Matt Arsenault [Mon, 9 Sep 2019 18:10:31 +0000 (18:10 +0000)]
AMDGPU: Use PatFrags to allow selecting custom nodes or intrinsics
This enables GlobalISel to handle various intrinsics. The custom node
pattern will be ignored, and the intrinsic will work. This will also
allow SelectionDAG to directly select the intrinsics, but as they are
all custom lowered to the nodes, this ends up leaving dead code in the
table.
Eventually either GlobalISel should add the equivalent of custom nodes
equivalent, or intrinsics should be directly used. These each have
different tradeoffs.
There are a few more to handle, but these are easy to handle
ones. Some others fail for other reasons.
[X86] Allow _MM_FROUND_CUR_DIRECTION and _MM_FROUND_NO_EXC to be used together on instructions that only support SAE and not embedded rounding.
Current for SAE instructions we only allow _MM_FROUND_CUR_DIRECTION(bit 2) or _MM_FROUND_NO_EXC(bit 3) to be used as the immediate passed to the inrinsics. But these instructions don't perform rounding so _MM_FROUND_CUR_DIRECTION is just sort of a default placeholder when you don't want to suppress exceptions. Using _MM_FROUND_NO_EXC by itself is really bit equivalent to (_MM_FROUND_NO_EXC | _MM_FROUND_TO_NEAREST_INT) since _MM_FROUND_TO_NEAREST_INT is 0. Since we aren't rounding on these instructions we should also accept (_MM_FROUND_CUR_DIRECTION | _MM_FROUND_NO_EXC) as equivalent to (_MM_FROUND_NO_EXC). icc allows this, but gcc does not.
The bitstream remark serializer landed in r367372.
This adds a bitstream remark parser that parser bitstream remark files
to llvm::remarks::Remark objects through the RemarkParser interface.
A few interesting things to point out:
* There are parsing helpers to parse the different types of blocks
* The main parsing helper allows us to parse remark metadata and open an
external file containing the encoded remarks
* This adds a dependency from the Remarks library to the BitstreamReader
library
* The testing strategy is to create a remark entry through YAML, parse
it, serialize it to bitstream, parse that back and compare the objects.
* There are close to no tests for malformed bitstream remarks, due to
the lack of textual format for the bitstream format.
* This adds a new C API for parsing bitstream remarks:
LLVMRemarkParserCreateBitstream.
* This bumps the REMARKS_API_VERSION to 1.
Simon Atanasyan [Mon, 9 Sep 2019 17:28:45 +0000 (17:28 +0000)]
[mips] Fix decoding of microMIPS JALX instruction
microMIPS jump and link exchange instruction stores a target in a
26-bits field. Despite other microMIPS JAL instructions these bits
are target address shifted right 2 bits [1]. The patch fixes the
JALX instruction decoding and uses 2-bit shift.
[1] MIPS Architecture for Programmers Volume II-B: The microMIPS32 Instruction Set
Matt Arsenault [Mon, 9 Sep 2019 17:25:35 +0000 (17:25 +0000)]
AMDGPU: Move MnemonicAlias out of instruction def hierarchy
Unfortunately MnemonicAlias defines a "Predicates" field just like an
instruction or pattern, with a somewhat different interpretation.
This ends up overriding the intended Predicates set by
PredicateControl on the pseudoinstruction defintions with an empty
list. This allowed incorrectly selecting instructions that should have
been rejected due to the SubtargetPredicate from patterns on the
instruction definition.
This does remove the divergent predicate from the 64-bit shift
patterns, which were already not used for the 32-bit shift, so I'm not
sure what the point was. This also removes a second, redundant copy of
the 64-bit divergent patterns.
[GlobalISel][AArch64] Handle tail calls with non-void return types
Just return once you emit the call, which is exactly what SelectionDAG does in
this situation.
Update call-translator-tail-call.ll.
