This adds assembly and bitcode support for `MDLocation`. The assembly
side is rather big, since this is the first `MDNode` subclass (that
isn't `MDTuple`). Part of PR21433.
(If you're wondering where the mountains of testcase updates are, we
don't need them until I update `DILocation` and `DebugLoc` to actually
use this class.)
Matt Arsenault [Tue, 13 Jan 2015 20:53:23 +0000 (20:53 +0000)]
R600: Implement getRecipEstimate
This requires a new hook to prevent expanding sqrt in terms
of rsqrt and reciprocal. v_rcp_f32, v_rsq_f32, and v_sqrt_f32 are
all the same rate, so this expansion would just double the number
of instructions and cycles.
Matt Arsenault [Tue, 13 Jan 2015 20:53:18 +0000 (20:53 +0000)]
R600: Implement getRsqrtEstimate
Only do for f32 since I'm unclear on both what this is expecting
for the refinement steps in terms of accuracy, and what
f64 instruction actually provides.
Add a new subclass of `UniquableMDNode`, `MDLocation`. This will be the
IR version of `DebugLoc` and `DILocation`. The goal is to rename this
to `DILocation` once the IR classes supersede the `DI`-prefixed
wrappers.
Matt Arsenault [Tue, 13 Jan 2015 19:46:48 +0000 (19:46 +0000)]
R600: Make cttz / ctlz cheap to speculate
Speculating things is generally good. SI+ has instructions for these
for 32-bit values. This is still probably better even with the expansion
for 64-bit values, although it is odd that this callback doesn't have
the size as a parameter.
Ulrich Weigand [Tue, 13 Jan 2015 19:43:45 +0000 (19:43 +0000)]
Use the integrated assembler as default on PowerPC
This was already done in clang, this commit now uses the integrated
assembler as default when using LLVM tools directly.
A number of test cases using inline asm had to be adapted, either by
updating the expected output, or by using -no-integrated-as (for such
tests that deliberately use an invalid instruction in inline asm).
Will Schmidt [Tue, 13 Jan 2015 18:17:08 +0000 (18:17 +0000)]
Update multiline.ll testcase to handle (ppc64le) .localentry directive
The ppc64le platform will emit a .localentry directive. This is triggering
a false-positive against a CHECK-NOT: .loc in multiline.ll.
Add a space "{{ }}" to the check-not line to allow for arguments, and
prevent .localentry from matching.
Hal Finkel [Tue, 13 Jan 2015 17:48:12 +0000 (17:48 +0000)]
[PowerPC] Add StackMap/PatchPoint support
This commit does two things:
1. Refactors PPCFastISel to use more of the common infrastructure for call
lowering (this lets us take advantage of this common code for lowering some
common intrinsics, stackmap/patchpoint among them).
2. Adds support for stackmap/patchpoint lowering. For the most part, this is
very similar to the support in the AArch64 target, with the obvious differences
(different registers, NOP instructions, etc.). The test cases are adapted
from the AArch64 test cases.
One difference of note is that the patchpoint call sequence takes 24 bytes, so
you can't use less than that (on AArch64 you can go down to 16). Also, as noted
in the docs, we take the patchpoint address to be the actual code address
(assuming the call is local in the TOC-sharing sense), which should yield
higher performance than generating the full cross-DSO indirect-call sequence
and is likely just as useful for JITed code (if not, we'll change it).
StackMaps and Patchpoints are still marked as experimental, and so this support
is doubly experimental. So go ahead and experiment!
Hal Finkel [Tue, 13 Jan 2015 17:48:07 +0000 (17:48 +0000)]
[StackMaps] Use CurrentFnSymForSize
When computing the call-site offset, use AP.CurrentFnSymForSize instead of
AP.CurrentFnSym. There should be no change for other targets, but this is
necessary for generating valid expressions for PPC64/ELF.
Hal Finkel [Tue, 13 Jan 2015 17:48:04 +0000 (17:48 +0000)]
[StackMaps] Mark in CallLoweringInfo when lowering a patchpoint
While, generally speaking, the process of lowering arguments for a patchpoint
is the same as lowering a regular indirect call, on some targets it may not be
exactly the same. Targets may not, for example, want to add additional register
dependencies that apply only to making cross-DSO calls through linker stubs,
may not want to load additional registers out of function descriptors, and may
not want to add additional side-effect-causing instructions that cannot be
removed later with the call itself being generated.
The PowerPC target will use this in a future commit (for all of the reasons
stated above).
