On AVX512F targets we'll produce an emulated sequence using 3 pmuludqs with shifts and adds. On AVX512DQ we'll use vpmulld.
Fixes PR37140.
llvm-svn: 330923
instructions.
These have special permission according to the x86 manual to read
unaligned memory, and this folding is done by ICC and GCC as well.
This corrects one of the issues identified in PR37246.
llvm-svn: 330896
comparison instructions (pcmp[ei]stri*).
These will help show improvements from fixes to PR37246.
I've not really covered the mask forms of this intrinsic as I don't have
as good of an intuition about the likely usage patterns there. Happy for
someone to extend this with tests covering the mask form.
llvm-svn: 330895
a preceding 'template' keyword.
We only diagnose in the dependent case (wherein we used to crash). Another bug
prevents the diagnostic from appearing in the non-template case.
llvm-svn: 330894
This reverts commit 023c8be90980e0180766196cba86f81608b35d38.
This patch triggers miscompile of zlib on PowerPC platform. Most likely it is
caused by some pre-backend PPC-specific pass, but we don't clearly know the
reason yet. So we temporally revert this patch with intention to return it
once the problem is resolved. See bug 37229 for details.
llvm-svn: 330893
std::isdigit can be overloaded, causing the template deduction to fail. Use
Clang's isDigit function which to avoid this. Switch the other calls for
consistency.
llvm-svn: 330887
Be defensive against a reentrant std::function::operator=(nullptr_t), in case
the held function object has a non-trivial destructor. Destroying the function
object in-place can lead to the destructor being called twice.
Patch by Duncan P. N. Exon Smith. C++03 support by Volodymyr Sapsai.
rdar://problem/32836603
Reviewers: EricWF, mclow.lists
Reviewed By: mclow.lists
Subscribers: cfe-commits, arphaman
Differential Revision: https://reviews.llvm.org/D34331
llvm-svn: 330885
If no data or instructions are emitted after a location directive, we
should clear the cv_loc when we change sections, or it will be emitted
at the beginning of the next section. This violates our invariant that
all .cv_loc directives belong to the same section. Add clearer
assertions for this.
llvm-svn: 330884
Summary:
When a symbol refers to a special section or a section that doesn't
exist, lld would fatal with "broken object file". This change gives a
different message for each scenario, and includes the name of the
file, name of the symbol, and the section being referred to.
Reviewers: pcc, ruiu
Reviewed By: ruiu
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D46090
llvm-svn: 330883
If 'A' is a C++ aggregate with a reference field of type 'C', in code like
A a = { C() };
C() is lifetime-extended by 'a'. The analyzer wasn't expecting this pattern and
crashing. Additionally, destructors aren't added in the CFG for this case,
so for now we shouldn't be inlining the constructor for C().
Differential Revision: https://reviews.llvm.org/D46037
llvm-svn: 330882
Supporting additional rules for parsing ISA string.
- RISC-V ISA strings must be lowercase.
E.g.: rv32IMC is not supported, rv32imc is correct.
- Multi-letter extensions are to be separated by a single
underscore '_'. The extension prefix counts as a letter.
This means extensions that start with 's', 'sx' and 'sx'
are all multi-letter.
E.g.:
xasb is a single non-standard extension named 'xasb'
xa_sb are two extensions, the non-standard user level extension
'xa', and the supervisor level extension 'sb'.
- Standard user-level extensions are specified following
a canonical order, according to Table 22.1 in
RISC-V User-Level ISA V2.2.
- Non-standard user-level 'x' extensions,
standard supervisor-level 's' extensions and
non-standard supervisor-level 'sx' extensions
are also specified following a canonical order according
to Table 22.1 in RISC-V User-Level ISA V2.2:
'x' extensions, follwed by 's' extensions and then 'sx' extensions.
- Extensions might have a version number.
Underscores may be used to separate ISA subset components to
improve readability and to provide disambiguation.
E.g.: rv32i2_m3_a1_f2_d2
- Version numbers are divided into major and minor numbers,
separated by a 'p'. If the minor version is 0, then 'p0' can
be omitted.
- Additional checks for dependent extensions and invalid
extensions combinations.
E.g.:
'e' requires rv32
'e' can't be combined with 'f' nor 'd'
'q' requires rv64
- TODO items have also been marked with comments in the code.
