bitcode support change for fast flags compatibility
Summary: The discussion and as per need, each vendor needs a way to keep the old fast flags and the new fast flags in the auto upgrade path of the IR upgrader. This revision addresses that issue.
[AArch64][GlobalISel] When copying from a gpr32 to an fpr16 reg, convert to fpr32 first.
This is a follow on commit to r[x] where we fix the other direction of copy.
For this case, after converting the source from gpr32 -> fpr32, we use a
subregister copy, which is essentially what EXTRACT_SUBREG does in SDAG land.
[LV] Use Demanded Bits and ValueTracking for reduction type-shrinking
The type-shrinking logic in reduction detection, although narrow in scope, is
also rather ad-hoc, which has led to bugs (e.g., PR35734). This patch modifies
the approach to rely on the demanded bits and value tracking analyses, if
available. We currently perform type-shrinking separately for reductions and
other instructions in the loop. Long-term, we should probably think about
computing minimal bit widths in a more complete way for the loops we want to
vectorize.
[LICM] update BlockColors after splitting predecessors
Update BlockColors after splitting predecessors. Do not allow splitting
EHPad for sinking when the BlockColors is not empty, so we can
simply assign predecessor's color to the new block.
[Dominators] Always recalculate postdominators when update yields different roots
Summary:
This patch makes postdominators always recalculate the tree when an update causes to change the tree roots.
As @dmgreen noticed in [[ https://reviews.llvm.org/D41298 | D41298 ]], the previous implementation was not conservative enough and it was possible to end up with a PostDomTree that was different than a freshly computed one.
The patch also compares postdominators with a freshly computed tree at the end of full verification to make sure we don't hit similar issues in the future.
This should (ideally) be also backported to 6.0 before the release, although I don't have any reports of this causing an observable error. It should be safe to do it even if it's late in the release, as the change only makes the current behavior more conservative.
[X86] Don't emit KTEST instructions unless only the Z flag is being used
Summary:
KTEST has weird flag behavior. The Z flag is set for all bits in the AND of the k-registers being 0, and the C flag is set for all bits being 1. All other flags are cleared.
We currently emit this instruction in EmitTEST and don't check the condition code. This can lead to strange things like using the S flag after a KTEST for a signed compare.
The domain reassignment pass can also transform TEST instructions into KTEST and is not protected against the flag usage either. For now I've disabled this part of the domain reassignment pass. I tried to comment out the checks in the mir test so that we could recover them later, but I couldn't figure out how to get that to work.
This patch moves the KTEST handling into LowerSETCC and now creates a ktest+x86setcc. I've chosen this approach because I'd like to add support for the C flag for all ones in a followup patch. To do that requires that I can rewrite the condition code going in the x86setcc to be different than the original SETCC condition code.
This fixes PR36182. I'll file a PR to fix domain reassignment once this goes in. Should this be merged to 6.0?
[X86] Use EDI for retpoline when no scratch regs are left
Summary:
Instead of solving the hard problem of how to pass the callee to the indirect
jump thunk without a register, just use a CSR. At a call boundary, there's
nothing stopping us from using a CSR to hold the callee as long as we save and
restore it in the prologue.
Also, add tests for this mregparm=3 case. I wrote execution tests for
__llvm_retpoline_push, but they never got committed as lit tests, either
because I never rewrote them or because they got lost in merge conflicts.
This allows the register name to be printed without the leading '%'.
This can be used for emitting calls to the retpoline thunks from inline
asm.
------------------------------------------------------------------------
[x86/retpoline] Make the external thunk names exactly match the names
that happened to end up in GCC.
This is really unfortunate, as the names don't have much rhyme or reason
to them. Originally in the discussions it seemed fine to rely on aliases
to map different names to whatever external thunk code developers wished
to use but there are practical problems with that in the kernel it turns
out. And since we're discovering this practical problems late and since
GCC has already shipped a release with one set of names, we are forced,
yet again, to blindly match what is there.
Somewhat rushing this patch out for the Linux kernel folks to test and
so we can get it patched into our releases.
