[DAGCombiner] improve throughput of shift+logic+shift
The motivating case for this is a long way from here:
https://bugs.llvm.org/show_bug.cgi?id=43146
...but I think this is where we have to start.
We need to canonicalize/optimize sequences of shift and logic to ease
pattern matching for things like bswap and improve perf in general.
But without the artificial limit of '!LegalTypes' (early combining),
there are a lot of test diffs, and not all are good.
In the minimal tests added for this proposal, x86 should have better
throughput in all cases. AArch64 is neutral for scalar tests because
it can fold shifts into bitwise logic ops.
There are 3 shift opcodes and 3 logic opcodes for a total of 9 possible patterns:
https://rise4fun.com/Alive/VlI
https://rise4fun.com/Alive/n1m
https://rise4fun.com/Alive/1Vn
Roman Lebedev [Sun, 1 Sep 2019 11:56:52 +0000 (11:56 +0000)]
[ConstantFolding] Fix 'undef' folding for @llvm.[us]{add,sub}.with.overflow ops (PR43188)
As we have already established/fixed in
https://bugs.llvm.org/show_bug.cgi?id=42209
https://reviews.llvm.org/D63065
https://reviews.llvm.org/rL363522
the InstSimplify handling for @llvm.with.overflow ops with undefs
is correct. Therefore if ConstantFolding produces different results,
then it is wrong.
This duplication of code hints at the need for some refactoring,
but for now address the brokenness of ConstantFolding by
copying the known-good handling from rL363522.
[TargetLowering] Fix Bugzilla ID 43183 to avoid soften comparison broken with constant inputs
Summary:
This fixes the bugzilla id 43183 which triggerd by the following commit:
[RISCV] Avoid generating AssertZext for LP64 ABI when lowering floating LibCall
[GlobalISel][NFC] Regression test cases for aarch64 legalizer (s128 sext+icmp).
There were legalizer asserts in aarch64 globalisel (in debug mode) with s128
sext+icmp before r367060 and r366943 landed. These are just a couple reduced
mir and ir regression tests that came from a build where these were encountered.
Craig Topper [Sat, 31 Aug 2019 23:52:25 +0000 (23:52 +0000)]
[X86] Replace some COPY_TO_REGCLASS from GR32/GR64 to VR128 in isel patterns with VMOVDI2PDIrr/VMOV64toPQIrr.
This is what the copies will eventually be turned into. We don't
use COPY_TO_REGCLASS for scalar_to_vector patterns. So we should
use the real instruction here too.
Simon Pilgrim [Sat, 31 Aug 2019 16:21:31 +0000 (16:21 +0000)]
[X86] EltsFromConsecutiveLoads - Don't confuse elt count with vector element count (PR43170)
EltsFromConsecutiveLoads was assuming that the number of input elts was the same as the number of elements in the output vector type when creating a zeroing shuffle, causing an assert when subvectors were being combined instead of just scalars.
Nandor Licker [Sat, 31 Aug 2019 15:00:38 +0000 (15:00 +0000)]
[Clang Interpreter] Initial patch for the constexpr interpreter
Summary:
This patch introduces the skeleton of the constexpr interpreter,
capable of evaluating a simple constexpr functions consisting of
if statements. The interpreter is described in more detail in the
RFC. Further patches will add more features.
James Molloy [Sat, 31 Aug 2019 10:46:16 +0000 (10:46 +0000)]
[DAGCombiner] Don't create illegal narrow stores
Narrowing stores when the target doesn't support the narrow version
forces the target to expand into a load-modify-store sequence, which
is highly suboptimal. The information narrowing throws away (legality
of the inverse transform) is hard to re-analyze. If the target doesn't
support a store of the narrow type, don't narrow even in pre-legalize
mode.
No test as this is DAGCombiner and depends on target bits.
Nikita Popov [Sat, 31 Aug 2019 09:58:37 +0000 (09:58 +0000)]
[CVP] Generate simpler code for elided with.overflow intrinsics
Use a { iN undef, i1 false } struct as the base, and only insert
the first operand, instead of using { iN undef, i1 undef } as the
base and inserting both. This is the same as what we do in InstCombine.
Restructured the code a little bit in preparation for adding
UMULFIXSAT. I think it will be easier to understand the code
if not interleaving the codegen for signed/unsigned/saturated
cases that much.
Bjorn Pettersson [Sat, 31 Aug 2019 09:01:16 +0000 (09:01 +0000)]
[LangRef] Update saturating examples for llvm.smul.fix.sat. NFC
Some saturation examples for llvm.smul.fix.sat were not showing
the correct result. I've adjusted the operands to make sure that
we actually trigger overflow in those examples.
