This adds support for constrained floating-point comparison intrinsics.
Specifically, we add:
declare <ty2>
@llvm.experimental.constrained.fcmp(<type> <op1>, <type> <op2>,
metadata <condition code>,
metadata <exception behavior>)
declare <ty2>
@llvm.experimental.constrained.fcmps(<type> <op1>, <type> <op2>,
metadata <condition code>,
metadata <exception behavior>)
The first variant implements an IEEE "quiet" comparison (i.e. we only
get an invalid FP exception if either argument is a SNaN), while the
second variant implements an IEEE "signaling" comparison (i.e. we get
an invalid FP exception if either argument is any NaN).
The condition code is implemented as a metadata string. The same set
of predicates as for the fcmp instruction is supported (except for the
"true" and "false" predicates).
These new intrinsics are mapped by SelectionDAG codegen onto two new
ISD opcodes, ISD::STRICT_FSETCC and ISD::STRICT_FSETCCS, again
representing quiet vs. signaling comparison operations. Otherwise
those nodes look like SETCC nodes, with an additional chain argument
and result as usual for strict FP nodes. The patch includes support
for the common legalization operations for those nodes.
The patch also includes full SystemZ back-end support for the new
ISD nodes, mapping them to all available SystemZ instruction to
fully implement strict semantics (scalar and vector).
Differential Revision: https://reviews.llvm.org/D69281
The file is intended to gather various VPlan transformations, not only
CFG related transforms. Actually, the only transformation there is not
CFG related.
Reviewers: Ayal, gilr, hsaito, rengolin
Reviewed By: gilr
Differential Revision: https://reviews.llvm.org/D70732
Summary:
This patch fixes an issue where the PPC MI peephole optimization pass incorrectly remove a vector swap.
Specifically, the pass can combine a splat/swap to a splat/copy. It uses `TargetRegisterInfo::lookThruCopyLike` to determine that the operands to the splat are the same. However, the current logic only compares the operands based on register numbers. In the case where the splat operands are ultimately feed from the same physical register, the pass can incorrectly remove a swap if the feed register for one of the operands has been clobbered.
This patch adds a check to ensure that the registers feeding are both virtual registers or the operands to the splat or swap are both the same register.
Here is an example in pseudo-MIR of what happens in the test cased added in this patch:
Before PPC MI peephole optimization:
```
%arg = XVADDDP %0, %1
$f1 = COPY %arg.sub_64
call double rint(double)
%res.first = COPY $f1
%vec.res.first = SUBREG_TO_REG 1, %res.first, %subreg.sub_64
%arg.swapped = XXPERMDI %arg, %arg, 2
$f1 = COPY %arg.swapped.sub_64
call double rint(double)
%res.second = COPY $f1
%vec.res.second = SUBREG_TO_REG 1, %res.second, %subreg.sub_64
%vec.res.splat = XXPERMDI %vec.res.first, %vec.res.second, 0
%vec.res = XXPERMDI %vec.res.splat, %vec.res.splat, 2
; %vec.res == [ %vec.res.second[0], %vec.res.first[0] ]
```
After optimization:
```
; ...
%vec.res.splat = XXPERMDI %vec.res.first, %vec.res.second, 0
; lookThruCopyLike(%vec.res.first) == lookThruCopyLike(%vec.res.second) == $f1
; so the pass replaces the swap with a copy:
%vec.res = COPY %vec.res.splat
; %vec.res == [ %vec.res.first[0], %vec.res.second[0] ]
```
As best as I can tell, this has occurred since r288152, which added support for lowering certain vector operations to direct moves in the form of a splat.
Committed for vddvss (Colin Samples). Thanks Colin for the patch!
Differential Revision: https://reviews.llvm.org/D69497
Scudo only supports building for android/linux/fuchsia, so require target_os to
be one of linux/fuchsia to do a stage2_unix scudo build. Android is already
covered by the stage2_android* toolchains below.
Differential Revision: https://reviews.llvm.org/D71131
Summary:
Add a new cl::callback attribute to Option.
This attribute specifies a callback function that is called when
an option is seen, and can be used to set other options, as in
option A implies option B. If the option is a `cl::list`, and
`cl::CommaSeparated` is also specified, the callback will fire
once for each value. This could be used to validate combinations
or selectively set other options.
