move IR-level optimization flags into their own struct
This is a preliminary step to using the IR-level floating-point fast-math-flags in the SDAG (D8900).
In this patch, we introduce the optimization flags as their own struct. As noted in the TODO comment,
we should eventually share this data between the IR passes and the backend.
We also switch the existing nsw / nuw / exact bit functionality of the BinaryWithFlagsSDNode class to
use the new struct.
The tradeoff is that instead of using the free but limited space of SDNode's SubclassData, we add a
data member to the subclass. This means we don't have to repeat all of the get/set methods per flag,
but we're potentially adding size to all nodes of this subclassi type.
In practice on 64-bit systems (measured on Linux and MacOS X), there is no size difference between an
SDNode and BinaryWithFlagsSDNode after this change: they're both 80 bytes. This means that we had at
least one free byte to play with due to struct alignment.
Reapply r235977 "[DebugInfo] Add debug locations to constant SD nodes"
[DebugInfo] Add debug locations to constant SD nodes
This adds debug location to constant nodes of Selection DAG and updates
all places that create constants to pass debug locations
(see PR13269).
Can't guarantee that all locations are correct, but in a lot of cases choice
is obvious, so most of them should be. At least all tests pass.
Tests for these changes do not cover everything, instead just check it for
SDNodes, ARM and AArch64 where it's easy to get incorrect locations on
constants.
This is not complete fix as FastISel contains workaround for wrong debug
locations, which drops locations from instructions on processing constants,
but there isn't currently a way to use debug locations from constants there
as llvm::Constant doesn't cache it (yet). Although this is a bit different
issue, not directly related to these changes.
[DebugInfo] Add debug locations to constant SD nodes
This adds debug location to constant nodes of Selection DAG and updates
all places that create constants to pass debug locations
(see PR13269).
Can't guarantee that all locations are correct, but in a lot of cases choice
is obvious, so most of them should be. At least all tests pass.
Tests for these changes do not cover everything, instead just check it for
SDNodes, ARM and AArch64 where it's easy to get incorrect locations on
constants.
This is not complete fix as FastISel contains workaround for wrong debug
locations, which drops locations from instructions on processing constants,
but there isn't currently a way to use debug locations from constants there
as llvm::Constant doesn't cache it (yet). Although this is a bit different
issue, not directly related to these changes.
David Blaikie [Tue, 28 Apr 2015 04:30:29 +0000 (04:30 +0000)]
[opaque pointer type] Encode the pointee type in the bitcode for 'cmpxchg'
As a space optimization, this instruction would just encode the pointer
type of the first operand and use the knowledge that the second and
third operands would be of the pointee type of the first. When typed
pointers go away, this assumption will no longer be available - so
encode the type of the second operand explicitly and rely on that for
the third.
Test case added to demonstrate the backwards compatibility concern,
which only comes up when the definition of the second operand comes
after the use (hence the weird basic block sequence) - at which point
the type needs to be explicitly encoded in the bitcode and the record
length changes to accommodate this.
Ahmed Bougacha [Tue, 28 Apr 2015 01:37:11 +0000 (01:37 +0000)]
[MC] Use LShr for constant evaluation of ">>" on ELF/arm64--darwin.
This matches other assemblers and is less unexpected (e.g. PR23227).
On ELF, I tried binutils gas v2.24 and nasm 2.10.09, and they both
agree on LShr. On COFF, I couldn't get my hands on an assembler yet,
so don't change the behavior. For now, don't change it on non-AArch64
Darwin either, as the other assembler is gas v1.38, which does an AShr.
Ahmed Bougacha [Tue, 28 Apr 2015 00:21:32 +0000 (00:21 +0000)]
[MC] Split MCBinaryExpr::Shr into LShr and AShr.
Defaulting to AShr without consulting the target MCAsmInfo isn't OK.
Add a flag to fix that. Keep it off for now: target migrations will
follow in separate commits.
Hans Wennborg [Mon, 27 Apr 2015 23:52:19 +0000 (23:52 +0000)]
Switch lowering: use uint32_t for weights everywhere
I previously thought switch clusters would need to use uint64_t in case
the weights of multiple cases overflowed a 32-bit int. It turns
out that the weights on a terminator instruction are capped to allow for
being added together, so using a uint32_t should be safe.
