We don't allow an initializer for LDS variables
and there is an early abort during instruction
selection. This patch legalizes them by ignoring
the init values. During assembly emission, proper
error reporting already exists for such instances.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D109901
On targets that do not support AGPR to AGPR copying directly, try to find the
defining accvgpr_write and propagate its source vgpr register to the copies
before register allocation so the source vgpr register does not get clobbered.
The postrapseudos pass also attempt to propagate the defining accvgpr_write but
if the register to propagate is clobbered, it will give up and create new
temporary vgpr registers instead.
Reviewed By: rampitec
Differential Revision: https://reviews.llvm.org/D108830
The fix applied in D23303 "LiveIntervalAnalysis: fix a crash in repairOldRegInRange"
was over-zealous. It would bail out when the end of the range to be
repaired was in the middle of the first segment of the live range of
Reg, which was always the case when the range contained a single def of
Reg.
This patch fixes it as suggested by Matthias Braun in post-commit review
on the original patch, and tests it by adding -early-live-intervals to
a selection of existing lit tests that now pass.
(Note that D23303 was originally applied to fix a crash in
SILoadStoreOptimizer, but that is now moot since D23814 updated
SILoadStoreOptimizer to run before scheduling so it no longer has to
update live intervals.)
Differential Revision: https://reviews.llvm.org/D110238
Use of output modifiers forces VOP3 encoding for a VOP2 mac/fmac
instruction, so we might as well convert it to the more flexible VOP3-
only mad/fma form.
With this change, the only way we should emit VOP3-encoded mac/fmac is
if regalloc chooses registers that require the VOP3 encoding, e.g. sgprs
for both src0 and src1. In all other cases the mac/fmac should either be
converted to mad/fma or shrunk to VOP2 encoding.
Differential Revision: https://reviews.llvm.org/D110156
And always print it.
This makes some LLVM diagnostics match up better with Clang's diagnostics.
Updated some AMDGPU uses of DiagnosticInfoResourceLimit and now we print
better diagnostics for those.
Reviewed By: dblaikie
Differential Revision: https://reviews.llvm.org/D110204
Normally, given that the DA results are kept consistent over the selection DAG, uniform comparisons get selected to S_CMP_* but divergent to V_CMP_*. Sometimes, for the sake of efficiency, SSA subgraphs may be converted to VALU to avoid repeatedly copying data back and forth. Hence we have to be able to sustain the correctness passing the i1 from VALU to SALU context and vice versa.
VALU operations only process the active lanes of the VGPR and ignore inactive ones.
Active lanes correspond to 1 bit in the EXEC mask register.
SALU represents i1 as just one bit but VALU as 64bits: 0/1 and 0/(0xffffffffffffffff & EXEC) respectively.
SALU uses one-bit conditional flag SCC but VALU - VCC that is a pair of 32-bit SGPRs
To expose SCC to the VALU context we need to convert the one-bit boolean value to the appropriate 64bit.
To return back to the SALU context we need to do the opposite.
To correctly convert 64bit VALU boolean to either 0 or 1 we need to filter out the bits corresponding to the inactive lanes.
Reviewed By: piotr
Differential Revision: https://reviews.llvm.org/D109900
When adding alias.scope and noalias metadata to a memcpy function,
the alias.scope and noalias metadata from the operands are merged.
The rule for merging alias.scope is to take the intersection of
the domains and the union of the scopes within those domains.
The rule for merging noalias is to take the intersection.
The bug is that AMDGPULowerModuleLDS was using concatenation for
both alias.scope and noalias. For example, when f1 and f2 are added
to the LDS structure and there is a memcpy(f2, f1, sizeof(f1)).
Then, concatenation creates noalias metadata for the memcpy that
includes both {f1, f2}. That means that the memcpy is assumed
not to alias a prior load of f2, which enables the optimizer to
remove a load of f2 that occurs after mempcy.
The function MDNode::getmostGenericAliasScope defines the semantics
for alias.scope. There is a function, combineMetadata in Local.cpp,
that uses intersect for noalias.
Differential Revision: https://reviews.llvm.org/D110049
When using instructions which have a MetadataAsValue argument
(e.g. some target-specific intrinsics) MD canonicalization strips
internal MDNodes with a single ConstantAsMetadata child. That
prevented IRTranslator from the proper translation of such a calls.