Also update dllimport.ll to show that we tail call here in GISel again. Add
-verify-machineinstrs to the GISel line too, to defend against verifier
failures.
Fangrui Song [Mon, 9 Sep 2019 16:45:17 +0000 (16:45 +0000)]
[yaml2obj] Simplify p_filesz/p_memsz computing
This fixes a bug as well. When "FileSize:" (p_filesz) is specified and
different from the actual value, the following code probably should not
use PHeader.p_filesz:
if (SHeader->sh_offset == PHeader.p_offset + PHeader.p_filesz)
PHeader.p_memsz += SHeader->sh_size;
David Green [Mon, 9 Sep 2019 16:35:49 +0000 (16:35 +0000)]
[ARM] Fix loads and stores for predicate vectors
These predicate vectors can usually be loaded and stored with a single
instruction, a VSTR_P0. However this instruction will store the entire P0
predicate, 16 bits, zeroextended to 32bits. Each lane of the the
v4i1/v8i1/v16i1 representing 4/2/1 bits.
As far as I understand, when llvm says "store this v4i1", it really does need
to store 4 bits (or 8, that being the size of a byte, with this bottom 4 as the
interesting bits). For example a bitcast from a v8i1 to a i8 is defined as a
store followed by a load, which is how the code is expanded.
So this instead lowers the v4i1/v8i1 load/store through some shuffles to get
the bits into the correct positions. This, as you might imagine, is not as
efficient as a single instruction. But I believe it is needed for correctness.
v16i1 equally should not load/store 32bits, only storing the 16bits of data.
Stack loads/stores are still using the VSTR_P0 (as can be seen by the test not
changing). This is fine as they are self-consistent, it is only "externally
observable loads/stores" (from our point of view) that need to be corrected.
Simon Tatham [Mon, 9 Sep 2019 15:17:26 +0000 (15:17 +0000)]
[ARM] Remove some spurious MVE reduction instructions.
The family of 'dual-accumulating' vector multiply-add instructions
(VMLADAV, VMLALDAV and VRMLALDAVH) can all operate on both signed and
unsigned integer types, and they all have an 'exchange' variant (with
an X in the name) that modifies which pairs of vector lanes in the two
inputs are multiplied together. But there's a clause in the spec that
says that the X variants //don't// operate on unsigned integer types,
only signed. You can have X, or unsigned, or neither, but not both.
We didn't notice that clause when we implemented the MC support for
these instructions, so LLVM believes that things like VMLADAVX.U8 do
exist, contradicting the spec. Here I fix that by conditioning them
out in Tablegen.
In order to do that, I've reversed the nesting order of the Tablegen
multiclasses for those instructions. Previously, the innermost
multiclass generated the X and not-X variants, and the one outside
that generated the A and not-A variants. Now X is done by the outer
multiclass, which allows me to bypass that one when I only want the
two not-X variants.
Changing the multiclass nesting order also changes the names of the
instruction ids unless I make a special effort not to. I decided that
while I was changing them anyway I'd make them look nicer; so now the
instructions have names like MVE_VMLADAVs32 or MVE_VMLADAVaxs32,
instead of cumbersome _noacc_noexch suffixes.
The corresponding multiply-subtract instructions are unaffected. Those
don't accept unsigned types at all, either in the spec or in LLVM.
James Molloy [Mon, 9 Sep 2019 13:17:55 +0000 (13:17 +0000)]
[DFAPacketizer] Reapply: Track resources for packetized instructions
Reapply with fix to reduce resources required by the compiler - use
unsigned[2] instead of std::pair. This causes clang and gcc to compile
the generated file multiple times faster, and hopefully will reduce
the resource requirements on Visual Studio also. This fix is a little
ugly but it's clearly the same issue the previous author of
DFAPacketizer faced (the previous tables use unsigned[2] rather uglily
too).
This patch allows the DFAPacketizer to be queried after a packet is formed to work out which
resources were allocated to the packetized instructions.