Reviewers: beanz, thomasfinch, MaskRay, thopre, serge-sans-paille
Reviewed By: beanz
Subscribers: llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70620
Before this change, the *InstPrinter.cpp files of each target where some
of the slowest objects to compile in all of LLVM. See this snippet produced by
ClangBuildAnalyzer:
https://reviews.llvm.org/P8171$96
Search for "InstPrinter", and see that it shows up in a few places.
Tablegen was emitting a large switch containing a sequence of operand checks,
each of which created many conditions and many BBs. Register allocation and
jump threading both did not scale well with such a large repetitive sequence of
basic blocks.
So, this change essentially turns those control flow structures into
data. The previous structure looked like:
switch (Opc) {
case TGT::ADD:
// check alias 1
if (MI->getOperandCount() == N && // check num opnds
MI->getOperand(0).isReg() && // check opnd 0
...
MI->getOperand(1).isImm() && // check opnd 1
AsmString = "foo";
break;
}
// check alias 2
if (...)
...
return false;
The new structure looks like:
OpToPatterns: Sorted table of opcodes mapping to pattern indices.
\->
Patterns: List of patterns. Previous table points to subrange of
patterns to match.
\->
Conds: The if conditions above encoded as a kind and 32-bit value.
See MCInstPrinter.cpp for the details of how the new data structures are
interpreted.
Here are some before and after metrics.
Time to compile AArch64InstPrinter.cpp:
0m29.062s vs. 0m2.203s
size of the obj:
3.9M vs. 676K
size of clang.exe:
97M vs. 96M
I have not benchmarked disassembly performance, but typically
disassemblers are bottlenecked on IO and string processing, not alias
matching, so I'm not sure it's interesting enough to be worth doing.
Reviewers: RKSimon, andreadb, xbolva00, craig.topper
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D70650
The xor'ing behaviour is only used for msvc/crt environments, when we're targeting
macho the guard load code doesn't know about the xor in the epilog. Disable xor'ing
when targeting win32-macho to be consistent.
Differential Revision: https://reviews.llvm.org/D71095
D53794 introduced code to perform the FP_TO_UINT expansion via FP_TO_SINT in a way that would never expose floating-point exceptions in the intermediate steps. Unfortunately, I just noticed there is still a way this can happen. As discussed in D53794, the compiler now generates this sequence:
// Sel = Src < 0x8000000000000000
// Val = select Sel, Src, Src - 0x8000000000000000
// Ofs = select Sel, 0, 0x8000000000000000
// Result = fp_to_sint(Val) ^ Ofs
The problem is with the Src - 0x8000000000000000 expression. As I mentioned in the original review, that expression can never overflow or underflow if the original value is in range for FP_TO_UINT. But I missed that we can get an Inexact exception in the case where Src is a very small positive value. (In this case the result of the sub is ignored, but that doesn't help.)
Instead, I'd suggest to use the following sequence:
// Sel = Src < 0x8000000000000000
// FltOfs = select Sel, 0, 0x8000000000000000
// IntOfs = select Sel, 0, 0x8000000000000000
// Result = fp_to_sint(Val - FltOfs) ^ IntOfs
In the case where the value is already in range of FP_TO_SINT, we now simply compute Val - 0, which now definitely cannot trap (unless Val is a NaN in which case we'd want to trap anyway).
In the case where the value is not in range of FP_TO_SINT, but still in range of FP_TO_UINT, the sub can never be inexact, as Val is between 2^(n-1) and (2^n)-1, i.e. always has the 2^(n-1) bit set, and the sub is always simply clearing that bit.
There is a slight complication in the case where Val is a constant, so we know at compile time whether Sel is true or false. In that scenario, the old code would automatically optimize the sub away, while this no longer happens with the new code. Instead, I've added extra code to check for this case and then just fall back to FP_TO_SINT directly. (This seems to catch even slightly more cases.)
Original version of the patch by Ulrich Weigand. X86 changes added by Craig Topper
Differential Revision: https://reviews.llvm.org/D67105
Summary:
musttail calls should not require allocating extra stack for arguments.
Updates to arguments passed in memory should happen in place before the
epilogue.
This bug was mostly a missed optimization, unless inalloca was used and
store to push conversion fired.
If a reserved call frame was used for an inalloca musttail call, the
call setup and teardown instructions would be deleted, and SP
adjustments would be inserted in the prologue and epilogue. You can see
these are removed from several test cases in this change.
