In some circumstances we can end up with setup costs that are very complex to
compute, even though the scevs are not very complex to create. This can also
lead to setupcosts that are calculated to be exactly -1, which LSR treats as an
invalid cost. This patch puts a limit on the recursion depth for setup cost to
prevent them taking too long.
Thanks to @reames for the report and test case.
Differential Revision: https://reviews.llvm.org/D60944
llvm-svn: 358958
This reverts r358910 (git commit 2b74466530)
While this patch *seems* trivial and safe and correct, it is not. The
copies are actually load bearing copies. You can observe this with MSan
or other ways of checking for use-after-destroy, but otherwise this may
result in ... difficult to debug inexplicable behavior.
I suspect the issue is that the debug location is used after the
original reference to it is removed. The metadata backing it gets
destroyed as its last references goes away, and then we reference it
later through these const references.
llvm-svn: 358940
Currently to opt in to debug_names in DWARFv5, the IR must contain
'nameTableKind: Default' which also enables debug_pubnames.
Instead, only allow one of {debug_names, apple_names, debug_pubnames,
debug_gnu_pubnames}.
nameTableKind: Default gives debug_names in DWARFv5 and greater,
debug_pubnames in v4 and earlier - and apple_names when tuning for lldb
on MachO.
nameTableKind: GNU always gives gnu_pubnames
llvm-svn: 358931
This was supposed to be NFC, but the change in SDLoc
definitions causes instruction scheduling changes.
There's nothing x86-specific in this code, and it can
likely be used from DAGCombiner's simplifyVBinOp().
llvm-svn: 358930
Summary:
- Only apply packed literal `op_sel_hi` skipping on operands requiring
packed literals. Even an instruction is `packed`, it may have operand
requiring non-packed literal, such as `v_dot2_f32_f16`.
Reviewers: rampitec, arsenm, kzhuravl
Subscribers: jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D60978
llvm-svn: 358922
If we have a store to a piece of memory which is known constant, then we know the store must be storing back the same value. As a result, the store (or memset, or memmove) must either be down a dead path, or a noop. In either case, it is valid to simply remove the store.
The motivating case for this involves a memmove to a buffer which is constant down a path which is dynamically dead.
Note that I'm choosing to implement the less aggressive of two possible semantics here. We could simply say that the store *is undefined*, and prune the path. Consensus in the review was that the more aggressive form might be a good follow on change at a later date.
Differential Revision: https://reviews.llvm.org/D60659
llvm-svn: 358919
In the process, use the existing masked.load combine which is slightly stronger, and handles a mix of zero and undef elements in the mask.
llvm-svn: 358913
These are inserted after branch relaxation, and for some reason it's
decided to put them in the long branch expansion block. It's probably
not great to rely on the source block address, so this should probably
be switched to being PC relative instead of relying on the block
address
llvm-svn: 358909
Back in August, r340525 introduced a dependency on the assumption
cache tracker in the ipsccp pass, but that commit missed a call to
INITIALIZE_PASS_DEPENDENCY, which leaves the assumption cache
improperly registered if SCCP is the only thing that pulls it in.
llvm-svn: 358903
Currently, we do not expose BPI to loop passes at all. In the old pass manager, we appear to have been ignoring the fact that LCSSA and/or LoopSimplify didn't preserve BPI, and making it available to the following loop passes anyways. In the new one, it's invalidated before running any loop pass if either LCSSA or LoopSimplify actually make changes. If they don't make changes, then BPI is valid and available. So, we go ahead and teach LCSSA and LoopSimplify how to preserve BPI for consistency between old and new pass managers.
This patch avoids an invalidation between the two requires in the following trivial pass pipeline:
opt -passes="requires<branch-prob>,loop(no-op-loop),requires<branch-prob>"
(when the input file is one which requires either LCSSA or LoopSimplify to canonicalize the loops)
Differential Revision: https://reviews.llvm.org/D60790
llvm-svn: 358901
to CallInst.
The issue was raised here: https://reviews.llvm.org/D60903#1472783
The function Instruction::updateProfWeight is only used for CallInst in
profile update. From the current interface, it is very easy to think that
the function can also be used for branch instruction. However, Branch
instruction does't need the scaling the function provides for
branch_weights and VP (value profile), in addition, scaling may introduce
inaccuracy for branch probablity.
The patch moves the function updateProfWeight from Instruction class to
CallInst to remove the confusion. The patch also changes the scaling of
branch_weights from a loop to a block because we know that ProfileData
for branch_weights of CallInst will only have two operands at most.