Hal Finkel [Tue, 13 Jan 2015 17:47:59 +0000 (17:47 +0000)]
[StackMaps] Allow the target to pre-process the live-out mask
Some targets, PowerPC for example, have pseudo-registers (such as that used to
represent the rounding mode), that don't have DWARF register numbers or a
register class. These are used only for internal dependency tracking, and
should not appear in the recorded live-outs. This adds a callback allowing the
target to pre-process the live-out mask in order to remove these kinds of
registers so that the StackMaps code does not complain about them and/or
attempt to include them in the output.
This will be used by the PowerPC target in a future commit.
Hal Finkel [Tue, 13 Jan 2015 17:47:54 +0000 (17:47 +0000)]
[PowerPC] Split the blr definition into BLR and BLR8
We really need a separate 64-bit version of this instruction so that it can be
marked as clobbering LR8 (instead of just LR). No change in functionality
(although the verifier might be slightly happier), however, it is required for
stackmap/patchpoint support. Thus, this will be covered by stackmap test cases
once those are added.
Hal Finkel [Tue, 13 Jan 2015 17:45:11 +0000 (17:45 +0000)]
[PowerPC] Add DWARF numbers for CA (XER), etc.
For registers that have DWARF numbers (like CA, which is really part of XER),
add them. Also, RM is not an SPR, and the declaration hack (where it is
declared as an SPR with an arbitrary number) is not needed, so just declare it
as a register.
NFC; although CA's register number will be needed when stackmap/patchpoint
support is added.
Erik Eckstein [Tue, 13 Jan 2015 11:37:51 +0000 (11:37 +0000)]
SLPVectorizer: Cache results from memory alias checking.
This speeds up the dependency calculations for blocks with many load/store/call instructions.
Beside the improved runtime, there is no functional change.
Chandler Carruth [Tue, 13 Jan 2015 11:36:43 +0000 (11:36 +0000)]
[PM] In the PassManager template, remove a pointless indirection through
a nested class template for the PassModel, and use the T-suffix for the
two typedefs to match the code in the AnalysisManager.
This is the last of the fairly fundamental code cleanups here. Will be
focusing on the printing of analyses next to finish that aspect off.
Chandler Carruth [Tue, 13 Jan 2015 11:31:43 +0000 (11:31 +0000)]
[PM] Remove the 'AnalysisManagerT' type parameter from numerous layers
of templates in the new pass manager.
The analysis manager is now itself just a template predicated on the IR
unit. This makes lots of the templates really trivial and more clear:
they are all parameterized on a single type, the IR unit's type.
Everything else is a function of that. To me, this is a really nice
cleanup of the APIs and removes a layer of 'magic' and 'indirection'
that really wasn't there and just got in the way of understanding what
is going on here.
Chandler Carruth [Tue, 13 Jan 2015 11:13:56 +0000 (11:13 +0000)]
[PM] Refactor the new pass manager to use a single template to implement
the generic functionality of the pass managers themselves.
In the new infrastructure, the pass "manager" isn't actually interesting
at all. It just pipelines a single chunk of IR through N passes. We
don't need to know anything about the IR or the passes to do this really
and we can replace the 3 implementations of the exact same functionality
with a single generic PassManager template, complementing the single
generic AnalysisManager template.
I've left typedefs in place to give convenient names to the various
obvious instantiations of the template.
With this, I think I've nuked almost all of the redundant logic in the
managers, and I think the overall design is actually simpler for having
single templates that clearly indicate there is no special logic here.
The logging is made somewhat more annoying by this change, but I don't
think the difference is worth having heavy-weight traits to help log
things.
Peter Zotov [Tue, 13 Jan 2015 09:47:59 +0000 (09:47 +0000)]
[OCaml] Use $CAMLORIGIN, an rpath-$ORIGIN-like mechanism in OCaml.
As a result, installations of LLVM in non-standard locations
will not require passing custom -ccopt -L flags when building
the binary, nor absolute paths would be embedded in the cma/cmxa
files. Additionally, the executables will not require changes
to LD_LIBRARY_PATH, although CAML_LD_LIBRARY_PATH still
has to be set for ocamlc without -custom.
See http://caml.inria.fr/mantis/view.php?id=6642.
Note that the patch is approved, but not merged yet.
It will be released in 4.03 and likely 4.02.
