This was reverted in 503deec218
because it caused gigantic increase (3x) in branch mispredictions
in certain benchmarks on certain CPU's,
see https://reviews.llvm.org/D84108#2227365.
It has since been investigated and here are the results:
https://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20200907/827578.html
> It's an amazingly severe regression, but it's also all due to branch
> mispredicts (about 3x without this). The code layout looks ok so there's
> probably something else to deal with. I'm not sure there's anything we can
> reasonably do so we'll just have to take the hit for now and wait for
> another code reorganization to make the branch predictor a bit more happy :)
>
> Thanks for giving us some time to investigate and feel free to recommit
> whenever you'd like.
>
> -eric
So let's just reland this.
Original commit message:
I've been looking at missed vectorizations in one codebase.
One particular thing that stands out is that some of the loops
reach vectorizer in a rather mangled form, with weird PHI's,
and some of the loops aren't even in a rotated form.
After taking a more detailed look, that happened because
the loop's headers were too big by then. It is evident that
SimplifyCFG's common code hoisting transform is at fault there,
because the pattern it handles is precisely the unrotated
loop basic block structure.
Surprizingly, `SimplifyCFGOpt::HoistThenElseCodeToIf()` is enabled
by default, and is always run, unlike it's friend, common code sinking
transform, `SinkCommonCodeFromPredecessors()`, which is not enabled
by default and is only run once very late in the pipeline.
I'm proposing to harmonize this, and disable common code hoisting
until //late// in pipeline. Definition of //late// may vary,
here currently i've picked the same one as for code sinking,
but i suppose we could enable it as soon as right after
loop rotation happens.
Experimentation shows that this does indeed unsurprizingly help,
more loops got rotated, although other issues remain elsewhere.
Now, this undoubtedly seriously shakes phase ordering.
This will undoubtedly be a mixed bag in terms of both compile- and
run- time performance, codesize. Since we no longer aggressively
hoist+deduplicate common code, we don't pay the price of said hoisting
(which wasn't big). That may allow more loops to be rotated,
so we pay that price. That, in turn, that may enable all the transforms
that require canonical (rotated) loop form, including but not limited to
vectorization, so we pay that too. And in general, no deduplication means
more [duplicate] instructions going through the optimizations. But there's still
late hoisting, some of them will be caught late.
As per benchmarks i've run {F12360204}, this is mostly within the noise,
there are some small improvements, some small regressions.
One big regression i saw i fixed in rG8d487668d09fb0e4e54f36207f07c1480ffabbfd, but i'm sure
this will expose many more pre-existing missed optimizations, as usual :S
llvm-compile-time-tracker.com thoughts on this:
http://llvm-compile-time-tracker.com/compare.php?from=e40315d2b4ed1e38962a8f33ff151693ed4ada63&to=c8289c0ecbf235da9fb0e3bc052e3c0d6bff5cf9&stat=instructions
* this does regress compile-time by +0.5% geomean (unsurprizingly)
* size impact varies; for ThinLTO it's actually an improvement
The largest fallout appears to be in GVN's load partial redundancy
elimination, it spends *much* more time in
`MemoryDependenceResults::getNonLocalPointerDependency()`.
Non-local `MemoryDependenceResults` is widely-known to be, uh, costly.
There does not appear to be a proper solution to this issue,
other than silencing the compile-time performance regression
by tuning cut-off thresholds in `MemoryDependenceResults`,
at the cost of potentially regressing run-time performance.
D84609 attempts to move in that direction, but the path is unclear
and is going to take some time.
If we look at stats before/after diffs, some excerpts:
* RawSpeed (the target) {F12360200}
* -14 (-73.68%) loops not rotated due to the header size (yay)
* -272 (-0.67%) `"Number of live out of a loop variables"` - good for vectorizer
* -3937 (-64.19%) common instructions hoisted
* +561 (+0.06%) x86 asm instructions
* -2 basic blocks
* +2418 (+0.11%) IR instructions
* vanilla test-suite + RawSpeed + darktable {F12360201}
* -36396 (-65.29%) common instructions hoisted
* +1676 (+0.02%) x86 asm instructions
* +662 (+0.06%) basic blocks
* +4395 (+0.04%) IR instructions
It is likely to be sub-optimal for when optimizing for code size,
so one might want to change tune pipeline by enabling sinking/hoisting
when optimizing for size.