Wei Mi [Sat, 31 Aug 2019 03:17:49 +0000 (03:17 +0000)]
Fix some errors introduced by rL370563 which were not exposed on my local machine.
1. zlib::compress accept &size_t but the param is an uint64_t.
2. Some systems don't have zlib installed. Don't use compression by default.
Wei Mi [Sat, 31 Aug 2019 02:27:26 +0000 (02:27 +0000)]
[SampleFDO] Add profile symbol list section to discriminate function being
cold versus function being newly added.
This is the second half of https://reviews.llvm.org/D66374.
Profile symbol list is the collection of function symbols showing up in
the binary which generates the current profile. It is used to discriminate
function being cold versus function being newly added. Profile symbol list
is only added for profile with ExtBinary format.
During profile use compilation, when profile-sample-accurate is enabled,
a function without profile will be regarded as cold only when it is
contained in that list.
Thomas Lively [Sat, 31 Aug 2019 00:12:29 +0000 (00:12 +0000)]
[WebAssembly] Add SIMD QFMA/QFMS
Summary:
Adds clang builtins and LLVM intrinsics for these experimental
instructions. They are not implemented in engines yet, but that is ok
because the user must opt into using them by calling the builtins.
[lit] Only set DYLD_LIBRARY_PATH for shared builds
In r370135 I committed a temporary workaround for the sanitized bot to
not set (DY)LD_LIBRARY_PATH when (DY)LD_INSERT_LIBRARIES was set.
Setting (DY)LD_LIBRARY_PATH is only necessary for (standalone)
shared-library builds, so a better solution is to only set the
environment variable when necessary.
Wei Mi [Fri, 30 Aug 2019 23:01:22 +0000 (23:01 +0000)]
[GVN] Verify value equality before doing phi translation for call instruction
This is an updated version of https://reviews.llvm.org/D66909 to fix PR42605.
Basically, current phi translatation translates an old value number to an new
value number for a call instruction based on the literal equality of call
expression, without verifying there is no clobber in between. This is incorrect.
To get a finegrain check, use MachineDependence analysis to do the job. However,
this is still not ideal. Although given a call instruction,
`MemoryDependenceResults::getCallDependencyFrom` returns identical call
instructions without clobber in between using MemDepResult with its DepType to
be `Def`. However, identical is too strict here and we want it to be relaxed a
little to consider phi-translation -- callee is the same, param operands can be
different. That means changing the semantic of `MemDepResult::Def` and I don't
know the potential impact.
So currently the patch is still conservative to only handle
MemDepResult::NonFuncLocal, which means the current call has no function local
clobber. If there is clobber, even if the clobber doesn't stand in between the
current call and the call with the new value, we won't do phi-translate.
Reid Kleckner [Fri, 30 Aug 2019 21:23:05 +0000 (21:23 +0000)]
[X86] Print register names in .seh_* directives
Also improve assembler parser register validation for .seh_ directives.
This requires moving X86-specific seh directive handling into the x86
backend, which addresses some assembler FIXMEs.
Reid Kleckner [Fri, 30 Aug 2019 20:46:39 +0000 (20:46 +0000)]
[Windows] Disable TrapUnreachable for Win64, add SEH_NoReturn
Users have complained llvm.trap produce two ud2 instructions on Win64,
one for the trap, and one for unreachable. This change fixes that.
TrapUnreachable was added and enabled for Win64 in r206684 (April 2014)
to avoid poorly understood issues with the Windows unwinder.
There seem to be two major things in play:
- the unwinder
- C++ EH, _CxxFrameHandler3 & co
The unwinder disassembles forward from the return address to scan for
epilogues. Inserting a ud2 had the effect of stopping the unwinder, and
ensuring that it ran the EH personality function for the current frame.
However, it's not clear what the unwinder does when the return address
happens to be the last address of one function and the first address of
the next function.
The Visual C++ EH personality, _CxxFrameHandler3, needs to figure out
what the current EH state number is. It does this by consulting the
ip2state table, which maps from PC to state number. This seems to go
wrong when the return address is the last PC of the function or catch
funclet.
I'm not sure precisely which system is involved here, but in order to
address these real or hypothetical problems, I believe it is enough to
insert int3 after a call site if it would otherwise be the last
instruction in a function or funclet. I was able to reproduce some
similar problems locally by arranging for a noreturn call to appear at
the end of a catch block immediately before an unrelated function, and I
confirmed that the problems go away when an extra trailing int3
instruction is added.
