Unfortunately the current call lowering code is built on top of the
legacy MVT/DAG based code. However, GlobalISel was not using it the
same way. In short, the DAG passes legalized types to the assignment
function, and GlobalISel was passing the original raw type if it was
simple.
I do believe the DAG lowering is conceptually broken since it requires
picking a type up front before knowing how/where the value will be
passed. This ends up being a problem for AArch64, which wants to pass
i1/i8/i16 values as a different size if passed on the stack or in
registers.
The argument type decision is split across 3 different places which is
hard to follow. SelectionDAG builder uses
getRegisterTypeForCallingConv to pick a legal type, tablegen gives the
illusion of controlling the type, and the target may have additional
hacks in the C++ part of the call lowering. AArch64 hacks around this
by not using the standard AnalyzeFormalArguments and special casing
i1/i8/i16 by looking at the underlying type of the original IR
argument.
I believe people have generally assumed the calling convention code is
processing the original types, and I've discovered a number of dead
paths in several targets.
x86 actually relies on the opposite behavior from AArch64, and relies
on x86_32 and x86_64 sharing calling convention code where the 64-bit
cases implicitly do not work on x86_32 due to using the pre-legalized
types.
AMDGPU targets without legal i16/f16 have always used a broken ABI
that promotes to i32/f32. GlobalISel accidentally fixed this to be the
ABI we should have, but this fixes it so we're using the worse ABI
that is compatible with the DAG. Ideally we would fix the DAG to match
the old GlobalISel behavior, but I don't wish to fight that battle.
A new native GlobalISel call lowering framework should let the target
process the incoming types directly.
CCValAssigns select a "ValVT" and "LocVT" but the meanings of these
aren't entirely clear. Different targets don't use them consistently,
even within their own call lowering code. My current belief is the
intent was "ValVT" is supposed to be the legalized value type to use
in the end, and and LocVT was supposed to be the ABI passed type
(which is also legalized).
With the default CCState::Analyze functions always passing the same
type for these arguments, these only differ when the TableGen part of
the lowering decide to promote the type from one legal type to
another. AArch64's i1/i8/i16 hack ends up inverting the meanings of
these values, so I had to add an additional hack to let the target
interpret how large the argument memory is.
Since targets don't consistently interpret ValVT and LocVT, this
doesn't produce quite equivalent code to the initial DAG
lowerings. I've opted to consistently interpret LocVT as the in-memory
size for stack passed values, and ValVT as the register type to assign
from that memory. We therefore produce extending loads directly out of
the IRTranslator, whereas the DAG would emit regular loads of smaller
values. This will also produce loads/stores that are wider than the
argument value if the allocated stack slot is larger (and there will
be undef padding bytes). If we had the optimizations to reduce
load/stores based on truncated values, this wouldn't produce a
different end result.
Since ValVT/LocVT are more consistently interpreted, we now will emit
more G_BITCASTS as requested by the CCAssignFn. For example AArch64
was directly assigning types to some physical vector registers which
according to the tablegen spec should have been casted to a vector
with a different element type.
This also moves the responsibility for inserting
G_ASSERT_SEXT/G_ASSERT_ZEXT from the target ValueHandlers into the
generic code, which is closer to how SelectionDAGBuilder works.
I had to xfail an x86 test since I don't see a quick way to fix it
right now (I filed bug 50035 for this). It's broken independently of
this change, and only triggers since now we end up with more ands
which hit the improperly handled selection pattern.
I also observed that FP arguments that need promotion (e.g. f16 passed
as f32) are broken, and use regular G_TRUNC and G_ANYEXT.
TLDR; the current call lowering infrastructure is bad and nobody has
ever understood how it chooses types.
Note, only src0 and src1 will be commuted if the isCommutable flag
is set. This patch does not change that, it just makes it possible
to commute src0 and src1 of some U/I/B vop3 instructions.
This patch revises d35d8da7d6.
It contains the commute opportunities excluding float insts
Reviewed By: rampitec
Differential Revision: https://reviews.llvm.org/D101474
Change-Id: I62938173d750453839f2457a3851661a29135faf
This merges more AMDGPU ABI lowering code into the generic call
lowering. Start cleaning up by factoring away more of the pack/unpack
logic into the buildCopy{To|From}Parts functions. These could use more
improvement, and the SelectionDAG versions are significantly more
complex, and we'll eventually have to emulate all of those cases too.
This is mostly NFC, but does result in some minor instruction
reordering. It also removes some of the limitations with mismatched
sizes the old code had. However, similarly to the merge on the input,
this is forcing gfx6/gfx7 to use the gfx8+ ABI (which is what we
actually want, but SelectionDAG is stuck using the weird emergent
ABI).
This also changes the load/store size for stack passed EVTs for
AArch64, which makes it consistent with the DAG behavior.
AMDGPU currently has a lot of pre-processing code to pre-split
argument types into 32-bit pieces before passing it to the generic
code in handleAssignments. This is a bit sloppy and also requires some
overly fancy iterator work when building the calls. It's better if all
argument marshalling code is handled directly in
handleAssignments. This handles more situations like decomposing large
element vectors into sub-element sized pieces.
This should mostly be NFC, but does change the generated code by
shifting where the initial argument packing instructions are placed. I
think this is nicer looking, since it now emits the packing code
directly after the relevant copies, rather than after the copies for
the remaining arguments.
This doubles down on gfx6/gfx7 using the gfx8+ ABI for 16-bit
types. This is ultimately the better option, but incompatible with the
DAG. Fixing this requires more work, especially for f16.
Summary:
pickNodeBidirectional tried to compare the best top candidate and the
best bottom candidate by examining TopCand.Reason and BotCand.Reason.
This is unsound because, after calling pickNodeFromQueue, Cand.Reason
does not reflect the most important reason why Cand was chosen. Rather
it reflects the most recent reason why it beat some other potential
candidate, which could have been for some low priority tie breaker
reason.
I have seen this cause problems where TopCand is a good candidate, but
because TopCand.Reason is ORDER (which is very low priority) it is
repeatedly ignored in favour of a mediocre BotCand. This is not how
bidirectional scheduling is supposed to work.
To fix this I changed the code to always compare TopCand and BotCand
directly, like the generic implementation of pickNodeBidirectional does.
This removes some uncommented AMDGPU-specific logic; if this logic turns
out to be important then perhaps it could be moved into an override of
tryCandidate instead.
Graphics shader benchmarking on gfx10 shows a lot more positive than
negative effects from this change.
Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB
Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D68338
There are few differences from the DAG handling. First, the DAG
handling uses a primitive selection pattern instead of custom
legalizing it. Because of this, this makes use of source modifiers
while the DAG does not.
Also instead of promoting f16, try to use the f16 log/exp. There's no
f16 fmul_legacy, so widen just for the multiply, although I'm not sure
that's the best solution.
Matt Arsenault