Reverse libLTO's default behaviour for preserving use-list order in
bitcode, and add API for controlling it. The default setting is now
`false` (don't preserve them), which is consistent with `clang`'s
default behaviour.
Users of libLTO should call `lto_codegen_should_embed_uselists(CG,true)`
prior to calling `lto_codegen_write_merged_modules()` whenever the
output file isn't part of the production workflow in order to reproduce
results with subsequent calls to `llc`.
(I haven't added tests since `llvm-lto` (the test tool for LTO) doesn't
support bitcode output, and even if it did: there isn't actually a good
way to test whether a tool has passed the flag. If the order is already
"natural" (if the order will already round-trip) then no use-list
directives are emitted at all. At some point I'll circle back to add
tests to `llvm-as` (etc.) that they actually respect the flag, at which
point I can somehow add a test here as well.)
Hans Wennborg [Mon, 27 Apr 2015 23:35:22 +0000 (23:35 +0000)]
Switch lowering: Take branch weight into account when ordering for fall-through
Previously, the code would try to put a fall-through case last,
even if that meant moving a case with much higher branch weight
further down the chain.
Ordering by branch weight is most important, putting a fall-through
block last is secondary.
LTO: Correct some doxygen comments about API availability
These look like copy/paste errors, and shouldn't have the "prior to"
qualifier. Each API was introduced at the given values of
`LTO_API_VERSION`. The "prior to" in other doxygen comments is because
I couldn't easily differentiate between versions 1 and 2 when I added
these comments.
Ahmed Bougacha [Mon, 27 Apr 2015 21:43:12 +0000 (21:43 +0000)]
[AArch64] Also combine vector selects fed by non-i1 SETCCs.
After legalization, scalar SETCC has an i32 result type on AArch64.
The i1 requirement seems too conservative, replace it with an assert.
This also means that we now can run after legalization. That should also
be fine, since the ops legalizer runs again after each combine, and
all types created all have the same sizes as the (legal) inputs.
Exposed by r235917; while there, robustize its tests (bsl also uses the
register it defines).
Ahmed Bougacha [Mon, 27 Apr 2015 21:01:20 +0000 (21:01 +0000)]
[AArch64] Don't assert when combining (v3f32 select (setcc f64)).
When the setcc has f64 operands, we can't build a vector setcc mask
to feed a vselect, because f64 doesn't divide v3f32 evenly.
Just bail out when that happens.
Chris Bieneman [Mon, 27 Apr 2015 20:45:35 +0000 (20:45 +0000)]
Fixes a hang that can occur if a signal comes in during malloc calls.
We need to dereference the signals mutex during handler registration so that we force its construction. This is to prevent the first use being during handling an actual signal because you can't safely allocate memory in a signal handler.
David Blaikie [Mon, 27 Apr 2015 19:58:56 +0000 (19:58 +0000)]
[opaque pointer type] encode the pointee type of global variables
Use a few extra bits in the const field (after widening it from a fixed
single bit) to stash the address space which is no longer provided by
the type (and an extra bit in there to specify that we're using that new
encoding).
Bill Schmidt [Mon, 27 Apr 2015 19:57:34 +0000 (19:57 +0000)]
[PPC64LE] Remove unnecessary swaps from lane-insensitive vector computations
This patch adds a new SSA MI pass that runs on little-endian PPC64
code with VSX enabled. Loads and stores of 4x32 and 2x64 vectors
without alignment constraints are accomplished for little-endian using
lxvd2x/xxswapd and xxswapd/stxvd2x. The existence of the additional
xxswapd instructions hurts performance in comparison with big-endian
code, but they are necessary in the general case to support correct
semantics.
However, the general case does not apply to most vector code. Many
vector instructions are lane-insensitive; they do not "care" which
lanes the parallel computations are performed within, provided that
the resulting data is stored into the correct locations. Thus this
pass looks for computations that perform only lane-insensitive
operations, and remove the unnecessary swaps from loads and stores in
such computations.
Future improvements will allow computations using certain
lane-sensitive operations to also be optimized in this manner, by
modifying the lane-sensitive operations to account for the permuted
order of the lanes. However, this patch only adds the infrastructure
to permit this; no lane-sensitive operations are optimized at this
time.
This code is heavily exercised by the various vectorizing applications
in the projects/test-suite tree. For the time being, I have only added
one simple test case to demonstrate what the pass is doing. Although
it is quite simple, it provides coverage for much of the code,
including the special case handling of copies and subreg-to-reg
operations feeding the swaps. I plan to add additional tests in the
future as I fill in more of the "special handling" code.