MSVC inserts int3 after every noreturn function call, but I believe it's
only necessary to do it if the call would be the last instruction. This
change inserts a pseudo instruction that expands to int3 if it is in the
last basic block of a function or funclet. I did what I could to run the
Microsoft compiler EH tests, and the ones I was able to run showed no
behavior difference before or after this change.
Sanjay Patel [Fri, 30 Aug 2019 20:15:01 +0000 (20:15 +0000)]
[Thumb2] tighten CHECK lines in test; NFC
The sequence between the function call and the asm start
may change without affecting what this test is looking for,
but we should have a better idea about what that sequence
looks like.
Puyan Lotfi [Fri, 30 Aug 2019 19:54:46 +0000 (19:54 +0000)]
[IFS][NFC] llvm-ifs: Fixing build errors for bots using GCC.
gcc produces the error:
error: specialization of
‘template<class T, class Enable> struct llvm::yaml::ScalarTraits’ in
different namespace
For all specializations outside of llvm::yaml. So I added llvm::yaml to these
specializations to fix the errors on the bots building with gcc (/usr/bin/c++).
James Molloy [Fri, 30 Aug 2019 19:50:49 +0000 (19:50 +0000)]
[DFAPacketizer] Allow namespacing of automata per-itinerary
The Hexagon itineraries are cunningly crafted such that functional units between
itineraries do not clash. Because all itineraries are bundled into the same DFA,
a functional unit index clash would cause an incorrect DFA to be generated.
A workaround for this is to ensure all itineraries declare the universe of all
possible functional units, but this isn't ideal for three reasons:
1) We only have a limited number of FUs we can encode in the packetizer, and
using the universe causes us to hit the limit without care.
2) Silent codegen faults are bad, and careful triage of the FU list shouldn't
be required.
3) Smooshing all itineraries into the same automaton allows combinations of
instruction classes that cannot exist, which bloats the table.
A simple solution is to allow "namespacing" packetizers.
Craig Topper [Fri, 30 Aug 2019 19:42:48 +0000 (19:42 +0000)]
[X86] Regenerate the test cases added in r370506.
Something weird happened with the v2i64/v2f64 test cases which
don't use broadcast. So they should already be hoisted, but
weren't in the version I submitted in r370506. This fixes that.
Not sure if something changed or I screwed up.
Craig Topper [Fri, 30 Aug 2019 19:26:06 +0000 (19:26 +0000)]
[X86] Add test caes for opportunities for machine LICM to unfold broadcasted constant pool loads.
MachineLICM is able to unfold loads to move an invariant load out
a loop, but X86 infrastructure currently lacks the ability to do
this when avx512 embedded broadcasting is used.
This test adds examples for the basic float point operations,
add, mul, and, or, and xor.
James Molloy [Fri, 30 Aug 2019 18:49:50 +0000 (18:49 +0000)]
[MachinePipeliner] Separate schedule emission, NFC
This is the first stage in refactoring the pipeliner and making it more
accessible for backends to override and control. This separates the logic and
state required to *emit* a scheudule from the logic that *computes* and
validates a schedule.
This will enable (a) new schedule emitters and (b) new modulo scheduling
implementations to coexist.
This tool merges interface stub files to produce a merged interface stub file
or a stub library. Currently it for stub library generation it can produce an
ELF .so stub file, or a TBD file (experimental). It will be used by the clang
-emit-interface-stubs compilation pipeline to merge and assemble the per-CU
stub files into a stub library.
Simon Pilgrim [Fri, 30 Aug 2019 17:58:55 +0000 (17:58 +0000)]
[TargetLowering] SimplifyDemandedBits ADD/SUB/MUL - correctly inherit SDNodeFlags from the original node.
Just disable NSW/NUW flags. This matches what we're already doing for the other situations for these nodes, it was just missed for the demanded constant case.
Noticed by inspection - confirmed in offline discussion with @spatel. I've checked we have test coverage in the x86 extract-bits.ll and extract-lowbits.ll tests
Craig Topper [Fri, 30 Aug 2019 17:35:08 +0000 (17:35 +0000)]
[X86] Pass v32i16/v64i8 in zmm registers on KNL target.
gcc and icc pass these types in zmm registers in zmm registers.
This patch implements a quick hack to override the register
type before calling convention handling to one that is legal.
Longer term we might want to do something similar to 256-bit
integer registers on AVX1 where we just split all the operations.
Evgeniy Stepanov [Fri, 30 Aug 2019 17:23:02 +0000 (17:23 +0000)]
MemTag: unchecked load/store optimization.
Summary:
MTE allows memory access to bypass tag check iff the address argument
is [SP, #imm]. This change takes advantage of this to demote uses of
tagged addresses to regular FrameIndex operands, reducing register
pressure in large functions.
