociate' is a lie at this point because
it's set to 'isFast()' which is more than the 'reassoc' FMF after
rL317488.
In D41286, we showed that this transform may be valid even with strict
math by brute force checking every 32-bit float result.
There's a potential problem here because we're replacing with a tan()
libcall rather than a hypothetical LLVM tan intrinsic. So we might
set errno when we should be guaranteed not to do that. But that's
independent of this change.
llvm-svn: 325247
The select may have been preventing a division by zero or INT_MIN/-1 so removing it might not be safe.
Fixes PR36362.
Differential Revision: https://reviews.llvm.org/D43276
llvm-svn: 325148
This keeps with our current usage of 'match' and is easier to see that
the optional NSW only applies in the non-constant operand case.
llvm-svn: 325140
Making a width of GEP Index, which is used for address calculation, to be one of the pointer properties in the Data Layout.
p[address space]:size:memory_size:alignment:pref_alignment:index_size_in_bits.
The index size parameter is optional, if not specified, it is equal to the pointer size.
Till now, the InstCombiner normalized GEPs and extended the Index operand to the pointer width.
It works fine if you can convert pointer to integer for address calculation and all registered targets do this.
But some ISAs have very restricted instruction set for the pointer calculation. During discussions were desided to retrieve information for GEP index from the Data Layout.
http://lists.llvm.org/pipermail/llvm-dev/2018-January/120416.html
I added an interface to the Data Layout and I changed the InstCombiner and some other passes to take the Index width into account.
This change does not affect any in-tree target. I added tests to cover data layouts with explicitly specified index size.
Differential Revision: https://reviews.llvm.org/D42123
llvm-svn: 325102
This replaces the bit-tracking based fold that did the same thing,
but it only worked for scalars and not directly.
There is no evidence in existing regression tests that the greater
power of bit-tracking was needed here, but we should be aware of
this potential loss of optimization.
llvm-svn: 325062
This is both a functional improvement for vectors and an
efficiency improvement for scalars. The existing code below
the new folds does the same thing for scalars, but in an
indirect and expensive way.
llvm-svn: 325048
Summary:
This change is part of step five in the series of changes to remove alignment argument from
memcpy/memmove/memset in favour of alignment attributes. In particular, this changes the
InstCombine pass to cease using the deprecated MemoryIntrinsic::getAlignment() method, and
instead we use the separate getSourceAlignment and getDestAlignment APIs to simplify
the source and destination alignment attributes separately.
Steps:
Step 1) Remove alignment parameter and create alignment parameter attributes for
memcpy/memmove/memset. ( rL322965, rC322964, rL322963 )
Step 2) Expand the IRBuilder API to allow creation of memcpy/memmove with differing
source and dest alignments. ( rL323597 )
Step 3) Update Clang to use the new IRBuilder API. ( rC323617 )
Step 4) Update Polly to use the new IRBuilder API. ( rL323618 )
Step 5) Update LLVM passes that create memcpy/memmove calls to use the new IRBuilder API,
and those that use use MemIntrinsicInst::[get|set]Alignment() to use [get|set]DestAlignment()
and [get|set]SourceAlignment() instead. ( rL323886, rL323891, rL324148, rL324273, rL324278,
rL324384, rL324395, rL324402, rL324626, rL324642, rL324653, rL324654, rL324773, rL324774,
rL324781, rL324784, rL324955 )
Step 6) Remove the single-alignment IRBuilder API for memcpy/memmove, and the
MemIntrinsicInst::[get|set]Alignment() methods.
Reference
http://lists.llvm.org/pipermail/llvm-dev/2015-August/089384.htmlhttp://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20151109/312083.html
Reviewers: majnemer, bollu, efriedma
Reviewed By: efriedma
Subscribers: efriedma, llvm-commits
Differential Revision: https://reviews.llvm.org/D42871
llvm-svn: 324960
This is similar to the instsimplify fold added with D42385
( rL323716 )
...but this can't be in instsimplify because we're creating/morphing
a different instruction.
llvm-svn: 324927
The related cases for (X * Y) / X were handled in rL124487.
https://rise4fun.com/Alive/6k9
The division in these tests is subsequently eliminated by existing instcombines
for 1/X.
llvm-svn: 324843
Refactor getLogBase2Vector into getLogBase2 to accept all scalars/vectors. Generalize from ConstantDataVector to support all constant vectors.
llvm-svn: 324603
This is the instcombine part of unsigned saturation canonicalization.