Reviewed By: mkazantsev
Differential Revision: https://reviews.llvm.org/D84108
This reverts commit 503deec218.
As disscussed in post-commit review starting with
https://reviews.llvm.org/D84108#2227365
while this appears to be mostly a win overall, especially code-size-wise,
this appears to shake //certain// code pattens in a way that is extremely
unfavorable for performance (+30% runtime regression)
on certain CPU's (i personally can't reproduce).
So until the behaviour is better understood, and a path forward is mapped,
let's back this out for now.
This reverts commit 1d51dc38d8.
this bug was causing miscompile.
now clang cant properly selfhost with -mllvm --enable-knowledge-retention
Reviewed By: jdoerfert, lebedev.ri
Differential Revision: https://reviews.llvm.org/D83507
This reverts commit babb59496b.
This test addition was queued up with some unrelated changes,
but it seems more likely that we need to fix something internal
to -memcpyopt. Also, I'm not sure if including target-specifc
attributes in a generic regression test dir will cause bot
problems.
Summary:
This patch takes the indices operands of `insertelement`/`insertvalue`
into account while generation of seed elements for `findBuildAggregate()`.
This function has kept the original order of `insert`s before.
Also this patch optimizes `findBuildAggregate()` preventing it from
redundant temporary vector allocations and its multiple reversing.
Fixes llvm.org/pr44067
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D83779
I've been looking at missed vectorizations in one codebase.
One particular thing that stands out is that some of the loops
reach vectorizer in a rather mangled form, with weird PHI's,
and some of the loops aren't even in a rotated form.
After taking a more detailed look, that happened because
the loop's headers were too big by then. It is evident that
SimplifyCFG's common code hoisting transform is at fault there,
because the pattern it handles is precisely the unrotated
loop basic block structure.
Surprizingly, `SimplifyCFGOpt::HoistThenElseCodeToIf()` is enabled
by default, and is always run, unlike it's friend, common code sinking
transform, `SinkCommonCodeFromPredecessors()`, which is not enabled
by default and is only run once very late in the pipeline.
I'm proposing to harmonize this, and disable common code hoisting
until //late// in pipeline. Definition of //late// may vary,
here currently i've picked the same one as for code sinking,
but i suppose we could enable it as soon as right after
loop rotation happens.
Experimentation shows that this does indeed unsurprizingly help,
more loops got rotated, although other issues remain elsewhere.
Now, this undoubtedly seriously shakes phase ordering.
This will undoubtedly be a mixed bag in terms of both compile- and
run- time performance, codesize. Since we no longer aggressively
hoist+deduplicate common code, we don't pay the price of said hoisting
(which wasn't big). That may allow more loops to be rotated,
so we pay that price. That, in turn, that may enable all the transforms
that require canonical (rotated) loop form, including but not limited to
vectorization, so we pay that too. And in general, no deduplication means
more [duplicate] instructions going through the optimizations. But there's still
late hoisting, some of them will be caught late.
As per benchmarks i've run {F12360204}, this is mostly within the noise,
there are some small improvements, some small regressions.
One big regression i saw i fixed in rG8d487668d09fb0e4e54f36207f07c1480ffabbfd, but i'm sure
this will expose many more pre-existing missed optimizations, as usual :S
llvm-compile-time-tracker.com thoughts on this:
http://llvm-compile-time-tracker.com/compare.php?from=e40315d2b4ed1e38962a8f33ff151693ed4ada63&to=c8289c0ecbf235da9fb0e3bc052e3c0d6bff5cf9&stat=instructions
* this does regress compile-time by +0.5% geomean (unsurprizingly)
* size impact varies; for ThinLTO it's actually an improvement
The largest fallout appears to be in GVN's load partial redundancy
elimination, it spends *much* more time in
`MemoryDependenceResults::getNonLocalPointerDependency()`.
Non-local `MemoryDependenceResults` is widely-known to be, uh, costly.
There does not appear to be a proper solution to this issue,
other than silencing the compile-time performance regression
by tuning cut-off thresholds in `MemoryDependenceResults`,
at the cost of potentially regressing run-time performance.
D84609 attempts to move in that direction, but the path is unclear
and is going to take some time.
