Summary: The specific number of records loaded depends on the number of kinds, but the difference between the lazy and not lazy cases does not.
Reviewers: modocache
Subscribers: llvm-commits, dexonsmith, steven_wu, hiraditya, mehdi_amini
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D71882
Summary: The specific number of records loaded depends on the number of kinds, but the difference between the lazy and not lazy cases does not.
Reviewers: modocache
Subscribers: llvm-commits, dexonsmith, steven_wu, hiraditya, mehdi_amini
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D71882
Summary: The specific number of records loaded depends on the number of kinds, but the difference between the lazy and not lazy cases does not.
Reviewers: modocache
Subscribers: mehdi_amini, hiraditya, steven_wu, dexonsmith, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D71730
Remove dead virtual functions from vtables with
replaceNonMetadataUsesWith, so that CGProfile metadata gets cleaned up
correctly.
Original commit message:
Currently, it is hard for the compiler to remove unused C++ virtual
functions, because they are all referenced from vtables, which are referenced
by constructors. This means that if the constructor is called from any live
code, then we keep every virtual function in the final link, even if there
are no call sites which can use it.
This patch allows unused virtual functions to be removed during LTO (and
regular compilation in limited circumstances) by using type metadata to match
virtual function call sites to the vtable slots they might load from. This
information can then be used in the global dead code elimination pass instead
of the references from vtables to virtual functions, to more accurately
determine which functions are reachable.
To make this transformation safe, I have changed clang's code-generation to
always load virtual function pointers using the llvm.type.checked.load
intrinsic, instead of regular load instructions. I originally tried writing
this using clang's existing code-generation, which uses the llvm.type.test
and llvm.assume intrinsics after doing a normal load. However, it is possible
for optimisations to obscure the relationship between the GEP, load and
llvm.type.test, causing GlobalDCE to fail to find virtual function call
sites.
The existing linkage and visibility types don't accurately describe the scope
in which a virtual call could be made which uses a given vtable. This is
wider than the visibility of the type itself, because a virtual function call
could be made using a more-visible base class. I've added a new
!vcall_visibility metadata type to represent this, described in
TypeMetadata.rst. The internalization pass and libLTO have been updated to
change this metadata when linking is performed.
This doesn't currently work with ThinLTO, because it needs to see every call
to llvm.type.checked.load in the linkage unit. It might be possible to
extend this optimisation to be able to use the ThinLTO summary, as was done
for devirtualization, but until then that combination is rejected in the
clang driver.
To test this, I've written a fuzzer which generates random C++ programs with
complex class inheritance graphs, and virtual functions called through object
and function pointers of different types. The programs are spread across
multiple translation units and DSOs to test the different visibility
restrictions.
I've also tried doing bootstrap builds of LLVM to test this. This isn't
ideal, because only classes in anonymous namespaces can be optimised with
-fvisibility=default, and some parts of LLVM (plugins and bugpoint) do not
work correctly with -fvisibility=hidden. However, there are only 12 test
failures when building with -fvisibility=hidden (and an unmodified compiler),
and this change does not cause any new failures for either value of
-fvisibility.
On the 7 C++ sub-benchmarks of SPEC2006, this gives a geomean code-size
reduction of ~6%, over a baseline compiled with "-O2 -flto
-fvisibility=hidden -fwhole-program-vtables". The best cases are reductions
of ~14% in 450.soplex and 483.xalancbmk, and there are no code size
increases.
I've also run this on a set of 8 mbed-os examples compiled for Armv7M, which
show a geomean size reduction of ~3%, again with no size increases.
I had hoped that this would have no effect on performance, which would allow
it to awlays be enabled (when using -fwhole-program-vtables). However, the
changes in clang to use the llvm.type.checked.load intrinsic are causing ~1%
performance regression in the C++ parts of SPEC2006. It should be possible to
recover some of this perf loss by teaching optimisations about the
llvm.type.checked.load intrinsic, which would make it worth turning this on
by default (though it's still dependent on -fwhole-program-vtables).