Move utility function that depends on codegen.
Fixes build with r324487 reapplied.
------------------------------------------------------------------------
In the rare case where the input contains rip-relative addressing with
immediate displacements, *and* the instruction ends with an immediate,
we encode the instruction in the wrong way:
movl $12345678, 0x400(%rdi) // all good, no rip-relative addr
movl %eax, 0x400(%rip) // all good, no immediate at the end of the instruction
movl $12345678, 0x400(%rip) // fails, encodes address as 0x3fc(%rip)
[LivePhysRegs] Fix handling of return instructions.
See D42509 for the original version of this.
Basically, there are two significant changes to behavior here:
- addLiveOuts always adds all pristine registers (even if a block has
no successors).
- addLiveOuts and addLiveOutsNoPristines always add all callee-saved
registers for return blocks (including conditional return blocks).
I cleaned up the functions a bit to make it clear these properties hold.
%0 will have a live interval with 3 value numbers (for the BB0, BB1 and
BB2 parts). Now SplitKit tries and succeeds in rematerializing the value
number in BB2 (This only works because it is a secondary split so
SplitKit is can trace this back to a single original def).
We need to recompute all live ranges affected by a value number that we
rematerialize. The case that we missed before is that when the value
that is rematerialized is at a join (Phi VNI) then we also have to
recompute liveness for the predecessor VNIs.
[DAGCombiner] When folding (insert_subvector undef, (bitcast (extract_subvector N1, Idx)), Idx) -> (bitcast N1) make sure that N1 has the same total size as the original output
We were only checking the element count, but not the total width. This could cause illegal bitcasts to be created if for example the output was 512-bits, but N1 is 256 bits, and the extraction size was 128-bits.
Summary: This was broken long ago in D12208, which failed to account for
the fact that 64-bit SPARC uses a stack bias of 2047, and it is the
*unbiased* value which should be aligned, not the biased one. This was
seen to be an issue with Rust.
[mips] Fix incorrect sign extension for fpowi libcall
PR36061 showed that during the expansion of ISD::FPOWI, that there
was an incorrect zero extension of the integer argument which for
MIPS64 would then give incorrect results. Address this with the
existing mechanism for correcting sign extensions.
[ARM] Allow the scheduler to clone a node with glue to avoid a copy CPSR ↔ GPR.
In Thumb 1, with the new ADDCARRY / SUBCARRY the scheduler may need to do
copies CPSR ↔ GPR but not all Thumb1 targets implement them.
The schedule can attempt, before attempting a copy, to clone the instructions
but it does not currently do that for nodes with input glue. In this patch we
introduce a target-hook to let the hook decide if a glued machinenode is still
eligible for copying. In this case these are ARM::tADCS and ARM::tSBCS .
As a follow-up of this change we should actually implement the copies for the
Thumb1 targets that do implement them and restrict the hook to the targets that
can't really do such copy as these clones are not ideal.
[x86] Make the retpoline thunk insertion a machine function pass.
Summary:
This removes the need for a machine module pass using some deeply
questionable hacks. This should address PR36123 which is a case where in
full LTO the memory usage of a machine module pass actually ended up
being significant.
We should revert this on trunk as soon as we understand and fix the
memory usage issue, but we should include this in any backports of
retpolines themselves.
Introduce the "retpoline" x86 mitigation technique for variant #2 of the speculative execution vulnerabilities disclosed today, specifically identified by CVE-2017-5715, "Branch Target Injection", and is one of the two halves to Spectre..
Summary:
First, we need to explain the core of the vulnerability. Note that this
is a very incomplete description, please see the Project Zero blog post
for details:
https://googleprojectzero.blogspot.com/2018/01/reading-privileged-memory-with-side.html
The basis for branch target injection is to direct speculative execution
of the processor to some "gadget" of executable code by poisoning the
prediction of indirect branches with the address of that gadget. The
gadget in turn contains an operation that provides a side channel for
reading data. Most commonly, this will look like a load of secret data
followed by a branch on the loaded value and then a load of some
predictable cache line. The attacker then uses timing of the processors
cache to determine which direction the branch took *in the speculative
execution*, and in turn what one bit of the loaded value was. Due to the
nature of these timing side channels and the branch predictor on Intel
processors, this allows an attacker to leak data only accessible to
a privileged domain (like the kernel) back into an unprivileged domain.
