This implements a more optimal algorithm for selecting a base constant in
constant hoisting. It not only takes into account the number of uses and the
cost of expressions in which constants appear, but now also the resulting
integer range of the offsets. Thus, the algorithm maximizes the number of uses
within an integer range that will enable more efficient code generation. On
ARM, for example, this will enable code size optimisations because less
negative offsets will be created. Negative offsets/immediates are not supported
by Thumb1 thus preventing more compact instruction encoding.
David Majnemer [Thu, 14 Jul 2016 06:58:42 +0000 (06:58 +0000)]
[InstCombine] Masked loads with undef masks can fold to normal loads
We were able to fold masked loads with an all-ones mask to a normal
load. However, we couldn't turn a masked load with a mask with mixed
ones and undefs into a normal load.
Summary:
In this patch we implement the following parts of XRay:
- Supporting a function attribute named 'function-instrument' which currently only supports 'xray-always'. We should be able to use this attribute for other instrumentation approaches.
- Supporting a function attribute named 'xray-instruction-threshold' used to determine whether a function is instrumented with a minimum number of instructions (IR instruction counts).
- X86-specific nop sleds as described in the white paper.
- A machine function pass that adds the different instrumentation marker instructions at a very late stage.
- A way of identifying which return opcode is considered "normal" for each architecture.
There are some caveats here:
1) We don't handle PATCHABLE_RET in platforms other than x86_64 yet -- this means if IR used PATCHABLE_RET directly instead of a normal ret, instruction lowering for that platform might do the wrong thing. We think this should be handled at instruction selection time to by default be unpacked for platforms where XRay is not availble yet.
2) The generated section for X86 is different from what is described from the white paper for the sole reason that LLVM allows us to do this neatly. We're taking the opportunity to deviate from the white paper from this perspective to allow us to get richer information from the runtime library.
Mehdi Amini [Thu, 14 Jul 2016 01:31:25 +0000 (01:31 +0000)]
[Scalarizer] PR28108: Skip over nullptr rather than crashing on it.
Summary:
In Scalarizer::gather we see if we already have a scattered form of Op,
and in that case use the new form.
In the particular case of PR28108, the found ValueVector SV has size 2,
where the first Value is nullptr, and the second is indeed a proper Value.
The nullptr then caused an assert to blow when we tried to do
cast<Instruction>(SV[I]).
With this patch we check SV[I] before doing the cast, and if it's nullptr
we just skip over it.
I don't know the Scalarizer well enough to know if this is the best fix
or if something should be done else where to prevent the nullptr from
being in the ValueVector at all, but at least this avoids the crash
and looking at the test case output it looks reasonable.
Adrian Prantl [Thu, 14 Jul 2016 00:41:18 +0000 (00:41 +0000)]
Synchronize LLVM and clang's ObjCDeclSpec::ObjCPropertyAttributeKind.
This adds Clang-specific DWARF constants for nullability and ObjC
class properties that are already generated by clang. This patch adds
dwarfdump support and a more comprehensive testcase.
Mehdi Amini [Wed, 13 Jul 2016 23:39:46 +0000 (23:39 +0000)]
Add EnableIPRA to TargetOptions, and move the cl::opt -enable-ipra to TargetMachine.cpp
Avoid exposing a cl::opt in a public header and instead promote this
option in the API.
Alternatively, we could land the cl::opt in CommandFlags.h so that
it is available to every tool, but we would still have to find an
option for clang.
Mehdi Amini [Wed, 13 Jul 2016 23:39:34 +0000 (23:39 +0000)]
[IPRA] Set callee saved registers to none for local function when IPRA is enabled.
IPRA try to optimize caller saved register by propagating register
usage information from callee to caller so it is beneficial to have
caller saved registers compare to callee saved registers when IPRA
is enabled. Please find more detailed explanation here
https://groups.google.com/d/msg/llvm-dev/XRzGhJ9wtZg/tjAJqb0eEgAJ.
This change makes local function do not have any callee preserved
register when IPRA is enabled. A simple test case is also added to
verify this change.
Davide Italiano [Wed, 13 Jul 2016 23:20:04 +0000 (23:20 +0000)]
[SCCP] Have the logic for replacing insts with constant in a single place.
The code was pretty much copy-pasted between SCCP and IPSCCP. The situation
became clearly worse after I introduced the support for folding structs in
SCCP. This commit is NFC as we currently (still) skip the replacement
step in IPSCCP, but I'll change this soon.