Reviewers: asb, kito-cheng
Reviewed By: asb
Subscribers: edward-jones, mgrang, zzheng, rbar, johnrusso, simoncook, jordy.potman.lists, sabuasal, niosHD, shiva0217, cfe-commits
Differential Revision: https://reviews.llvm.org/D45284
llvm-svn: 330880
When I merged the 2 codepaths that return an OS type, I hade
checked that the places accepting 'iphoneos' would also accept
'ios', but then I got it backwards and return 'iphoneos'.
We use this value to build triples, and there 'iphoneos' is
invalid.
This also makes the test slightly simpler.
llvm-svn: 330877
Normally the analyzer begins path-sensitive analysis from functions within
the main file, even though the path is allowed to go through any functions
within the translation unit.
When a recent version of WebKit is compiled, the "unified sources" technique
is used, that assumes #including multiple code files into a single main file.
Such file would have no functions defined in it, so the analyzer wouldn't be
able to find any entry points for path-sensitive analysis.
This patch pattern-matches unified file names that are similar to those
used by WebKit and allows the analyzer to find entry points in the included
code files. A more aggressive/generic approach is being planned as well.
Differential Revision: https://reviews.llvm.org/D45839
llvm-svn: 330876
This makes it possible to reverse a filtered range. For example, here's
a way to visit memory accesses in a BasicBlock in reverse order:
auto MemInsts = reverse(make_filter_range(BB, [](Instruction &I) {
return isa<StoreInst>(&I) || isa<LoadInst>(&I);
}));
for (auto &MI : MemInsts)
...
To implement this functionality, I factored out forward iteration
functionality into filter_iterator_base, and added a specialization of
filter_iterator_impl which supports bidirectional iteration. Thanks to
Tim Shen, Zachary Turner, and others for suggesting this design and
providing feedback! This version of the patch supersedes the original
(https://reviews.llvm.org/D45792).
This was motivated by a problem we encountered in D45657: we'd like to
visit the non-debug-info instructions in a BasicBlock in reverse order.
Testing: check-llvm, check-clang
Differential Revision: https://reviews.llvm.org/D45853
llvm-svn: 330875
Before this patch:
Symbol 56
Defined 80
Undefined 56
SharedSymbol 88
LazyArchive 72
LazyObject 56
With this patch
Symbol 48
Defined 72
Undefined 48
SharedSymbol 80
LazyArchive 64
LazyObject 48
The result is that peak allocation when linking chromium (according to
heaptrack) goes from 578 to 568 MB.
llvm-svn: 330874
Fixed directory separators in tests to be compatible with both
Windows and !Windows.
This reverts commit aa423850afa4c16a53c4c492fe254dcad3d5a53e.
llvm-svn: 330873
If -no-canonical-prefixes isn't used, the clang executable name used
is the one of the actual executable, not the name of the symlink that
the user invoked.
In these cases, the target prefix was overridden based on the clang
executable name. (On the other hand the implicit -target option
that such a symlink adds, is added as an actual command line parameter
in tools/driver/driver.cop, before resolving the symlink and finding
the actual clang executable.
Use the original ClangNameParts (set from argv[0] in
tools/driver/driver.cpp) if it seems to be initialized propery.
All existing tests of this feature used -no-canonical-prefixes
(possibly because it also makes the driver look in the directory
of the symlink instead of the directory of the executable); add
another one that uses --config-user-dir= to specify the directory
instead. (For actual users of such symlinks, outisde of the test
suite, the directory is probably the same for both.)
This makes this feature work more like what the documentation
describes.
Differential Revision: https://reviews.llvm.org/D45964
llvm-svn: 330871
Algorithmically compute the 'x20' SDIV/UDIV vector costs - this is necessary for PR36550 when DIV costs will be driven from the scheduler models.
llvm-svn: 330870
lit is picking up a stale executable in the unittests tree, which is
failing on Windows.
To simplify the CMake and avoid problems like this in the future, now we
always compile the test, but the test exits successfully when plugins
are not enabled.
llvm-svn: 330867
Add the options -polly-codegen-trace-stmts and
-polly-codegen-trace-scalars. When enabled, adds a call to the
beginning of every generated statement that prints the executed
statement instance. With -polly-codegen-trace-scalars, it also prints
the value of all scalars that are used in the statement, and PHIs
defined in the beginning of the statement.
Differential Revision: https://reviews.llvm.org/D45743
llvm-svn: 330864
Follow up to r330840 and r330849.
It seems that s390 is also not distinguishing illegal WRITE and READ memory
access.