MO_TAGGED target flag is used to signal that the FrameIndex operand
refers to memory that might be tagged, and needs to be handled with
care. Such operand must be lowered to [SP, #imm] directly, without a
scratch register.
The transformation pass attempts to predict when the offset will be
out of range and disable the optimization.
AArch64RegisterInfo::eliminateFrameIndex has an escape hatch in case
this prediction has been wrong, but it is quite inefficient and should
be avoided.
Craig Topper [Fri, 30 Aug 2019 16:05:57 +0000 (16:05 +0000)]
[X86] Merge X86InstrInfo::loadRegFromAddr/storeRegToAddr into their only call site.
I'm looking at unfolding broadcast loads on AVX512 which will
require refactoring this code to select broadcast opcodes instead
of regular load/stores in some cases. Merging them to avoid
further complicating their interfaces.
[Attributor] Use existing function information for the call site
Summary:
Instead of recomputing information for call sites we now use the
function information directly. This is always valid and once we have
call site specific information we can improve here.
This patch also bootstraps attributes that are created on-demand through
an initial update call. Information that is known will then directly be
available in the new attribute without causing an iteration delay.
The tests show how this improves the iteration count.
George Rimar [Fri, 30 Aug 2019 13:39:22 +0000 (13:39 +0000)]
[yaml2obj][obj2yaml] - Use a single "Other" field instead of "Other", "Visibility" and "StOther".
Currenly we can encode the 'st_other' field of symbol using 3 fields.
'Visibility' is used to encode STV_* values.
'Other' is used to encode everything except the visibility, but it can't handle arbitrary values.
'StOther' is used to encode arbitrary values when 'Visibility'/'Other' are not helpfull enough.
'st_other' field is used to encode symbol visibility and platform-dependent
flags and values. Problem to encode it is that it consists of Visibility part (STV_* values)
which are enumeration values and the Other part, which is different and inconsistent.
For MIPS the Other part contains flags for all STO_MIPS_* values except STO_MIPS_MIPS16.
(Like comment in ELFDumper says: "Someones in their infinite wisdom decided to make
STO_MIPS_MIPS16 flag overlapped with other ST_MIPS_xxx flags."...)
And for PPC64 the Other part might actually encode any value.
This patch implements custom logic for handling the st_other and removes
'Visibility' and 'StOther' fields.
Here is an example of a new YAML style this patch allows:
Simon Pilgrim [Fri, 30 Aug 2019 12:22:06 +0000 (12:22 +0000)]
[DAGCombine] visitMULHS - use getScalarValueSizeInBits() to make safe for vector types.
This is hidden behind a (scalar-only) isOneConstant(N1) check at the moment, but once we get around to adding vector support we need to ensure we're dealing with the scalar bitwidth, not the total.
Bjorn Pettersson [Fri, 30 Aug 2019 11:23:10 +0000 (11:23 +0000)]
[CodeGen] Introduce MachineBasicBlock::replacePhiUsesWith helper and use it. NFC
Summary:
Found a couple of places in the code where all the PHI nodes
of a MBB is updated, replacing references to one MBB by
reference to another MBB instead.
This patch simply refactors the code to use a common helper
(MachineBasicBlock::replacePhiUsesWith) for such PHI node
updates.
Roman Lebedev [Fri, 30 Aug 2019 09:51:23 +0000 (09:51 +0000)]
[LoopIdiomRecognize] BCmp loop idiom recognition
Summary:
@mclow.lists brought up this issue up in IRC.
It is a reasonably common problem to compare some two values for equality.
Those may be just some integers, strings or arrays of integers.
In C, there is `memcmp()`, `bcmp()` functions.
In C++, there exists `std::equal()` algorithm.
One can also write that function manually.
libstdc++'s `std::equal()` is specialized to directly call `memcmp()` for
various types, but not `std::byte` from C++2a. https://godbolt.org/z/mx2ejJ
libc++ does not do anything like that, it simply relies on simple C++'s
`operator==()`. https://godbolt.org/z/er0Zwf (GOOD!)
So likely, there exists a certain performance opportunities.