If we look at stats before/after diffs, some excerpts:
* RawSpeed (the target) {F12360200}
* -14 (-73.68%) loops not rotated due to the header size (yay)
* -272 (-0.67%) `"Number of live out of a loop variables"` - good for vectorizer
* -3937 (-64.19%) common instructions hoisted
* +561 (+0.06%) x86 asm instructions
* -2 basic blocks
* +2418 (+0.11%) IR instructions
* vanilla test-suite + RawSpeed + darktable {F12360201}
* -36396 (-65.29%) common instructions hoisted
* +1676 (+0.02%) x86 asm instructions
* +662 (+0.06%) basic blocks
* +4395 (+0.04%) IR instructions
It is likely to be sub-optimal for when optimizing for code size,
so one might want to change tune pipeline by enabling sinking/hoisting
when optimizing for size.
Reviewed By: mkazantsev
Differential Revision: https://reviews.llvm.org/D84108
This patch adds simplification for pattern:
```
if (cond)
/ \
... ...
\ /
p = phi [true] [false]
...
br p, succ_1, succ_2
```
If we can prove that top block's branches dominate respective
inputs of a block that has a Phi with constant inputs, we can
use the branch condition (maybe inverted) instead of Phi.
This will make proofs of implication for further jump threading
more transparent.
Differential Revision: https://reviews.llvm.org/D81375
Reviewed By: xbolva00
Summary:
NOTE: There is a mailing list discussion on this: http://lists.llvm.org/pipermail/llvm-dev/2019-December/137632.html
Complemantary to the assumption outliner prototype in D71692, this patch
shows how we could simplify the code emitted for an alignemnt
assumption. The generated code is smaller, less fragile, and it makes it
easier to recognize the additional use as a "assumption use".
As mentioned in D71692 and on the mailing list, we could adopt this
scheme, and similar schemes for other patterns, without adopting the
assumption outlining.
Reviewers: hfinkel, xbolva00, lebedev.ri, nikic, rjmccall, spatel, jdoerfert, sstefan1
Reviewed By: jdoerfert
Subscribers: thopre, yamauchi, kuter, fhahn, merge_guards_bot, hiraditya, bollu, rkruppe, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D71739
Assume bundle can have more than one entry with the same name,
but at least AlignmentFromAssumptionsPass::extractAlignmentInfo() uses
getOperandBundle("align"), which internally assumes that it isn't the
case, and happily crashes otherwise.
Minimal reduced reproducer: run `opt -alignment-from-assumptions` on
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"
%0 = type { i64, %1*, i8*, i64, %2, i32, %3*, i8* }
%1 = type opaque
%2 = type { i8, i8, i16 }
%3 = type { i32, i32, i32, i32 }
; Function Attrs: nounwind
define i32 @f(%0* noalias nocapture readonly %arg, %0* noalias %arg1) local_unnamed_addr #0 {
bb:
call void @llvm.assume(i1 true) [ "align"(%0* %arg, i64 8), "align"(%0* %arg1, i64 8) ]
ret i32 0
}
; Function Attrs: nounwind willreturn
declare void @llvm.assume(i1) #1
attributes #0 = { nounwind "reciprocal-estimates"="none" }
attributes #1 = { nounwind willreturn }
This is what we'd have with -mllvm -enable-knowledge-retention
This reverts commit c95ffadb24.
binop i1 (cmp Pred (ext X, Index0), C0), (cmp Pred (ext X, Index1), C1)
-->
vcmp = cmp Pred X, VecC
ext (binop vNi1 vcmp, (shuffle vcmp, Index1)), Index0
This is a larger pattern than the existing extractelement folds because we can't
reasonably vectorize the sub-patterns with constants based on cost model calcs
(it doesn't usually make sense to replace a single extracted scalar op with
constant operand with a vector op).
I salvaged as much of the existing logic as I could, but there might be better
ways to share and reduce code.
The motivating case from PR43745:
https://bugs.llvm.org/show_bug.cgi?id=43745
...is the special case of a 2-way reduction. We tried to get SLP to handle that
particular pattern in D59710, but that caused crashing and regressions.
This patch is more general, but hopefully safer.
