Summary: This patch introduces a new heuristic for guiding operand reordering. The new "look-ahead" heuristic can look beyond the immediate predecessors. This helps break ties when the immediate predecessors have identical opcodes (see lit test for examples).
Reviewers: RKSimon, ABataev, dtemirbulatov, Ayal, hfinkel, rnk
Reviewed By: RKSimon, dtemirbulatov
Subscribers: xbolva00, Carrot, hiraditya, phosek, rnk, rcorcs, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D60897
We have a vector compare reduction problem seen in PR39665 comment 2:
https://bugs.llvm.org/show_bug.cgi?id=39665#c2
Or slightly reduced here:
define i1 @cmp2(<2 x double> %a0) {
%a = fcmp ogt <2 x double> %a0, <double 1.0, double 1.0>
%b = extractelement <2 x i1> %a, i32 0
%c = extractelement <2 x i1> %a, i32 1
%d = and i1 %b, %c
ret i1 %d
}
SLP would not attempt to turn this into a vector reduction because there is an
artificial lower limit on that transform. We can not completely remove that limit
without inducing regressions though, so this patch just hacks an extra attempt at
creating a 2-way reduction to the end of the analysis.
As shown in the test file, we are still not getting some of the motivating cases,
so follow-on patches will be needed to solve those cases.
Differential Revision: https://reviews.llvm.org/D59710
"[SLP] Generalization of stores vectorization."
"[SLP] Fix -Wunused-variable. NFC"
"[SLP] Vectorize jumbled stores."
As they're causing significant (10-30x) compile time regressions on
vectorizable code.
The primary cause of the compile-time regression is f228b5371647f471853c5fb3e6719823a42fe451.
This reverts commits:
f228b5371647f471853c5fb3e6719823a42fe451
5503455ccb3f5fcedced158332c016c8d3a7fa81
21d498c9c0f32dcab5bc89ac593aa813b533b43a
Summary:
If the GEP instructions are going to be vectorized, the indices in those
GEP instructions must be of the same type. Otherwise, the compiler may
crash when trying to build the vector constant.
Reviewers: RKSimon, spatel
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D69627
Summary:
Patch adds support for vectorization of the jumbled stores. The value
operands are vectorized and then shuffled in the right order before
store.
Reviewers: RKSimon, spatel, hfinkel, mkuper
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D43339
Summary:
Patch adds support for vectorization of the jumbled stores. The value
operands are vectorized and then shuffled in the right order before
store.
Reviewers: RKSimon, spatel, hfinkel, mkuper
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D43339
Stores are vectorized with maximum vectorization factor of 16. Patch
tries to improve the situation and use maximal vectorization factor.
Reviewers: spatel, RKSimon, mkuper, hfinkel
Differential Revision: https://reviews.llvm.org/D43582
The 1st attempt at this modified the cost model in a bad way to avoid the vectorization,
but that caused problems for other users (the loop vectorizer) of the cost model.
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 cost-independent bailout with a conservative pattern match for a
multi-instruction sequence that can probably be reduced later.
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: 375025
In loop-vectorize, interleave count and vector factor depend on target register number. Currently, it does not
estimate different register pressure for different register class separately(especially for scalar type,
float type should not be on the same position with int type), so it's not accurate. Specifically,
it causes too many times interleaving/unrolling, result in too many register spills in loop body and hurting performance.
So we need classify the register classes in IR level, and importantly these are abstract register classes,
and are not the target register class of backend provided in td file. It's used to establish the mapping between
the types of IR values and the number of simultaneous live ranges to which we'd like to limit for some set of those types.
For example, POWER target, register num is special when VSX is enabled. When VSX is enabled, the number of int scalar register is 32(GPR),
float is 64(VSR), but for int and float vector register both are 64(VSR). So there should be 2 kinds of register class when vsx is enabled,
and 3 kinds of register class when VSX is NOT enabled.
It runs on POWER target, it makes big(+~30%) performance improvement in one specific bmk(503.bwaves_r) of spec2017 and no other obvious degressions.
Differential revision: https://reviews.llvm.org/D67148
llvm-svn: 374634
We failed to account for the target register width (max vector factor)
when vectorizing starting from GEPs. This causes vectorization to
proceed to obviously illegal widths as in:
https://bugs.llvm.org/show_bug.cgi?id=43578
For x86, this also means that SLP can produce rogue AVX or AVX512
code even when the user specifies a narrower vector width.
