We were using v_readlane_b32 with the lane set to zero, but this won't
work if thread 0 is not active.
Differential Revision: http://reviews.llvm.org/D19745
llvm-svn: 268295
Summary:
When we restore an SGPR value from scratch, we first load it into a
temporary VGPR and then use v_readlane_b32 to copy the value from the
VGPR back into an SGPR.
We weren't setting the kill flag on the VGPR in the v_readlane_b32
instruction, so the register scavenger wasn't able to re-use this
temp value later.
I wasn't able to create a lit test for this.
Reviewers: arsenm
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D19744
llvm-svn: 268287
Summary:
This includes a hazard recognizer implementation to replace some of
the hazard handling we had during frame index elimination.
Reviewers: arsenm
Subscribers: qcolombet, arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D18602
llvm-svn: 268143
This resolves more frame indexes early and folds
the immediate offsets into the scratch mubuf instructions.
This cleans up a lot of the mess that's currently emitted,
such as emitting add 0s and repeatedly initializing the same
register to 0 when spilling.
llvm-svn: 266508
Summary:
For GL_ARB_compute_shader we need to support workgroup sizes of at least 1024. However, if we want to allow large workgroup sizes, we may need to use less registers, as we have to run more waves per SIMD.
This patch adds an attribute to specify the maximum work group size the compiled program needs to support. It defaults, to 256, as that has no wave restrictions.
Reducing the number of registers available is done similarly to how the registers were reserved for chips with the sgpr init bug.
Reviewers: mareko, arsenm, tstellarAMD, nhaehnle
Subscribers: FireBurn, kerberizer, llvm-commits, arsenm
Differential Revision: http://reviews.llvm.org/D18340
Patch By: Bas Nieuwenhuizen
llvm-svn: 266337
Summary:
When we are spilling SGPRs to scratch memory, we usually don't have
free SGPRs to do the address calculation, so we need to re-use the
ScratchOffset register for the calculation.
Reviewers: arsenm
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D18917
llvm-svn: 266244
Tests added along with implemented feature.
Note that there is a small leftover of unecessary MI sheduling issue
(more info in the review). CodeGen/AMDGPU/salu-to-valu.ll updated to fix
the false regression.
TODO: Support for TTMP quads, comma-separated syntax in "[]" and more.
Differential Revision: http://reviews.llvm.org/D17825
llvm-svn: 266205
Summary: This makes it possible to insert nops at the end of blocks.
Reviewers: arsenm
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D18549
llvm-svn: 265678
We can statically decide whether or not a register pressure set is for
SGPRs or VGPRs, so we don't need to re-compute this information in
SIRegisterInfo::getRegPressureSetLimit().
Differential Revision: http://reviews.llvm.org/D14805
llvm-svn: 264126
Summary:
They correspond to BUFFER_LOAD/STORE_FORMAT_XYZW and will be used by Mesa
to implement the GL_ARB_shader_image_load_store extension.
The intention is that for llvm.amdgcn.buffer.load.format, LLVM will decide
whether one of the _X/_XY/_XYZ opcodes can be used (similar to image sampling
and loads). However, this is not currently implemented.
For llvm.amdgcn.buffer.store, LLVM cannot decide to use one of the "smaller"
opcodes and therefore the intrinsic is overloaded. Currently, only the v4f32
is actually implemented since GLSL also only has a vec4 variant of the store
instructions, although it's conceivable that Mesa will want to be smarter
about this in the future.
BUFFER_LOAD_FORMAT_XYZW is already exposed via llvm.SI.vs.load.input, which
has a legacy name, pretends not to access memory, and does not capture the
full flexibility of the instruction.
Reviewers: arsenm, tstellarAMD, mareko
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D17277
llvm-svn: 263140
Summary:
This is necessary for when we run out of VGPRs and can no
longer use v_{read,write}_lane for spilling SGPRs.
Reviewers: arsenm
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D17592
llvm-svn: 262732
Summary:
This allows us to use virtual registers when we need extra registers
for inserting spill instructions in SIRegisterInfo:eliminateFrameIndex().
Once all the frame indices have been eliminated, the
PrologEpilogueInserter does an extra pass over the program to replace
all virtual registers with physical ones.
This allows us to make more efficient use of our emergency spill slots,
so we only need to create one.
Reviewers: arsenm
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D17591
llvm-svn: 262728
This was hardcoded to the static private size, but this
would be missing the offset and additional size for someday
when we have dynamic sizing.
Also stops always initializing flat_scratch even when unused.
