Legalization algorithm is complicated by two facts:
1) While regular instructions should be possible to legalize in
an isolated, per-instruction, context-free manner, legalization
artifacts can only be eliminated in pairs, which could be deeply, and
ultimately arbitrary nested: { [ () ] }, where which paranthesis kind
depicts an artifact kind, like extend, unmerge, etc. Such structure
can only be fully eliminated by simple local combines if they are
attempted in a particular order (inside out), or alternatively by
repeated scans each eliminating only one innermost pair, resulting in
O(n^2) complexity.
2) Some artifacts might in fact be regular instructions that could (and
sometimes should) be legalized by the target-specific rules. Which
means failure to eliminate all artifacts on the first iteration is
not a failure, they need to be tried as instructions, which may
produce more artifacts, including the ones that are in fact regular
instructions, resulting in a non-constant number of iterations
required to finish the process.
I trust the recently introduced termination condition (no new artifacts
were created during as-a-regular-instruction-retrial of artifacts not
eliminated on the previous iteration) to be efficient in providing
termination, but only performing the legalization in full if and only if
at each step such chains of artifacts are successfully eliminated in
full as well.
Which is currently not guaranteed, as the artifact combines are applied
only once and in an arbitrary order that has to do with the order of
creation or insertion of artifacts into their worklist, which is a no
particular order.
In this patch I make a small change to the artifact combiner, making it
to re-insert into the worklist immediate (modulo a look-through copies)
artifact users of each vreg that changes its definition due to an
artifact combine.
Here the first scan through the artifacts worklist, while not
being done in any guaranteed order, only needs to find the innermost
pair(s) of artifacts that could be immediately combined out. After that
the process follows def-use chains, making them shorter at each step, thus
combining everything that can be combined in O(n) time.
Reviewers: volkan, aditya_nandakumar, qcolombet, paquette, aemerson, dsanders
Reviewed By: aditya_nandakumar, paquette
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D71448
This reverts commit 69fcfb7d3597e0cdb5554b4e672e9032b411b167.
As shown in the test I attached to this commit, the change I reverted
causes a problem with "zext(cc1) - zext(cc2)". It commuted
the operands to the sub and used different logic to select the addc/subc
instruction:
sub zext (setcc), x => addcarry 0, x, setcc
sub sext (setcc), x => subcarry 0, x, setcc
... but that is bogus. I believe it is not possible to fold those commuted
patterns into any form of addcarry or subcarry. It may have worked as
intended before "AMDGPU: Change boolean content type to 0 or 1" because
the setcc was considered to be -1 rather than 1.
Differential Revision: https://reviews.llvm.org/D70978
Change-Id: If2139421aa6c935cbd1d925af58fe4a4aa9e8f43
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:
Catch the (admittedly unusual) case where SIFoldOperands attempts to fold 2
constant operands into the same SALU operation, with neither operand able to be
encoded as an inline constant.
Change-Id: Ibc48d662c9ffd8bbacd154976b0b1c257ace0927
Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, tpr, t-tye, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70896
Summary:
Pre gfx9 we need to scavenge a 64-bit SGPR to use as the carry out for an Add.
If only one SGPR was available this crashed when trying to scavenge another
32bit SGPR to materialize the offset.
Instead, reuse a 32-bit SGPR from the carry out as the offset register.
Also prefer to use vcc for the unused carry out when it is available.
Reviewers: arsenm, rampitec
Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70614
Summary:
I was seeing some failures on a test with slightly different instruction
ordering. Adding in some DAG directives solved the issue.
Change-Id: If5a3d3969055fb19279943bd45161bb70a3dabce
Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, tpr, t-tye, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70531
This reapplies: 8ff85ed905a7306977d07a5cd67ab4d5a56fafb4
Original commit message:
As a follow-up to my initial mail to llvm-dev here's a first pass at the O1 described there.
This change doesn't include any change to move from selection dag to fast isel
and that will come with other numbers that should help inform that decision.
There also haven't been any real debuggability studies with this pipeline yet,
this is just the initial start done so that people could see it and we could start
tweaking after.
