As noted in D80236, moving the pass in the pipeline exposed this
shortcoming. Extra work to recalculate the alias results showed
up as a compile-time slowdown.
There are 2 known problem patterns shown in the test diffs here:
vector horizontal ops (an x86 specialization) and vector reductions.
SLP has greater ability to match and fold those than vector-combine,
so let SLP have first chance at that.
This is a quick fix while we continue to improve vector-combine and
possibly canonicalize to reduction intrinsics.
In the longer term, we should improve matching of these patterns
because if they were created in the "bad" forms shown here, then we
would miss optimizing them.
I'm not sure what is happening with alias analysis on the addsub test.
The old pass manager now shows an extra line for that, and we see an
improvement that comes from SLP vectorizing a store. I don't know
what's missing with the new pass manager to make that happen.
Strangely, I can't reproduce the behavior if I compile from C++ with
clang and invoke the new PM with "-fexperimental-new-pass-manager".
Differential Revision: https://reviews.llvm.org/D80236
Unlike SelectionDAGBuilder, IRTranslator omits the unconditional
branch in fallthrough cases. Confusingly, the control flow pseudos
function in the opposite way the intrinsics are used, and the branch
targets always need to be swapped. We're inverting the target blocks,
so we need to figure out the old fallthrough block and insert a branch
to the original unconditional branch target.
This only affects assembly and -filetype=asm codegen of PAL metadata.
Differential Revision: https://reviews.llvm.org/D78860
Change-Id: I7b822e1917bf7b403486820d31afc483be207652
Promote alloca to vector before SROA and loop unroll. If we manage
to eliminate allocas before unroll we may choose to unroll less.
Differential Revision: https://reviews.llvm.org/D80386
It was set in total vector size while the idea was to limit
a number of instructions. Now it started to work with doubles
and thresholds needs to be updated.
Differential Revision: https://reviews.llvm.org/D80322
Even though series of cmd/cndmask can produce quite a lot of
code that is still better than a loop. In case of doubles we
would even produce two loops.
Differential Revision: https://reviews.llvm.org/D80032
Relying on any MachineFunction state in the MachineFunctionInfo
constructor is hazardous, because the construction time is unclear and
determined by the first use. The function may be only partially
constructed, which is part of why we have many of these hacky string
attributes to track what we need for ABI lowering.
For SelectionDAG, all stack objects are created up-front before
calling convention lowering so stack objects are visible at
construction time. For GlobalISel, none of the IR function has been
visited yet and the allocas haven't been added to the MachineFrameInfo
yet. This should fix failing to set flat_scratch_init in GlobalISel
when needed.
This pass really needs to be turned into some kind of analysis, but I
haven't found a nice way use one here.
This should be directly implied from the register class, and there's
no need to special case live ins here. This was getting the wrong
answer for the queue ptr argument in callable functions, since it's
not an explicit IR argument and is always uniform.
Fixes not using scalar loads for the aperture in addrspacecast
lowering, and any other places that use implicit SGPR arguments.
This was assumed to be a simple move, and interpreting the immediate
modifier operand as a materialized immediate. Apparently the SDWA pass
never produces these, but GlobalISel does emit these for some vector
shuffles.
When the callee requires a dynamic stack realignment,
it is not possible to correcty access the incoming
stack arguments using the stack pointer. We reserve a
base pointer in such cases to access the function arguments
inside the callee. The base pointer will hold the incoming
stack pointer value before any kind of delta added to it.
Reviewed By: arsenm, scott.linder
Differential Revision: https://reviews.llvm.org/D78811
Along the lines of D77454 and D79968. Unlike loads and stores, the
default alignment is getPrefTypeAlign, to match the existing handling in
various places, including SelectionDAG and InstCombine.
Differential Revision: https://reviews.llvm.org/D80044
Summary:
When spilling in the entry function we should be able to borrow
StackPtrOffsetReg as a last resort. This restores behaviour
removed in D75138, and fixes failures when shaders use all
SGPRs, VGPRs and spill in the entry function.
Reviewers: scott.linder, arsenm, tpr
Reviewed By: scott.linder, arsenm
Subscribers: qcolombet, foad, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, t-tye, hiraditya, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D79776
For IR generated by a compiler, this is really simple: you just take the
datalayout from the beginning of the file, and apply it to all the IR
later in the file. For optimization testcases that don't care about the
datalayout, this is also really simple: we just use the default
datalayout.
The complexity here comes from the fact that some LLVM tools allow
overriding the datalayout: some tools have an explicit flag for this,
some tools will infer a datalayout based on the code generation target.
Supporting this properly required plumbing through a bunch of new
machinery: we want to allow overriding the datalayout after the
datalayout is parsed from the file, but before we use any information
from it. Therefore, IR/bitcode parsing now has a callback to allow tools
to compute the datalayout at the appropriate time.
Not sure if I covered all the LLVM tools that want to use the callback.
(clang? lli? Misc IR manipulation tools like llvm-link?). But this is at
least enough for all the LLVM regression tests, and IR without a
datalayout is not something frontends should generate.