Differential revision: https://reviews.llvm.org/D63932
llvm-svn: 375094
Currently, it is hard for the compiler to remove unused C++ virtual
functions, because they are all referenced from vtables, which are referenced
by constructors. This means that if the constructor is called from any live
code, then we keep every virtual function in the final link, even if there
are no call sites which can use it.
This patch allows unused virtual functions to be removed during LTO (and
regular compilation in limited circumstances) by using type metadata to match
virtual function call sites to the vtable slots they might load from. This
information can then be used in the global dead code elimination pass instead
of the references from vtables to virtual functions, to more accurately
determine which functions are reachable.
To make this transformation safe, I have changed clang's code-generation to
always load virtual function pointers using the llvm.type.checked.load
intrinsic, instead of regular load instructions. I originally tried writing
this using clang's existing code-generation, which uses the llvm.type.test
and llvm.assume intrinsics after doing a normal load. However, it is possible
for optimisations to obscure the relationship between the GEP, load and
llvm.type.test, causing GlobalDCE to fail to find virtual function call
sites.
The existing linkage and visibility types don't accurately describe the scope
in which a virtual call could be made which uses a given vtable. This is
wider than the visibility of the type itself, because a virtual function call
could be made using a more-visible base class. I've added a new
!vcall_visibility metadata type to represent this, described in
TypeMetadata.rst. The internalization pass and libLTO have been updated to
change this metadata when linking is performed.
This doesn't currently work with ThinLTO, because it needs to see every call
to llvm.type.checked.load in the linkage unit. It might be possible to
extend this optimisation to be able to use the ThinLTO summary, as was done
for devirtualization, but until then that combination is rejected in the
clang driver.
To test this, I've written a fuzzer which generates random C++ programs with
complex class inheritance graphs, and virtual functions called through object
and function pointers of different types. The programs are spread across
multiple translation units and DSOs to test the different visibility
restrictions.
I've also tried doing bootstrap builds of LLVM to test this. This isn't
ideal, because only classes in anonymous namespaces can be optimised with
-fvisibility=default, and some parts of LLVM (plugins and bugpoint) do not
work correctly with -fvisibility=hidden. However, there are only 12 test
failures when building with -fvisibility=hidden (and an unmodified compiler),
and this change does not cause any new failures for either value of
-fvisibility.
On the 7 C++ sub-benchmarks of SPEC2006, this gives a geomean code-size
reduction of ~6%, over a baseline compiled with "-O2 -flto
-fvisibility=hidden -fwhole-program-vtables". The best cases are reductions
of ~14% in 450.soplex and 483.xalancbmk, and there are no code size
increases.
I've also run this on a set of 8 mbed-os examples compiled for Armv7M, which
show a geomean size reduction of ~3%, again with no size increases.
I had hoped that this would have no effect on performance, which would allow
it to awlays be enabled (when using -fwhole-program-vtables). However, the
changes in clang to use the llvm.type.checked.load intrinsic are causing ~1%
performance regression in the C++ parts of SPEC2006. It should be possible to
recover some of this perf loss by teaching optimisations about the
llvm.type.checked.load intrinsic, which would make it worth turning this on
by default (though it's still dependent on -fwhole-program-vtables).
Differential revision: https://reviews.llvm.org/D63932
llvm-svn: 374539
This patch contains the basic functionality for reporting potentially
incorrect usage of __builtin_expect() by comparing the developer's
annotation against a collected PGO profile. A more detailed proposal and
discussion appears on the CFE-dev mailing list
(http://lists.llvm.org/pipermail/cfe-dev/2019-July/062971.html) and a
prototype of the initial frontend changes appear here in D65300
We revised the work in D65300 by moving the misexpect check into the
LLVM backend, and adding support for IR and sampling based profiles, in
addition to frontend instrumentation.
We add new misexpect metadata tags to those instructions directly
influenced by the llvm.expect intrinsic (branch, switch, and select)
when lowering the intrinsics. The misexpect metadata contains
information about the expected target of the intrinsic so that we can
check against the correct PGO counter when emitting diagnostics, and the
compiler's values for the LikelyBranchWeight and UnlikelyBranchWeight.