The goal is simple: avoid generating code which contains an indirect
branch that could have its prediction poisoned by an attacker. In many
cases, the compiler can simply use directed conditional branches and
a small search tree. LLVM already has support for lowering switches in
this way and the first step of this patch is to disable jump-table
lowering of switches and introduce a pass to rewrite explicit indirectbr
sequences into a switch over integers.
However, there is no fully general alternative to indirect calls. We
introduce a new construct we call a "retpoline" to implement indirect
calls in a non-speculatable way. It can be thought of loosely as
a trampoline for indirect calls which uses the RET instruction on x86.
Further, we arrange for a specific call->ret sequence which ensures the
processor predicts the return to go to a controlled, known location. The
retpoline then "smashes" the return address pushed onto the stack by the
call with the desired target of the original indirect call. The result
is a predicted return to the next instruction after a call (which can be
used to trap speculative execution within an infinite loop) and an
actual indirect branch to an arbitrary address.
On 64-bit x86 ABIs, this is especially easily done in the compiler by
using a guaranteed scratch register to pass the target into this device.
For 32-bit ABIs there isn't a guaranteed scratch register and so several
different retpoline variants are introduced to use a scratch register if
one is available in the calling convention and to otherwise use direct
stack push/pop sequences to pass the target address.
This "retpoline" mitigation is fully described in the following blog
post: https://support.google.com/faqs/answer/7625886
We also support a target feature that disables emission of the retpoline
thunk by the compiler to allow for custom thunks if users want them.
These are particularly useful in environments like kernels that
routinely do hot-patching on boot and want to hot-patch their thunk to
different code sequences. They can write this custom thunk and use
`-mretpoline-external-thunk` *in addition* to `-mretpoline`. In this
case, on x86-64 thu thunk names must be:
```
__llvm_external_retpoline_r11
```
or on 32-bit:
```
__llvm_external_retpoline_eax
__llvm_external_retpoline_ecx
__llvm_external_retpoline_edx
__llvm_external_retpoline_push
```
And the target of the retpoline is passed in the named register, or in
the case of the `push` suffix on the top of the stack via a `pushl`
instruction.
There is one other important source of indirect branches in x86 ELF
binaries: the PLT. These patches also include support for LLD to
generate PLT entries that perform a retpoline-style indirection.
The only other indirect branches remaining that we are aware of are from
precompiled runtimes (such as crt0.o and similar). The ones we have
found are not really attackable, and so we have not focused on them
here, but eventually these runtimes should also be replicated for
retpoline-ed configurations for completeness.
For kernels or other freestanding or fully static executables, the
compiler switch `-mretpoline` is sufficient to fully mitigate this
particular attack. For dynamic executables, you must compile *all*
libraries with `-mretpoline` and additionally link the dynamic
executable and all shared libraries with LLD and pass `-z retpolineplt`
(or use similar functionality from some other linker). We strongly
recommend also using `-z now` as non-lazy binding allows the
retpoline-mitigated PLT to be substantially smaller.
When manually apply similar transformations to `-mretpoline` to the
Linux kernel we observed very small performance hits to applications
running typical workloads, and relatively minor hits (approximately 2%)
even for extremely syscall-heavy applications. This is largely due to
the small number of indirect branches that occur in performance
sensitive paths of the kernel.
When using these patches on statically linked applications, especially
C++ applications, you should expect to see a much more dramatic
performance hit. For microbenchmarks that are switch, indirect-, or
virtual-call heavy we have seen overheads ranging from 10% to 50%.