This change had some sort of weird effects for certain CodeGen
regression tests: if the datalayout is overridden with a datalayout with
a different program or stack address space, we now parse IR based on the
overridden datalayout, instead of the one written in the file (or the
default one, if none is specified). This broke a few AVR tests, and one
AMDGPU test.
Outside the CodeGen tests I mentioned, the test changes are all just
fixing CHECK lines and moving around datalayout lines in weird places.
Differential Revision: https://reviews.llvm.org/D78403
Enable clausing of memory loads on gfx10 by adding a new pass to insert
the s_clause instructions that mark the start of each hard clause.
Differential Revision: https://reviews.llvm.org/D79792
SelectMOVRELOffset prevents peeling of a constant from an index
if final base could be negative. isBaseWithConstantOffset() succeeds
if a value is an "add" or "or" operator. In case of "or" it shall
be an add-like "or" which never changes a sign of the sum given a
non-negative offset. I.e. we can safely allow peeling if operator is
an "or".
Differential Revision: https://reviews.llvm.org/D79898
We can produce such vectors in the Promote Alloca pass,
but we are unable to use movrel to operate it and lower
via scratch. Making it legal makes SI_INDIRECT patterns
work.
There is more work to do in subsequent changes:
1. We initialize m0 twice to access each dword. It shall
be possible to only do it once and increment base register
number instead.
2. We also need v16i64/v16f64 but these first need to be
added to tablegen.
Differential Revision: https://reviews.llvm.org/D79808
Summary:
ConstantExprs involving operations on <1 x Ty> could translate into MIR
that failed to verify with:
*** Bad machine code: Reading virtual register without a def ***
The problem was that translate(const Constant &C, Register Reg) had
recursive calls that passed the same Reg in for the translation of a
subexpression, but without updating VMap for the subexpression first as
translate(const Constant &C, Register Reg) expects.
Fix this by using the same translateCopy helper function that we use for
translating Instructions. In some cases this causes extra G_COPY
MIR instructions to be generated.
Fixes https://bugs.llvm.org/show_bug.cgi?id=45576
Reviewers: arsenm, volkan, t.p.northover, aditya_nandakumar
Subscribers: jvesely, wdng, nhaehnle, rovka, hiraditya, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D78378
When splitting loads in RegBankSelect G_EXTRACT_VECTOR_ELT were being added
which could not be selected. Since invoking the legalizer will generate
instructions that split and combine wide loads, we can remove the redundant
repair instructions which are added by RegBankSelect.
Differential Revision: https://reviews.llvm.org/D75547
allocas in LLVM IR have a specified alignment. When that alignment is
specified, the alloca has at least that alignment at runtime.
If the specified type of the alloca has a higher preferred alignment,
SelectionDAG currently ignores that specified alignment, and increases
the alignment. It does this even if it would trigger stack realignment.
I don't think this makes sense, so this patch changes that.
I was looking into this for SVE in particular: for SVE, overaligning
vscale'ed types is extra expensive because it requires realigning the
stack multiple times, or using dynamic allocation. (This currently isn't
implemented.)
I updated the expected assembly for a couple tests; in particular, for
arg-copy-elide.ll, the optimization in question does not increase the
alignment the way SelectionDAG normally would. For the rest, I just
increased the specified alignment on the allocas to match what
SelectionDAG was inferring.
Differential Revision: https://reviews.llvm.org/D79532
If there are no available lanes in a reserved VGPR, no free SGPR, and no unused CSR
VGPR when trying to save the FP it needs to be spilled to memory as a last
resort. This can be done in the prolog/epilog if we manually add the spill
and manage exec.
Differential Revision: https://reviews.llvm.org/D79610
Just do not touch loads and stores which are already vector.
Previously pass was just unable to see these loads and stores
because these were hidden bitcasts.
Differential Revision: https://reviews.llvm.org/D79738
Currently this code exists in widenScalar for G_MERGE_VALUE
sources. I'm not sure if the existing expansion in widenScalar should
be removed or not. The widenScalar variant tries to extend to the
requested size, but this just uses the original bitwidth.
G_BITCAST can be lowered with a pair of G_UNMERGE_VALUES and
G_MERGE_VALUES with different types, but G_UNMERGE_VALUES of a vector
can also be implemented with a bitcast to a scalar, which introduces
the possibility for infinite loops. Try to eliminate an illegal source
register type in the artifact combiner to avoid this from happening.
Avoids infinite looping in the legalizer in a future patch which
allows lowering G_UNMERGE_VALUES of a vector source with a G_BITCAST.
Check the address space first before searching for the object
definition to save compile time. As an added bonus, this will now
treat casts to constant addrspace as constant.
We also seemed to be missing targeted tests for this, so add a few
missing other cases too.
This is mostly useful if alloca element type is not integer
and then casted to an integer for load or store. We now can
vectorize an [i32] alloca but cannot do so for [float].
There also a separate patch needed to properly lower 64 bit
types after they vectorized. At the moment these are lowered
via scratch anyway.
Differential Revision: https://reviews.llvm.org/D79641
If the SimplifyMultipleUseDemandedBits calls BITCASTs that peek through back to the original type then we can remove the BITCASTs entirely.
Differential Revision: https://reviews.llvm.org/D79572
