Since the beginning, the offset of a frame index has been consistently
interpreted backwards. It was treating it as an offset from the
scratch wave offset register as a frame register. The correct
interpretation is the offset from the SP on entry to the function,
before the prolog. Frame index elimination then should select either
SP or another register as an FP.
Treat the scratch wave offset on kernel entry as the pre-incremented
SP. Rely more heavily on the standard hasFP and frame pointer
elimination logic, and clean up the private reservation code. This
saves a copy in most callee functions.
The kernel prolog emission code is still kind of a mess relying on
checking the uses of physical registers, which I would prefer to
eliminate.
Currently selection directly emits MUBUF instructions, which require
using a reference to some register. Use the register chosen for SP,
and then ignore this later. This should probably be cleaned up to use
pseudos that don't refer to any specific base register until frame
index elimination.
Add a workaround for shaders using large numbers of SGPRs. I'm not
sure these cases were ever working correctly, since as far as I can
tell the logic for figuring out which SGPR is the scratch wave offset
doesn't match up with the shader input initialization in the shader
programming guide.
llvm-svn: 362661
The proposal in D62498 showed that x86 would benefit from vector
store splitting, but that may conflict with the generic DAG
combiner's store merging transforms.
Add memory type to the existing TLI hook that enables the merging
transforms, so we can limit those changes to scalars only for x86.
llvm-svn: 362507
This matches APInt's versions of these functions, and there is no need for these to be size_t.
(as well as __builtin_clzll())
Differential Revision: https://reviews.llvm.org/D60823
llvm-svn: 362503
This is something of a workaround, and the state of stack realignment
controls is kind of a mess. Ideally, we would be able to specify the
stack is infinitely aligned on entry to a kernel.
TargetFrameLowering provides multiple controls which apply at
different points. The StackRealignable field is used during
SelectionDAG, and for some reason distinct from this
hook. StackAlignment is a single field not dependent on the
function. It would probably be better to make that dependent on the
calling convention, and the maximum value for kernels.
Currently this doesn't really change anything, since the frame
lowering mostly does its own thing. This helps avoid regressions in a
future change which will rely more heavily on hasFP.
llvm-svn: 362447
For some reason multiple places need to do this, and the variant the
loop unroller and inliner use was not handling it.
Also, introduce a new wrapper to be slightly more precise, since on
AMDGPU some addrspacecasts are free, but not no-ops.
llvm-svn: 362436
AMDGPU uses multiplier 9 for the inline cost. It is taken into account
everywhere except for inline hint threshold. As a result we are penalizing
functions with the inline hint making them less probable to be inlined
than those without the hint. Defaults are 225 for a normal function and
325 for a function with an inline hint. Currently we have effective
threshold 225 * 9 = 2025 for normal functions and just 325 for those with
the hint. That is fixed by this patch.
Differential Revision: https://reviews.llvm.org/D62707
llvm-svn: 362239
With LLPC, previous investigation has suggested that si-scheduler
interacts badly with SiFormMemoryClauses on an XNACK target in some
games.
That needs further investigation in the future. In the meantime, this
commit adds a target-specific attribute to allow us to disable
SIFormMemoryClauses by setting it to 1 on a per-function basis for LLPC
to use.
Differential Revision: https://reviews.llvm.org/D62572
Change-Id: Ia0ca12ce79093cbbe86caded723ffb13384ede92
llvm-svn: 362127
Summary:
- There's a regression due to the cross-block RC assignment. Use the
proper way to derive the output register RC in inline asm.
Reviewers: rampitec, alex-t
Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, dstuttard, tpr, t-tye, eraman, hiraditya, llvm-commits, yaxunl
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D62537
llvm-svn: 361868
If the only VGPRs used for SGPR spilling were not CSRs, this was
enabling all laness and immediately restoring exec. This is the usual
situation in leaf functions.
llvm-svn: 361848
Summary:
- Don't treat the use of a scalar register as `vreg_1` an VGPR usage.
Otherwise, that promotes that scalar register into vector one, which
breaks the assumption that scalar register holds the lane mask.