Let's compare performance of naive `std::equal()` (no `memcmp()`) with one that
is using `memcmp()` (in this case, compiled with modified compiler). {F8768213}
template <typename T>
static void CustomArguments(benchmark::internal::Benchmark* b) {
const size_t L2SizeBytes = []() {
for (const benchmark::CPUInfo::CacheInfo& I :
benchmark::CPUInfo::Get().caches) {
if (I.level == 2) return I.size;
}
return 0;
}();
// What is the largest range we can check to always fit within given L2 cache?
const size_t MaxLen = L2SizeBytes / /*total bufs*/ 2 /
/*maximal elt size*/ sizeof(T) / /*safety margin*/ 2;
b->RangeMultiplier(2)->Range(1, MaxLen)->Complexity(benchmark::oN);
}
BENCHMARK_TEMPLATE(BM_bcmp, uint8_t, InequalHalfway)
->Apply(CustomArguments<uint8_t>);
BENCHMARK_TEMPLATE(BM_bcmp, uint16_t, InequalHalfway)
->Apply(CustomArguments<uint16_t>);
BENCHMARK_TEMPLATE(BM_bcmp, uint32_t, InequalHalfway)
->Apply(CustomArguments<uint32_t>);
BENCHMARK_TEMPLATE(BM_bcmp, uint64_t, InequalHalfway)
->Apply(CustomArguments<uint64_t>);
```
{F8768210}
```
$ ~/src/googlebenchmark/tools/compare.py --no-utest benchmarks build-{old,new}/test/llvm-bcmp-bench
RUNNING: build-old/test/llvm-bcmp-bench --benchmark_out=/tmp/tmpb6PEUx
2019-04-25 21:17:11
Running build-old/test/llvm-bcmp-bench
Run on (8 X 4000 MHz CPU s)
CPU Caches:
L1 Data 16K (x8)
L1 Instruction 64K (x4)
L2 Unified 2048K (x4)
L3 Unified 8192K (x1)
Load Average: 0.65, 3.90, 4.14
---------------------------------------------------------------------------------------------------
Benchmark Time CPU Iterations UserCounters...
---------------------------------------------------------------------------------------------------
<...>
BM_bcmp<uint8_t, Identical>/512000 432131 ns 432101 ns 1613 bytes_read/iteration=1000k bytes_read/sec=2.20706G/s eltcnt=825.856M eltcnt/sec=1.18491G/s
BM_bcmp<uint8_t, Identical>_BigO 0.86 N 0.86 N
BM_bcmp<uint8_t, Identical>_RMS 8 % 8 %
<...>
BM_bcmp<uint16_t, Identical>/256000 161408 ns 161409 ns 4027 bytes_read/iteration=1000k bytes_read/sec=5.90843G/s eltcnt=1030.91M eltcnt/sec=1.58603G/s
BM_bcmp<uint16_t, Identical>_BigO 0.67 N 0.67 N
BM_bcmp<uint16_t, Identical>_RMS 25 % 25 %
<...>
BM_bcmp<uint32_t, Identical>/128000 81497 ns 81488 ns 8415 bytes_read/iteration=1000k bytes_read/sec=11.7032G/s eltcnt=1077.12M eltcnt/sec=1.57078G/s
BM_bcmp<uint32_t, Identical>_BigO 0.71 N 0.71 N
BM_bcmp<uint32_t, Identical>_RMS 42 % 42 %
<...>
BM_bcmp<uint64_t, Identical>/64000 50138 ns 50138 ns 10909 bytes_read/iteration=1000k bytes_read/sec=19.0209G/s eltcnt=698.176M eltcnt/sec=1.27647G/s
BM_bcmp<uint64_t, Identical>_BigO 0.84 N 0.84 N
BM_bcmp<uint64_t, Identical>_RMS 27 % 27 %
<...>
BM_bcmp<uint8_t, InequalHalfway>/512000 192405 ns 192392 ns 3638 bytes_read/iteration=1000k bytes_read/sec=4.95694G/s eltcnt=1.86266G eltcnt/sec=2.66124G/s
BM_bcmp<uint8_t, InequalHalfway>_BigO 0.38 N 0.38 N
BM_bcmp<uint8_t, InequalHalfway>_RMS 3 % 3 %
<...>
BM_bcmp<uint16_t, InequalHalfway>/256000 127858 ns 127860 ns 5477 bytes_read/iteration=1000k bytes_read/sec=7.45873G/s eltcnt=1.40211G eltcnt/sec=2.00219G/s
BM_bcmp<uint16_t, InequalHalfway>_BigO 0.50 N 0.50 N
BM_bcmp<uint16_t, InequalHalfway>_RMS 0 % 0 %
<...>
BM_bcmp<uint32_t, InequalHalfway>/128000 49140 ns 49140 ns 14281 bytes_read/iteration=1000k bytes_read/sec=19.4072G/s eltcnt=1.82797G eltcnt/sec=2.60478G/s
BM_bcmp<uint32_t, InequalHalfway>_BigO 0.40 N 0.40 N
BM_bcmp<uint32_t, InequalHalfway>_RMS 18 % 18 %
<...>
BM_bcmp<uint64_t, InequalHalfway>/64000 32101 ns 32099 ns 21786 bytes_read/iteration=1000k bytes_read/sec=29.7101G/s eltcnt=1.3943G eltcnt/sec=1.99381G/s
BM_bcmp<uint64_t, InequalHalfway>_BigO 0.50 N 0.50 N
BM_bcmp<uint64_t, InequalHalfway>_RMS 1 % 1 %
RUNNING: build-new/test/llvm-bcmp-bench --benchmark_out=/tmp/tmpQ46PP0
2019-04-25 21:19:29
Running build-new/test/llvm-bcmp-bench
Run on (8 X 4000 MHz CPU s)
CPU Caches:
L1 Data 16K (x8)
L1 Instruction 64K (x4)
L2 Unified 2048K (x4)
L3 Unified 8192K (x1)
Load Average: 1.01, 2.85, 3.71
---------------------------------------------------------------------------------------------------
Benchmark Time CPU Iterations UserCounters...