Differential Revision: https://reviews.llvm.org/D60911
llvm-svn: 358900
This patch adds support for BigBitWidth -> SmallBitWidth bitcasts, splitting the DemandedBits/Elts accordingly.
The AMDGPU backend needed an extra (srl (and x, c1 << c2), c2) -> (and (srl(x, c2), c1) combine to encourage BFE creation, I investigated putting this in DAGCombine but it caused a lot of noise on other targets - some improvements, some regressions.
The X86 changes are all definite wins.
Differential Revision: https://reviews.llvm.org/D60462
llvm-svn: 358887
This reverts commit 7bf4d7c07f2fac862ef34c82ad0fef6513452445.
After thinking about this more, this isn't right, the range is not exact
in the same sense as makeExactICmpRegion(). This needs a separate
function.
llvm-svn: 358876
Following D60632 makeGuaranteedNoWrapRegion() always returns an
exact nowrap region. Rename the function accordingly. This is in
line with the naming of makeExactICmpRegion().
llvm-svn: 358875
Section atoms are not sorted, so we need to scan the whole section to find the
start address.
No test case: Found by inspection, and any reproduction would depend on pointer
ordering.
llvm-svn: 358865
llvm-undname used to put '\x' in front of every pair of nibbles, but
u"\xD7\xFF" produces a string with 6 bytes: \xD7 \0 \xFF \0 (and \0\0). Correct
for a single character (plus terminating \0) is u\xD7FF instead.
Now, wchar_t, char16_t, and char32_t strings roundtrip from source to
clang-cl (and cl.exe) and then llvm-undname.
(...at least as long as it's not a string like L"\xD7FF" L"foo" which
gets demangled as L"\xD7FFfoo", where the compiler then considers the
"f" as part of the hex escape. That seems ok.)
Also add a comment saying that the "almost-valid" char32_t string I
added in my last commit is actually produced by compilers.
llvm-svn: 358857
If a unsigned with all 4 bytes non-0 was passed to outputHex(), there
were two off-by-ones in it:
- Both MaxPos and Pos left space for the final \0, which left the buffer
one byte to small. Set MaxPos to 16 instead of 15 to fix.
- The `assert(Pos >= 0);` was after a `Pos--`, move it up one line.
Since valid Unicode codepoints are <= 0x10ffff, this could never really
happen in practice.
Found by oss-fuzz.
llvm-svn: 358856
Add support for uadd_sat and friends to ConstantRange, so we can
handle uadd.sat and friends in LVI. The implementation is forwarding
to the corresponding APInt methods with appropriate bounds.
One thing worth pointing out here is that the handling of wrapping
ranges is not maximally accurate. A simple example is that adding 0
to a wrapped range will return a full range, rather than the original
wrapped range. The tests also only check that the non-wrapping
envelope is correct and minimal.
Differential Revision: https://reviews.llvm.org/D60946
llvm-svn: 358855
ConstantRanges have an annoying special case: If upper and lower are
the same, it can be either an empty or a full set. When constructing
constant ranges nearly always a full set is intended, but this still
requires an explicit check in many places.
This revision adds a getNonEmpty() constructor that disambiguates this
case: If upper and lower are the same, a full set is created.
Differential Revision: https://reviews.llvm.org/D60947
llvm-svn: 358854
This does two main things, firstly adding some at least basic addressing modes
for i64 types, and secondly treats floats and doubles sensibly when there is no
fpu. The floating point change can help codesize in some cases, especially with
D60294.
Most backends seems to not consider the exact VT in isLegalAddressingMode,
instead switching on type size. That is now what this does when the target does
not have an fpu (as the float data will be loaded using LDR's). i64's currently
use the address range of an LDRD (even though they may be legalised and loaded
with an LDR). This is at least better than marking them all as illegal
addressing modes.
I have not attempted to do much with vectors yet. That will need changing once
MVE is added.
Differential Revision: https://reviews.llvm.org/D60677
llvm-svn: 358845
The error check required FDEs to refer to the most recent CIE, but the eh-frame
spec allows them to refer to any previously seen CIE. This patch removes the
offending check.
llvm-svn: 358840
- Don't assert when a string looks like a u32 string to the heuristic
but doesn't have a length that's 0 mod 4. Instead, classify those
as u16 with embedded \0 chars. Found by oss-fuzz.