In the future we should stop emitting this unless flat instructions
are used to access private memory. For example this will initialize
it almost always on VI because flat is used for global access.
llvm-svn: 260658
Summary:
It's possible to have resource descriptors and samplers stored in
VGPRs, either by a VMEM instruction or in the case of samplers,
floating-point calculations. When this happens, we need to use
v_readfirstlane to copy these values back to sgprs.
Reviewers: mareko, arsenm
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D17102
llvm-svn: 260599
Summary:
This fixes a crash where subsequent spills would be unable to scavenge
a register. In particular, it fixes a crash in piglit's
spec@glsl-1.50@execution@geometry@max-input-components (the test still
has a shader that fails to compile because of too many SGPR spills, but
at least it doesn't crash any more).
This is a candidate for the release branch.
Reviewers: arsenm, tstellarAMD
Subscribers: qcolombet, arsenm
Differential Revision: http://reviews.llvm.org/D16558
llvm-svn: 260427
Summary:
It is off by default, but can be used
with --misched=si
Patch by: Axel Davy
Reviewers: arsenm, tstellarAMD, nhaehnle
Subscribers: nhaehnle, solenskiner, arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D11885
llvm-svn: 257609
Summary:
Multi-dword constant loads generated unnecessary moves from SGPRs into VGPRs,
increasing the code size and VGPR pressure. These moves are now folded away.
Note that this lack of operand folding was not a problem for VMEM loads,
because COPY nodes from VReg_Nnn to VGPR32 are eliminated by the register
coalescer.
Some tests are updated, note that the fsub.ll test explicitly checks that
the move is elided.
With the IR generated by current Mesa, the changes are obviously relatively
minor:
7063 shaders in 3531 tests
Totals:
SGPRS: 351872 -> 352560 (0.20 %)
VGPRS: 199984 -> 200732 (0.37 %)
Code Size: 9876968 -> 9881112 (0.04 %) bytes
LDS: 91 -> 91 (0.00 %) blocks
Scratch: 1779712 -> 1767424 (-0.69 %) bytes per wave
Wait states: 295164 -> 295337 (0.06 %)
Totals from affected shaders:
SGPRS: 65784 -> 66472 (1.05 %)
VGPRS: 38064 -> 38812 (1.97 %)
Code Size: 1993828 -> 1997972 (0.21 %) bytes
LDS: 42 -> 42 (0.00 %) blocks
Scratch: 795648 -> 783360 (-1.54 %) bytes per wave
Wait states: 54026 -> 54199 (0.32 %)
Reviewers: tstellarAMD, arsenm, mareko
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D15875
llvm-svn: 257074
Summary:
Somehow, I first interpreted the docs as saying space for xnack_mask is only
reserved when XNACK is enabled via SH_MEM_CONFIG. I felt uneasy about this and
went back to actually test what is happening, and it turns out that xnack_mask
is always reserved at least on Tonga and Carrizo, in the sense that flat_scr
is always fixed below the SGPRs that are used to implement xnack_mask, whether
or not they are actually used.
I confirmed this by writing a shader using inline assembly to tease out the
aliasing between flat_scratch and regular SGPRs. For example, on Tonga, where
we fix the number of SGPRs to 80, s[74:75] aliases flat_scratch (so
xnack_mask is s[76:77] and vcc is s[78:79]).
This patch changes both the calculation of the total number of SGPRs and the
various register reservations to account for this.
It ought to be possible to use the gap left by xnack_mask when the feature
isn't used, but this patch doesn't try to do that. (Note that the same applies
to vcc.)
Note that previously, even before my earlier change in r256794, the SGPRs that
alias to xnack_mask could end up being used as well when flat_scr was unused
and the total number of SGPRs happened to fall on the right alignment
(e.g. highest regular SGPR being used s29 and VCC used would lead to number
of SGPRs being 32, where s28 and s29 alias with xnack_mask). So if there
were some conflict due to such aliasing, we should have noticed that already.
Reviewers: arsenm, tstellarAMD
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D15898
llvm-svn: 257073
Summary:
Enabling this feature will account for the two SGPRs used by the hardware
to store the XNACK_MASK physically.
The hardware only requires this reservation when the XNACK feature is
explicitly enabled. At some point, HSA will probably want to do that, but
it does increase SGPR register pressure, so leave it disabled by default
for now (but do add a small test).
Reviewers: arsenm, tstellarAMD
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D15869
llvm-svn: 256794
Summary:
The comment explains it: emitError does not necessarily exit the compilation
process, and then using NoRegister leads to assertions later on.