Backend patches already commited:
https://reviews.llvm.org/D37510https://reviews.llvm.org/D37534
It converts unsigned saturated subtraction patterns to forms recognized
by the backend:
(a > b) ? a - b : 0 -> ((a > b) ? a : b) - b)
(b < a) ? a - b : 0 -> ((a > b) ? a : b) - b)
(b > a) ? 0 : a - b -> ((a > b) ? a : b) - b)
(a < b) ? 0 : a - b -> ((a > b) ? a : b) - b)
((a > b) ? b - a : 0) -> - ((a > b) ? a : b) - b)
((b < a) ? b - a : 0) -> - ((a > b) ? a : b) - b)
((b > a) ? 0 : b - a) -> - ((a > b) ? a : b) - b)
((a < b) ? 0 : b - a) -> - ((a > b) ? a : b) - b)
Patch by Yulia Koval!
Differential Revision: https://reviews.llvm.org/D41480
llvm-svn: 324255
There was a logic hole in D42739 / rL324014 because we're not accounting for select and phi
instructions that might have repeated operands. This is likely a source of an infinite loop.
I haven't manufactured a test case to prove that, but it should be safe to speculatively limit
this transform to binops while we try to create that test.
llvm-svn: 324252
This, in instcombine, allows conversions to i8/i16/i32 (very
common cases) even if the resulting type is not legal according
to the data layout. This can often open up extra combine
opportunities.
Differential Revision: https://reviews.llvm.org/D42424
llvm-svn: 324174
This is the enhancement suggested in D42536 to fix a shortcoming in
regular InstCombine's canEvaluate* functionality.
When we have multiple uses of a value, but they're all in one instruction, we can
allow that expression to be narrowed or widened for the same cost as a single-use
value.
AFAICT, this can only matter for multiply: sub/and/or/xor/select would be simplified
away if the operands are the same value; add becomes shl; shifts with a variable shift
amount aren't handled.
Differential Revision: https://reviews.llvm.org/D42739
llvm-svn: 324014
This, in instcombine, allows conversions to i8/i16/i32 (very
common cases) even if the resulting type is not legal according
to the data layout. This can often open up extra combine
opportunities.
Differential Revision: https://reviews.llvm.org/D42424
llvm-svn: 323951
A cast from A to B is eliminable if its result is casted to C, and if
the pair of casts could just be expressed as a single cast. E.g here,
%c1 is eliminable:
%c1 = zext i16 %A to i32
%c2 = sext i32 %c1 to i64
InstCombine optimizes away eliminable casts. This patch teaches it to
insert a dbg.value intrinsic pointing to the final result, so that local
variables pointing to the eliminable result are preserved.
Differential Revision: https://reviews.llvm.org/D42566
llvm-svn: 323570
This is guarded by shouldChangeType(), so the tests show that
we don't do the fold if the narrower type is not legal. Note
that there is a proposal (D42424) that would change the results
for the specific cases shown in these tests. That difference is
also discussed in PR35792:
https://bugs.llvm.org/show_bug.cgi?id=35792
Alive proofs for the cases handled here as well as the bitwise
logic binops that we should already do better on:
https://rise4fun.com/Alive/c97https://rise4fun.com/Alive/Lc5Ehttps://rise4fun.com/Alive/kdf
llvm-svn: 323437
...when the shift is known to not overflow with the matching
signed-ness of the division.
This closes an optimization gap caused by canonicalizing mul
by power-of-2 to shl as shown in PR35709:
https://bugs.llvm.org/show_bug.cgi?id=35709
Patch by Anton Bikineev!
Differential Revision: https://reviews.llvm.org/D42032
llvm-svn: 323068
Three (or more) operand getelementptrs could plausibly also be handled, but
handling only two-operand fits in easily with the existing BinaryOperator
handling.