Mehdi Amini [Tue, 13 Jan 2015 07:07:13 +0000 (07:07 +0000)]
Peephole opt needs optimizeSelect() to keep track of newly created MIs
Peephole optimizer is scanning a basic block forward. At some point it
needs to answer the question "given a pointer to an MI in the current
BB, is it located before or after the current instruction".
To perform this, it keeps a set of the MIs already seen during the scan,
if a MI is not in the set, it is assumed to be after.
It means that newly created MIs have to be inserted in the set as well.
This commit passes the set as an argument to the target-dependent
optimizeSelect() so that it can properly update the set with the
(potentially) newly created MIs.
The functions {pred,succ,use,user}_{begin,end} exist, but many users
have to check *_begin() with *_end() by hand to determine if the
BasicBlock or User is empty. Fix this with a standard *_empty(),
demonstrating a few usecases.
ARM: prepare prefix parsing for improved AAELF support
AAELF specifies a number of ELF specific relocation types which have custom
prefixes for the symbol reference. Switch the parser to be more table driven
with an idea of file formats for which they apply. NFC.
Chandler Carruth [Tue, 13 Jan 2015 02:51:47 +0000 (02:51 +0000)]
[PM] Fold all three analysis managers into a single AnalysisManager
template.
This consolidates three copies of nearly the same core logic. It adds
"complexity" to the ModuleAnalysisManager in that it makes it possible
to share a ModuleAnalysisManager across multiple modules... But it does
so by deleting *all of the code*, so I'm OK with that. This will
naturally make fixing bugs in this code much simpler, etc.
The only down side here is that we have to use 'typename' and 'this->'
in various places, and the implementation is lifted into the header.
I'll take that for the code size reduction.
The convenient names are still typedef-ed and used throughout so that
users can largely ignore this aspect of the implementation.
The follow-up change to this will do the exact same refactoring for the
PassManagers. =D
It turns out that the interesting different code is almost entirely in
the adaptors. At the end, that should be essentially all that is left.
Reid Kleckner [Tue, 13 Jan 2015 01:51:34 +0000 (01:51 +0000)]
Rename llvm.recoverframeallocation to llvm.framerecover
This name is less descriptive, but it sort of puts things in the
'llvm.frame...' namespace, relating it to frameallocate and
frameaddress. It also avoids using "allocate" and "allocation" together.
Reid Kleckner [Tue, 13 Jan 2015 00:48:10 +0000 (00:48 +0000)]
Add the llvm.frameallocate and llvm.recoverframeallocation intrinsics
These intrinsics allow multiple functions to share a single stack
allocation from one function's call frame. The function with the
allocation may only perform one allocation, and it must be in the entry
block.
Functions accessing the allocation call llvm.recoverframeallocation with
the function whose frame they are accessing and a frame pointer from an
active call frame of that function.
These intrinsics are very difficult to inline correctly, so the
intention is that they be introduced rarely, or at least very late
during EH preparation.
IR: Remove an invalid assertion when replacing resolved operands
This adds back the testcase from r225738, and adds to it. Looks like we
need both sides for now (the assertion was incorrect both ways, and
although it seemed reasonable (when written correctly) it wasn't
particularly important).
Chandler Carruth [Tue, 13 Jan 2015 00:36:47 +0000 (00:36 +0000)]
[PM] Re-clang-format much of this code as the code has changed some and
so has clang-format. Notably, this fixes a bunch of formatting in the
CGSCC pass manager side of things that has been improved in clang-format
recently.
Revert "IR: Fix an inverted assertion when replacing resolved operands"
This reverts commit r225738. Maybe the assertion is just plain wrong,
but this version fails on WAY more bots. I'll make sure both ways work
in a follow-up but I want to get bots green in the meantime.
Hans Wennborg [Tue, 13 Jan 2015 00:07:31 +0000 (00:07 +0000)]
Release merge script: don't actually commit the merge
Instead, just present the command for committing it. This way,
the user can test the merge locally, resolve conflicts, etc.
before committing, which seems much safer to me.
Adrian Prantl [Tue, 13 Jan 2015 00:04:06 +0000 (00:04 +0000)]
Debug Info: Move support for constants into DwarfExpression.
Move the declaration of DebugLocDwarfExpression into DwarfExpression.h
because it needs to be accessed from AsmPrinterDwarf.cpp and DwarfDebug.cpp
Chandler Carruth [Mon, 12 Jan 2015 22:53:31 +0000 (22:53 +0000)]
[PM] Sink the reference vs. value decision for IR units out of the
templated interface.
So far, every single IR unit I can come up with has address-identity.