FMA_W_CHAIN is used when lowering fdiv f32. Prefer to select it to fmac
if there are no source modifiers, just like we do for other mad/mac and
fma/fmac cases.
Differential Revision: https://reviews.llvm.org/D110074
v_fmac with source modifiers forces VOP3 encoding, but it is strictly
better to use the VOP3-only v_fma instead, because $dst and $src2 are
not tied so it gives the register allocator more freedom and avoids a
copy in some cases.
This is the same strategy we already use for v_mad vs v_mac and
v_fma_legacy vs v_fmac_legacy.
Differential Revision: https://reviews.llvm.org/D110070
This is motivated by an pathological compile time issue during unmerge combining.
We should be able to use the AVF to do simplification. However AMDGPU
has a lot of codegen changes which I'm not sure how to evaluate.
Differential Revision: https://reviews.llvm.org/D109748
For artifacts excluding G_TRUNC/G_SEXT, which have IR counterparts, we don't
seem to have debug users of defs. However, in the legalizer we're always calling
MachineInstr::eraseFromParentAndMarkDBGValuesForRemoval() which is expensive.
In some rare cases, this contributes significantly to unreasonably long compile
times when we have lots of artifact combiner activity.
To verify this, I added asserts to that function when it actually replaced a debug
use operand with undef for these artifacts. On CTMark with both -O0 and -Os and
debug info enabled, I didn't see a single case where it triggered.
In my measurements I saw around a 0.5% geomean compile-time improvement on -g -O0
for AArch64 with this change.
Differential Revision: https://reviews.llvm.org/D109750
Add eraseInstr(s) utility functions. Before deleting an instruction
collects its use instructions. After deletion deletes use instructions
that became trivially dead.
This patch clears all dead instructions in existing legalizer mir tests.
Differential Revision: https://reviews.llvm.org/D109154
Mostly this fixes cases where !noalias or !alias.scope were passed
a scope rather than a scope list. In some cases I opted to drop
the metadata entirely instead, because it is not really relevant
to the test.
In https://reviews.llvm.org/D100481, forceful inline of all non-kernel
functions using lds was disabled since AMDGPULowerModuleLDS pass now handles
static lds. However that pass does not handle extern lds so non-kernel
functions using extern lds must sill be inline.
Reviewed By: hsmhsm, arsenm
Differential Revision: https://reviews.llvm.org/D109773
The fmul is a canonicalizing operation, and fneg is not so this would
break denormals that need flushing and also would not quiet signaling
nans. Fold to fsub instead, which is also canonicalizing.
This simple heuristic uses the estimated live range length combined
with the number of registers in the class to switch which heuristic to
use. This was taking the raw number of registers in the class, even
though not all of them may be available. AMDGPU heavily relies on
dynamically reserved numbers of registers based on user attributes to
satisfy occupancy constraints, so the raw number is highly misleading.
There are still a few problems here. In the original testcase that
made me notice this, the live range size is incorrect after the
scheduler rearranges instructions, since the instructions don't have
the original InstrDist offsets. Additionally, I think it would be more
appropriate to use the number of disjointly allocatable registers in
the class. For the AMDGPU register tuples, there are a large number of
registers in each tuple class, but only a small fraction can actually
be allocated at the same time since they all overlap with each
other. It seems we do not have a query that corresponds to the number
of independently allocatable registers. Relatedly, I'm still debugging
some allocation failures where overlapping tuples seem to not be
handled correctly.
The test changes are mostly noise. There are a handful of x86 tests
that look like regressions with an additional spill, and a handful
that now avoid a spill. The worst looking regression is likely
test/Thumb2/mve-vld4.ll which introduces a few additional
spills. test/CodeGen/AMDGPU/soft-clause-exceeds-register-budget.ll
shows a massive improvement by completely eliminating a large number
of spills inside a loop.
Use GCNHazardRecognizer in postra sched.
Updated tests for the new schedules.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D109536
Change-Id: Ia86ba2ae168f12fb34b4d8efdab491f84d936cde
The attributor can determine that some indirect calls do not require
special inputs. The special inputs will still be present in the ABI,
so we need to allocate the registers and pass undefs.
This reverts commit 98f47131228cc73b5308e3cc6fd70375808594e3.