Differential Revision: https://reviews.llvm.org/D39958
llvm-svn: 322930
I was comparing the demanded-bits implementations between InstCombine
and TargetLowering as part of investigating questions in D42088 and
noticed that this was wrong in IR. We were losing all of the prior
known bits when we got back to the 'zext'.
llvm-svn: 322662
Summary:
In preparation for https://reviews.llvm.org/D41675 this NFC changes this
prototype of MemIntrinsicInst::setAlignment() to accept an unsigned instead
of a Constant.
llvm-svn: 322403
parent function
Ideally we should merge the attributes from the functions somehow, but
this is obviously an improvement over taking random attributes from the
caller which will trip up the verifier if they're nonsensical for an
unary intrinsic call.
llvm-svn: 322284
Because of potential UB (known bits conflicts with an llvm.assume),
we have to check rather than assert here because InstSimplify doesn't
kill the compare:
https://bugs.llvm.org/show_bug.cgi?id=35846
llvm-svn: 322104
There is precedence for factorization transforms in instcombine for FP ops with fast-math.
We also have similar logic in foldSPFofSPF().
It would take more work to add this to reassociate because that's specialized for binops,
and min/max are not binops (or even single instructions). Also, I don't have evidence that
larger min/max trees than this exist in real code, but if we find that's true, we might
want to reorganize where/how we do this optimization.
In the motivating example from https://bugs.llvm.org/show_bug.cgi?id=35717 , we have:
int test(int xc, int xm, int xy) {
int xk;
if (xc < xm)
xk = xc < xy ? xc : xy;
else
xk = xm < xy ? xm : xy;
return xk;
}
This patch solves that problem because we recognize more min/max patterns after rL321672
https://rise4fun.com/Alive/Qjnehttps://rise4fun.com/Alive/3yg
Differential Revision: https://reviews.llvm.org/D41603
llvm-svn: 321998
In the minimal case, this won't remove instructions, but it still improves
uses of existing values.
In the motivating example from PR35834, it does remove instructions, and
sets that case up to be optimized by something like D41603:
https://reviews.llvm.org/D41603
llvm-svn: 321936
Besides the bug of omitting the inverse transform of max(~a, ~b) --> ~min(a, b),
the use checking and operand creation were off. We were potentially creating
repeated identical instructions of existing values. This led to infinite
looping after I added the extra folds.
By using the simpler m_Not matcher and not creating new 'not' ops for a and b,
we avoid that problem. It's possible that not using IsFreeToInvert() here is
more limiting than the simpler matcher, but there are no tests for anything
more exotic. It's also possible that we should relax the use checking further
to handle a case like PR35834:
https://bugs.llvm.org/show_bug.cgi?id=35834
...but we can make that a follow-up if it is needed.
llvm-svn: 321882
We want to do this for 2 reasons:
1. Value tracking does not recognize the ashr variant, so it would fail to match for cases like D39766.
2. DAGCombiner does better at producing optimal codegen when we have the cmp+sel pattern.
More detail about what happens in the backend:
1. DAGCombiner has a generic transform for all targets to convert the scalar cmp+sel variant of abs
into the shift variant. That is the opposite of this IR canonicalization.
2. DAGCombiner has a generic transform for all targets to convert the vector cmp+sel variant of abs
into either an ABS node or the shift variant. That is again the opposite of this IR canonicalization.
3. DAGCombiner has a generic transform for all targets to convert the exact shift variants produced by #1 or #2
into an ISD::ABS node. Note: It would be an efficiency improvement if we had #1 go directly to an ABS node
when that's legal/custom.
4. The pattern matching above is incomplete, so it is possible to escape the intended/optimal codegen in a
variety of ways.
a. For #2, the vector path is missing the case for setlt with a '1' constant.
b. For #3, we are missing a match for commuted versions of the shift variants.
5. Therefore, this IR canonicalization can only help get us to the optimal codegen. The version of cmp+sel
produced by this patch will be recognized in the DAG and converted to an ABS node when possible or the
shift sequence when not.
