This patch reverts part of r362750 / D62650, which stopped
LiveDebugVariables from trimming leading variable location ranges down
to only covering those instructions that are in scope. I've observed some
circumstances where the number of DBG_VALUEs in a function can be
amplified in an un-necessary way, to cover more instructions that are
out of scope, leading to very slow compile times. Trimming the range
of instructions that the variables cover solves the slow compile times.
The specific problem that r362750 tries to fix is addressed by the
assignment to RStart that I've added. Any variable location that begins
at the first instruction of a block will now be considered to begin at the
start of the block. While these sound the same, the have different
SlotIndexes, and the register allocator may shoehorn additional
instructions in between the two. The test added in the past
(wrong_debug_loc_after_regalloc.ll) still works with this modification.
live-debug-variables.ll has a range trimmed to not cover the prologue of
the function, while dbg-addr-dse.ll has a DBG_VALUE sink past one
instruction with no DebugLoc, which is expected behaviour.
Differential Revision: https://reviews.llvm.org/D73691
Summary:
After following Simon's suggestion about additional testing posted at
https://reviews.llvm.org/D73906, I found several more places that
need to be updated.
Reviewers: simon_tatham, dmgreen, ostannard, eli.friedman
Reviewed By: simon_tatham
Subscribers: merge_guards_bot, kristof.beyls, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73963
This is a bug noted in the recent D72733 and seen
in the similar transform just above the changed source code.
I added tests with illegal types and zexts to show the bug -
we could transform legal phi ops to illegal, etc. I did not add
tests with trunc because we won't see any diffs on those patterns.
That is because InstCombiner::SliceUpIllegalIntegerPHI() appears to
do those transforms independently of datalayout. It can also create
more casts than are present in existing code.
There are some existing regression tests that do not include a
datalayout that would be altered by this fix. I assumed that the
lack of a datalayout in those regression files is an oversight, so
I added the minimal layout (make i32 legal) necessary to preserve
behavior on those tests.
Differential Revision: https://reviews.llvm.org/D73907
Summary:
This patch changes the underlying type of the ARM::ArchExtKind
enumeration to uint64_t and adjusts the related code.
The goal of the patch is to prepare the code base for a new
architecture extension.
Reviewers: simon_tatham, eli.friedman, ostannard, dmgreen
Reviewed By: dmgreen
Subscribers: merge_guards_bot, kristof.beyls, hiraditya, cfe-commits, llvm-commits, pbarrio
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D73906
Ensure that OptLevelChanger::SavedFastISel is initialized in the constructor.
This should be NFC - as the equivalent 'same opt level' early-out is used in the destructor as well, so SavedFastISel is only actually referenced in the general case.
Differential Revision: https://reviews.llvm.org/D73875
This reverts commit 8413116bf10402eef12f556cb9d80b08faeb9890.
this seems to be causing crashes while compiling ncurses.
```
$ ./bin/llc bugpoint-reduced-simplified.ll
LLVM ERROR: Cannot emit physreg copy instruction
```
Here are the crashers: https://gist.github.com/kadircet/918f5bb97a2afe048cb875490edba46e
executing with an llc compiled at 904d54de9ba9f71e937b24e04ad5941281cd50b7 works fine.
Under MVE, we do not have any lowering for fminimum, which a
vector_reduce_fmin without NoNan will be expanded into. As with the
other recent patches, force this to expand in the pre-isel pass. Note
that Neon lowering would be OK because the scalar fminimum uses the
vector VMIN instruction, but is probably better to just rely on the
scalar operations, which is what is done here.
Also fixes what appears to be the reversal of INF vs -INF in the
vector_reduce_fmin widening code.
Adds the global (cl::opt) GVNOption enable-load-in-loop-pre in order
to control whether the optimization will be performed if the load
is part of a loop.
Patch by Hendrik Greving!
Differential Revision: https://reviews.llvm.org/D73804
This code matches (zext (trunc (setcc_carry))) -> (and (setcc_carry), 1)
but the code never checks what type we're truncating too. An and
mask of 1 would only make sense if the trunc was to MVT::i1, but
we didn't check for that.
I believe this code is a leftover from when i1 was a legal type.