---------------------------------------------------------------------------------------------------
<...>
BM_bcmp<uint8_t, Identical>/512000 18593 ns 18590 ns 37565 bytes_read/iteration=1000k bytes_read/sec=51.2991G/s eltcnt=19.2333G eltcnt/sec=27.541G/s
BM_bcmp<uint8_t, Identical>_BigO 0.04 N 0.04 N
BM_bcmp<uint8_t, Identical>_RMS 37 % 37 %
<...>
BM_bcmp<uint16_t, Identical>/256000 18950 ns 18948 ns 37223 bytes_read/iteration=1000k bytes_read/sec=50.3324G/s eltcnt=9.52909G eltcnt/sec=13.511G/s
BM_bcmp<uint16_t, Identical>_BigO 0.08 N 0.08 N
BM_bcmp<uint16_t, Identical>_RMS 34 % 34 %
<...>
BM_bcmp<uint32_t, Identical>/128000 18627 ns 18627 ns 37895 bytes_read/iteration=1000k bytes_read/sec=51.198G/s eltcnt=4.85056G eltcnt/sec=6.87168G/s
BM_bcmp<uint32_t, Identical>_BigO 0.16 N 0.16 N
BM_bcmp<uint32_t, Identical>_RMS 35 % 35 %
<...>
BM_bcmp<uint64_t, Identical>/64000 18855 ns 18855 ns 37458 bytes_read/iteration=1000k bytes_read/sec=50.5791G/s eltcnt=2.39731G eltcnt/sec=3.3943G/s
BM_bcmp<uint64_t, Identical>_BigO 0.32 N 0.32 N
BM_bcmp<uint64_t, Identical>_RMS 33 % 33 %
<...>
BM_bcmp<uint8_t, InequalHalfway>/512000 9570 ns 9569 ns 73500 bytes_read/iteration=1000k bytes_read/sec=99.6601G/s eltcnt=37.632G eltcnt/sec=53.5046G/s
BM_bcmp<uint8_t, InequalHalfway>_BigO 0.02 N 0.02 N
BM_bcmp<uint8_t, InequalHalfway>_RMS 29 % 29 %
<...>
BM_bcmp<uint16_t, InequalHalfway>/256000 9547 ns 9547 ns 74343 bytes_read/iteration=1000k bytes_read/sec=99.8971G/s eltcnt=19.0318G eltcnt/sec=26.8159G/s
BM_bcmp<uint16_t, InequalHalfway>_BigO 0.04 N 0.04 N
BM_bcmp<uint16_t, InequalHalfway>_RMS 29 % 29 %
<...>
BM_bcmp<uint32_t, InequalHalfway>/128000 9396 ns 9394 ns 73521 bytes_read/iteration=1000k bytes_read/sec=101.518G/s eltcnt=9.41069G eltcnt/sec=13.6255G/s
BM_bcmp<uint32_t, InequalHalfway>_BigO 0.08 N 0.08 N
BM_bcmp<uint32_t, InequalHalfway>_RMS 30 % 30 %
<...>
BM_bcmp<uint64_t, InequalHalfway>/64000 9499 ns 9498 ns 73802 bytes_read/iteration=1000k bytes_read/sec=100.405G/s eltcnt=4.72333G eltcnt/sec=6.73808G/s
BM_bcmp<uint64_t, InequalHalfway>_BigO 0.16 N 0.16 N
BM_bcmp<uint64_t, InequalHalfway>_RMS 28 % 28 %
Comparing build-old/test/llvm-bcmp-bench to build-new/test/llvm-bcmp-bench
Benchmark Time CPU Time Old Time New CPU Old CPU New
---------------------------------------------------------------------------------------------------------------------------------------
<...>
BM_bcmp<uint8_t, Identical>/512000 -0.9570 -0.9570 432131 18593 432101 18590
<...>
BM_bcmp<uint16_t, Identical>/256000 -0.8826 -0.8826 161408 18950 161409 18948
<...>
BM_bcmp<uint32_t, Identical>/128000 -0.7714 -0.7714 81497 18627 81488 18627
<...>
BM_bcmp<uint64_t, Identical>/64000 -0.6239 -0.6239 50138 18855 50138 18855
<...>
BM_bcmp<uint8_t, InequalHalfway>/512000 -0.9503 -0.9503 192405 9570 192392 9569
<...>
BM_bcmp<uint16_t, InequalHalfway>/256000 -0.9253 -0.9253 127858 9547 127860 9547
<...>
BM_bcmp<uint32_t, InequalHalfway>/128000 -0.8088 -0.8088 49140 9396 49140 9394
<...>
BM_bcmp<uint64_t, InequalHalfway>/64000 -0.7041 -0.7041 32101 9499 32099 9498
```
What can we tell from the benchmark?