- Print embedded nul bytes as \0 instead of \x00.
llvm-svn: 358835
Knowing the address/symbolnum field values makes it easier to identify the
unsupported relocation, and provides enough information for the full bit
pattern of the relocation to be reconstructed.
llvm-svn: 358833
Summary:
JITLink is a jit-linker that performs the same high-level task as RuntimeDyld:
it parses relocatable object files and makes their contents runnable in a target
process.
JITLink aims to improve on RuntimeDyld in several ways:
(1) A clear design intended to maximize code-sharing while minimizing coupling.
RuntimeDyld has been developed in an ad-hoc fashion for a number of years and
this had led to intermingling of code for multiple architectures (e.g. in
RuntimeDyldELF::processRelocationRef) in a way that makes the code more
difficult to read, reason about, extend. JITLink is designed to isolate
format and architecture specific code, while still sharing generic code.
(2) Support for native code models.
RuntimeDyld required the use of large code models (where calls to external
functions are made indirectly via registers) for many of platforms due to its
restrictive model for stub generation (one "stub" per symbol). JITLink allows
arbitrary mutation of the atom graph, allowing both GOT and PLT atoms to be
added naturally.
(3) Native support for asynchronous linking.
JITLink uses asynchronous calls for symbol resolution and finalization: these
callbacks are passed a continuation function that they must call to complete the
linker's work. This allows for cleaner interoperation with the new concurrent
ORC JIT APIs, while still being easily implementable in synchronous style if
asynchrony is not needed.
To maximise sharing, the design has a hierarchy of common code:
(1) Generic atom-graph data structure and algorithms (e.g. dead stripping and
| memory allocation) that are intended to be shared by all architectures.
|
+ -- (2) Shared per-format code that utilizes (1), e.g. Generic MachO to
| atom-graph parsing.
|
+ -- (3) Architecture specific code that uses (1) and (2). E.g.
JITLinkerMachO_x86_64, which adds x86-64 specific relocation
support to (2) to build and patch up the atom graph.
To support asynchronous symbol resolution and finalization, the callbacks for
these operations take continuations as arguments:
using JITLinkAsyncLookupContinuation =
std::function<void(Expected<AsyncLookupResult> LR)>;
using JITLinkAsyncLookupFunction =
std::function<void(const DenseSet<StringRef> &Symbols,
JITLinkAsyncLookupContinuation LookupContinuation)>;
using FinalizeContinuation = std::function<void(Error)>;
virtual void finalizeAsync(FinalizeContinuation OnFinalize);
In addition to its headline features, JITLink also makes other improvements:
- Dead stripping support: symbols that are not used (e.g. redundant ODR
definitions) are discarded, and take up no memory in the target process
(In contrast, RuntimeDyld supported pointer equality for weak definitions,
but the redundant definitions stayed resident in memory).
- Improved exception handling support. JITLink provides a much more extensive
eh-frame parser than RuntimeDyld, and is able to correctly fix up many
eh-frame sections that RuntimeDyld currently (silently) fails on.
- More extensive validation and error handling throughout.
This initial patch supports linking MachO/x86-64 only. Work on support for
other architectures and formats will happen in-tree.
Differential Revision: https://reviews.llvm.org/D58704
llvm-svn: 358818
Summary:
If you pass two 1024 bit vectors in IR with AVX2 on Windows 64. Both vectors will be split in four 256 bit pieces. The four pieces of the first argument will be passed indirectly using 4 gprs. The second argument will get passed via pointers in memory.
The PartOffsets stored for the second argument are all in terms of its original 1024 bit size. So the PartOffsets for each piece are 32 bytes apart. So if we consider it for copy elision we'll only load an 8 byte pointer, but we'll move the address 32 bytes. The stack object size we create for the first part is probably wrong too.
This issue was encountered by ISPC. I'm working on getting a reduce test case, but wanted to go ahead and get feedback on the fix.
Reviewers: rnk
Reviewed By: rnk
Subscribers: dbabokin, llvm-commits, hiraditya
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D60801
llvm-svn: 358817
Summary:
Teach CorrelatedValuePropagation to also handle sub instructions in addition to add. Relatively simple since makeGuaranteedNoWrapRegion already understood sub instructions. Only subtle change is which range is passed as "Other" to that function, since sub isn't commutative.