Our normal lowering for ISD::SETCC uses X86ISD::SUB to enable
CSE unless the RHS is 0. optimizeCompareInstr called by the peephole
pass can turn subs with unused results into cmps to clean this up.
This commit makes other places that create X86ISD::CMP have the
same behavior.
We were creating two with different operand orders, and then only
using one of them.
Instead just swap the operands when needed and create a single node.
This code was incorrectly emitting extra bytes into arbitrary parts of
the object file when it was meant to be hashing them to compute the DWO
ID.
Follow-up patch(es) will refactor this API somewhat to make such bugs
harder to introduce, hopefully.
Broadwell was missing half the gather instructions. Both models
had some mixups in the resource costs and number of uops.
I've updated here based on what I think the original IACA source
says with some cross checking against the microcode.
I'm not sure about latency as the IACA source I have doesn't have
that information. So I'm using the latency from uops.info.
I plan to update Skylake models as well, but I'll do that in a
separate patch.
Differential Revision: https://reviews.llvm.org/D73844
This ports the existing case for G_XOR from `getTestBitOperand` in
AArch64ISelLowering into GlobalISel.
The idea is to flip between TBZ and TBNZ while walking through G_XORs.
Let's say we have
```
tbz (xor x, c), b
```
Let's say the `b`-th bit in `c` is 1. Then
- If the `b`-th bit in `x` is 1, the `b`-th bit in `(xor x, c)` is 0.
- If the `b`-th bit in `x` is 0, then the `b`-th bit in `(xor x, c)` is 1.
So, then
```
tbz (xor x, c), b == tbnz x, b
```
Let's say the `b`-th bit in `c` is 0. Then
- If the `b`-th bit in `x` is 1, the `b`-th bit in `(xor x, c)` is 1.
- If the `b`-th bit in `x` is 0, then the `b`-th bit in `(xor x, c)` is 0.
So, then
```
tbz (xor x, c), b == tbz x, b
```
Differential Revision: https://reviews.llvm.org/D73929
This implements the following optimization:
```
(tbz (shl x, c), b) -> (tbz x, b-c)
```
Which appears in `getTestBitOperand` in AArch64ISelLowering.cpp.
If we test bit `b` of `shl x, c`, we can fold away the `shl` by looking `c` bits
to the right of `b` in `x` when this fits in the type. So, we can just test the
`b-c`th bit.
Differential Revision: https://reviews.llvm.org/D73924
Summary:
The AIX assembler .space directive can't take a second non-zero argument to fill
with. But LLVM emitFill currently assumes it can. We add a flag to the AsmInfo
to check if non-zero fill is supported, and if we can't zerofill non-zero values
we just splat the .byte directives.
Reviewers: stevewan, sfertile, DiggerLin, jasonliu, Xiangling_L
Reviewed By: jasonliu
Subscribers: Xiangling_L, wuzish, nemanjai, hiraditya, kbarton, jsji, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73554
Given
```
tb(n)z (and x, m), b
```
Where the `b`-th bit of `m` is 1,
```
tb(n)z (and x, m), b == tb(n)z x, b
```
So, we can walk past a `G_AND` in this case.
Also add test/CodeGen/AArch64/GlobalISel/opt-fold-and-tbz-tbnz.mir to test this.
Differential Revision: https://reviews.llvm.org/D73790
convertPtrAddToAdd improved overall code size and quality by a significant amount,
but on -O0 we generate some cross-class copies due to the fact that we emitted
G_PTRTOINT and G_INTTOPTR around the G_ADD. Unfortunately at -O0 we don't run any
register coalescing, so these cross class copies end up escaping as moves, and
we ended up regressing 3 benchmarks on CTMark (though still a winner overall).
This patch changes the lowering to instead directly emit the G_ADD into the
destination register, and then force changes the dest LLT to s64 from p0. This
should be ok, as all uses of the register should now be selected and therefore
the LLT doesn't matter for the users. It does however matter for the importer
patterns, which will fail to select a G_ADD if there's a p0 LLT.
I'm not able to get rid of the G_PTRTOINT on the source yet however. We can't
use the same trick of breaking the type system since that could break the
selection of the defining instruction. Thus with -O0 we still end up with a
cross class copy on source.
