The logic in RISCVMatInt would previously produce lui+addiw on RV64
whenever a 32-bit integer must be materialised and the Hi20 and Lo12
parts are non-zero. However, sometimes addi can be used equivalently
(whenever the sign extension behaviour of addiw would be a no-op). This
patch moves to using addiw only when necessary. Although there is
absolutely no advantage in terms of compressibility or performance, this
has the following advantages:
* It's more consistent with logic used elsewhere in the backend. For
instance, RISCVOptWInstrs will try to convert addiw to addi on the basis
it reduces test diffs vs RV32.
* This matches the lowering GCC does in its codegen path. Unlike LLVM,
GCC seems to have different expansion logic for the assembler vs
codegen. For codegen it will use lui+addi if possible, but expanding
`li` in the assembler will always produces lui+addiw as LLVM did prior
to this commit. As someone who has been looking at a lot of gcc vs clang
diffs lately, reducing unnecessary divergence is of at least some value.
* As the diff for fold-mem-offset.ll shows, we can fold memory offsets
in more cases when addi is used. Memory offset folding could be taught
to recognise when the addiw could be replaced with an addi, but that
seems unnecessary when we can simply change the logic in RISCVMatInt.
As pointed out by @topperc during review, making this change without
modifying RISCVOptWInstrs risks introducing some cases where we fail to
remove a sext.w that we removed before. I've incorporated a patch based
on a suggestion from Craig that avoids it, and also adds appropriate
RISCVOptWInstrs test cases.
The initial patch description noted that the main motivation was to
avoid unnecessary differences both for RV32/RV64 and when comparing GCC,
but noted that very occasionally we see a benefit from memory offset
folding kicking in when it didn't before. Looking at the dynamic
instruction count difference for SPEC benchmarks targeting rva22u64 and
it shows we actually get a meaningful
~4.3% reduction in dynamic icount for 519.lbm_r. Looking at the data
more closely, the codegen difference is in `LBM_performStreamCollideTRT`
which as a function accounts for ~98% for dynamically executed
instructions and the codegen diffs appear to be a knock-on effect of the
address merging reducing register pressure right from function entry
(for instance, we get a big reduction in dynamically executed loads in
that function).
Below is the icount data (rva22u64 -O3, no LTO):
```
Benchmark Baseline This PR Diff (%)
============================================================
500.perlbench_r 174116601991 174115795810 -0.00%
502.gcc_r 218903280858 218903215788 -0.00%
505.mcf_r 131208029185 131207692803 -0.00%
508.namd_r 217497594322 217497594297 -0.00%
510.parest_r 289314486153 289313577652 -0.00%
511.povray_r 30640531048 30640765701 0.00%
519.lbm_r 95897914862 91712688050 -4.36%
520.omnetpp_r 134641549722 134867015683 0.17%
523.xalancbmk_r 281462762992 281432092673 -0.01%
525.x264_r 379776121941 379535558210 -0.06%
526.blender_r 659736022025 659738387343 0.00%
531.deepsjeng_r 349122867552 349122867481 -0.00%
538.imagick_r 238558760552 238558753269 -0.00%
541.leela_r 406578560612 406385135260 -0.05%
544.nab_r 400997131674 400996765827 -0.00%
557.xz_r 130079522194 129945515709 -0.10%
```
The instcounting setup I use doesn't have good support for drilling down
into functions from outside the linked executable (e.g. libc). The
difference in omnetpp all seems to come from there, and does not reflect
any degradation in codegen quality.
I can confirm with the current version of the PR there is no change in
the number of static sext.w across all the SPEC 2017 benchmarks
(rva22u64 O3)
Co-authored-by: Craig Topper <craig.topper@sifive.com>
Tests have been re-generated with recent scheduler changes.
Original message:
SelectionDAG will not reassociate adds to the end of a chain if
there are multiple users of later additions. This prevents isel
from folding the immediate into a load/store address.
One easy way to see this is accessing an array in a struct with
two different indices. An ADDI will be used to get to the start
of the array then 2 different SHXADD instructions will be used to
add the scaled indices. Finally the SHXADD will be used by different
load instructions. We can remove the ADDI by folding the offset into
each load.
This patch adds a new pass that analyzes how an ADDI constant
propagates through address arithmetic. If the arithmetic is only
used by a load/store and the offset is small enough, we can adjust
the load/store offset and remove the ADDI.