Test updates: Outside of the newpm tests most of the updates are coming from either
optimization passes not run anymore (and without a compelling argument at the moment)
that were largely used for canonicalization in clang.
Original post:
http://lists.llvm.org/pipermail/llvm-dev/2019-April/131494.html
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D65410
This reverts commit c9ddb02659e3ece7a0d9d6b4dac7ceea4ae46e6d.
This change doesn't include any change to move from selection dag to fast isel
and that will come with other numbers that should help inform that decision.
There also haven't been any real debuggability studies with this pipeline yet,
this is just the initial start done so that people could see it and we could start
tweaking after.
Test updates: Outside of the newpm tests most of the updates are coming from either
optimization passes not run anymore (and without a compelling argument at the moment)
that were largely used for canonicalization in clang.
Original post:
http://lists.llvm.org/pipermail/llvm-dev/2019-April/131494.html
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D65410
Summary:
The waitcnt pass can overflow the counters when the number of outstanding events
for a type exceed the capacity of the counter. This can lead to inefficient
insertion of waitcnts, or to waitcnt instructions with max values for each type.
The last situation can cause an instruction which when disassembled appears to
be an illegal waitcnt without an operand.
In these cases we should add a wait for the 'counter maximum' - 1, and update the
waitcnt brackets accordingly.
Reviewers: rampitec, arsenm
Reviewed By: rampitec
Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70418
Summary:
Add a function attribute to allow the target specific default loop unroll threshold
to be specified on a per-function basis. This allows a front-end to give guidance
where it has insight that is not available to the back-end, while still allowing the
target specific heuristics to also have an effect.
Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D68873
Hostcall is a service that allows a kernel to submit requests to the
host using shared buffers, and block until a response is
received. This will eventually replace the shared buffer currently
used for printf, and repurposes the same hidden kernel argument. This
change introduces a new ValueKind in the HSA metadata to represent the
hostcall buffer.
Differential Revision: https://reviews.llvm.org/D70038
Start moving towards treating this as a property of the calling
convention, and not the subtarget. The default denormal mode should
not be part of the subtarget, and be moved into a separate function
attribute.
This patch is still NFC. The denormal mode remains as a subtarget
feature for now, but make the necessary changes to switch to using an
attribute.
Summary:
Most of IR instructions got better code size estimations after commit 47a5c36b.
So default parameters values should be updated to improve inlining and
unrolling for the target.
Reviewers: rampitec, arsenm
Reviewed By: rampitec
Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, zzheng, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70391
The usage of target boolean checks is overly inflexible, since sext
and zext of a compare are equally cheap. The choice is arbitrary, but
using 0/1 to some degree is the choice of lower resistance since
that's what most targets use. This enables a few combines that don't
bother to support ZeroOrNegativeOneBooleanContent.
Previously this would default to 256, not the maximum supported size
of 1024. Using a maximum lower than the hardware maximum requires
language runtimes to enforce this limit for correctness, which no
language has correctly done. Switch the default to the conservatively
correct maximum, and force frontends to opt-in to the more optimal 256
default maximum.
I don't really understand why the changes in occupancy-levels.ll
increased the computed occupancy, which I expected to decrease. I'm
not sure if these tests should be forcing the old maximum.
In MachineCopyPropagation, when propagating the source of a copy into
the operand of a later instruction, bail if a destination overlaps
(partly defines) the copy source. If the instruction where the
substitution is happening is also a copy, allowing the propagation
confuses the tracking mechanism.
Differential Revision: https://reviews.llvm.org/D69953
Change-Id: Ic570754f878f2d91a4a50a9bdcf96fbaa240726d
The combine G_UNMERGE_VALUES with G_CONCAT_VECTORS used to only be performed
when the result type of the G_UNMERGE_VALUES was a vector type.
In other words, we were expecting that the G_UNMERGE_VALUES was effectively
the exact opposite of the G_CONCAT_VECTORS.
Lift that constraint by allowing any G_UNMERGE_VALUES to be combined
with any G_CONCAT_VECTORS (as long as the size of the different pieces
that we merge/unmerge match).
