I don't see an ideal solution to these 2 related, potentially large, perf regressions:
https://bugs.llvm.org/show_bug.cgi?id=42708https://bugs.llvm.org/show_bug.cgi?id=43146
We decided that load combining was unsuitable for IR because it could obscure other
optimizations in IR. So we removed the LoadCombiner pass and deferred to the backend.
Therefore, preventing SLP from destroying load combine opportunities requires that it
recognizes patterns that could be combined later, but not do the optimization itself (
it's not a vector combine anyway, so it's probably out-of-scope for SLP).
Here, we add a scalar cost model adjustment with a conservative pattern match and cost
summation for a multi-instruction sequence that can probably be reduced later.
This should prevent SLP from creating a vector reduction unless that sequence is
extremely cheap.
In the x86 tests shown (and discussed in more detail in the bug reports), SDAG combining
will produce a single instruction on these tests like:
movbe rax, qword ptr [rdi]
or:
mov rax, qword ptr [rdi]
Not some (half) vector monstrosity as we currently do using SLP:
vpmovzxbq ymm0, dword ptr [rdi + 1] # ymm0 = mem[0],zero,zero,..
vpsllvq ymm0, ymm0, ymmword ptr [rip + .LCPI0_0]
movzx eax, byte ptr [rdi]
movzx ecx, byte ptr [rdi + 5]
shl rcx, 40
movzx edx, byte ptr [rdi + 6]
shl rdx, 48
or rdx, rcx
movzx ecx, byte ptr [rdi + 7]
shl rcx, 56
or rcx, rdx
or rcx, rax
vextracti128 xmm1, ymm0, 1
vpor xmm0, xmm0, xmm1
vpshufd xmm1, xmm0, 78 # xmm1 = xmm0[2,3,0,1]
vpor xmm0, xmm0, xmm1
vmovq rax, xmm0
or rax, rcx
vzeroupper
ret
Differential Revision: https://reviews.llvm.org/D67841
llvm-svn: 373833
We do indeed already get it right in some cases, but only transitively,
with one-use restrictions. Since we only need to produce a single
comparison, it makes sense to match the pattern directly:
https://rise4fun.com/Alive/kPg
llvm-svn: 373802
Initially (D65380) i believed that if we have rightshift-trunc-rightshift,
we can't do any folding. But as it usually happens, i was wrong.
https://rise4fun.com/Alive/GEwhttps://rise4fun.com/Alive/gN2O
In https://bugs.llvm.org/show_bug.cgi?id=43564 we happen to have
this very sequence, of two right shifts separated by trunc.
And "just" so that happens, we apparently can fold the pattern
if the total shift amount is either 0, or it's equal to the bitwidth
of the innermost widest shift - i.e. if we are left with only the
original sign bit. Which is exactly what is wanted there.
llvm-svn: 373801
A set of function attributes is required in any function that uses constrained
floating point intrinsics. None of our tests use these attributes.
This patch fixes this.
These tests have been tested against the IR verifier changes in D68233.
Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand
Approved by: andrew.w.kaylor
Differential Revision: https://reviews.llvm.org/D67925
llvm-svn: 373761
Without this we can encounter link errors or incorrect behaviour
at runtime as a result of the wrong function being referenced.
Differential Revision: https://reviews.llvm.org/D67945
llvm-svn: 373678
As we have previously estabilished, `sub` is an outcast,
and should be considered non-canonical iff it can be converted to `add`.
It can be converted to `add` if it's second operand can be negated.
So far we mostly only do that for constants and negation itself,
but we should be more through.
llvm-svn: 373597
https://rise4fun.com/Alive/8BY - valid for lshr+trunc+variable sext
https://rise4fun.com/Alive/7jk - the variable sext can be redundant
https://rise4fun.com/Alive/Qslu - 'exact'-ness of first shift can be preserver
https://rise4fun.com/Alive/IF63 - without trunc we could view this as
more general "drop redundant mask before right-shift",
but let's handle it here for now
https://rise4fun.com/Alive/iip - likewise, without trunc, variable sext can be redundant.
There's more patterns for sure - e.g. we can have 'lshr' as the final shift,
but that might be best handled by some more generic transform, e.g.
"drop redundant masking before right-shift" (PR42456)
I'm singling-out this sext patch because you can only extract
high bits with `*shr` (unlike abstract bit masking),
and i *know* this fold is wanted by existing code.