Note that CorrelatedValuePropagation::processAddSub is still hidden behind a default-off flag as IndVarSimplify hasn't yet been fixed to strip the added nsw/nuw flags and causes a miscompile. (PR31181)
Reviewers: sanjoy, apilipenko, nikic
Reviewed By: nikic
Subscribers: hiraditya, jfb, jdoerfert, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D60036
llvm-svn: 358816
This is very minor issue. The returned section index is only used by
DWARFDebugLine as an llvm::upper_bound input and the use case shouldn't
cause any behavioral change.
llvm-svn: 358814
Fix for https://bugs.llvm.org/show_bug.cgi?id=41477. On the x32 ABI
with stack probing a dynamic alloca will result in a WIN_ALLOCA_32
with a 32-bit size. The current implementation tries to copy it into
RAX, resulting in a physreg copy error. Fix this by copying to EAX
instead. Also fix incorrect opcodes or registers used in subs.
llvm-svn: 358807
The MOVZX doesn't require an immediate to be encoded at all. Though it does use
a 2 byte opcode so its the same size as a 1 byte immediate. But it has a
separate source and dest register so can help avoid copies.
llvm-svn: 358805
There's one slight regression in here because we don't check that the immediate
already allowed movzx before the shift. I'll fix that next.
llvm-svn: 358804
Exactly the same as G_FCEIL, G_FABS, etc.
Add tests for the fp16/nofp16 behaviour, update arm64-vfloatintrinsics, etc.
Differential Revision: https://reviews.llvm.org/D60895
llvm-svn: 358799
My understanding is that once BuildMI has been called we can't fallback
to SelectionDAG.
This change moves the fallback for when getRegForValue() fails for
that target of an indirect call. This was failing in -fPIC mode when
the callee is GlobalValue.
Add a test case that tickles this.
Differential Revision: https://reviews.llvm.org/D60908
llvm-svn: 358793
This is a follow-up to r291037+r291258, which used null debug locations
to prevent jumpy line tables.
Using line 0 locations achieves the same effect, but works better for
crash attribution because it preserves the right inline scope.
Differential Revision: https://reviews.llvm.org/D60913
llvm-svn: 358791
VK_SABS is part of the SymLoc bitfield in the variant kind which should
be compared for equality, not by checking the VK_SABS bit.
As far as I know, the existing code happened to produce the correct
results in all cases, so this is just a cleanup.
Patch by Stephen Crane.
Differential Revision: https://reviews.llvm.org/D60596
llvm-svn: 358788
Summary:
This emits labels around heapallocsite calls and S_HEAPALLOCSITE debug
info in codeview. Currently only changes FastISel, so emitting labels still
needs to be implemented in SelectionDAG.
Reviewers: hans, rnk
Subscribers: aprantl, hiraditya, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D60800
llvm-svn: 358783
Summary:
Make the flags in LICM + MemorySSA tuning options in the old and new
pass managers.
Subscribers: mehdi_amini, jlebar, Prazek, george.burgess.iv, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D60490
llvm-svn: 358772
This instruction is legalized in the same way as G_FSIN, G_FCOS, G_FLOG10, etc.
Update legalize-pow.mir and arm64-vfloatintrinsics.ll to reflect the change.
Differential Revision: https://reviews.llvm.org/D60218
llvm-svn: 358764
Summary:
Trying to add the plumbing necessary to add tuning options to the new pass manager.
Testing with the flags for loop vectorize.
Reviewers: chandlerc
Subscribers: sanjoy, mehdi_amini, jlebar, steven_wu, dexonsmith, dang, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D59723
llvm-svn: 358763
These are general queries, so they should not die when given
a degenerate input like an all undef mask. Callers should be
able to deal with an op that will eventually be simplified away.
llvm-svn: 358761
Gold and ld on Linux already support saving stats, but the
infrastructure is missing on Darwin. Unfortunately it seems like the
configuration from lib/LTO/LTO.cpp is not used.
This patch adds a new LTOStatsFile option and adds plumbing in Clang to
use it on Darwin, similar to the way remarks are handled.
Currnetly the handling of LTO flags seems quite spread out, with a bunch
of duplication. But I am not sure if there is an easy way to improve
that?
Reviewers: anemet, tejohnson, thegameg, steven_wu
Reviewed By: steven_wu
Differential Revision: https://reviews.llvm.org/D60516
llvm-svn: 358753
Summary:
The basic idea here is to make it possible to use
MachineInstr::mayAlias also when the MachineInstr
is const (or the "Other" MachineInstr is const).
The addition of const in MachineInstr::mayAlias
then rippled down to the need for adding const
in several other places, such as
TargetTransformInfo::getMemOperandWithOffset.