Code size improvements on -O0:
Program baseline new diff
test-suite :: CTMark/Bullet/bullet.test 965520 949164 -1.7%
test-suite...TMark/7zip/7zip-benchmark.test 1069456 1052600 -1.6%
test-suite...ark/tramp3d-v4/tramp3d-v4.test 1213692 1199804 -1.1%
test-suite...:: CTMark/sqlite3/sqlite3.test 421680 419736 -0.5%
test-suite...-typeset/consumer-typeset.test 837076 833380 -0.4%
test-suite :: CTMark/lencod/lencod.test 799712 796976 -0.3%
test-suite...:: CTMark/ClamAV/clamscan.test 688264 686132 -0.3%
test-suite :: CTMark/kimwitu++/kc.test 1002344 999648 -0.3%
test-suite...Mark/mafft/pairlocalalign.test 422296 421768 -0.1%
test-suite :: CTMark/SPASS/SPASS.test 656792 656532 -0.0%
Geomean difference -0.6%
Differential Revision: https://reviews.llvm.org/D73910
Start using a new strategy with a combination of merge and unmerges.
This allows scalarizing before lowering, which in cases like
<2 x s128> avoids producing giant illegal shifts.
RegAllocGreedy uses a fairly compile time intensive splitting heuristic
called region splitting. This heuristic was disabled via another heuristic
when it is likely that it won't be worth the compile time. The only way
to control this other heuristic was via a command line option (huge-size-for-split).
This commit gives more control on this heuristic by making it overridable
by the target using a target hook in TargetRegisterInfo called
shouldRegionSplitForVirtReg.
The default implementation of this hook keeps the heuristic as it was
before this patch.
ClangBuildAnalyzer results show that a lot of time is spent
instantiating AnalysisManager::getResultImpl across the code base:
**** Templates that took longest to instantiate:
50445 ms: llvm::AnalysisManager<llvm::Function>::getResultImpl (412 times, avg 122 ms)
47797 ms: llvm::AnalysisManager<llvm::Function>::getResult<llvm::TargetLibraryAnalysis> (389 times, avg 122 ms)
46894 ms: std::tie<const unsigned long long, const bool> (2452 times, avg 19 ms)
43851 ms: llvm::BumpPtrAllocatorImpl<llvm::MallocAllocator, 4096, 4096>::Allocate (3228 times, avg 13 ms)
33911 ms: std::tie<const unsigned int, const unsigned int, const unsigned int, const unsigned int> (897 times, avg 37 ms)
33854 ms: std::tie<const unsigned long long, const unsigned long long> (1897 times, avg 17 ms)
27886 ms: std::basic_string<char, std::char_traits<char>, std::allocator<char> >::basic_string (11156 times, avg 2 ms)
I mentioned this result to @chandlerc, and he suggested this direction.
AnalysisManager is already explicitly instantiated, and getResultImpl
doesn't need to be inlined. Move the definition to an Impl header, and
include that header in files that explicitly instantiate
AnalysisManager. There are only four (real) IR units:
- function
- module
- loop
- cgscc
Looking at a specific transform (ArgumentPromotion.cpp), here are three
compilations before & after this change:
BEFORE:
$ for i in $(seq 3) ; do ./ccit.bat ; done
peak memory: 258.15MB
real: 0m6.297s
peak memory: 257.54MB
real: 0m5.906s
peak memory: 257.47MB
real: 0m6.219s
AFTER:
$ for i in $(seq 3) ; do ./ccit.bat ; done
peak memory: 235.35MB
real: 0m5.454s
peak memory: 234.72MB
real: 0m5.235s
peak memory: 234.39MB
real: 0m5.469s
The 20MB of memory saved seems real, and the time improvement seems like
it is there.
Reviewed By: MaskRay
Differential Revision: https://reviews.llvm.org/D73817
Summary:
Method appendLoopsToWorklist is duplicate in LoopUnroll and in the
LoopPassManager as an internal method. Make it an utility.
Reviewers: dmgreen, chandlerc, fedor.sergeev, yamauchi
Subscribers: mehdi_amini, hiraditya, zzheng, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73569
Summary:
* Most of the simplifications in SimplifyShuffleVectorInst depend on the
concrete value of, or the length of the mask vector. For scalable
vectors, this cannot be known at compile time.