This is particularly important for targets that do their own bundle packing - it's not
sufficient to know simply that instructions can share a packet; which slots are used is
also required for encoding.
This extends the emitter to emit a side-table containing resource usage diffs for each
state transition. The packetizer maintains a set of all possible resource states in its
current state. After packetization is complete, all remaining resource states are
possible packetization strategies.
The sidetable is only ~500K for Hexagon, but the extra tracking is disabled by default
(most uses of the packetizer like MachinePipeliner don't care and don't need the extra
maintained state).
Summary:
This tests inlining size thresholds, but relies on the output of running
the full O2 pipeline, making it brittle against changes in unrelated
passes.
Only run the inlining pass and set thresholds on the test RUN line
instead.
Simon Pilgrim [Mon, 9 Sep 2019 12:33:22 +0000 (12:33 +0000)]
Revert rL371198 from llvm/trunk: [DFAPacketizer] Track resources for packetized instructions
This patch allows the DFAPacketizer to be queried after a packet is formed to work out which
resources were allocated to the packetized instructions.
This is particularly important for targets that do their own bundle packing - it's not
sufficient to know simply that instructions can share a packet; which slots are used is
also required for encoding.
This extends the emitter to emit a side-table containing resource usage diffs for each
state transition. The packetizer maintains a set of all possible resource states in its
current state. After packetization is complete, all remaining resource states are
possible packetization strategies.
The sidetable is only ~500K for Hexagon, but the extra tracking is disabled by default
(most uses of the packetizer like MachinePipeliner don't care and don't need the extra
maintained state).
Differential Revision: https://reviews.llvm.org/D66936
........
Reverted as this is causing "compiler out of heap space" errors on MSVC 2017/19 NDEBUG builds
David Green [Mon, 9 Sep 2019 10:46:25 +0000 (10:46 +0000)]
[ARM] Prevent generating NEON stack accesses under MVE.
We should not be generating Neon stack loads/stores even for these large
registers.
No test here because my understanding is we will only generate these QQPR regs
for intrinsics and VLDn's. The tests will follow once those are available.
Tim Northover [Mon, 9 Sep 2019 10:04:23 +0000 (10:04 +0000)]
GlobalISel: add combiner to form indexed loads.
Loosely based on DAGCombiner version, but this part is slightly simpler in
GlobalIsel because all address calculation is performed by G_GEP. That makes
the inc/dec distinction moot so there's just pre/post to think about.
No targets can handle it yet so testing is via a special flag that overrides
target hooks.
George Rimar [Mon, 9 Sep 2019 09:43:03 +0000 (09:43 +0000)]
[lib/ObjectYAML] - Improve and cleanup error reporting in ELFState<ELFT> class.
The aim of this patch is to refactor how we handle and report error.
I suggest to use the same approach we use in LLD: delayed error reporting.
For that I introduced 'HasError' flag which triggers when we report an error.
Now we do not exit instantly on any error. The benefits are:
1) There are no more 'exit(1)' calls in the library code.
2) Code was simplified significantly in a few places.
3) It is now possible to print multiple errors instead of only one.
Also, I changed the messages to be lower case and removed a full stop.
Sam Parker [Mon, 9 Sep 2019 08:39:14 +0000 (08:39 +0000)]
[ARM][ParallelDSP] Fix for sext input
The incoming accumulator value can be discovered through a sext, in
which case there will be a mismatch between the input and the result.
So sign extend the accumulator input if we're performing a 64-bit mac.
Kai Luo [Mon, 9 Sep 2019 02:32:42 +0000 (02:32 +0000)]
[MachineCopyPropagation] Remove redundant copies after TailDup via machine-cp
Summary:
After tailduplication, we have redundant copies. We can remove these
copies in machine-cp if it's safe to, i.e.
```
$reg0 = OP ...
... <<< No read or clobber of $reg0 and $reg1
$reg1 = COPY $reg0 <<< $reg0 is killed
...