6. In the following examples with this patch applied, we may get conditional moves rather than the shift
produced by the generic DAGCombiner transforms. The conditional move is created using a target-specific
decision for any given target. Whether it is optimal or not for a particular subtarget may be up for debate.
define i32 @abs_shifty(i32 %x) {
%signbit = ashr i32 %x, 31
%add = add i32 %signbit, %x
%abs = xor i32 %signbit, %add
ret i32 %abs
}
define i32 @abs_cmpsubsel(i32 %x) {
%cmp = icmp slt i32 %x, zeroinitializer
%sub = sub i32 zeroinitializer, %x
%abs = select i1 %cmp, i32 %sub, i32 %x
ret i32 %abs
}
define <4 x i32> @abs_shifty_vec(<4 x i32> %x) {
%signbit = ashr <4 x i32> %x, <i32 31, i32 31, i32 31, i32 31>
%add = add <4 x i32> %signbit, %x
%abs = xor <4 x i32> %signbit, %add
ret <4 x i32> %abs
}
define <4 x i32> @abs_cmpsubsel_vec(<4 x i32> %x) {
%cmp = icmp slt <4 x i32> %x, zeroinitializer
%sub = sub <4 x i32> zeroinitializer, %x
%abs = select <4 x i1> %cmp, <4 x i32> %sub, <4 x i32> %x
ret <4 x i32> %abs
}
> $ ./opt -instcombine shiftyabs.ll -S | ./llc -o - -mtriple=x86_64 -mattr=avx
> abs_shifty:
> movl %edi, %eax
> negl %eax
> cmovll %edi, %eax
> retq
>
> abs_cmpsubsel:
> movl %edi, %eax
> negl %eax
> cmovll %edi, %eax
> retq
>
> abs_shifty_vec:
> vpabsd %xmm0, %xmm0
> retq
>
> abs_cmpsubsel_vec:
> vpabsd %xmm0, %xmm0
> retq
>
> $ ./opt -instcombine shiftyabs.ll -S | ./llc -o - -mtriple=aarch64
> abs_shifty:
> cmp w0, #0 // =0
> cneg w0, w0, mi
> ret
>
> abs_cmpsubsel:
> cmp w0, #0 // =0
> cneg w0, w0, mi
> ret
>
> abs_shifty_vec:
> abs v0.4s, v0.4s
> ret
>
> abs_cmpsubsel_vec:
> abs v0.4s, v0.4s
> ret
>
> $ ./opt -instcombine shiftyabs.ll -S | ./llc -o - -mtriple=powerpc64le
> abs_shifty:
> srawi 4, 3, 31
> add 3, 3, 4
> xor 3, 3, 4
> blr
>
> abs_cmpsubsel:
> srawi 4, 3, 31
> add 3, 3, 4
> xor 3, 3, 4
> blr
>
> abs_shifty_vec:
> vspltisw 3, -16
> vspltisw 4, 15
> vsubuwm 3, 4, 3
> vsraw 3, 2, 3
> vadduwm 2, 2, 3
> xxlxor 34, 34, 35
> blr
>
> abs_cmpsubsel_vec:
> vspltisw 3, -16
> vspltisw 4, 15
> vsubuwm 3, 4, 3
> vsraw 3, 2, 3
> vadduwm 2, 2, 3
> xxlxor 34, 34, 35
> blr
>
Differential Revision: https://reviews.llvm.org/D40984
llvm-svn: 320921
Summary:
Passing AliasAnalysis results instead of nullptr appears to work just fine.
A couple new-pass-manager tests updated to align with new order of analyses.
Reviewers: chandlerc, spatel, craig.topper
Reviewed By: chandlerc
Subscribers: mehdi_amini, eraman, llvm-commits
Differential Revision: https://reviews.llvm.org/D41203
llvm-svn: 320687
Summary:
If we have pattern `store (load(bitcast(select (cmp(V1, V2), &V1,
&V2)))), bitcast)`, but the load is used in other instructions, it leads
to looping in InstCombiner. Patch adds additional check that all users
of the load instructions are stores and then replaces all uses of load
instruction by the new one with new type.
Reviewers: RKSimon, spatel, majnemer
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D41072
llvm-svn: 320525
Summary:
If we have pattern `store (load(bitcast(select (cmp(V1, V2), &V1,
&V2)))), bitcast)`, but the load is used in other instructions, it leads
to looping in InstCombiner. Patch adds additional check that all users
of the load instructions are stores and then replaces all uses of load
instruction by the new one with new type.
Reviewers: RKSimon, spatel, majnemer
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D41072
llvm-svn: 320510
Chad Rosier