Differential Revision: https://reviews.llvm.org/D69288
Summary:
G_GEP is rather poorly named. It's a simple pointer+scalar addition and
doesn't support any of the complexities of getelementptr. I therefore
propose that we rename it. There's a G_PTR_MASK so let's follow that
convention and go with G_PTR_ADD
Reviewers: volkan, aditya_nandakumar, bogner, rovka, arsenm
Subscribers: sdardis, jvesely, wdng, nhaehnle, hiraditya, jrtc27, atanasyan, arphaman, Petar.Avramovic, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D69734
readlane and writelane instructions are not allowed to use m0 as the
data operand, so spilling them is tricky and would require an
intermediate SGPR to spill it. Constrain the virtual register class in
this caes to disallow the inline spiller from folding the m0 operand
directly into the spill instruction.
I copied this hack from AArch64 which has the same problem for $sp.
Summary:
VCCZBugHandledSet was used to make sure we don't apply the same
workaround more than once to a single cbranch instruction, but it's not
necessary because the workaround involves inserting an s_waitcnt
instruction, which is enough for subsequent iterations to detect that no
further workaround is necessary.
Also beef up the test case to check that the workaround was only applied
once. I have also manually verified that the test still passes even if I
hack the big do-while loop in runOnMachineFunction to run a minimum of
five iterations.
Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D69621
Summary:
This is used on AMDGPU for rounding from v3f64 (which is illegal) to
v3f32 (which is legal).
Subscribers: jvesely, nhaehnle, tpr, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D69339
When TableGen is inferring register classes from contexts, it uses a
sorting function based on the number of registers in the class. Since
this was being treated as an alias of VGPR_32, they had exactly the
same size. The sort used wasn't a stable sort, and even if it were, I
believe the tie breaker would effectively end up being the
alphabetical ordering of the class name. There appear to be issues
trying to use an empty set of registers, so add only one so this will
always sort to the end.
Also add a comment explaining how VReg_1 is a dirty hack for
SelectionDAG.
This does end up changing the behavior of i1 with inline asm and VGPR
constraints, but the existing behavior was was already nonsensical and
inconsistent. It should probably be disallowed anyway.
Fixes bug 43699
Summary:
An outstanding load with same destination sgpr as call could cause PC to be
updated with junk value on return.
Reviewers: arsenm, rampitec
Reviewed By: arsenm
Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D69474
There is a minor flaw in the implementation of function lowerPhis.
This function replaces values of regclass Vreg_1 (boolean values)
involved in PHIs into an SGPR. Currently it iterates over the MBBs
and performs an inplace lowering of PHIs and fails to lower any
incoming value that itself is another PHI of Vreg_1 regclass.
The failure occurs only when the MBB where the incoming PHI value
belongs is not visited/lowered yet.
To fix this problem, collect all Vreg_1 PHIs upfront and then
perform the lowering.
Differential Revision: https://reviews.llvm.org/D69182
This makes the DAG behavior consistent with IR's extractelement after:
rGb32e4664a715
https://bugs.llvm.org/show_bug.cgi?id=42689
I've tried to maintain test intent for WebAssembly.
The AMDGPU test is trying to test for crashing or other bad behavior,
but I'm not sure if that's possible after this change.
That used to fail in the last testcase function because after
%0:sreg_64.sub0 was folded into %3:sreg_32_xm0_xexec COPY, it
was further folded into S_STORE_DWORD_IMM. Its legal effective
subreg class is SReg_32 while instruction expects more restricted
SReg_32_XM0_EXEC. However, SIInstrInfo::isLegalRegOperand()
passed the legality check and it was caught in the verifier.
Borrowed code from the verifier to check for RC legality.
Differential Revision: https://reviews.llvm.org/D69445
Custom lower this to a target instruction with the merge operands. I
think it might be better to directly select this and emit a
REG_SEQUENCE, but this would be more work since it would require
splitting the tablegen patterns for these cases from the other
atomics.
Summary:
In loadSRsrcFromVGPR, if MBB is the same as Succ, Remiander is not the immediate dominator of Succ.
Reviewer:
arsenm
Differential Revision:
https://reviews.llvm.org/D69358