<RET>
```
will be transformed to
```
$reg1 = OP ...
...
<RET>
```
This patch decodes target and faux shuffles with getTargetShuffleInputs - a reduced version of resolveTargetShuffleInputs that doesn't resolve SM_SentinelZero cases, so we can correctly remove zero vectors if they aren't demanded.
[X86] Remove call to getZeroVector from materializeVectorConstant. Add isel patterns for zero vectors with all types.
The change to avx512-vec-cmp.ll is a regression, but should be
easy to fix. It occurs because the getZeroVector call was
canonicalizing both sides to the same node, then SimplifySelect
was able to simplify it. But since only called getZeroVector
on some VTs this isn't a robust way to combine this.
The change to vector-shuffle-combining-ssse3.ll is more
instructions, but removes a constant pool load so its unclear
if its a regression or not.
Roman Lebedev [Sun, 8 Sep 2019 20:14:15 +0000 (20:14 +0000)]
[InstSimplify] simplifyUnsignedRangeCheck(): if we know that X != 0, handle more cases (PR43246)
Summary:
This is motivated by D67122 sanitizer check enhancement.
That patch seemingly worsens `-fsanitize=pointer-overflow`
overhead from 25% to 50%, which strongly implies missing folds.
In this particular case, given
```
char* test(char& base, unsigned long offset) {
return &base + offset;
}
```
it will end up producing something like
https://godbolt.org/z/LK5-iH
which after optimizations reduces down to roughly
```
define i1 @t0(i8* nonnull %base, i64 %offset) {
%base_int = ptrtoint i8* %base to i64
%adjusted = add i64 %base_int, %offset
%non_null_after_adjustment = icmp ne i64 %adjusted, 0
%no_overflow_during_adjustment = icmp uge i64 %adjusted, %base_int
%res = and i1 %non_null_after_adjustment, %no_overflow_during_adjustment
ret i1 %res
}
```
Without D67122 there was no `%non_null_after_adjustment`,
and in this particular case we can get rid of the overhead:
Here we add some offset to a non-null pointer,
and check that the result does not overflow and is not a null pointer.
But since the base pointer is already non-null, and we check for overflow,
that overflow check will already catch the null pointer,
so the separate null check is redundant and can be dropped.
Alive proofs:
https://rise4fun.com/Alive/WRzq
There are more patterns of "unsigned-add-with-overflow", they are not handled here,
but this is the main pattern, that we currently consider canonical,
so it makes sense to handle it.
[InstCombine] fold extract+insert into identity shuffle
This is similar to the existing fold for splats added with:
rL365379
If we can adjust the shuffle mask to include another element
in an identity mask (if it changes vector length, that's an
extract/insert subvector operation in the backend), then that
can eliminate extractelement/insertelement pairs in IR.
All targets are expected to lower shuffles with identity masks
efficiently.
Roman Lebedev [Sun, 8 Sep 2019 17:50:40 +0000 (17:50 +0000)]
[NFC][InstSimplify] Some tests for dropping null check after uadd.with.overflow of non-null (PR43246)
https://rise4fun.com/Alive/WRzq
Name: C <= Y && Y != 0 --> C <= Y iff C != 0
Pre: C != 0
%y_is_nonnull = icmp ne i64 %y, 0
%no_overflow = icmp ule i64 C, %y
%r = and i1 %y_is_nonnull, %no_overflow
=>
%r = %no_overflow
Name: C <= Y || Y != 0 --> Y != 0 iff C != 0
Pre: C != 0
%y_is_nonnull = icmp ne i64 %y, 0
%no_overflow = icmp ule i64 C, %y
%r = or i1 %y_is_nonnull, %no_overflow
=>
%r = %y_is_nonnull
Name: C > Y || Y == 0 --> C > Y iff C != 0
Pre: C != 0
%y_is_null = icmp eq i64 %y, 0
%overflow = icmp ugt i64 C, %y
%r = or i1 %y_is_null, %overflow
=>
%r = %overflow
Name: C > Y && Y == 0 --> Y == 0 iff C != 0
Pre: C != 0
%y_is_null = icmp eq i64 %y, 0
%overflow = icmp ugt i64 C, %y
%r = and i1 %y_is_null, %overflow
=>
%r = %y_is_null
David Stenberg [Sun, 8 Sep 2019 14:22:06 +0000 (14:22 +0000)]
[DebugInfo][X86] Describe call site values for zero-valued imms
Summary:
Add zero-materializing XORs to X86's describeLoadedValue() hook in order
to produce call site values.