We use these branch weight values to determine when to emit the
diagnostic to the user.
A future patch should address the comment at the top of
LowerExpectIntrisic.cpp to hoist the LikelyBranchWeight and
UnlikelyBranchWeight values into a shared space that can be accessed
outside of the LowerExpectIntrinsic pass. Once that is done, the
misexpect metadata can be updated to be smaller.
In the long term, it is possible to reconstruct portions of the
misexpect metadata from the existing profile data. However, we have
avoided this to keep the code simple, and because some kind of metadata
tag will be required to identify which branch/switch/select instructions
are influenced by the use of llvm.expect
Patch By: paulkirth
Differential Revision: https://reviews.llvm.org/D66324
llvm-svn: 371635
This reverts commit r371584. It introduced a dependency from compiler-rt
to llvm/include/ADT, which is problematic for multiple reasons.
One is that it is a novel dependency edge, which needs cross-compliation
machinery for llvm/include/ADT (yes, it is true that right now
compiler-rt included only header-only libraries, however, if we allow
compiler-rt to depend on anything from ADT, other libraries will
eventually get used).
Secondly, depending on ADT from compiler-rt exposes ADT symbols from
compiler-rt, which would cause ODR violations when Clang is built with
the profile library.
llvm-svn: 371598
This patch contains the basic functionality for reporting potentially
incorrect usage of __builtin_expect() by comparing the developer's
annotation against a collected PGO profile. A more detailed proposal and
discussion appears on the CFE-dev mailing list
(http://lists.llvm.org/pipermail/cfe-dev/2019-July/062971.html) and a
prototype of the initial frontend changes appear here in D65300
We revised the work in D65300 by moving the misexpect check into the
LLVM backend, and adding support for IR and sampling based profiles, in
addition to frontend instrumentation.
We add new misexpect metadata tags to those instructions directly
influenced by the llvm.expect intrinsic (branch, switch, and select)
when lowering the intrinsics. The misexpect metadata contains
information about the expected target of the intrinsic so that we can
check against the correct PGO counter when emitting diagnostics, and the
compiler's values for the LikelyBranchWeight and UnlikelyBranchWeight.
We use these branch weight values to determine when to emit the
diagnostic to the user.
A future patch should address the comment at the top of
LowerExpectIntrisic.cpp to hoist the LikelyBranchWeight and
UnlikelyBranchWeight values into a shared space that can be accessed
outside of the LowerExpectIntrinsic pass. Once that is done, the
misexpect metadata can be updated to be smaller.
In the long term, it is possible to reconstruct portions of the
misexpect metadata from the existing profile data. However, we have
avoided this to keep the code simple, and because some kind of metadata
tag will be required to identify which branch/switch/select instructions
are influenced by the use of llvm.expect
Patch By: paulkirth
Differential Revision: https://reviews.llvm.org/D66324
llvm-svn: 371584
This patch contains the basic functionality for reporting potentially
incorrect usage of __builtin_expect() by comparing the developer's
annotation against a collected PGO profile. A more detailed proposal and
discussion appears on the CFE-dev mailing list
(http://lists.llvm.org/pipermail/cfe-dev/2019-July/062971.html) and a
prototype of the initial frontend changes appear here in D65300
We revised the work in D65300 by moving the misexpect check into the
LLVM backend, and adding support for IR and sampling based profiles, in
addition to frontend instrumentation.
We add new misexpect metadata tags to those instructions directly
influenced by the llvm.expect intrinsic (branch, switch, and select)
when lowering the intrinsics. The misexpect metadata contains
information about the expected target of the intrinsic so that we can
check against the correct PGO counter when emitting diagnostics, and the
compiler's values for the LikelyBranchWeight and UnlikelyBranchWeight.
We use these branch weight values to determine when to emit the
diagnostic to the user.
A future patch should address the comment at the top of
LowerExpectIntrisic.cpp to hoist the LikelyBranchWeight and
UnlikelyBranchWeight values into a shared space that can be accessed
outside of the LowerExpectIntrinsic pass. Once that is done, the
misexpect metadata can be updated to be smaller.