Summary:
Previously it would say we had an invariant load if any of the memory
operands were invariant. But the load should be invariant only if *all*
the memory operands are invariant.
No testcase because this has proven to be very difficult to tickle in
practice. As just one example, ARM's ldrd instruction, which loads 64
bits into two 32-bit regs, is theoretically affected by this. But when
it's produced, it loses its memoperands' invariance bits!
MIRParser: Move MachineFunction reference into PFS; NFC
Code cleanup: The PerFunctionMIParsingState is per function, moving a
reference into PFS we can avoid passing around the MachineFunction in an
extra parameter most of the time.
Also change most signatures to consistently pass PFS reference first.
[llvm-cov] Use a thread pool to speed up report generation (NFC)
It's safe to print out source coverage views using multiple threads when
using the -output-dir mode of the `llvm-cov show` sub-command.
While testing this on my development machine, I observed that the speed
up is roughly linear with the number of available cores. Avg. time for
`llvm-cov show ./llvm-as -show-line-counts-or-regions`:
1 thread: 7.79s user 0.33s system 98% cpu 8.228 total
4 threads: 7.82s user 0.34s system 283% cpu 2.880 total
Extended LoadStoreVectorizer to vectorize subchains.
Summary:
LSV used to abort vectorizing a chain for interleaved load/store accesses that alias.
Allow a valid prefix of the chain to be vectorized, mark just the prefix and retry vectorizing the remaining chain.
Lang Hames [Wed, 13 Jul 2016 21:13:05 +0000 (21:13 +0000)]
[Object] Change Archive::child_iterator for better interop with Error/Expected.
See http://reviews.llvm.org/D22079
Changes the Archive::child_begin and Archive::children to require a reference
to an Error. If iterator increment fails (because the archive header is
damaged) the iterator will be set to 'end()', and the error stored in the
given Error&. The Error value should be checked by the user immediately after
the loop. E.g.:
Error Err;
for (auto &C : A->children(Err)) {
// Do something with archive child C.
}
// Check the error immediately after the loop.
if (Err)
return Err;
Failure to check the Error will result in an abort() when the Error goes out of
scope (as guaranteed by the Error class).
Teresa Johnson [Wed, 13 Jul 2016 20:19:09 +0000 (20:19 +0000)]
Define a module map entry for ProfileData.
As per Richard Smith, this should help avoid a modules bug exposed
by my r275216 commit:
http://lab.llvm.org:8011/builders/clang-x86_64-linux-selfhost-modules/builds/17560
[ADT] Add LLVM_MARK_AS_BITMASK_ENUM, used to enable bitwise operations on enums without static_cast.
Summary: Normally when you do a bitwise operation on an enum value, you
get back an instance of the underlying type (e.g. int). But using this
macro, bitwise ops on your enum will return you back instances of the
enum. This is particularly useful for enums which represent a
combination of flags.
Suppose you have a function which takes an int and a set of flags. One
way to do this would be to take two numeric params:
enum SomeFlags { F1 = 1, F2 = 2, F3 = 4, ... };
void Fn(int Num, int Flags);
void foo() {
Fn(42, F2 | F3);
}
But now if you get the order of arguments wrong, you won't get an error.
You might try to fix this by changing the signature of Fn so it accepts
a SomeFlags arg:
The LLVM_MARK_AS_BITMASK_ENUM macro enables overloads for bitwise
operators on SomeFlags. Critically, these operators return the enum
type, not its underlying type, so you don't need any static_casts.
An advantage of this solution over the previously-proposed BitMask class
[0, 1] is that we don't need any wrapper classes -- we can operate
directly on the enum itself.
The approach here is somewhat similar to OpenOffice's typed_flags_set
[2]. But we skirt the need for a wrapper class (and a good deal of
complexity) by judicious use of enable_if. We SFINAE on the presence of
a particular enumerator (added by the LLVM_MARK_AS_BITMASK_ENUM macro)
instead of using a traits class so that it's impossible to use the enum
before the overloads are present. The solution here also seamlessly
works across multiple namespaces.
[InstCombine] extend vector select matching for non-splat constants
In D21740, we discussed trying to make this a more general matcher. However, I didn't see a clean
way to handle the regular m_Not cases and these non-splat vector patterns, so I've opted for the
direct approach here. If there are other potential uses of areInverseVectorBitmasks(), we could
move that helper function to a higher level.