That is, when two units of IR are both active in LLVM, their addresses
will be distinct of the IR is distinct. This is clearly true for
Modules, Functions, BasicBlocks, and Instructions. It turns out that the
only practical way to make the CGSCC stuff work the way we want is to
make it true for SCCs as well. I expect this pattern to continue.
When first designing the pass manager code, I kept this dimension of
freedom in the type parameters, essentially allowing for a wrapper-type
whose address did not form identity. But that really no longer makes
sense and is making the code more complex or subtle for no gain. If we
ever have an actual use case for this, we can figure out what makes
sense then and there. It will be better because then we will have the
actual example in hand.
While the simplifications afforded in this patch are fairly small
(mostly sinking the '&' out of many type parameters onto a few
interfaces), it would have become much more pronounced with subsequent
changes. I have a sequence of changes that will completely remove the
code duplication that currently exists between all of the pass managers
and analysis managers. =] Should make things much cleaner and avoid bug
fixing N times for the N pass managers.
Simon Pilgrim [Mon, 12 Jan 2015 22:38:08 +0000 (22:38 +0000)]
[X86][SSE] Minor regression fix for r225551
r225551 vector byte shuffle optimization caused an assertion as fully zeroable vectors can be produced under certain circumstances. This fix drops the assert and returns a zero vector where the assert would have failed.
AsmParser: Use subclass API instead of MDNode wrappers, NFC
Use subclass API instead of the wrappers in `MDNode` in the assembly
parser. This will make the code easier to follow once we have multiple
subclasses.
Adrian Prantl [Mon, 12 Jan 2015 22:19:22 +0000 (22:19 +0000)]
Debug info: Factor out the creation of DWARF expressions from AsmPrinter
into a new class DwarfExpression that can be shared between AsmPrinter
and DwarfUnit.
This is the first step towards unifying the two entirely redundant
implementations of dwarf expression emission in DwarfUnit and AsmPrinter.
Almost no functional change — Testcases were updated because asm comments
that used to be on two lines now appear on the same line, which is
actually preferable.
AsmParser: Use normal references for metadata attachments, NFC
Remove special parsing logic for metadata attachments. Now that
`DebugLoc` is stored normally (since the metadata/value split), we don't
need this special forward referencing logic.
IR: Prepare for a new UniquableMDNode subclass, NFC
Add generic dispatch for the parts of `UniquableMDNode` that cast to
`MDTuple`. This makes adding other subclasses (like PR21433's
`MDLocation`) easier.
IR: Stop erasing MDNodes from uniquing sets during teardown
Stop erasing `MDNode`s from the uniquing sets in `LLVMContextImpl`
during teardown (in particular, during
`UniquableMDNode::~UniquableMDNode()`). Although it's currently
feasible, there isn't any clear benefit and it may not be feasible for
other subclasses (which don't explicitly store the lookup hash).
Ahmed Bougacha [Mon, 12 Jan 2015 20:31:30 +0000 (20:31 +0000)]
[X86] Also create+widen FMIN/FMAX nodes for v2f32.
This happens in the HINT benchmark, where the SLP-vectorizer created
v2f32 fcmp/select code. The "correct" solution would have been to
teach the vectorizer cost model that v2f32 isn't legal (because really,
it isn't), but if we can vectorize we might as well do so.
We legalize these v2f32 FMIN/FMAX nodes by widening to v4f32 later on.
v3f32 were already widened to v4f32 by the generic unroll-and-build-vector
legalization.
Move creation logic for `MDTuple`s down where it belongs. Once there
are a few more subclasses, these functions really won't make much sense
here (the `friend` relationship was already awkward). For now, leave
the `MDNode` versions around, but have it forward down.
IR: Split GenericMDNode into MDTuple and UniquableMDNode
Split `GenericMDNode` into two classes (with more descriptive names).
- `UniquableMDNode` will be a common subclass for `MDNode`s that are
sometimes uniqued like constants, and sometimes 'distinct'.
This class gets the (short-lived) RAUW support and related API.
- `MDTuple` is the basic tuple that has always been returned by
`MDNode::get()`. This is as opposed to more specific nodes to be
added soon, which have additional fields, custom assembly syntax,
and extra semantics.
This class gets the hash-related logic, since other sublcasses of
`UniquableMDNode` may need to hash based on other fields.
To keep this diff from getting too big, I've added casts to `MDTuple`
that won't really scale as new subclasses of `UniquableMDNode` are
added, but I'll clean those up incrementally.
projects / llvm / log