In the long term, it is possible to reconstruct portions of the
misexpect metadata from the existing profile data. However, we have
avoided this to keep the code simple, and because some kind of metadata
tag will be required to identify which branch/switch/select instructions
are influenced by the use of llvm.expect
Patch By: paulkirth
Differential Revision: https://reviews.llvm.org/D66324
llvm-svn: 371484
For background of BPF CO-RE project, please refer to
http://vger.kernel.org/bpfconf2019.html
In summary, BPF CO-RE intends to compile bpf programs
adjustable on struct/union layout change so the same
program can run on multiple kernels with adjustment
before loading based on native kernel structures.
In order to do this, we need keep track of GEP(getelementptr)
instruction base and result debuginfo types, so we
can adjust on the host based on kernel BTF info.
Capturing such information as an IR optimization is hard
as various optimization may have tweaked GEP and also
union is replaced by structure it is impossible to track
fieldindex for union member accesses.
Three intrinsic functions, preserve_{array,union,struct}_access_index,
are introducted.
addr = preserve_array_access_index(base, index, dimension)
addr = preserve_union_access_index(base, di_index)
addr = preserve_struct_access_index(base, gep_index, di_index)
here,
base: the base pointer for the array/union/struct access.
index: the last access index for array, the same for IR/DebugInfo layout.
dimension: the array dimension.
gep_index: the access index based on IR layout.
di_index: the access index based on user/debuginfo types.
For example, for the following example,
$ cat test.c
struct sk_buff {
int i;
int b1:1;
int b2:2;
union {
struct {
int o1;
int o2;
} o;
struct {
char flags;
char dev_id;
} dev;
int netid;
} u[10];
};
static int (*bpf_probe_read)(void *dst, int size, const void *unsafe_ptr)
= (void *) 4;
#define _(x) (__builtin_preserve_access_index(x))
int bpf_prog(struct sk_buff *ctx) {
char dev_id;
bpf_probe_read(&dev_id, sizeof(char), _(&ctx->u[5].dev.dev_id));
return dev_id;
}
$ clang -target bpf -O2 -g -emit-llvm -S -mllvm -print-before-all \
test.c >& log
The generated IR looks like below:
...
define dso_local i32 @bpf_prog(%struct.sk_buff*) #0 !dbg !15 {
%2 = alloca %struct.sk_buff*, align 8
%3 = alloca i8, align 1
store %struct.sk_buff* %0, %struct.sk_buff** %2, align 8, !tbaa !45
call void @llvm.dbg.declare(metadata %struct.sk_buff** %2, metadata !43, metadata !DIExpression()), !dbg !49
call void @llvm.lifetime.start.p0i8(i64 1, i8* %3) #4, !dbg !50
call void @llvm.dbg.declare(metadata i8* %3, metadata !44, metadata !DIExpression()), !dbg !51
%4 = load i32 (i8*, i32, i8*)*, i32 (i8*, i32, i8*)** @bpf_probe_read, align 8, !dbg !52, !tbaa !45
%5 = load %struct.sk_buff*, %struct.sk_buff** %2, align 8, !dbg !53, !tbaa !45
%6 = call [10 x %union.anon]* @llvm.preserve.struct.access.index.p0a10s_union.anons.p0s_struct.sk_buffs(
%struct.sk_buff* %5, i32 2, i32 3), !dbg !53, !llvm.preserve.access.index !19
%7 = call %union.anon* @llvm.preserve.array.access.index.p0s_union.anons.p0a10s_union.anons(
[10 x %union.anon]* %6, i32 1, i32 5), !dbg !53
%8 = call %union.anon* @llvm.preserve.union.access.index.p0s_union.anons.p0s_union.anons(
%union.anon* %7, i32 1), !dbg !53, !llvm.preserve.access.index !26
%9 = bitcast %union.anon* %8 to %struct.anon.0*, !dbg !53
%10 = call i8* @llvm.preserve.struct.access.index.p0i8.p0s_struct.anon.0s(
%struct.anon.0* %9, i32 1, i32 1), !dbg !53, !llvm.preserve.access.index !34
%11 = call i32 %4(i8* %3, i32 1, i8* %10), !dbg !52
%12 = load i8, i8* %3, align 1, !dbg !54, !tbaa !55
%13 = sext i8 %12 to i32, !dbg !54
call void @llvm.lifetime.end.p0i8(i64 1, i8* %3) #4, !dbg !56
ret i32 %13, !dbg !57
}
!19 = distinct !DICompositeType(tag: DW_TAG_structure_type, name: "sk_buff", file: !3, line: 1, size: 704, elements: !20)
!26 = distinct !DICompositeType(tag: DW_TAG_union_type, scope: !19, file: !3, line: 5, size: 64, elements: !27)
!34 = distinct !DICompositeType(tag: DW_TAG_structure_type, scope: !26, file: !3, line: 10, size: 16, elements: !35)
Note that @llvm.preserve.{struct,union}.access.index calls have metadata llvm.preserve.access.index
attached to instructions to provide struct/union debuginfo type information.