In the case where the stack frame was not reserved, i.e. call frame
optimization fires and turns argument stores into pushes, then the
imbalanced call frame setup instructions created for inalloca calls
become a problem. They remain in the instruction stream, resulting in a
call setup that allocates zero bytes (expected for inalloca), and a call
teardown that deallocates the inalloca pack. This deallocation was
unbalanced, leading to subsequent crashes.
Reviewers: hans
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D71097
Summary:
Sample profile loader of AutoFDO tries to replay previous inlining using context sensitive profile. The replay only repeats inlining if the call site block is hot. As a result it punts inlining of small functions, some of which can be beneficial for size, and will still be inlined by CSGCC inliner later. The oscillation between sample profile loader's inlining and regular CGSSC inlining cause unnecessary loss of context-sensitive profile. It doesn't have much impact for inline decision itself, but it negatively affects post-inline profile quality as CGSCC inliner have to scale counts which is not as accurate as the original context sensitive profile, and bad post-inline profile can misguide code layout.
This change added regular Inline Cost calculation for sample profile loader, so we can inline small functions upfront under switch -sample-profile-inline-size. In addition -sample-profile-cold-inline-threshold is added so we can tune the separate size threshold - currently the default is chosen to be the same as regular inliner's cold call-site threshold.
Reviewers: wmi, davidxl
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70750
GCC says:
.../llvm/lib/DebugInfo/GSYM/FunctionInfo.cpp:195:12:
error: ‘InfoType’ is not a class, namespace, or enumeration
case InfoType::EndOfList:
^
Presumably, GCC thinks InfoType is a variable here. Work around it by
using the name IT as is done above.
This reverts commit a7d90af1be48234ce583e00fb16e33633d44ae38.
This revision came back as the root-cause for crashes in internal
ARM-IOS apps.
Reproducer in https://bugs.llvm.org/show_bug.cgi?id=44231.
This is a follow-up to D70607 where we made any
extract element on SLM more costly than default. But that is
pessimistic for extract from element 0 because that corresponds
to x86 movd/movq instructions. These generally have >1 cycle
latency, but they are probably implemented as single uop
instructions.
Note that no vectorization tests are affected by this change.
Also, no targets besides SLM are affected because those are
falling through to the default cost of 1 anyway. But this will
become visible/important if we add more specializations via cost
tables.
Differential Revision: https://reviews.llvm.org/D71023
Summary:
Split off of D67120.
Add the profile guided size optimization instrumentation / queries in the code
gen or target passes. This doesn't enable the size optimizations in those passes
yet as they are currently disabled in shouldOptimizeForSize (for non-IR pass
queries).
Reviewers: davidxl
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D71072
Current tail duplication integrated in bb layout is designed to increase the fallthrough from a BB's predecessor to its successor, but we have observed cases that duplication doesn't increase fallthrough, or it brings too much size overhead.
To overcome these two issues in function canTailDuplicateUnplacedPreds I add two checks:
make sure there is at least one duplication in current work set.
the number of duplication should not exceed the number of successors.
The modification in hasBetterLayoutPredecessor fixes a bug that potential predecessor must be at the bottom of a chain.
Differential Revision: https://reviews.llvm.org/D64376
SUMMARY:
if the size of Csect is zero, the Csect do not need write any data into sections
for example, the TOC Csect has zero size, it do not need invoke a
Asm.writeSectionData(W.OS, Csect.MCCsect, Layout);
Reviewers: daltenty
Subscribers: rupprecht, seiyai,hiraditya
Differential Revision: https://reviews.llvm.org/D71120
Summary:
The immediate forms of the MVE VQSHL instruction have MC names like
`MVE_VSLIimms8` and `MVE_VSLIimmu32`. Those names are confusing,
because VSLI is a completely different shift instruction with no
semantic relation to VQSHL. But it just happens to be defined
immediately before VQSHL in `ARMInstrMVE.td`, so this looks like a
copy-paste error. Renamed the ids to match the instruction name.
Reviewers: ostannard, dmgreen, MarkMurrayARM, miyuki
Reviewed By: miyuki
Subscribers: kristof.beyls, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D71114
SUMMARY:
in the patch https://reviews.llvm.org/D66969 . we need a test case to verify the out text section of the xcoffobject file is correct or not.
but we do not have llvm disassembly tools to dump the xcoffobjectfile . since we commit the patch https://reviews.llvm.org/D70255, we have tools for it. we create this test case for it.