I have had to change the defs logic in collectCallSiteParameters() a bit
to be able to describe the XORs. The XORs implicitly define $eflags,
which would cause them to never be considered, due to a guard condition
that I->getNumDefs() is one. I have changed that condition so that we
now only consider instructions where a forwarded register overlaps with
the instruction's single explicit define. We still need to collect the implicit
defines of other forwarded registers to remove them from the work list.
I'm not sure how to move towards supporting instructions with multiple
explicit defines, cases where forwarded register are implicitly defined,
and/or cases where an instruction produces values for multiple forwarded
registers. Perhaps the describeLoadedValue() hook should take a register
argument, and we then leave it up to the hook to describe the loaded
value in that register? I have not yet encountered a situation where
that would be necessary though.
David Stenberg [Sun, 8 Sep 2019 14:05:10 +0000 (14:05 +0000)]
[NFC] Make the describeLoadedValue() hook return machine operand objects
Summary:
This changes the ParamLoadedValue pair which the describeLoadedValue()
hook returns so that MachineOperand objects are returned instead of
pointers.
When describing call site values we may need to describe operands which
are not part of the instruction. One such example is zero-materializing
XORs on x86, which I have implemented support for in a child revision.
Instead of having to return a pointer to an operand stored somewhere
outside the instruction, start returning objects directly instead, as
that simplifies the code.
The MachineOperand class only holds POD members, and on x86-64 it is 32
bytes large. That combined with copy elision means that the overhead of
returning a machine operand object from the hook does not become very
large.
I benchmarked this on a 8-thread i7-8650U machine with 32 GB RAM. The
benchmark consisted of building a clang 8.0 binary configured with:
Simon Pilgrim [Sun, 8 Sep 2019 11:46:21 +0000 (11:46 +0000)]
[X86][SSE] Add support for <64 x i1> bool reduction
This generalizes the existing <32 x i1> pre-AVX2 split code to support reductions from <64 x i1> as well, we can probably generalize to any larger pow2 case in the future if the (unlikely) need ever arises.
We still need to tweak combineBitcastvxi1 to improve AVX512F codegen as its assumes vXi1 types should be handled on the mask registers even when they aren't legal.
[X86] Make getZeroVector return floating point vectors in their native type on SSE2 and later.
isel used to require zero vectors to be canonicalized to a single
type to minimize the number of patterns needed to match. This is
no longer required.
I plan to do this to integers too, but floating point was simpler
to start with. Integer has a complication where v32i16/v64i8 aren't
legal when the other 512-bit integer types are.
Summary:
Similar to the previous prefer-256-bit flag. We might want to
enable this by default some CPUs. This just starts the initial
work to implement and prove that it effects TTI's vector width.
Simon Pilgrim [Sat, 7 Sep 2019 16:13:57 +0000 (16:13 +0000)]
[X86] Avoid uses of getZextValue(). NFCI.
Use getAPIntValue() directly - this is mainly a best practice style issue to help prevent fuzz tests blowing up when a i12345 (or whatever) is generated.
Use getConstantOperandVal/getConstantOperandAPInt wrappers where possible.