The AArch64 test in ext-trunc.ll appears to be better using the
narrower width. I'm not exactly sure what getelementptr.ll is trying
to do, but it's testing with "-slp-threshold=-18", so I'm not worried
about those diffs. The x86 test is an over-reduction from SPEC h264;
this patch appears to restore the perf loss caused by SLP when using
-march=haswell.
Differential Revision: https://reviews.llvm.org/D68667
llvm-svn: 374183
Also Revert "[LoopVectorize] Fix non-debug builds after rL374017"
This reverts commit 9f41deccc0e648a006c9f38e11919f181b6c7e0a.
This reverts commit 18b6fe07bcf44294f200bd2b526cb737ed275c04.
The patch is breaking PowerPC internal build, checked with author, reverting
on behalf of him for now due to timezone.
llvm-svn: 374091
In loop-vectorize, interleave count and vector factor depend on target register number. Currently, it does not
estimate different register pressure for different register class separately(especially for scalar type,
float type should not be on the same position with int type), so it's not accurate. Specifically,
it causes too many times interleaving/unrolling, result in too many register spills in loop body and hurting performance.
So we need classify the register classes in IR level, and importantly these are abstract register classes,
and are not the target register class of backend provided in td file. It's used to establish the mapping between
the types of IR values and the number of simultaneous live ranges to which we'd like to limit for some set of those types.
For example, POWER target, register num is special when VSX is enabled. When VSX is enabled, the number of int scalar register is 32(GPR),
float is 64(VSR), but for int and float vector register both are 64(VSR). So there should be 2 kinds of register class when vsx is enabled,
and 3 kinds of register class when VSX is NOT enabled.
It runs on POWER target, it makes big(+~30%) performance improvement in one specific bmk(503.bwaves_r) of spec2017 and no other obvious degressions.
Differential revision: https://reviews.llvm.org/D67148
llvm-svn: 374017
This reverts SVN r373833, as it caused a failed assert "Non-zero loop
cost expected" on building numerous projects, see PR43582 for details
and reproduction samples.
llvm-svn: 373882
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
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
Summary:
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: 372626
The static analyzer is warning about potential null dereferences of dyn_cast<> results - in these cases we can safely use cast<> directly as we know that these cases should all be the correct type, which is why its working atm and anyway cast<> will assert if they aren't.
llvm-svn: 371973
This is a fix for:
https://bugs.llvm.org/show_bug.cgi?id=33958
It seems universally true that we would not want to transform this kind of
sequence on any target, but if that's not correct, then we could view this
as a target-specific cost model problem. We could also white-list ConstantInt,
ConstantFP, etc. rather than blacklist Global and ConstantExpr.
Differential Revision: https://reviews.llvm.org/D67362
llvm-svn: 371931
Summary:
This is the first change to enable the TLI to be built per-function so
that -fno-builtin* handling can be migrated to use function attributes.
See discussion on D61634 for background. This is an enabler for fixing
handling of these options for LTO, for example.
This change should not affect behavior, as the provided function is not
yet used to build a specifically per-function TLI, but rather enables
that migration.
Most of the changes were very mechanical, e.g. passing a Function to the
legacy analysis pass's getTLI interface, or in Module level cases,
adding a callback. This is similar to the way the per-function TTI
analysis works.
There was one place where we were looking for builtins but not in the
context of a specific function. See FindCXAAtExit in
lib/Transforms/IPO/GlobalOpt.cpp. I'm somewhat concerned my workaround
could provide the wrong behavior in some corner cases. Suggestions
welcome.
Reviewers: chandlerc, hfinkel
Subscribers: arsenm, dschuff, jvesely, nhaehnle, mehdi_amini, javed.absar, sbc100, jgravelle-google, eraman, aheejin, steven_wu, george.burgess.iv, dexonsmith, jfb, asbirlea, gchatelet, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D66428
llvm-svn: 371284
We can avoid repetitive calls getSameOpcode() for already known tree elements by keeping MainOp and AltOp in TreeEntry.