However, real-world workloads exhibit substantially lower performance
impact. Notably, techniques such as PGO and ThinLTO dramatically reduce
the impact of hot indirect calls (by speculatively promoting them to
direct calls) and allow optimized search trees to be used to lower
switches. If you need to deploy these techniques in C++ applications, we
*strongly* recommend that you ensure all hot call targets are statically
linked (avoiding PLT indirection) and use both PGO and ThinLTO. Well
tuned servers using all of these techniques saw 5% - 10% overhead from
the use of retpoline.
We will add detailed documentation covering these components in
subsequent patches, but wanted to make the core functionality available
as soon as possible. Happy for more code review, but we'd really like to
get these patches landed and backported ASAP for obvious reasons. We're
planning to backport this to both 6.0 and 5.0 release streams and get
a 5.0 release with just this cherry picked ASAP for distros and vendors.
This patch is the work of a number of people over the past month: Eric, Reid,
Rui, and myself. I'm mailing it out as a single commit due to the time
sensitive nature of landing this and the need to backport it. Huge thanks to
everyone who helped out here, and everyone at Intel who helped out in
discussions about how to craft this. Also, credit goes to Paul Turner (at
Google, but not an LLVM contributor) for much of the underlying retpoline
design.
Teach ValueMapper to use ODR uniqued types when available
Summary:
This is exposed during ThinLTO compilation, when we import an alias by
creating a clone of the aliasee. Without this fix the debug type is
unnecessarily cloned and we get a duplicate, undoing the uniquing.
[X86] Teach Intel syntax InstPrinter to print lock prefixes that have been parsed from the asm parser.
The asm parser puts the lock prefix in the MCInst flags so we need to check that in addition to TSFlags. This matches what the ATT printer does.
------------------------------------------------------------------------
When RAFast sees liveins in on a basic block, it uses that information
to initialize the availability of the registers. The called
method uses an instruction as one of its argument and in the liveins
case, RAFast was dereferencing MBB::begin which can be MBB::end for
empty basic block.
Change the API of definePhysReg to use MachineBasicBlock::iterator
instead of MachineInstr so that we don't dereference an
invalid iterator while making the call.
[CVP] Replace incoming values from unreachable blocks with undef.
This is an attempt of fixing PR35807.
Due to the non-standard definition of dominance in LLVM, where uses in
unreachable blocks are dominated by anything, you can have, in an
unreachable block:
%patatino = OP1 %patatino, CONSTANT
When `SimplifyInstruction` receives a PHI where an incoming value is of
the aforementioned form, in some cases, loops indefinitely.
What I propose here instead is keeping track of the incoming values
from unreachable blocks, and replacing them with undef. It fixes this
case, and it seems to be good regardless (even if we can't prove that
the value is constant, as it's coming from an unreachable block, we
can ignore it).
[ValueTracking] add recursion depth param to matchSelectPattern
We're getting bug reports:
https://bugs.llvm.org/show_bug.cgi?id=35807
https://bugs.llvm.org/show_bug.cgi?id=35840
https://bugs.llvm.org/show_bug.cgi?id=36045
...where we blow up the stack in value tracking because other passes are sending
in selects that have an operand that is itself the select.
We don't currently have a reliable way to avoid analyzing dead code that may take
non-standard forms, so bail out when things go too far.
This mimics the recursion depth limitations in other parts of value tracking.
Unfortunately, this pushes the underlying problems for other passes (jump-threading,
simplifycfg, correlated-propagation) into hiding. If someone wants to uncover those
again, the first draft of this patch on Phab would do that (it would assert rather
than bail out).
Apparently checking the pass structure isn't enough to ensure that we don't fall
back to FastISel, as it's set up as part of the SelectionDAGISel.
------------------------------------------------------------------------
[AArch64][GlobalISel] Fall back during AArch64 isel if we have a volatile load.
The tablegen imported patterns for sext(load(a)) don't check for single uses
of the load or delete the original after matching. As a result two loads are
left in the generated code. This particular issue will be fixed by adding
support for a G_SEXTLOAD opcode in future.
There are however other potential issues around this that wouldn't be fixed by
a G_SEXTLOAD, so until we have a proper solution we don't try to handle volatile
loads at all in the AArch64 selector.