Reviewers: hfinkel
Reviewed By: hfinkel
Subscribers: hfinkel, MatzeB, arsenm, jvesely, nhaehnle, hiraditya, javed.absar, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D60856
llvm-svn: 358744
Pending instructions that may have been blocked from being available by the HazardRecognizer may no longer may not be blocked any more when an instruction is scheduled; pending instructions should be re-checked in this case.
This is primarily aimed at VLIW targets with large parallelism and esoteric constraints.
No testcase as no in-tree targets have this behavior.
Differential revision: https://reviews.llvm.org/D60861
llvm-svn: 358743
removeUsers uses a work list to collect indirect users and call remove()
on those functions. However it has a bug (`if (!Visited.insert(UU).second)`).
Actually, we don't have to collect indirect users.
After the merge of F and G, G's callers will be considered (added to
Deferred). If G's callers can be merged, G's callers' callers will be
considered.
Update the test unnamed-addr-reprocessing.ll to make it clear we can
still merge indirect callers.
llvm-svn: 358741
Summary:
Ignore edges to non-SUnits (e.g. ExitSU) when checking
for low latency instructions.
When calling the function isLowLatencyInstruction(),
an ExitSU could be on the list of successors, not necessarily
a regular SU. In other places in the code there is a check
"Succ->NodeNum >= DAGSize" to prevent further processing of
ExitSU as "Succ->getInstr()" is NULL in such a case.
Also, 8 out of 9 cases of "SUnit *Succ = SuccDep.getSUnit())"
has the guard, so it is clearly an omission here.
Change-Id: Ica86f0327c7b2e6bcb56958e804ea6c71084663b
Reviewers: nhaehnle
Reviewed By: nhaehnle
Subscribers: MatzeB, arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, javed.absar, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D60864
llvm-svn: 358740
code to `CallBase`.
This patch focuses on the legacy PM, call graph, and some of inliner and legacy
passes interacting with those APIs from `CallSite` to the new `CallBase` class.
No interesting changes.
Differential Revision: https://reviews.llvm.org/D60412
llvm-svn: 358739
Summary:
There are two places where we create a HandleSDNode in address matching in order to handle the case where N is changed by CSE. But if we end up not matching, we fall back to code at the bottom of the switch that really would like N to point to something that wasn't CSEd away. So we should make sure we copy the handle back to N on any paths that can reach that code.
This appears to be the true reason we needed to check DELETED_NODE in the negation matching. In pr32329.ll we had two subtracts back to back. We recursed through the first subtract, and onto the second subtract. The second subtract called matchAddressRecursively on its LHS which caused that subtract to CSE. We ultimately failed the match and ended up in the default code. But N was pointing at the old node that had been deleted, but the default code didn't know that and took it as the base register. Then we unwound back to the first subtract and tried to access this bogus base reg requiring the check for deleted node. With this patch we now use the CSE result as the base reg instead.
matchAdd has been broken since sometime in 2015 when it was pulled out of the switch into a helper function. The assignment to N at the end was still there, but N was passed by value and not by reference so the update didn't go anywhere.
Reviewers: niravd, spatel, RKSimon, bkramer
Reviewed By: niravd
Subscribers: llvm-commits, hiraditya
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D60843
llvm-svn: 358735
Another attempt to land the changes in debug line header to prevent duplicate
files in Dwarf 5. I rolled back my previous commit because of a mistake in
generating the object file in a test. Meanwhile, I addressed some offline
comments and changed the implementation; the largest difference is that
MCDwarfLineTableHeader does not keep DwarfVersion but gets it as a parameter. I
also merged the patch to fix two lld tests that will strt to fail into this
patch.
Original Commit:
https://reviews.llvm.org/D59515
Original Message:
Motivation: In previous dwarf versions, file name indexes started from 1, and
the primary source file was not explicit. Dwarf 5 standard (6.2.4) prescribes
the primary source file to be explicitly given an entry with an index number 0.
The current implementation honors the specification by just duplicating the
main source file, once with index number 0, and later maybe with another
index number. While this is compliant with the letter of the standard, the
duplication causes problems for consumers of this information such as lldb.
(Some files are duplicated, where only some of them have a line table although
all refer to the same file)
With this change, dwarf 5 debug line section files always start from 0, and
the zeroth entry is not duplicated whenever possible. This requires different
handling of dwarf 4 and dwarf 5 during generation (e.g. when a function returns
an index zero for a file name, it signals an error in dwarf 4, but not in dwarf
5) However, I think the minor complication is worth it, because it enables all
consumers (lldb, gdb, dwarfdump, objdump, and so on) to treat all files in the
file name list homogenously.
llvm-svn: 358732
Change two costly udiv() calls to lshr(1)*RHS + left-shift + plus
On one 64-bit umul_ov benchmark, I measured an obvious improvement: 12.8129s -> 3.6257s
Note, there may be some value to special case 64-bit (the most common
case) with __builtin_umulll_overflow().