Two existing tests were affected, because they expected the swaps to
be present, but they are now removed.
[Hexagon] Use constant extenders to fix up hardware loops
Use a loop instruction with a constant extender for a hardware
loop instruction that is too far away from the start of the loop.
This is cheaper than changing the SA register value.
[mips] [IAS] Improve warning for using AT with .set noat.
Summary:
Changed the warning message to show the current value of $at, similar to what clang does for typedef's, and renamed warnIfAssemblerTemporary to a more descriptive name.
I also changed the type of variables which store registers from int to unsigned, updated the relevant test and tried to make the related comments clearer.
Reapply "[mips][FastISel] Implement shift ops for Mips fast-isel.""
This reapplies r235194, which was reverted in r235495 because it was causing a
failure in our out-of-tree buildbots for MIPS. With the sign-extension patch
in r235718, this patch doesn't cause any problem any more.
[MC] [IAS] Add support for the \@ .macro pseudo-variable.
Summary:
When used, it is substituted with the number of .macro instantiations we've done up to that point in time.
So if this is the 1st time we've instantiated a .macro (any .macro, regardless of name), \@ will instantiate to 0, if it's the 2nd .macro instantiation, it will instantiate to 1 etc.
It can only be used inside a .macro definition, an .irp definition or an .irpc definition (those last 2 uses are undocumented).
Constfold insertelement to undef when index is out-of-bounds
Summary:
This patch adds constant folding of insertelement instruction to undef value when index operand is constant and is not less than vector size or is undef.
InstCombine does not support this case, but I'm happy to add it there also if this change is accepted.
Test Plan: Unittests and regression tests for ConstProp pass.
Philip Reames [Sun, 26 Apr 2015 22:15:18 +0000 (22:15 +0000)]
Add two new items to PerformanceTips
1) Turns out we're not great at recognizing redundant checks when one is a != and the other is an ==. This is a bug, but it's one that matters to frontend authors.
2) Frontends shouldn't use intrinsics unless strictly neccessary. This has been pretty widely proven by this point and is good to document.
Philip Reames [Sun, 26 Apr 2015 22:00:34 +0000 (22:00 +0000)]
Make the message associated with a fatal error slightly more helpful
Looking into 23095, my best guess is that the CodeGen library itself isn't getting linked and initialized properly. To make this slightly more obvious to consumers of LLVM, emit a different error message if we can tell that the registry is empty vs you've simply happened to name a collector which hasn't been registered.
Philip Reames [Sun, 26 Apr 2015 19:48:03 +0000 (19:48 +0000)]
[RewriteStatepointsForGC] Exclude constant values from being considered live at a safepoint
There can be various constant pointers in the IR which do not get relocated at a safepoint. One example is the address of a global variable. Another example is a pointer created via inttoptr. Note that the optimizer itself likes to create such inttoptrs when locally propagating constants through dynamically dead code.
To deal with this, we need to exclude uses of constants from contributing to the liveness of a safepoint which might reach that use. At some later date, it might be worth exploring what could be done to support the relocation of various special types of "constants", but that's future work.
Philip Reames [Sun, 26 Apr 2015 19:41:23 +0000 (19:41 +0000)]
Don't Place Entry Safepoints Before the llvm.frameescape() Intrinsic
llvm.frameescape() intrinsic is not a real call. The intrinsic can only exist in the entry block. Inserting a gc.statepoint() before llvm.frameescape() may split the entry block, and push the intrinsic out of the entry block.
[x86] instcombine more cases of insertps into a shufflevector
This is a follow-on to D8833 (insertps optimization when the zero mask is not used).
In this patch, we check for the case where the zmask is used, but both input vectors
to the insertps intrinsic are the same operand or the zmask overrides the destination
lane. This lets us replace the 2nd shuffle input operand with the zero vector.
IR: Add assembly/bitcode support for function metadata attachments
Add serialization support for function metadata attachments (added in
r235783). The syntax is:
define @foo() !attach !0 {
Metadata attachments are only allowed on functions with bodies. Since
they come before the `{`, they're not really part of the body; since
they require a body, they're not really part of the header. In
`LLParser` I gave them a separate function called from `ParseDefine()`,
`ParseOptionalFunctionMetadata()`.