I don't believe there is much to review here,
so i'm gonna opt into post-review mode here.
https://bugs.llvm.org/show_bug.cgi?id=43523
llvm-svn: 373542
bcopy is still widely used mainly for network apps. Sadly, LLVM has no optimizations for bcopy, but there are some for memmove.
Since bcopy == memmove, it is profitable to transform bcopy to memmove and use current optimizations for memmove for free here.
llvm-svn: 373537
This is intended to be similar to the constant folding results from
D67446
and earlier, but not all operands are constant in these tests, so the
responsibility for folding is left to InstSimplify.
Differential Revision: https://reviews.llvm.org/D67721
llvm-svn: 373455
This patch implements a variation of a well known techniques for JIT compilers - we have an implementation in tree as LoopPredication - but with an interesting twist. This version does not assume the ability to execute a path which wasn't taken in the original program (such as a guard or widenable.condition intrinsic). The benefit is that this works for arbitrary IR from any frontend (including C/C++/Fortran). The tradeoff is that it's restricted to read only loops without implicit exits.
This builds on SCEV, and can thus eliminate the loop varying portion of the any early exit where all exits are understandable by SCEV. A key advantage is that fixing deficiency exposed in SCEV - already found one while writing test cases - will also benefit all of full redundancy elimination (and most other loop transforms).
I haven't seen anything in the literature which quite matches this. Given that, I'm not entirely sure that keeping the name "loop predication" is helpful. Anyone have suggestions for a better name? This is analogous to partial redundancy elimination - since we remove the condition flowing around the backedge - and has some parallels to our existing transforms which try to make conditions invariant in loops.
Factoring wise, I chose to put this in IndVarSimplify since it's a generally applicable to all workloads. I could split this off into it's own pass, but we'd then probably want to add that new pass every place we use IndVars. One solid argument for splitting it off into it's own pass is that this transform is "too good". It breaks a huge number of existing IndVars test cases as they tend to be simple read only loops. At the moment, I've opted it off by default, but if we add this to IndVars and enable, we'll have to update around 20 test files to add side effects or disable this transform.
Near term plan is to fuzz this extensively while off by default, reflect and discuss on the factoring issue mentioned just above, and then enable by default. I also need to give some though to supporting widenable conditions in this framing.
Differential Revision: https://reviews.llvm.org/D67408
llvm-svn: 373351
Expand the simplification of special cases of `log()` to include `log2()`
and `log10()` as well as intrinsics and more types.
Differential revision: https://reviews.llvm.org/D67199
llvm-svn: 373261
With this patch, compiler generated profile variables will have its own COMDAT
name for ELF format, which syncs the behavior with COFF. Tested with clang
PGO bootstrap. This shows a modest reduction in object sizes in ELF format.
Differential Revision: https://reviews.llvm.org/D68041
llvm-svn: 373241
"Captured" and "relevant to Stack Protector" are not the same thing.
This reverts commit f29366b1f594f48465c5a2754bcffac6d70fd0b1.
aka r363169.
Differential Revision: https://reviews.llvm.org/D67842
llvm-svn: 373216
If we happen to have the same div in two basic blocks,
and in one of those we also happen to have the rem part,
we'd match the div-rem pair, but the wrong ones.
So let's drop overly-ambiguous assert.
Fixes https://bugs.llvm.org/show_bug.cgi?id=43500
llvm-svn: 373167
Initially SLP vectorizer replaced all going-to-be-vectorized
instructions with Undef values. It may break ScalarEvaluation and may
cause a crash.
Reworked SLP vectorizer so that it does not replace vectorized
instructions by UndefValue anymore. Instead vectorized instructions are
marked for deletion inside if BoUpSLP class and deleted upon class
destruction.
Reviewers: mzolotukhin, mkuper, hfinkel, RKSimon, davide, spatel
Subscribers: RKSimon, Gerolf, anemet, hans, majnemer, llvm-commits, sanjoy
Differential Revision: https://reviews.llvm.org/D29641
llvm-svn: 373166
profile symbol list.
Currently many existing users using profile-sample-accurate want to reduce
code size as much as possible. Their use cases are different from the scenario
profile symbol list tries to handle -- the major motivation of adding profile
symbol list is to get the major memory/code size saving without introduce
performance regression. So to keep the behavior of profile-sample-accurate
unchanged, we think decoupling these two things and using a new flag to
control the handling of profile symbol list may be better.
When profile-sample-accurate and the new flag profile-accurate-for-symsinlist
are both present, since profile-sample-accurate is a user assertion we let it
have a higher precedence.