For &ctx->u[5].dev.dev_id,
. The "%6 = ..." represents struct member "u" with index 2 for IR layout and index 3 for DI layout.
. The "%7 = ..." represents array subscript "5".
. The "%8 = ..." represents union member "dev" with index 1 for DI layout.
. The "%10 = ..." represents struct member "dev_id" with index 1 for both IR and DI layout.
Basically, traversing the use-def chain recursively for the 3rd argument of bpf_probe_read() and
examining all preserve_*_access_index calls, the debuginfo struct/union/array access index
can be achieved.
The intrinsics also contain enough information to regenerate codes for IR layout.
For array and structure intrinsics, the proper GEP can be constructed.
For union intrinsics, replacing all uses of "addr" with "base" should be enough.
The test case ThinLTO/X86/lazyload_metadata.ll is adjusted to reflect the
new addition of the metadata.
Signed-off-by: Yonghong Song <yhs@fb.com>
Differential Revision: https://reviews.llvm.org/D61810
llvm-svn: 365423
An abstract call site is a wrapper that allows to treat direct,
indirect, and callback calls the same. If an abstract call site
represents a direct or indirect call site it behaves like a stripped
down version of a normal call site object. The abstract call site can
also represent a callback call, thus the fact that the initially
called function (=broker) may invoke a third one (=callback callee).
In this case, the abstract call side hides the middle man, hence the
broker function. The result is a representation of the callback call,
inside the broker, but in the context of the original instruction that
invoked the broker.
Again, there are up to three functions involved when we talk about
callback call sites. The caller (1), which invokes the broker
function. The broker function (2), that may or may not invoke the
callback callee. And finally the callback callee (3), which is the
target of the callback call.
The abstract call site will handle the mapping from parameters to
arguments depending on the semantic of the broker function. However,
it is important to note that the mapping is often partial. Thus, some
arguments of the call/invoke instruction are mapped to parameters of
the callee while others are not. At the same time, arguments of the
callback callee might be unknown, thus "null" if queried.
This patch introduces also !callback metadata which describe how a
callback broker maps from parameters to arguments. This metadata is
directly created by clang for known broker functions, provided through
source code attributes by the user, or later deduced by analyses.
For motivation and additional information please see the corresponding
talk (slides/video)
https://llvm.org/devmtg/2018-10/talk-abstracts.html#talk20
as well as the LCPC paper
http://compilers.cs.uni-saarland.de/people/doerfert/par_opt_lcpc18.pdf
Differential Revision: https://reviews.llvm.org/D54498
llvm-svn: 351627
The current llvm.mem.parallel_loop_access metadata has a problem in that
it uses LoopIDs. LoopID unfortunately is not loop identifier. It is
neither unique (there's even a regression test assigning the some LoopID
to multiple loops; can otherwise happen if passes such as LoopVersioning
make copies of entire loops) nor persistent (every time a property is
removed/added from a LoopID's MDNode, it will also receive a new LoopID;
this happens e.g. when calling Loop::setLoopAlreadyUnrolled()).