The v2f64 test with SSE2 surprised me - the cost model accounting looks like this:
OldCost = 0 (free extract of f64 at index 0) + 1 (extract of f64 at index 1) + 2 (scalar fcmps) + 1 (and of bools) = 4
NewCost = 2 (vector fcmp) + 1 (shuffle) + 1 (vector 'and') + 1 (extract of bool) = 5
Differential Revision: https://reviews.llvm.org/D82474
Summary:
NOTE: There is a mailing list discussion on this: http://lists.llvm.org/pipermail/llvm-dev/2019-December/137632.html
Complemantary to the assumption outliner prototype in D71692, this patch
shows how we could simplify the code emitted for an alignemnt
assumption. The generated code is smaller, less fragile, and it makes it
easier to recognize the additional use as a "assumption use".
As mentioned in D71692 and on the mailing list, we could adopt this
scheme, and similar schemes for other patterns, without adopting the
assumption outlining.
Reviewers: hfinkel, xbolva00, lebedev.ri, nikic, rjmccall, spatel, jdoerfert, sstefan1
Reviewed By: jdoerfert
Subscribers: yamauchi, kuter, fhahn, merge_guards_bot, hiraditya, bollu, rkruppe, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D71739
This is the integer sibling to D81491.
(a[0] + a[1] + a[2] + a[3]) - (b[0] + b[1] + b[2] +b[3]) -->
(a[0] - b[0]) + (a[1] - b[1]) + (a[2] - b[2]) + (a[3] - b[3])
Removing the "experimental" from these intrinsics is likely
not too far away.
This saves creating/destroying a builder every time we
perform some transform.
The tests show instruction ordering diffs resulting from
always inserting at the root instruction now, but those
should be benign.
(a[0] + a[1] + a[2] + a[3]) - (b[0] + b[1] + b[2] +b[3]) -->
(a[0] - b[0]) + (a[1] - b[1]) + (a[2] - b[2]) + (a[3] - b[3])
This should be the last step in solving PR43953:
https://bugs.llvm.org/show_bug.cgi?id=43953
We started emitting reduction intrinsics with:
D80867/ rGe50059f6b6b3
So it's a relatively easy pattern match now to re-order those ops.
Also, I have not seen any complaints for the switch to intrinsics
yet, so I'll propose to remove the "experimental" tag from the
intrinsics soon.
Differential Revision: https://reviews.llvm.org/D81491
scalarizeBinop currently folds
vec_bo((inselt VecC0, V0, Index), (inselt VecC1, V1, Index))
->
inselt(vec_bo(VecC0, VecC1), scl_bo(V0,V1), Index)
This patch extends this to account for cases where one of the vec_bo operands is already all-constant and performs similar cost checks to determine if the scalar binop with a constant still makes sense:
vec_bo((inselt VecC0, V0, Index), VecC1)
->
inselt(vec_bo(VecC0, VecC1), scl_bo(V0,extractelt(V1,Index)), Index)
Fixes PR42174
Differential Revision: https://reviews.llvm.org/D80885
Motivating examples are seen in the PhaseOrdering tests based on:
https://bugs.llvm.org/show_bug.cgi?id=43953#c2 - if we have
intrinsics there, some pass can fold them.
The intrinsics are still named "experimental" at this point, but
if there is no fallout from this patch, that will be a good
indicator that it is safe to finalize them.
Differential Revision: https://reviews.llvm.org/D80867
As discussed in D80236 - this test (like all PhaseOrdering tests?)
was intended to show that there is no difference with the new
pass manager, but the 'opt' command requires extra parameters
to make that happen.
There are 2 known problem patterns shown in the test diffs here:
vector horizontal ops (an x86 specialization) and vector reductions.
SLP has greater ability to match and fold those than vector-combine,
so let SLP have first chance at that.
This is a quick fix while we continue to improve vector-combine and
possibly canonicalize to reduction intrinsics.
In the longer term, we should improve matching of these patterns
because if they were created in the "bad" forms shown here, then we
would miss optimizing them.
I'm not sure what is happening with alias analysis on the addsub test.
The old pass manager now shows an extra line for that, and we see an
improvement that comes from SLP vectorizing a store. I don't know
what's missing with the new pass manager to make that happen.
Strangely, I can't reproduce the behavior if I compile from C++ with
clang and invoke the new PM with "-fexperimental-new-pass-manager".