In bitcode, I'm using the same `METADATA_ATTACHMENT` record used by
instructions. Instruction metadata attachments are included in a
special "attachment" block at the end of a `Function`. The attachment
records are laid out like this:
InstID (KindID MetadataID)+
Note that these records always have an odd number of fields. The new
code takes advantage of this to recognize function attachments (which
don't need an instruction ID):
(KindID MetadataID)+
This means we can use the same attachment block already used for
instructions.
Verifier: Function metadata attachments require a body
Add a verifier check that only functions with bodies have metadata
attachments. This should help catch bugs in frontends and
transformation passes. Part of PR23340.
Add IR support for `Metadata` attachments. Assembly and bitcode support
will follow shortly, but for now we just have unit tests. This is part
of PR23340.
[DAGCombiner] Fix the type used in canFoldInAddressingMode to account for the
right scaling.
In the function canFoldInAddressingMode, VT is computed as the type of the
destination/source of a LOAD/STORE operations, instead of the memory type of the
operation.
On targets with a scaling factor on the offset of the LOAD/STORE operations, the
function may return false for actually valid cases. This may then prevent the
selection of profitable pre or post indexed load/store operations, and instead
select pre or post indexed load/store for unprofitable cases.
Patch by Francois de Ferriere <francois.de-ferriere@st.com>!
AsmWriter: Parameterize the syntactic separator for attachments
Parameterize the separator for attachments, since `Function` metadata
attachments (PR23340) aren't going to use a `,` (comma). No real
functionality change.
AsmWriter: Only collect attachment names once per module
Collect metadata names once per `AssemblyWriter` instead of every time
we need to print some attachments. Just a drive-by; this caught my eye
while I was refactoring the code in r235772.
Hans Wennborg [Fri, 24 Apr 2015 20:57:56 +0000 (20:57 +0000)]
SimplifyCFG: Correctly handle switch lookup tables which fully cover the input type and use bit tests to check for holes
When using bit tests for hole checks, we call AddPredecessorToBlock to give the
phi node a value from the bit test block. This would break if we've
previously called removePredecessor on the default destination because the
switch is fully covered.
IR: Use a bitmask to access GlobalObject subclass data
Make room for more than just `Function::isMaterializable()` in the
`GlobalObject` subclass data bitfield. Since we're treating it like a
bitfield, change `Function::Function()` to zero-out the whole thing.
IR: Extract set logic from Instruction attachments, NFC
Extract the set logic for metadata attachments from `Instruction` so it
can be reused for `Function` (PR23340).
This data structure makes a `SmallVector<>` look (a little) like a map,
just doing the bare minimum to support the `Instruction` (and soon,
`Function`) metadata API.
[SEH] Implement GetExceptionCode in __except blocks
This introduces an intrinsic called llvm.eh.exceptioncode. It is lowered
by copying the EAX value live into whatever basic block it is called
from. Obviously, this only works if you insert it late during codegen,
because otherwise mid-level passes might reschedule it.
IR: Use remove_if for Instruction::dropUnknownMetadata()
Technically the operations are different -- the old logic moved items
from the back into the opened-up slots, instead of the usual
`remove_if()` logic of a slow and a fast iterator -- but unless a
profile tells us otherwise I prefer the simpler logic here. Regardless,
there shouldn't be an observable function change.
David Blaikie [Fri, 24 Apr 2015 19:32:54 +0000 (19:32 +0000)]
[opaque pointer type] Add textual IR support for explicit type parameter to the invoke instruction
Same as r235145 for the call instruction - the justification, tradeoffs,
etc are all the same. The conversion script worked the same without any
false negatives (after replacing 'call' with 'invoke').
Chris Bieneman [Fri, 24 Apr 2015 17:09:20 +0000 (17:09 +0000)]
[CMake] Fix for PR 23328: LLVM_OPTIMIZED_TABLEGEN broken
In CMake dependencies can be filenames or targets, and targets can't be filenames. The Ninja generator handles filename dependencies because it generates targets for every output file from a command. For example:
With the Ninja generator this generates a target foo.txt, but with the Makefile generator it doesn't. This is probably because Ninja explicitly requires these hard dependency ties, and Make just behaves oddly in general.
To fix this we need to make the tablegen actions depend on a target rather than a filename.
projects / llvm / log