Since most loop transformation passes change the loop attributes (even
if it just to mark that a loop should not be processed again as
llvm.loop.isvectorized does, for the versioned and unversioned loop),
the parallel access information is lost for any subsequent pass.
This patch unlinks LoopIDs and parallel accesses.
llvm.mem.parallel_loop_access metadata on instruction is replaced by
llvm.access.group metadata. llvm.access.group points to a distinct
MDNode with no operands (avoiding the problem to ever need to add/remove
operands), called "access group". Alternatively, it can point to a list
of access groups. The LoopID then has an attribute
llvm.loop.parallel_accesses with all the access groups that are parallel
(no dependencies carries by this loop).
This intentionally avoid any kind of "ID". Loops that are clones/have
their attributes modifies retain the llvm.loop.parallel_accesses
attribute. Access instructions that a cloned point to the same access
group. It is not necessary for each access to have it's own "ID" MDNode,
but those memory access instructions with the same behavior can be
grouped together.
The behavior of llvm.mem.parallel_loop_access is not changed by this
patch, but should be considered deprecated.
Differential Revision: https://reviews.llvm.org/D52116
llvm-svn: 349725
Summary:
Currently the block frequency analysis is an approximation for irreducible
loops.
The new irreducible loop metadata is used to annotate the irreducible loop
headers with their header weights based on the PGO profile (currently this is
approximated to be evenly weighted) and to help improve the accuracy of the
block frequency analysis for irreducible loops.
This patch is a basic support for this.
Reviewers: davidxl
Reviewed By: davidxl
Subscribers: mehdi_amini, llvm-commits, eraman
Differential Revision: https://reviews.llvm.org/D39028
llvm-svn: 317278
This patch adds a new kind of metadata that indicates the possible callees of
indirect calls.
Differential Revision: https://reviews.llvm.org/D37354
llvm-svn: 315944
This is an ELF-specific thing that adds SHF_LINK_ORDER to the global's section
pointing to the metadata argument's section. The effect of that is a reverse dependency
between sections for the linker GC.
!associated does not change the behavior of global-dce. The global
may also need to be added to llvm.compiler.used.
Since SHF_LINK_ORDER is per-section, !associated effectively enables
fdata-sections for the affected globals, the same as comdats do.
Differential Revision: https://reviews.llvm.org/D29104
llvm-svn: 298157
CFI is using intrinsics that takes MDString as arguments, and this
was broken during lazy-loading of metadata.
Differential Revision: https://reviews.llvm.org/D28916
llvm-svn: 292641
Summary:
Without this, we're stressing the RAUW of unique nodes,
which is a costly operation. This is intended to limit
the number of RAUW, and is very effective on the total
link-time of opt with ThinLTO, before:
real 4m4.587s user 15m3.401s sys 0m23.616s
after:
real 3m25.261s user 12m22.132s sys 0m24.152s
Reviewers: tejohnson, pcc
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D28751
llvm-svn: 292420
Summary:
This is a relatively simple scheme: we use the index emitted in the
bitcode to avoid loading all the global metadata. Instead we load
the index with their position in the bitcode so that we can load each
of them individually. Materializing the global metadata block in this
condition only triggers loading the named metadata, and the ones
referenced from there (transitively). When materializing a function,
metadata from the global block are loaded lazily as they are
referenced.
Two main current limitations are:
1) Global values other than functions are not materialized on demand,
so we need to eagerly load METADATA_GLOBAL_DECL_ATTACHMENT records
(and their transitive dependencies).
2) When we load a single metadata, we don't recurse on the operands,
instead we use a placeholder or a temporary metadata. Unfortunately
tepmorary nodes are very expensive. This is why we don't have it
always enabled and only for importing.
These two limitations can be lifted in a subsequent improvement if
needed.
With this change, the total link time of opt with ThinLTO and Debug
Info enabled is going down from 282s to 224s (~20%).
Reviewers: pcc, tejohnson, dexonsmith
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D28113
llvm-svn: 291027
David Herzka