This pass is placed before MachineCSE to allow cleanups if some
instructions become common after removing offsets from their inputs.
This pass gives ~3% improvement on dynamic instruction count on
541.leela_r and 544.nab_r from SPEC2017 for the train data set. There's
a ~1% improvement on 557.xz_r.
SelectionDAG will not reassociate adds to the end of a chain if
there are multiple users of later additions. This prevents isel
from folding the immediate into a load/store address.
One easy way to see this is accessing an array in a struct with
two different indices. An ADDI will be used to get to the start
of the array then 2 different SHXADD instructions will be used to
add the scaled indices. Finally the SHXADD will be used by different
load instructions. We can remove the ADDI by folding the offset into
each load.
This patch adds a new pass that analyzes how an ADDI constant
propagates through address arithmetic. If the arithmetic is only
used by a load/store and the offset is small enough, we can adjust
the load/store offset and remove the ADDI.
This pass is placed before MachineCSE to allow cleanups if some
instructions become common after removing offsets from their inputs.
This pass gives ~3% improvement on dynamic instruction count on
541.leela_r and 544.nab_r from SPEC2017 for the train data set. There's
a ~1% improvement on 557.xz_r.
This reverts commit 9cc8442a2b438962883bbbfd8ff62ad4b1a2b95d.
This reverts commit 859c871184bdfdebb47b5c7ec5e59348e0534e0b.
A performance regression was reported on the original review. There appears
to have been an unexpected interaction here. Reverting during investigation.
This change introduces a default schedule model for the RISCV target
which leaves everything unchanged except the MicroOpBufferSize. The
default value of this flag in NoSched is 0. Both configurations
represent in order cores (i.e. no reorder window), the difference
between them comes down to whether heuristics other than latency are
allowed to apply. (Implementation details below)
I left the processor models which explicitly set MicroOpBufferSize=0
unchanged in this patch, but strongly suspect we should change those
too. Honestly, I think the LLVM wide default for this flag should be
changed, but don't have the energy to manage the updates for all
targets.
Implementation wise, the effect of this change is that schedule units
which are ready to run *except that* one of their predecessors may not
have completed yet are added to the Available list, not the Pending one.
The result of this is that it becomes possible to chose to schedule a
node before it's ready cycle if the heuristics prefer. This is
essentially chosing to insert a resource stall instead of e.g.
increasing register pressure.
Note that I was initially concerned there might be a correctness aspect
(as in some kind of exposed pipeline design), but the generic scheduler
doesn't seem to know how to insert noop instructions. Without that, a
program wouldn't be guaranteed to schedule on an exposed pipeline
depending on the program and schedule model in question.
The effect of this is that we sometimes prefer register pressure in
codegen results. This is mostly churn (or small wins) on scalar because
we have many more registers, but is of major importance on vector -
particularly high LMUL - because we effectively have many fewer
registers and the relative cost of spilling is much higher. This is a
significant improvement on high LMUL code quality for default rva23u
configurations - or any non -mcpu vector configuration for that matter.
Fixes#107532
This is based on other targets like PPC/AArch64 and some experiments.
This PR will only enable bidirectional scheduling and tracking register
pressure.
Disclaimer: I haven't tested it on many cores, maybe we should make
some options being features. I believe downstreams must have tried
this before, so feedbacks are welcome.
AArch64 has had it enabled since late November, so hopefully the main
issues have been resolved.
I see a small reduction in dynamic instruction count on every benchmark
in specint2017. The best improvement was 0.3% so nothing amazing.
This uses the recently introduced sink-and-fold support in MachineSink.
https://reviews.llvm.org/D152828
This enables folding ADDI into load/store addresses.
Enabling by default will be a separate PR.
Like what has been done in AArch64 (D125335).
We enable this under `-O2` to show the codegen diffs here but we
may only do this under `-O3` like AArch64.
There are two cases that we may produce these eliminable copies:
1. ISel of `FrameIndex`. Like `rvv/fixed-vectors-calling-conv.ll`.
2. Tail duplication. Like `select-optimize-multiple.ll`.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D144535
add can always be compressed to c.add if one of the sources is the
same as the destination.
The same is not true for c.addw where the registers need to be x8-x15.
We have custom isel that tries to select the Lo12 bits using a
separate ADDI that can later folded into the load/store address
by the post-isel peephole.