Fixes/works around PR36018.
------------------------------------------------------------------------
As detailed in rL317463, PSHUFB (like most variable shuffle instructions) uses Op[0] for the source vector and Op[1] for the shuffle index vector, VPERMV works in reverse which is probably where the confusion comes from.
[PowerPC] Zero-extend the compare operand for ATOMIC_CMP_SWAP
Part of the fix for https://bugs.llvm.org/show_bug.cgi?id=35812.
This patch ensures that the compare operand for the atomic compare and swap
is properly zero-extended to 32 bits if applicable.
A follow-up commit will fix the extension for the SETCC node generated when
expanding an ATOMIC_CMP_SWAP_WITH_SUCCESS. That will complete the bug fix.
[Dominators] Visit affected node candidates found at different root levels
Summary:
This patch attempts to fix the DomTree incremental insertion bug found here [[ https://bugs.llvm.org/show_bug.cgi?id=35969 | PR35969 ]] .
When performing an insertion into a piece of unreachable CFG, we may find the same not at different levels. When this happens, the node can turn out to be affected when we find it starting from a node with a lower level in the tree. The level at which we start visitation affects if we consider a node affected or not.
This patch tracks the lowest level at which each node was visited during insertion and allows it to be visited multiple times, if it can cause it to be considered affected.
[cmake] Include LLVM_LIBXML2_ENABLED in LLVMConfig.cmake, PR36006
Include the LLVM_LIBXML2_ENABLED cache variable in LLVMConfig.cmake
in order to make it available for other LLVM packages to query. This
is necessary to fix stand-alone testing of LLD.
[X86] Don't mutate shuffle arguments after early-out for AVX512
The match* functions have the annoying behavior of modifying its inputs.
Save and restore the inputs, just in case the early out for AVX512 is
hit. This is still not great and its only a matter of time this kind of
bug happens again, but I couldn't come up with a better pattern without
rewriting significant chunks of this code. Fixes PR35977.
------------------------------------------------------------------------
[X86] When legalizing (v64i1 select i8, v64i1, v64i1) make sure not to introduce bitcasts to i64 in 32-bit mode
We legalize selects of masks with scalar conditions using a bitcast to an integer type. But if we are in 32-bit mode we can't convert v64i1 to i64. So instead split the v64i1 to v32i1 and concat it back together. Each half will then be legalized by bitcasting to i32 which is fine.
The test case is a little indirect. If we have the v64i1 select in IR it will get legalized by legalize vector ops which has a run of type legalization after it. That type legalization run is able to fix this i64 bitcast. So in order to avoid that we need a build_vector of a splat which legalize vector ops will ignore. Legalize DAG will then turn that into a select via LowerBUILD_VECTORvXi1. And the select will get legalized. In this case there is no type legalizer run to cleanup the bitcast.
This fixes pr35972.
------------------------------------------------------------------------
[DAG] Teach BaseIndexOffset to correctly handle with indexed operations
BaseIndexOffset address analysis incorrectly ignores offsets folded
into indexed memory operations causing potential errors in alias
analysis of pre-indexed operations.
The work order was changed in r228186 from SCC order
to RPO with an arbitrary sorting function. The sorting
function attempted to move inner loop nodes earlier. This
was was apparently relying on an assumption that every block
in a given loop / the same loop depth would be seen before
visiting another loop. In the broken testcase, a block
outside of the loop was encountered before moving onto
another block in the same loop. The testcase would then
structurize such that one blocks unconditional successor
could never be reached.
Revert to plain RPO for the analysis phase. This fixes
detecting edges as backedges that aren't really.
The processing phase does use another visited set, and
I'm unclear on whether the order there is as important.
An arbitrary order doesn't work, and triggers some infinite
loops. The reversed RPO list seems to work and is closer
to the order that was used before, minus the arbitary
custom sorting.
A few of the changed tests now produce smaller code,
and a few are slightly worse looking.