[CodeGen] Handle SMULFIXSAT with scale zero in TargetLowering::expandFixedPointMul
Summary:
Normally TargetLowering::expandFixedPointMul would handle
SMULFIXSAT with scale zero by using an SMULO to compute the
product and determine if saturation is needed (if overflow
happened). But if SMULO isn't custom/legal it falls through
and uses the same technique, using MULHS/SMUL_LOHI, as used
for non-zero scales.
Problem was that when checking for overflow (handling saturation)
when not using MULO we did not expect to find a zero scale. So
we ended up in an assertion when doing
APInt::getLowBitsSet(VTSize, Scale - 1)
This patch fixes the problem by adding a new special case for
how saturation is computed when scale is zero.
Summary:
Add an intrinsic that takes 2 unsigned integers with
the scale of them provided as the third argument and
performs fixed point multiplication on them. The
result is saturated and clamped between the largest and
smallest representable values of the first 2 operands.
This is a part of implementing fixed point arithmetic
in clang where some of the more complex operations
will be implemented as intrinsics.
[X86] Fix pshuflw formation from repeated shuffle mask (PR43230)
Fix for https://bugs.llvm.org/show_bug.cgi?id=43230.
When creating PSHUFLW from a repeated shuffle mask, we have to apply
the checks to the repeated mask, not the original one. For the test
case from PR43230 the inspected part of the original mask is all undef.
This addresses the issue mentioned on D19867. When we simplify
with.overflow instructions in CVP, we leave behind extractvalue
of insertvalue sequences that LVI no longer understands. This
means that we can not simplify any instructions based on the
with.overflow anymore (until some over pass like InstCombine
cleans them up).
This patch extends LVI extractvalue handling by calling
SimplifyExtractValueInst (which doesn't do anything more than
constant folding + looking through insertvalue) and using the block
value of the simplification.
A possible alternative would be to do something similar to
SimplifyIndVars, where we instead directly try to replace
extractvalue users of the with.overflow. This would need some
additional structural changes to CVP, as it's currently not legal
to remove anything but the current instruction -- we'd have to
introduce a worklist with instructions scheduled for deletion or similar.
[DwarfExpression] Disallow some rewrites to avoid undefined behavior
Summary:
The value operand in DW_OP_plus_uconst/DW_OP_constu value can be
large (it uses uint64_t as representation internally in LLVM).
This means that in the uint64_t to int conversions, previously done
by DwarfExpression::addMachineRegExpression, could lose information.
Also, the negation done in "-Offset" was undefined behavior in case
Offset was exactly INT_MIN.
To avoid the above problems, we now avoid transformation like
[Reg, DW_OP_plus_uconst, Offset] --> [DW_OP_breg, Offset]
and
[Reg, DW_OP_constu, Offset, DW_OP_plus] --> [DW_OP_breg, Offset]
when Offset > INT_MAX.
And we avoid to transform
[Reg, DW_OP_constu, Offset, DW_OP_minus] --> [DW_OP_breg,-Offset]
when Offset > INT_MAX+1.
The patch also adjusts DwarfCompileUnit::constructVariableDIEImpl
to make sure that "DW_OP_constu, Offset, DW_OP_minus" is used
instead of "DW_OP_plus_uconst, Offset" when creating DIExpressions
with negative frame index offsets.
Notice that this might just be the tip of the iceberg. There
are lots of fishy handling related to these constants. I think both
DIExpression::appendOffset and DIExpression::extractIfOffset may
trigger undefined behavior for certain values.
Replicate the change "[Alignment][NFC] Use Align with TargetLowering::setMinFunctionAlignment"
on AVR to avoid a breakage.
See r371200 / https://reviews.llvm.org/D67229
Summary:
This patch introduces initial `AAValueSimplify` which simplifies a value in a context.
example
- (for function returned) If all the return values are the same and constant, then we can replace callsite returned with the constant.
- If an internal function takes the same value(constant) as an argument in the callsite, then we can replace the argument with that constant.
projects / llvm / log