Differential Revision: https://reviews.llvm.org/D80236
This eliminates a use of 'B', so it can enable follow-on transforms
as well as improve analysis/codegen.
The PhaseOrdering test was added for D61726, and that shows
the limits of instcombine vs. real reassociation. We would
need to run some form of CSE to collapse that further.
The intermediate variable naming here is intentional because
there's a test at llvm/test/Bitcode/value-with-long-name.ll
that would break with the usual nameless value. I'm not sure
how to improve that test to be more robust.
The naming may also be helpful to debug regressions if this
change exposes weaknesses in the reassociation pass for example.
This is split off from D79799 - where I was proposing to fully iterate
over a function until there are no more transforms. I suspect we are
still going to want to do something like that eventually.
But we can achieve the same gains much more efficiently on the current
set of regression tests just by reversing the order that we visit the
instructions.
This may also reduce the motivation for D79078, but we are still not
getting the optimal pattern for a reduction.
In D74183 clang started emitting alignment for sret parameters
unconditionally. This caused a 1.5% compile-time regression on
tramp3d-v4. The reason is that we now generate many instance of IR like
%ptrint = ptrtoint %class.GuardLayers* %guards_m to i64
%maskedptr = and i64 %ptrint, 3
%maskcond = icmp eq i64 %maskedptr, 0
tail call void @llvm.assume(i1 %maskcond)
to preserve the alignment information during inlining. Based on IR
analysis, these assumptions also regress optimization. The attached
phase ordering test case illustrates two issues: One are instruction
count based optimization heuristics, which are affected by the four
additional instructions of the assumption. The other is blocking of
SROA due to ptrtoint casts (PR45763).
We already encountered the same problem in Rust, where we (unlike
Clang) generally prefer to emit alignment information absolutely
everywhere it is available. We were only able to do this after
hardcoding -preserve-alignment-assumptions-during-inlining=false,
because we were seeing significant optimization and compile-time
regressions otherwise.
This patch disables -preserve-alignment-assumptions-during-inlining
by default, because we should not be punishing people for adding
more alignment annotations.
Once the assume bundle work shakes out and we can represent (and use)
alignment assumptions using assume bundles, it should be possible to
re-enable this with reduced overhead.
Differential Revision: https://reviews.llvm.org/D76886
bitcast (shuf V, MaskC) --> shuf (bitcast V), MaskC'
This is the widen shuffle elements enhancement to D76727.
It builds on the analysis and simplifications in
D77881 and rG6a7e958a423e.
The phase ordering tests show that we can simplify inverse
shuffles across a binop in both directions (widen/narrow or
narrow/widen) now.
There's another potential transform visible in some of the
remaining TODOs - move a bitcasted operand of a shuffle
after the shuffle.
Differential Revision: https://reviews.llvm.org/D78371
As discussed in D76983, that patch can turn a chain of insert/extract
with scalar trunc ops into bitcast+extract and existing instcombine
vector transforms end up creating a shuffle out of that (see the
PhaseOrdering test for an example). Currently, that process requires
at least this sequence: -instcombine -early-cse -instcombine.
Before D76983, the sequence of insert/extract would reach the SLP
vectorizer and become a vector trunc there.
Based on a small sampling of public targets/types, converting the
shuffle to a trunc is better for codegen in most cases (and a
regression of that form is the reason this was noticed). The trunc is
clearly better for IR-level analysis as well.
This means that we can induce "spontaneous vectorization" without
invoking any explicit vectorizer passes (at least a vector cast op
may be created out of scalar casts), but that seems to be the right
choice given that we started with a chain of insert/extract, and the
backend would expand back to that chain if a target does not support
the op.
Differential Revision: https://reviews.llvm.org/D77299
We got some of the potential optimizations with D76727 and D76844.
There are 2 likely enhancements that we could add to -vector-combine
to get most of the remaining cases:
1. Allow bitcasted shuffle mask narrowing (widen the elements).
2. Combine shuffle-of-shuffle into a single shuffle.
This is already partly handled by the x86 backend, but the
tests here show that we still miss some of the potential
combines.
These are versions of a function that regressed with:
rGf2fbdf76d8d0
That particular problem occurs with an instcombine-simplifycfg-instcombine
sequence, but we can show that it exists within instcombine only with
other variations of the pattern.
Roman Lebedev