------------------------------------------------------------------------
StructurizeCFG: xfail one of the testcases from r321751
It fails with -verify-region-info. This seems to be a issue
with RegionInfo itself which existed before.
------------------------------------------------------------------------
RegionInfo: Use report_fatal_error instead of llvm_unreachable
Otherwise when using -verify-region-info in a release build the
error won't be emitted.
------------------------------------------------------------------------
PeepholeOptimizer: Do not form PHI with subreg arguments
When replacing a PHI the PeepholeOptimizer currently takes the register
class of the register at the first operand. This however is not correct
if this argument has a subregister index.
As there is currently no API to query the register class resulting from
applying a subregister index to all registers in a class, we can only
abort in these cases and not perform the transformation.
This changes findNextSource() to require the end of all copy chains to
not use a subregister if there is any PHI in the chain. I had to rewrite
the overly complicated inner loop there to have a good place to insert
the new check.
This fixes https://llvm.org/PR33071 (aka rdar://32262041)
TargetLoweringBase: The ios simulator has no bzero function.
Make sure I really get back to the beahvior before my rewrite in r321035
which turned out not to be completely NFC as I changed the behavior for
the ios simulator environment.
------------------------------------------------------------------------
[SLP] Fix PR35777: Incorrect handling of aggregate values.
Summary:
Fixes the bug with incorrect handling of InsertValue|InsertElement
instrucions in SLP vectorizer. Currently, we may use incorrect
ExtractElement instructions as the operands of the original
InsertValue|InsertElement instructions.
[LV] Don't call recordVectorLoopValueForInductionCast for newly-created IV from a trunc.
Summary:
This method is supposed to be called for IVs that have casts in their use-def
chains that are completely ignored after vectorization under PSE. However, for
truncates of such IVs the same InductionDescriptor is used during
creation/widening of both original IV based on PHINode and new IV based on
TruncInst.
This leads to unintended second call to recordVectorLoopValueForInductionCast
with a VectorLoopVal set to the newly created IV for a trunc and causes an
assert due to attempt to store new information for already existing entry in the
map. This is wrong and should not be done.
Fix crash when linking metadata with ODR type uniquing
Summary:
With DebugTypeODRUniquing enabled, during IR linking debug metadata
in the destination module may be reached from the source module.
This means that ConstantAsMetadata nodes (e.g. on DITemplateValueParameter)
may contain a value the destination module. When trying to map such
metadata nodes, we will attempt to map a GV already in the dest module.
linkGlobalValueProto will end up with a source GV that is the same as
the dest GV as well as the new GV. Trying to access the TypeMap for the
source GV type, which is actually a dest GV type, hits an assertion
since it appears that we have mapped into the source module (because the
type is the value not a key into the map).
Detect that we don't need to access the TypeMap in this case, since
there is no need to create a bitcast from the new GV to the source GV
type as they GV are the same.
Alex Bradbury [Wed, 3 Jan 2018 13:46:21 +0000 (13:46 +0000)]
[ARM][NFC] Avoid recreating MCSubtargetInfo in ARMAsmBackend
After D41349, we can now directly access MCSubtargetInfo from
createARM*AsmBackend. This patch makes use of this, avoiding the need to
create a fresh MCSubtargetInfo (which was previously always done with a blank
CPU and feature string). Given the total size of the change remains pretty
tiny and we're removing the old explicit destructor, I changed the STI field
to a reference rather than a pointer.
Hal Finkel [Wed, 3 Jan 2018 11:35:09 +0000 (11:35 +0000)]
[TableGen] Add support of Intrinsics with multiple returns
This change deals with intrinsics with multiple outputs, for example load
instrinsic with address updated.
DAG selection for Instrinsics could be done either through source code or
tablegen. Handling all intrinsics in source code would introduce a huge chunk
of repetitive code if we have a large number of intrinsic that return multiple
values (see NVPTX as an example). While intrinsic class in tablegen supports
multiple outputs, tablegen only supports Intrinsics with zero or one output on
TreePattern. This appears to be a simple bug in tablegen that is fixed by this
change.
projects / llvm / log