Differential Revision: https://reviews.llvm.org/D64700
llvm-svn: 369315
Summary:
The scheduler's dependence graph gets the use-def dependencies by accessing the operands of the instructions in a bundle. However, buildTree_rec() may change the order of the operands in TreeEntry, and the scheduler is currently not aware of this. This is not causing any functional issues currently, because reordering is restricted to the operands of a single instruction. Once we support operand reordering across multiple TreeEntries, as shown here: http://www.llvm.org/devmtg/2019-04/slides/Poster-Porpodas-Supernode_SLP.pdf , the scheduler will need to get the correct operands from TreeEntry and not from the individual instructions.
In short, this patch:
- Connects the scheduler's bundle with the corresponding TreeEntry. It introduces new TE and Lane fields in ScheduleData.
- Moves the location where the operands of the TreeEntry are initialized. This used to take place in newTreeEntry() setting one operand at a time, but is now moved pre-order just before the recursion of buildTree_rec(). This is required because the scheduler needs to access both operands of the TreeEntry in tryScheduleBundle().
- Updates the scheduler to access the instruction operands through the TreeEntry operands instead of accessing the instruction operands directly.
Reviewers: ABataev, RKSimon, dtemirbulatov, Ayal, dorit, hfinkel
Reviewed By: ABataev
Subscribers: hiraditya, llvm-commits, lebedev.ri, rcorcs
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D62432
llvm-svn: 369131
cppcheck + MSVC analyzer both over zealously warn that we might dereference a null Bundle pointer - add an assertion to check for null to silence the warning, plus its a good idea to check that we succeeded in finding a schedule bundle anyway....
llvm-svn: 369094
Now that we've moved to C++14, we no longer need the llvm::make_unique
implementation from STLExtras.h. This patch is a mechanical replacement
of (hopefully) all the llvm::make_unique instances across the monorepo.
llvm-svn: 369013
This patch introduces the DAG version of SimplifyMultipleUseDemandedBits, which attempts to peek through ops (mainly and/or/xor so far) that don't contribute to the demandedbits/elts of a node - which means we can do this even in cases where we have multiple uses of an op, which normally requires us to demanded all bits/elts. The intention is to remove a similar instruction - SelectionDAG::GetDemandedBits - once SimplifyMultipleUseDemandedBits has matured.
The InstCombine version of SimplifyMultipleUseDemandedBits can constant fold which I haven't added here yet, and so far I've only wired this up to some basic binops (and/or/xor/add/sub/mul) to demonstrate its use.
We do see a couple of regressions that need to be addressed:
AMDGPU unsigned dot product codegen retains an AND mask (for ZERO_EXTEND) that it previously removed (but otherwise the dotproduct codegen is a lot better).
X86/AVX2 has poor handling of vector ANY_EXTEND/ANY_EXTEND_VECTOR_INREG - it prematurely gets converted to ZERO_EXTEND_VECTOR_INREG.
The code owners have confirmed its ok for these cases to fixed up in future patches.
Differential Revision: https://reviews.llvm.org/D63281
llvm-svn: 366799
cast<CallInst> shouldn't return null and we dereference the pointer in a lot of other places, causing both MSVC + cppcheck to warn about dereferenced null pointers
llvm-svn: 366793
As there are some reported miscompiles with AVX512 and performance regressions
in Eigen. Verified with the original committer and testcases will be forthcoming.
This reverts commit r364964.
llvm-svn: 366154
For a given set of live values, the spill cost will always be the
same for each call. Compute the cost once and multiply it by the
number of calls.
(I'm not sure this spill cost modeling makes sense if there are
multiple calls, as the spill cost will likely be shared across
calls in that case. But that's how it currently works.)
llvm-svn: 365552
Summary: This patch introduces a new heuristic for guiding operand reordering. The new "look-ahead" heuristic can look beyond the immediate predecessors. This helps break ties when the immediate predecessors have identical opcodes (see lit test for an example).
Reviewers: RKSimon, ABataev, dtemirbulatov, Ayal, hfinkel, rnk
Reviewed By: RKSimon, dtemirbulatov
Subscribers: hiraditya, phosek, rnk, rcorcs, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D60897
llvm-svn: 364964
Summary: This patch introduces a new heuristic for guiding operand reordering. The new "look-ahead" heuristic can look beyond the immediate predecessors. This helps break ties when the immediate predecessors have identical opcodes (see lit test for an example).
Reviewers: RKSimon, ABataev, dtemirbulatov, Ayal, hfinkel, rnk
Reviewed By: RKSimon, dtemirbulatov
Subscribers: rnk, rcorcs, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D60897
llvm-svn: 364478