Differential Revision: https://reviews.llvm.org/D60669
llvm-svn: 358730
We would previously drop the COMDAT on the thunk we generated when replacing a
function body with the forwarding thunk. This would result in a function that
may have been multiply emitted and multiply merged to be emitted with the same
name without the COMDAT. This is a hard error with PE/COFF where the COMDAT is
used for the deduplication of Value Witness functions for Swift.
llvm-svn: 358728
This adds legalization for G_SEXT, G_ZEXT, and G_ANYEXT for v8s8s.
We were falling back on G_ZEXT in arm64-vabs.ll before, preventing us from
selecting the @llvm.aarch64.neon.sabd.v8i8 intrinsic.
This adds legalizer support for those 3, which gives us selection via the
importer. Update the relevant tests (legalize-ext.mir, select-int-ext.mir) and
add a GISel line to arm64-vabs.ll.
Differential Revision: https://reviews.llvm.org/D60881
llvm-svn: 358715
Prior to this patch, each basic block listed in the extrack-blocks-file
would be extracted to a different function.
This patch adds the support for comma separated list of basic blocks
to form group.
When the region formed by a group is not extractable, e.g., not single
entry, all the blocks of that group are left untouched.
Let us see this new format in action (comments are not part of the
file format):
;; funcName bbName[,bbName...]
foo bb1 ;; Extract bb1 in its own function
foo bb2,bb3 ;; Extract bb2,bb3 in their own function
bar bb1,bb4 ;; Extract bb1,bb4 in their own function
bar bb2 ;; Extract bb2 in its own function
Assuming all regions are extractable, this will create one function and
thus one call per region.
Differential Revision: https://reviews.llvm.org/D60746
llvm-svn: 358701
combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together.
This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356).
This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit.
My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great.
Differential Revision: https://reviews.llvm.org/D60375#inline-539623
llvm-svn: 358692
The bug is that I didn't check whether the operand of the invariant_loads were themselves invariant. I don't know how this got missed in the patch and review. I even had an unreduced test case locally, and I remember handling this case, but I must have lost it in one of the rebases. Oops.
llvm-svn: 358688
The purpose of this patch is to eliminate a pass ordering dependence between LoopPredication and LICM. To understand the purpose, consider the following snippet of code inside some loop 'L' with IV 'i'
A = _a.length;
guard (i < A)
a = _a[i]
B = _b.length;
guard (i < B);
b = _b[i];
...
Z = _z.length;
guard (i < Z)
z = _z[i]
accum += a + b + ... + z;
Today, we need LICM to hoist the length loads, LoopPredication to make the guards loop invariant, and TrivialUnswitch to eliminate the loop invariant guard to establish must execute for the next length load. Today, if we can't prove speculation safety, we'd have to iterate these three passes 26 times to reduce this example down to the minimal form.
Using the fact that the array lengths are known to be invariant, we can short circuit this iteration. By forming the loop invariant form of all the guards at once, we remove the need for LoopPredication from the iterative cycle. At the moment, we'd still have to iterate LICM and TrivialUnswitch; we'll leave that part for later.
As a secondary benefit, this allows LoopPred to expose peeling oppurtunities in a much more obvious manner. See the udiv test changes as an example. If the udiv was not hoistable (i.e. we couldn't prove speculation safety) this would be an example where peeling becomes obviously profitable whereas it wasn't before.
A couple of subtleties in the implementation:
- SCEV's isSafeToExpand guarantees speculation safety (i.e. let's us expand at a new point). It is not a precondition for expansion if we know the SCEV corresponds to a Value which dominates the requested expansion point.
- SCEV's isLoopInvariant returns true for expressions which compute the same value across all iterations executed, regardless of where the original Value is located. (i.e. it can be in the loop) This implies we have a speculation burden to prove before expanding them outside loops.
- invariant_loads and AA->pointsToConstantMemory are two cases that SCEV currently does not handle, but meets the SCEV definition of invariance. I plan to sink this part into SCEV once this has baked for a bit.
Differential Revision: https://reviews.llvm.org/D60093
llvm-svn: 358684
David Green