There is an open question as to which is of these forms should be considered the canonical IR:
%sel = select <4 x i1> <i1 true, i1 false, i1 false, i1 true>, <4 x i32> %a, <4 x i32> %b
%shuf = shufflevector <4 x i32> %a, <4 x i32> %b, <4 x i32> <i32 0, i32 5, i32 6, i32 3>
Marek Olsak [Wed, 13 Jul 2016 17:35:15 +0000 (17:35 +0000)]
AMDGPU/SI: Emit the number of SGPR and VGPR spills
Summary:
v2: don't count SGPRs spilled to scratch twice
I think this is sufficient. It doesn't count private memory usage, which
happens often and uses scratch but isn't technically a spill. The private
memory usage can be computed by:
[scratch_per_thread - vgpr_spills - a random multiple of SGPR spills].
The fact SGPR spills add very high numbers to the scratch size make that
computation a guessing game, but I don't have a solution to that.
Teresa Johnson [Wed, 13 Jul 2016 16:35:56 +0000 (16:35 +0000)]
[ThinLTO/gold] Enable symbol resolution in distributed backend case
While testing a follow-on change to enable index-based symbol resolution
and internalization in the distributed backends, I realized that a test
case change I made in r275247 was only required because we were not
analyzing symbols in the claimed files in thinlto-index-only mode.
In the fixed test case there should be no internalization because we are
linking in -shared mode, so f() is in fact exported, which is detected
properly when we analyze symbols in thinlto-index-only mode. Note that
this is not (yet) a correctness issue (because we are not yet performing
the index-based linkage optimizations in the distributed backends -
that's coming in a follow-on patch).
We know that pcmp produces all-ones/all-zeros bitmasks, so we can use that behavior to avoid unnecessary constant loading.
One could argue that load+and is actually a better solution for some CPUs (Intel big cores) because shifts don't have the
same throughput potential as load+and on those cores, but that should be handled as a CPU-specific later transformation if
it ever comes up. Removing the load is the more general x86 optimization. Note that the uneven usage of vpbroadcast in the
test cases is filed as PR28505:
https://llvm.org/bugs/show_bug.cgi?id=28505
Adam Nemet [Wed, 13 Jul 2016 05:01:48 +0000 (05:01 +0000)]
[BFI] Add new LazyBFI analysis pass
Summary:
This is necessary for D21771. In order to add the hotness attribute to
optimization remarks we need BFI to be available in all passes that emit
optimization remarks.
However we don't want to pay for computing BFI unless the hotness
attribute is requested.
This is achieved by making BFI lazy at the very high-level through a new
analysis pass -- BFI is not calculated unless requested.
I am adding a test to check the laziness under D21771 where the first
user of the analysis is added.
This achieves the same result as previously by using line wrapping. This allows
us to have one keyword per line which makes adding a new keyword significantly
easier, especially if they are inserted in a lexicographical sort order as you
no longer need to reflow the content around it.
This only does the keywords as that is the group which changes more often.
Teresa Johnson [Wed, 13 Jul 2016 03:42:41 +0000 (03:42 +0000)]
[ThinLTO/gold] ThinLTO internalization fixes
Internalization was missing cases where we originally had a local symbol
that was promoted eagerly but not actually exported. This is because we
were only internalizing the set of global (non-local) symbols that were
PREVAILAING_DEF_IRONLY. Instead, collect the set of global symbols that
are referenced outside of a single IR file, and skip internalization for
those.
[X86] Remove some seemingly unnecessary patterns that supported vector zext/sext with 256-bit source types producing a 256-bit result.
These patterns just extracted the source down to 128-bits to use the instructions. AVX512 seems to have blindly copied them over for VLX, but did not create similar patterns for 512-bit sources. So I'm hoping the backend can't actually produce these cases.
The expandAddRecExprLiterally function incorrectly transforms
`[Start + Step * X]` into `Step * [Start + X]` instead of the correct
transform of `[Step * X] + Start`.
This caused https://github.com/JuliaLang/julia/issues/14704#issuecomment-174126219
due to what appeared to be sufficiently complicated loop interactions.
Patch by Jameson Nash (jameson@juliacomputing.com).
Tim Northover [Tue, 12 Jul 2016 22:23:42 +0000 (22:23 +0000)]
GlobalISel: freeze reserved regs after IRTranslator.
We can freeze the registers after the MachineFrameInfo has been configured (by
telling it about calls, inline asm, ...). This doesn't happen at all yet, but
will be part of IR translation.
projects / llvm / log