* Performance of naive equality check somewhat improves with element size,
maxing out at eltcnt/sec=1.58603G/s for uint16_t, or bytes_read/sec=19.0209G/s
for uint64_t. I think, that instability implies performance problems.
* Performance of `memcmp()`-aware benchmark always maxes out at around
bytes_read/sec=51.2991G/s for every type. That is 2.6x the throughput of the
naive variant!
* eltcnt/sec metric for the `memcmp()`-aware benchmark maxes out at
eltcnt/sec=27.541G/s for uint8_t (was: eltcnt/sec=1.18491G/s, so 24x) and
linearly decreases with element size.
For uint64_t, it's ~4x+ the elements/second.
* The call obvious is more pricey than the loop, with small element count.
As it can be seen from the full output {F8768210}, the `memcmp()` is almost
universally worse, independent of the element size (and thus buffer size) when
element count is less than 8.
So all in all, bcmp idiom does indeed pose untapped performance headroom.
This diff does implement said idiom recognition. I think a reasonable test
coverage is present, but do tell if there is anything obvious missing.
Now, quality. This does succeed to build and pass the test-suite, at least
without any non-bundled elements. {F8768216} {F8768217}
This transform fires 91 times:
```
$ /build/test-suite/utils/compare.py -m loop-idiom.NumBCmp result-new.json
Tests: 1149
Metric: loop-idiom.NumBCmp
Program result-new
MultiSourc...Benchmarks/7zip/7zip-benchmark 79.00
MultiSource/Applications/d/make_dparser 3.00
SingleSource/UnitTests/vla 2.00
MultiSource/Applications/Burg/burg 1.00
MultiSourc.../Applications/JM/lencod/lencod 1.00
MultiSource/Applications/lemon/lemon 1.00
MultiSource/Benchmarks/Bullet/bullet 1.00
MultiSourc...e/Benchmarks/MallocBench/gs/gs 1.00
MultiSourc...gs-C/TimberWolfMC/timberwolfmc 1.00
MultiSourc...Prolangs-C/simulator/simulator 1.00
```
The size changes are:
I'm not sure what's going on with SingleSource/UnitTests/vla.test yet, did not look.
```
$ /build/test-suite/utils/compare.py -m size..text result-{old,new}.json --filter-hash
Tests: 1149
Same hash: 907 (filtered out)
Remaining: 242
Metric: size..text
Program result-old result-new diff
test-suite...ingleSource/UnitTests/vla.test 753.00 833.00 10.6%
test-suite...marks/7zip/7zip-benchmark.test 1001697.00 966657.00 -3.5%
test-suite...ngs-C/simulator/simulator.test 32369.00 32321.00 -0.1%
test-suite...plications/d/make_dparser.test 89585.00 89505.00 -0.1%
test-suite...ce/Applications/Burg/burg.test 40817.00 40785.00 -0.1%
test-suite.../Applications/lemon/lemon.test 47281.00 47249.00 -0.1%
test-suite...TimberWolfMC/timberwolfmc.test 250065.00 250113.00 0.0%
test-suite...chmarks/MallocBench/gs/gs.test 149889.00 149873.00 -0.0%
test-suite...ications/JM/lencod/lencod.test 769585.00 769569.00 -0.0%
test-suite.../Benchmarks/Bullet/bullet.test 770049.00 770049.00 0.0%
test-suite...HMARK_ANISTROPIC_DIFFUSION/128 NaN NaN nan%
test-suite...HMARK_ANISTROPIC_DIFFUSION/256 NaN NaN nan%
test-suite...CHMARK_ANISTROPIC_DIFFUSION/64 NaN NaN nan%
test-suite...CHMARK_ANISTROPIC_DIFFUSION/32 NaN NaN nan%
test-suite...ENCHMARK_BILATERAL_FILTER/64/4 NaN NaN nan%
Geomean difference nan%
result-old result-new diff
count 1.000000e+01 10.00000 10.000000
mean 3.152090e+05 311695.40000 0.006749
std 3.790398e+05 372091.42232 0.036605
min 7.530000e+02 833.00000 -0.034981
25% 4.243300e+04 42401.00000 -0.000866
50% 1.197370e+05 119689.00000 -0.000392
75% 6.397050e+05 639705.00000 -0.000005
max 1.001697e+06 966657.00000 0.106242
```
I don't have timings though.