Reviewers: daltenty,hubert.reinterpretcast,
Differential Revision: https://reviews.llvm.org/D70719
Summary:
When trying to calculate the offsets for the jump table entries
we fail to take into account the block alignment, which could be
greater than 4 bytes. This led to cases where the jump table
offset was too big to fit in a byte.
Reviewers: t.p.northover, sdesmalen, ostannard
Reviewed By: ostannard
Subscribers: ostannard, kristof.beyls, hiraditya, llvm-commits
Committed on behalf of David Sherwood (david-arm)
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70533
This introduce a new helper which is used to parse the SHT_GNU_versym section.
LLVM/GNU styles implementations now use it to share the logic.
Differential revision: https://reviews.llvm.org/D71054
InnerLoopVectorizer's code called during VPlan execution still relies on
original IR's def-use relations to decide which vector code to generate,
limiting VPlan transformations ability to modify def-use relations and still
have ILV generate the vector code.
This commit moves GEP operand queries controlling how GEPs are widened to a
dedicated recipe and extracts GEP widening code to its own ILV method taking
those recorded decisions as arguments. This reduces ingredient def-use usage by
ILV as a step towards full VPlan-based def-use relations.
Differential revision: https://reviews.llvm.org/D69067
One of CodeGenPrepare's optimizations is to duplicate address calculations
into basic blocks, so that as much information as possible can be folded
into memory addressing operands. This is great -- but the dbg.value
variable location intrinsics are not updated in the same way. This can lead
to dbg.values referring to address computations in other blocks that will
never be encoded into the DAG, while duplicate address computations are
performed locally that could be used by the dbg.value. Some of these (such
as non-constant-offset GEPs) can't be salvaged past.
Fix this by, whenever we duplicate an address computation into a block,
looking for dbg.value users of the original memory address in the same
block, and redirecting those to the local computation.
Differential Revision: https://reviews.llvm.org/D58403
After my recent commit daee549 the following test case is failing:
CodeGen/X86/vector-constrained-fp-intrinsics.ll
Not sure why I didn't catch this earlier, seems to be affected
by other changes that came in recently.
Fixed by regerenating the test again. Sorry for the disruption!
Summary:
Adds intrinsics for the following:
* cmphs, cmphi
* cmpge, cmpgt
* cmpeq, cmpne
* cmplt, cmple
* cmplo, cmpls
Includes a minor change to `TLI.getMemValueType` that fixes a crash due to the
scalable flag being dropped.
Reviewers: sdesmalen, efriedma, rengolin, rovka, dancgr, huntergr
Reviewed By: efriedma
Subscribers: tschuett, kristof.beyls, hiraditya, rkruppe, psnobl, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70889
This patch implements the following changes:
1) SelectionDAGBuilder::visitConstrainedFPIntrinsic currently treats
each constrained intrinsic like a global barrier (e.g. a function call)
and fully serializes all pending chains. This is actually not required;
it is allowed for constrained intrinsics to be reordered w.r.t one
another or (nonvolatile) memory accesses. The MI-level scheduler already
allows for that flexibility, so it makes sense to allow it at the DAG
level as well.
This patch therefore changes the way chains for constrained intrisincs
are created, and handles them basically like load operations are handled.
This has the effect that constrained intrinsics are no longer serialized
against one another or (nonvolatile) loads. They are still serialized
against stores, but that seems hard to change with the current DAG chain
setup, and it also doesn't seem to be a big problem preventing DAG
2) The OPC_CheckFoldableChainNode check requires that each of the
intermediate nodes in a multi-node pattern match only has a single use.
This check tends to fail if those intermediate nodes are strict operations
as those have a chain output that typically indeed has another use.
However, we don't really need to consider chains here at all, since they
will all be rewritten anyway by UpdateChains later. Other parts of the
matcher therefore already ignore chains, but this hasOneUse check doesn't.
This patch replaces hasOneUse by a custom test that verifies there is no
more than one use of any non-chain output value.
In theory, this change could affect code unrelated to strict FP nodes,
but at least on SystemZ I could not find any single instance of that
happening
3) The SystemZ back-end currently does not allow matching multiply-and-
extend operations (32x32 -> 64bit or 64x64 -> 128bit FP multiply) for
strict FP operations. This was not possible in the past due to the
problems described under 1) and 2) above.
With those issues fixed, it is now possible to fully support those
instructions in strict mode as well, and this patch does so.
Differential Revision: https://reviews.llvm.org/D70913