Without any (theoretical/practical) guarantee that all the allocas within
*entry* basic block are clustered together at the beginning of the block,
this patch is doomed to fail. Hence reverting it.
Drop the legacy version in AMDGPUAnnotateKernelFeatures. This has the
side effect of now respecting the linkage, and not changing externally
visible functions.
Previously we assumed all callable functions did not need any
implicitly passed inputs, and added attributes to functions to
indicate when they were necessary. Requiring attributes for
correctness is pretty ugly, and it makes supporting indirect and
external calls more complicated.
This inverts the direction of the attributes, so an undecorated
function is assumed to need all implicit imputs. This enables
AMDGPUAttributor by default to mark when functions are proven to not
need a given input. This strips the equivalent functionality from the
legacy AMDGPUAnnotateKernelFeatures pass.
However, AMDGPUAnnotateKernelFeatures is not fully removed at this
point although it should be in the future. It is still necessary for
the two hacky amdgpu-calls and amdgpu-stack-objects attributes, which
would be better served by a trivial analysis on the IR during
selection. Additionally, AMDGPUAnnotateKernelFeatures still
redundantly handles the uniform-work-group-size attribute to be
removed in a future commit.
At this point when not using -amdgpu-fixed-function-abi, we are still
modifying the ABI based on these newly negated attributes. In the
future, this option will be removed and the locations for implicit
inputs will always be fixed. We will then use the new attributes to
avoid passing the values when unnecessary.
This test is very verbose and appears generated by a script.
Make it truly autogenerated for easy updates.
NFC
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D109530
Change-Id: I1352b17b6d13ab9c5650dbe95ef0da97f71f1930
Legalizing G_MUL for non-standard types (like i33) generated an error. Putting
minScalar and maxScalar instead of clampScalar. Also using new rule, instead
of widening to the next power of 2, widen to the next multiple of the passed
argument (32 in this case), so instead of widening i65 to i128, we widen it to
i96.
Patch by: Mateja Marjanovic
Differential Revision: https://reviews.llvm.org/D109228
Add implementation for the legalization of G_ROTL and G_ROTR machine
instructions. They are very similar to funnel shift instructions, the only
difference is funnel shifts have 3 operands, whereas rotate instructions have
two operands, the first being the register that is being rotated and the second
being the number of shifts. The legalization of G_ROTL/G_ROTR is just lowering
them into funnel shift instructions if they are legal.
Patch by: Mateja Marjanovic
Differential Revision: https://reviews.llvm.org/D105347
Legalize G_MEMCPY, G_MEMMOVE, G_MEMSET and G_MEMCPY_INLINE.
Corresponding intrinsics are replaced by a loop that uses loads/stores in
AMDGPULowerIntrinsics pass unless their length is a constant lower then
MemIntrinsicExpandSizeThresholdOpt (default 1024). Any G_MEM* instruction that
reaches legalizer should have a const length argument and should be expanded
into appropriate number of loads + stores.
Differential Revision: https://reviews.llvm.org/D108357
Given a select_cc producing a constant and a invertion of the constant
for a comparison more than zero, we can produce an xor with ashr
instead, which produces smaller code. The ashr either sets all bits or
clear all bits depending on if the value is negative. This is then xor'd
with the constant to optionally negate the value.
https://alive2.llvm.org/ce/z/DTFaBZ
This includes a OneUseCheck on the Cmp, which seems to make thinks a
little worse and will be removed in a followup.
Differential Revision: https://reviews.llvm.org/D109149
Pulled out of D109149, this folds set_cc seteq (ashr X, BW-1), -1 ->
set_cc setlt X, 0 to prevent some regressions later on when folding
select_cc setgt X, -1, C, ~C -> xor (ashr X, BW-1), C
Differential Revision: https://reviews.llvm.org/D109214
Extend SILoadStoreOptimizer to merge into DWORDX8 variant of S_BUFFER_LOAD.
Merging into DWORDX2 and DWORDX4 variants is handled already.
Differential Revision: https://reviews.llvm.org/D108909
Currently, the truncate selection dag node is expanded as a bitwise AND plus compare to 1. This change enables scalar comparison in the pattern if the truncate node is uniform.
Reviewed By: rampitec
Differential Revision: https://reviews.llvm.org/D108925