Differential Revision: https://reviews.llvm.org/D68047
llvm-svn: 373133
We start with two separate sext's, but EarlyCSE runs before InstCombine,
so when we get them, they are a single sext, and we just ignore that.
Likewise, if we had a single sext, we don't do anything there.
llvm-svn: 373115
Summary:
This is valid for any `sext` bitwidth pair:
```
Processing /tmp/opt.ll..
----------------------------------------
%signed = sext %y
%r = shl %x, %signed
ret %r
=>
%unsigned = zext %y
%r = shl %x, %unsigned
ret %r
%signed = sext %y
Done: 2016
Optimization is correct!
```
(This isn't so for funnel shifts, there it's illegal for e.g. i6->i7.)
Main motivation is the C++ semantics:
```
int shl(int a, char b) {
return a << b;
}
```
ends as
```
%3 = sext i8 %1 to i32
%4 = shl i32 %0, %3
```
https://godbolt.org/z/0jgqUq
which is, as this shows, too pessimistic.
There is another problem here - we can only do the fold
if sext is one-use. But we can trivially have cases
where several shifts have the same sext shift amount.
This should be resolved, later.
Reviewers: spatel, nikic, RKSimon
Reviewed By: spatel
Subscribers: efriedma, hiraditya, nlopes, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D68103
llvm-svn: 373106
exits"
Get a better approach in https://reviews.llvm.org/D68107 to solve the problem.
Revert the initial patch and will commit the new one soon.
This reverts commit rL372990.
llvm-svn: 373044
Summary:
This patch extends the current capabilities in loop fusion to fuse guarded loops
(as defined in https://reviews.llvm.org/D63885). The patch adds the necessary
safety checks to ensure that it safe to fuse the guarded loops (control flow
equivalent, no intervening code, and same guard conditions). It also provides an
alternative method to perform the actual fusion of guarded loops. The mechanics
to fuse guarded loops are slightly different then fusing non-guarded loops, so I
opted to keep them separate methods. I will be cleaning this up in later
patches, and hope to converge on a single method to fuse both guarded and
non-guarded loops, but for now I think the review will be easier to keep them
separate.
Reviewers: jdoerfert, Meinersbur, dmgreen, etiotto, Whitney
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D65464
llvm-svn: 373018
For a runtime loop if we can compute its trip count upperbound:
Don't unroll if:
1. loop is not guaranteed to run either zero or upperbound iterations; and
2. trip count upperbound is less than UnrollMaxUpperBound
Unless user or TTI asked to do so.
If unrolling, limit unroll factor to loop's trip count upperbound.
Differential Revision: https://reviews.llvm.org/D62989
Change-Id: I6083c46a9d98b2e22cd855e60523fdc5a4929c73
llvm-svn: 373017
The test case here previously infinite looped. Only one element from the GEP is used so SimplifyDemandedVectorElts would replace the other lanes in each index with undef leading to the first index being <0, undef, undef, undef>. But there's a GEP transform that tries to replace an index into a 0 sized type with a zero index. But the zero index check only works on ConstantInt 0 or ConstantAggregateZero so it would turn the index back to zeroinitializer. Resulting in a loop.
The fix is to use m_Zero() to allow a vector of zeroes and undefs.
Differential Revision: https://reviews.llvm.org/D67977
llvm-svn: 373000
for extreme large case.
We had a case that a single loop which has 4000 exits and the average number
of predecessors of each exit is > 1000, and we found compiling the case spent
a significant amount of time on checking whether a loop has dedicated exits.
This patch adds a limit for the iterations to the check. With the patch, the
time to compile our testcase reduced from 1000s to 200s (clang release build).
Differential Revision: https://reviews.llvm.org/D67359
llvm-svn: 372990
Summary:
FlattenCFG merges two 'if' basicblocks by inserting one basicblock
to another basicblock. The inserted basicblock can have a successor
that contains a PHI node whoes incoming basicblock is the inserted
basicblock. Since the existing code does not handle it, it becomes
a badref.
if (cond1)
statement
if (cond2)
statement
successor - contains PHI node whose predecessor is cond2
-->
if (cond1 || cond2)
statement
(BB for cond2 was deleted)
successor - contains PHI node whose predecessor is cond2 --> bad ref!
Author: Jaebaek Seo
Reviewers: asbirlea, kuhar, tstellar, chandlerc, davide, dexonsmith
Reviewed By: kuhar
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D68032
llvm-svn: 372989