This generates incorrect code, of course, but the user should know to not use
the result when an error has been emitted.
It would be nice to have a test-case for this inside the LLVM repository,
but llc exits on error. shader-db tests trigger the underlying issue at least
on Tonga.
Reviewers: arsenm, tstellarAMD, mareko
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D15826
llvm-svn: 256757
Summary:
The method insertNOPs expected the number of wait states to be passed as
parameter, while eliminateFrameIndex passed the immediate argument for the
S_NOP, leading to an off-by-one error. Rename the method to make the
meaning of its parameter clearer. The number of 4 / 5 wait states (which
is what the method has always _tried_ to do according to the comment) is
correct according to the hardware docs.
I stumbled upon this while trying to track down the cause of
https://bugs.freedesktop.org/show_bug.cgi?id=93264. While clearly needed,
this patch unfortunately does not fix that bug...
Reviewers: arsenm, tstellarAMD
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D15542
llvm-svn: 255906
If we know we have stack objects, we reserve the registers
that the private buffer resource and wave offset are passed
and use them directly.
If not, reserve the last 5 SGPRs just in case we need to spill.
After register allocation, try to pick the next available registers
instead of the last SGPRs, and then insert copies from the inputs
to the reserved registers in the progloue.
This also only selectively enables all of the input registers
which are really required instead of always enabling them.
llvm-svn: 254331
It does not work because of emergency stack slots.
This pass was supposed to eliminate dummy registers for the
spill instructions, but the register scavenger can introduce
more during PrologEpilogInserter, so some would end up
left behind if they were needed.
The potential for spilling the scratch resource descriptor
and offset register makes doing something like this
overly complicated. Reserve registers to use for the resource
descriptor and use them directly in eliminateFrameIndex.
Also removes creating another scratch resource descriptor
when directly selecting scratch MUBUF instructions.
The choice of which registers are reserved is temporary.
For now it attempts to pick the next available registers
after the user and system SGPRs.
llvm-svn: 254329
Summary:
This returns a pointer to the dispatch packet, which can be used to load
information about the kernel dispach.
Reviewers: arsenm
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D14898
llvm-svn: 254116
This reverts commit r252565.
This also includes the revert of the commit mentioned below in order to
avoid breaking tests in AMDGPU:
Revert "AMDGPU: Set isAllocatable = 0 on VS_32/VS_64"
This reverts commit r252674.
llvm-svn: 252956
For some reason VS_32 ends up factoring into the pressure heuristics
even though we should never see a virtual register with this class.
When SGPRs are reserved for register spilling, this for some reason
triggers reg-crit scheduling.
Setting isAllocatable = 0 may help with this since that seems to remove
it from the default implementation's generated table.
llvm-svn: 252321
This wasn't doing anything useful. They weren't explicitly used
anywhere, and the RegScavenger ignores reserved registers.
This for some reason caused a random scheduling change in the test.
Getting the check lines to pass is too frustrating, and there's probably
not too much value in checking the vector case's operands N times.
llvm-svn: 250794
This was the slowest target custom pass and was spending 80%
of the time in getMinimalPhysRegClass which was called
for every register operand.
Try to use the statically known register class when possible from
the instruction's MCOperandInfo. There are a few pseudo instructions
which are not well behaved with unknown register classes which still
require the expensive physical register class search.
There are a few other possibilities for making this even faster,
such as not inspecting implicit operands. For now those are checked
because it is technically possible to have a scalar load into
exec or vcc which can be implicitly used.
llvm-svn: 249079
Allow a target to do something other than search for copies
that will avoid cross register bank copies.
Implement for SI by only rewriting the most basic copies,
so it should look through anything like a subregister extract.
I'm not entirely satisified with this because it seems like
eliminating a reg_sequence that isn't fully used should work
generically for all targets without them having to override
something. However, it seems to be tricky to have a simple
implementation of this without rewriting to invalid kinds
of subregister copies on some targets.
I'm not sure if there is currently a generic way to easily check
if a subregister index would be valid for the current use.
The current set of TargetRegisterInfo::get*Class functions don't
quite behave like I would expect (e.g. getSubClassWithSubReg
returns the maximal register class rather than the minimal), so
I'm not sure how to make the generic test keep searching if
SrcRC:SrcSubReg is a valid replacement for DefRC:DefSubReg. Making
the default implementation to check for simple copies breaks
a variety of ARM and x86 tests by producing illegal subregister uses.
The ARM tests are not actually changed since it should still be using
the same sharesSameRegisterFile implementation, this just relaxes
them to not check for specific registers.
llvm-svn: 248478