- The issue is triggered in a complicated case, where if the uses of
that (lane mask) scalar register is legalized firstly before its
definition, e.g., due to the mismatch block placement and its
topological order or loop. In that cases, the legalization of PHI
introduces the use of that scalar register as `vreg_1`.
Reviewers: rampitec, nhaehnle, arsenm, alex-t
Subscribers: kzhuravl, jvesely, wdng, dstuttard, tpr, t-tye, hiraditya, llvm-commits, yaxunl
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D62492
llvm-svn: 361847
1a8b2ea611cf4ca7cb09562e0238cfefa27c05b5 Divergence driven ISel. Assign register class for cross block values according to the divergence.
llvm-svn: 361770
Details: To make instruction selection really divergence driven it is necessary to assign
the correct register classes to the cross block values beforehand. For the divergent targets
same value type requires different register classes dependent on the value divergence.
Reviewers: rampitec, nhaehnle
Differential Revision: https://reviews.llvm.org/D59990
This commit was reverted because of the build failure.
The reason was mlformed patch.
Build failure fixed.
llvm-svn: 361741
This matches countLeadingOnes() and countTrailingOnes(), and
APInt's countLeadingZeros() and countTrailingZeros().
(as well as __builtin_clzll())
llvm-svn: 361724
This was skipping GetUnderlyingObject for nonprivate addresses, but an
alloca could also be found through an addrspacecast if it's flat.
llvm-svn: 361649
Details: To make instruction selection really divergence driven it is necessary to assign
the correct register classes to the cross block values beforehand. For the divergent targets
same value type requires different register classes dependent on the value divergence.
Reviewers: rampitec, nhaehnle
Differential Revision: https://reviews.llvm.org/D59990
llvm-svn: 361644
We were assuming a much larger possible per-wave visible stack
allocation than is possible:
faa3ae5138/src/core/runtime/amd_gpu_agent.cpp (L70)
Based on this, we can assume the high 15 bits of a frame index or sret
are 0. The frame index value is the per-lane offset, so the maximum
frame index value is MAX_WAVE_SCRATCH / wavesize.
Remove the corresponding subtarget feature and option that made
this configurable.
llvm-svn: 361541
Keep it optional in cases this is ever needed in some global
context. Currently it's only used for getting an upper bound inline
asm code size.
For AMDGPU, gfx10 increases the maximum instruction size to
20-bytes. This avoids penalizing older subtargets when estimating code
size, and making some annoying branch relaxation test adjustments.
llvm-svn: 361405
Unfortunately the way SIInsertSkips works is backwards, and is
required for correctness. r338235 added handling of some special cases
where skipping is mandatory to avoid side effects if no lanes are
active. It conservatively handled asm correctly, but the same logic
needs to apply to calls.
Usually the call sequence code is larger than the skip threshold,
although the way the count is computed is really broken, so I'm not
sure if anything was likely to really hit this.
llvm-svn: 361202
A std::array is implemented as a template with an array
inside a struct. Older versions of clang, like 3.6,
require an extra set of curly braces around std::array
initializations to avoid warnings.
The C++ language was changed regarding this by CWG 1270.
So more modern tool chains does not complaing even if
leaving out one level of braces.
llvm-svn: 361171
Summary:
Avoid introducing hazard mitigation when lgkmcnt is reduced to 0.
Clarify code comments to explain assumptions made for this hazard
mitigation. Expand and correct test cases to cover variants of
s_waitcnt.
Reviewers: nhaehnle, rampitec
Subscribers: arsenm, kzhuravl, jvesely, wdng, yaxunl, dstuttard, tpr, t-tye, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D62058
llvm-svn: 361124
This is ported from the custom AMDGPU DAG implementation. I think this
is a better default expansion than what the DAG currently uses, at
least if the target has CTLZ.
This implements the signed version in terms of the unsigned
conversion, which is implemented with bit operations. SelectionDAG has
several other implementations that should eventually be ported
depending on what instructions are legal.
llvm-svn: 361081
Matt Arsenault