Add s390x to XFAIL for the tests.
llvm-svn: 330863
- Add "amdgpu-waitcnt-forcezero" to force all waitcnt instrs to be emitted as s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
- Add debug counters to control force emit of s_waitcnt instrs; debug counters:
si-insert-waitcnts-forceexp: force emit s_waitcnt expcnt(0) instrs
si-insert-waitcnts-forcevm: force emit s_waitcnt lgkmcnt(0) instrs
si-insert-waitcnts-forcelgkm: force emit s_waitcnt vmcnt(0) instrs
- Add some debug statements
Note that a variant of this patch was previously committed/reverted.
Differential Revision: https://reviews.llvm.org/D45888
llvm-svn: 330862
Passing the features in random order will lead to unpredictable results
when some of the features are related (like the architecture-version
features on ARM).
It might be possible to fix this particular case in the ARM target code,
to avoid adding overlapping target features. But we should probably be
sorting in any case: the behavior shouldn't depend on StringMap's
hashing algorithm.
Differential Revision: https://reviews.llvm.org/D46030
llvm-svn: 330861
Debug var, expr and loc were only supported for non-fixed stack objects.
This patch adds the following fields to the "fixedStack:" entries, and
renames the ones from "stack:" to:
* debug-info-variable
* debug-info-expression
* debug-info-location
Differential Revision: https://reviews.llvm.org/D46032
llvm-svn: 330859
The current statement domain derivation algorithm does not (always)
consider that different exit blocks of a loop can have different
conditions to be reached.
From the code
for (int i = n; ; i-=2) {
if (i <= 0) goto even;
if (i <= 1) goto odd;
A[i] = i;
}
even:
A[0] = 42;
return;
odd:
A[1] = 21;
return;
Polly currently derives the following domains:
Stmt_even_critedge
Domain :=
[n] -> { Stmt_even_critedge[] };
Stmt_odd
Domain :=
[n] -> { Stmt_odd[] : (1 + n) mod 2 = 0 and n > 0 };
while the domain for the odd case is correct, Stmt_even is assumed to be
executed unconditionally, which is obviously wrong. While projecting out
the loop dimension in `adjustDomainDimensions`, it does not consider
that there are other exit condition that have matched before.
I don't know a how to fix this without changing a lot of code. Therefore
This patch rejects loops with multiple exist blocks to fix the
miscompile of test-suite's uuencode.
The odd condition is transformed by LLVM to
%cmp1 = icmp eq i64 %indvars.iv, 1
such that the project_out in adjustDomainDimensions() indeed only
matches for odd n (using this condition only, we'd have an infinite loop
otherwise).
The even condition manifests as
%cmp = icmp slt i64 %indvars.iv, 3
Because buildDomainsWithBranchConstraints() does not consider other exit
conditions, it has to assume that the induction variable will eventually
be lower than 3 and taking this exit.
IMHO we need to reuse the algorithm that determines the number of
iterations (addLoopBoundsToHeaderDomain) to determine which exit
condition applies first. It has to happen in
buildDomainsWithBranchConstraints() because the result will need to
propagate to successor BBs. Currently addLoopBoundsToHeaderDomain() just
look for union of all backedge conditions (which means leaving not the
loop here). The patch in llvm.org/PR35465 changes it to look for exit
conditions instead. This is required because there might be other exit
conditions that do not alternatively go back to the loop header.
Differential Revision: https://reviews.llvm.org/D45649
llvm-svn: 330858
Summary:
This adds `__scudo_print_stats` as an interface function to display the Primary
and Secondary allocator statistics for Scudo.
Reviewers: alekseyshl, flowerhack
Reviewed By: alekseyshl
Subscribers: delcypher, llvm-commits, #sanitizers
Differential Revision: https://reviews.llvm.org/D46016
llvm-svn: 330857
This is necessary in order to get a working C++ compiler on Darwin
since Clang expects libc++ headers to be part of the toolchain.
Differential Revision: https://reviews.llvm.org/D46075
llvm-svn: 330855
Imports in a wasm module can have custom module name. This change
adds the module name to the WasmSymbol structure so that the linker
can preserve this module name.
This is needed to fix: https://bugs.llvm.org/show_bug.cgi?id=37168
Differential Revision: https://reviews.llvm.org/D45797
llvm-svn: 330854
Previously we only formed MUL_IMM when we split a constant. This blocked load folding on those cases. We should also form MUL_IMM for 3/5/9 to favor LEA over load folding.
Differential Revision: https://reviews.llvm.org/D46040
llvm-svn: 330850
Craig Topper