And now to the code. The basic idea is to completely replace the whole loop.
If we can't fully kill it, don't transform.
I have left one or two comments in the code, so hopefully it can be understood.
Also, there is a few TODO's that i have left for follow-ups:
* widening of `memcmp()`/`bcmp()`
* step smaller than the comparison size
* Metadata propagation
* more than two blocks as long as there is still a single backedge?
* ???
Summary:
The internal `Builder` is private, which means there is
currently no way to set the debuginfo locations for `SCEVExpander`.
This only adds the wrappers, but does not use them anywhere.
David Stenberg [Fri, 30 Aug 2019 09:06:50 +0000 (09:06 +0000)]
[LiveDebugValues] Insert entry values after bundles
Summary:
Change LiveDebugValues so that it inserts entry values after the bundle
which contains the clobbering instruction. Previously it would insert
the debug value after the bundle head using insertAfter(), breaking the
bundle.
Dmitri Gribenko [Fri, 30 Aug 2019 08:21:55 +0000 (08:21 +0000)]
[ADT] Removed VariadicFunction
Summary:
It is not used. It uses macro-based unrolling instead of variadic
templates, so it is not idiomatic anymore, and therefore it is a
questionable API to keep "just in case".
Martin Storsjo [Fri, 30 Aug 2019 06:56:33 +0000 (06:56 +0000)]
[LLD] [COFF] Support merging resource object files
Extend WindowsResourceParser to support using a ResourceSectionRef for
loading resources from an object file.
Only allow merging resource object files in mingw mode; keep the
existing error on multiple resource objects in link mode.
If there only is one resource object file and no .res resources,
don't parse and recreate the .rsrc section, but just link it in without
inspecting it. This allows users to produce any .rsrc section (outside
of what the parser supports), just like before. (I don't have a specific
need for this, but it reduces the risk of this new feature.)
Separate out the .rsrc section chunks in InputFiles.cpp, and only include
them in the list of section chunks to link if we've determined that there
only was one single resource object. (We need to keep other chunks from
those object files, as they can legitimately contain other sections as
well, in addition to .rsrc section chunks.)
Martin Storsjo [Fri, 30 Aug 2019 06:56:02 +0000 (06:56 +0000)]
[WindowsResource] Remove use of global variables in WindowsResourceParser
Instead of updating a global variable counter for the next index of
strings and data blobs, pass along a reference to actual data/string
vectors and let the TreeNode insertion methods add their data/strings to
the vectors when a new entry is needed.
Additionally, if the resource tree had duplicates, that were ignored
with -force:multipleres in lld, we no longer store all versions of the
duplicated resource data, now we only keep the one that actually ends
up referenced.
Martin Storsjo [Fri, 30 Aug 2019 06:55:49 +0000 (06:55 +0000)]
[COFF] Add a ResourceSectionRef method for getting resource contents
This allows llvm-readobj to print the contents of each resource
when printing resources from an object file or executable, like it
already does for plain .res files.
This requires providing the whole COFFObjectFile to ResourceSectionRef.
This supports both object files and executables. For executables,
the DataRVA field is used as is to look up the right section.
For object files, ideally we would need to complete linking of them
and fix up all relocations to know what the DataRVA field would end up
being. In practice, the only thing that makes sense for an RVA field
is an ADDR32NB relocation. Thus, find a relocation pointing at this
field, verify that it has the expected type, locate the symbol it
points at, look up the section the symbol points at, and read from the
right offset in that section.
This works both for GNU windres object files (which use one single
.rsrc section, with all relocations against the base of the .rsrc
section, with the original value of the DataRVA field being the
offset of the data from the beginning of the .rsrc section) and
cvtres object files (with two separate .rsrc$01 and .rsrc$02 sections,
and one symbol per data entry, with the original pre-relocated DataRVA
field being set to zero).
projects / llvm / log