This patch disables this if the load/store already had a non-zero
offset. A non-zero offset implies that CodeGenPrepare split several
large offsets used by different loads and stores into a common large
offset and multiple small offsets that could be folded. Folding more
of the lo12 bits changes this common offset by increasing the small
offsets. While this can save an instruction to materialize the common
offset, it can also prevent the small offsets from fitting in a
compressed load/store instruction.
Removing this also simplifies the last piece needed to fold the custom
isel for add into SelectAddrRegImm and remove the post-isel peephole.
If the imm is out of range for an ADDI, we will materialize it in
a register using multiple instructions. If the ADD is used by a
load/store, doPeepholeLoadStoreADDI can try to pull an ADDI from
the constant materialization into the load/store offset. This only
works if the ADD has a single use, otherwise the peephole would have
to rebuild multiple nodes.
This patch instead tries to solve the problem when the add is selected.
We check that the add is only used by loads/stores and if it is
we will select it to (ADDI (ADD X, Imm-Lo12), Lo12). This will enable
the simple case in doPeepholeLoadStoreADDI that can bypass an ADDI
used as a pointer. As a result we can remove the more complicated
peephole from doPeepholeLoadStoreADDI.
Reviewed By: reames
Differential Revision: https://reviews.llvm.org/D126576
When looking for memory uses,
reassociationCanBreakAddressingModePattern should check uses of
the outer ADD rather than the inner ADD. We want to know if the
two ops we're reassociating are used by a load/store.
In practice, the existing check usually works because CodeGenPrepare
will make one of the load/stores have an offset of 0 relative to
split GEP. That will make the inner add have a memory use.
To test this, I've manually split the GEPs so there is no 0 offset
store.
This issue was recently discussed in the original review D60294.
Reviewed By: luismarques
Differential Revision: https://reviews.llvm.org/D124644
Function calls and compare instructions tend to cause sext.w
instructions to be inserted. If we make good use of W instructions,
these operations can often end up being redundant. We don't always
detect these during SelectionDAG due to things like phis. There also
some cases caused by failure to turn extload into sextload in
SelectionDAG. extload selects to LW allowing later sext.ws to become
redundant.
This patch adds a pass that examines the input of sext.w instructions trying
to determine if it is already sign extended. Either by finding a
W instruction, other instructions that produce a sign extended result,
or looking through instructions that propagate sign bits. It uses
a worklist and visited set to search as far back as necessary.
Reviewed By: asb, kito-cheng
Differential Revision: https://reviews.llvm.org/D116397
Add an alias of `addi [x], zero, imm` to generate pseudo
instruction li, which makes assembly mush more readable.
For existed tests, users can update them by running script
`llvm/utils/update_llc_test_checks.py`.
Reviewed By: asb
Differential Revision: https://reviews.llvm.org/D112692
We already do this for non-constants RHS. This just removes the
special case. I believe the special case may have been needed
because the ANY_EXTEND of a constant used to create zero extended
constants, but we recently changed that to produce sign extended
constants.
D107658 is needed to prevent some regressions.
Reviewed By: luismarques
Differential Revision: https://reviews.llvm.org/D107697
Summary: Removes CFI CFA directives that could incorrectly propagate
beyond the basic block they were inteded for. Specifically it removes
the epilogue CFI directives. See the branch_and_tail_call test for an
example of the issue. Should fix the stack unwinding issues caused by
the incorrect directives.
Reviewers: asb, lenary, shiva0217
Reviewed By: lenary
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D69723
Summary: Removes CFI CFA directives that could incorrectly propagate
beyond the basic block they were inteded for. Specifically it removes
the epilogue CFI directives. See the branch_and_tail_call test for an
example of the issue. Should fix the stack unwinding issues caused by
the incorrect directives.
Reviewers: asb, lenary, shiva0217
Reviewed By: lenary
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D69723
Most of the test changes are trivial instruction reorderings and differing
register allocations, without any obvious performance impact.
Differential Revision: https://reviews.llvm.org/D66973
llvm-svn: 372106
Some GEPs were not being split, presumably because that split would just be
undone by the DAGCombiner. Not performing those splits can prevent important
optimizations, such as preventing the element indices / member offsets from
being (partially) folded into load/store instruction immediates. This patch:
- Makes the splits also occur in the cases where the base address and the GEP
are in the same BB.
- Ensures that the DAGCombiner doesn't reassociate them back again.
Differential Revision: https://reviews.llvm.org/D60294
llvm-svn: 363544