** for these tests, detect if the vector is scalable before attempting
the transformation
* The functions ShuffleVectorInst::getMaskValue and
ShuffleVectorInst::getShuffleMask access the value of the constant mask.
However, since the length of the mask is unknown at compile time, these
function do not work for scalable vectors. Add asserts to ensure that
the input mask is not scalable
Reviewers: efriedma, sdesmalen, apazos, chrisj, huihuiz
Reviewed By: efriedma
Subscribers: tschuett, hiraditya, rkruppe, psnobl, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73555
Adds a replaceOperand() helper, which is like Instruction.setOperand()
but adds the old operand to the worklist. This reduces the amount of
missing or incorrect worklist management.
This only applies the helper to a relatively small subset of
setOperand() calls in InstCombine, namely those of the pattern
`I.setOperand(); return &I;`, where it is most obviously applicable.
Differential Revision: https://reviews.llvm.org/D73803
This renames Worklist.AddDeferred() to Worklist.add() and
Worklist.Add() to Worklist.push(). The intention here is that
Worklist.add() should be the go-to method for explicit worklist
management, while the raw Worklist.push() is mostly for
InstCombine internals. I will then migrate uses of Worklist.push()
to Worklist.add() in followup changes.
As suggested by spatel on D73411 I'm also changing the remaining
method names to lowercase first character, in line with current
coding standards.
Differential Revision: https://reviews.llvm.org/D73745
Followup to D73135. If the target doesn't have hard float (default
for ARM), then we assert when trying to soften the result of vector
reduction intrinsics. This patch marks these for expansion as well.
(A bit odd to use vectors on a target without hard float ... but
that's where you end up if you expose target-independent vector types.)
Differential Revision: https://reviews.llvm.org/D73854
Summary:
A recent change to enable more importing of global variables with
references exposed some efficiency issues with export computation.
See D73724 for more information and detailed analysis.
The first was specific to variable importing. The code was marking every
copy of a referenced value (from possibly thousands of files in the case
of linkonce_odr) as exported, and we only need to mark the copy in the
module containing the variable def being imported as exported. The
reason is that this is tracking what values are newly exported as a
result of importing. Anything that was defined in another module and
simply used in the exporting module is already exported, and would have
been identified by the caller (e.g. the LTO API implementations).
The second issue is that the code was re-adding previously exported
values (along with all references). It is easy to identify when a
variable was already imported into the same module (via the
import list insert call return value), and we already did this for
function importing. However, what we weren't doing for either function
or variable importing was avoiding a re-insertion when it was previously
exported into a different importing module. The reason we couldn't do
this is there was no way of telling from the export list whether it was
previously inserted there because its definition was exported (in which
case we already marked all its references as exported) from when it was
inserted there because it was referenced by another exported value (in
which case we haven't yet inserted its own references).
To address this we can restructure the way the export list is
constructed. This patch only adds the actual imported definitions
(variable or function) to the export list for its module during the
import computation. After import computation is complete, where we were
already post-processing the export list we go ahead and add all
references made by those exported values to the export list.
These changes speed up the thin link not only with constant variable
importing enabled, but also without (due to the efficiency improvement
in function importing).
Some thin link user time measurements for one large application, average
of 5 runs:
With constant variable importing enabled:
- without this patch: 479.5s
- with this patch: 74.6s
Without constant variable importing enabled:
- without this patch: 80.6s
- with this patch: 70.3s
Note I have not re-enabled constant variable importing here, as I would
like to do additional compile time measurements with these fixes first.
Reviewers: evgeny777
Subscribers: mehdi_amini, inglorion, hiraditya, dexonsmith, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73851
As detailed on PR43463, this fixes a static analyzer null dereference warning by sinking Changed = true into the if() blocks where the MIB is actually created.
I did a quick check that suggested that one of those if() blocks is always guaranteed to be hit (so we could change it to if-else), but this seems like a safer approach
Differential Revision: https://reviews.llvm.org/D73883
We have to be careful in SimplifyDemandedBits with loads in case we attempt to combine back to a constant (which then gets turned into a constant pool load again), but we can at least set the upper KnownBits for a ZEXTLoad to zero.
