Function `replaceMathCmpWithIntrinsic` artificially limits the scope
of the optimization, setting a requirement of two instructions be in
the same block, due to two reasons:
- usage of DT for more general check is costly in terms of compile time;
- risk of creating a new value that lives through multiple blocks.
Because of this, two semantically equivalent tests may be or not be the
subject of this opt depending on where the binary operation is located.
See `test/CodeGen/X86/usub_inc_iv.ll` for motivation
There is one important particular case where this limitation is too strict:
it is when the binary operation is the increment of the induction variable.
As result, the application of this opt becomes fragile and highly reliant on
where other passes decide to place IV increment. In most cases, they place
it in the end of the latch block, killing the opt opportunity (when in fact it
does not matter where to insert the actual instruction).
This patch handles this particular case separately.
- The detector does not use dom tree and has constant cost;
- The value of IV or IV.next lives through all loop in any case, so this should not
create a new unexpected long-living value.
As result, the transform becomes more robust. It also seems to lead to
better code generation in some cases (see `test/CodeGen/X86/lsr-loop-exit-cond.ll`).
Differential Revision: https://reviews.llvm.org/D96119
Reviewed By: spatel, reames
Avoid doing the following combine for vector types:
```
copysign(x, fp_extend(y)) -> copysign(x, y)
copysign(x, fp_round(y)) -> copysign(x, y)
```
That combine seemed to impede the selection of vector instruction and cause
a mess in some circumstances.
Differential Revision: https://reviews.llvm.org/D96037
Fold shufps(hop(x,y),hop(z,w)) -> permute(hop(x,z)) - this is very similar to the equivalent unpack fold.
I did start trying to convert foldShuffleOfHorizOp to handle generic shuffle masks but we're relying on a lot of special cases at the moment.
This reverts commit 0fc1738eb75d613b9e16143b83e7cb80512e84eb.
The test passes (unexpectedly, due to the XFAIL: *) when x86 isn't
the default triple (such as on an arm machine).
Now that PR48908 has been dealt with, we can handle v4f64 permute cases by extracting the low/high lane VPERMILPD masks and creating a new mask based on which lanes are referenced by the VPERM2F128 mask.
GNU ld>=2.36 supports mixed SHF_LINK_ORDER and non-SHF_LINK_ORDER sections in an
output section, so we can set SHF_LINK_ORDER if -fbinutils-version=2.36 or above.
If -fno-function-sections or older binutils, drop unique ID for -fno-unique-section-names.
The users can just specify -fbinutils-version=2.36 or above to allow GC with both GNU ld and LLD.
(LLD does not support garbage collection of non-group non-SHF_LINK_ORDER .gcc_except_table sections.)
This patch adds a pass to replace calls to vector intrinsics
(i.e., LLVM intrinsics operating on vector operands) with
calls to a vector library.
Currently, calls to LLVM intrinsics are only replaced with
calls to vector libraries when scalar calls to intrinsics are
vectorized by the Loop- or SLP-Vectorizer.
With this pass, it is now possible to replace calls to LLVM
intrinsics already operating on vector operands, e.g., if
such code was generated by MLIR. For the replacement,
information from the TargetLibraryInfo, e.g., as specified
via -vector-library is used.
Differential Revision: https://reviews.llvm.org/D95373
The test shows how "usub with overflow" opt gets or does not get applied
to increment of IV depending on its code placement, while two cases are
semantically equivalent.
These tests demonstrate that LSR does not insert IV increment
into the latch block (as it supposes to) when it can use an
existing Phi as IV rather than creating a new LSR IV.
loadRegFromStackSlot()/storeRegToStackSlot() can generate aligned access
instructions for stack slots even if the stack is unaligned, based on the
assumption that the stack can be realigned.
However, this doesn't work for fixed slots, which are e.g. used for
spilling XMM registers in a non-leaf function with
`__attribute__((preserve_all))`.
When compiling such code with `-mstack-alignment=8`, this causes general
protection faults.
Fix it by only considering stack realignment for non-fixed slots.
Note that this changes the output of three existing tests which spill AVX
registers, since AVX requires higher alignment than the ABI provides on
stack frame entry.
Reviewed By: rnk, jyknight
Differential Revision: https://reviews.llvm.org/D73126
The upstream callers (the vectorizers) were fixed with:
bbed5f2f8a04 ( D95690 )
77adbe6a8c71
We should remove this pass entirely now that reduction
legalization/lowering is expected to work just as well,
but we need to confirm that the shuffle ops do not
regress (for x86 in particular).
This should be the last step needed to close:
https://llvm.org/PR23116
Extends D95779 to permit insertion into float/doubles vectors while avoiding a lot of aliased memory traffic.
The scalar value is already on the simd unit, so we only need to transfer and splat the index value, then perform the select.
SSE4 codegen is a little bulky due to the tied register requirements of (non-VEX) BLENDPS/PD but the extra moves are cheap so shouldn't be an actual problem.
Differential Revision: https://reviews.llvm.org/D95866
FixupStatepoints pass does not take into account the undef use
it skips may have a tied def. So when defs are handled pass
considers that tied-use should be spilled and triggers an assert.
FixupStatepoints should skip undef def as well.
Reviewers: reames, dantrushin
Reviewed By: dantrushin
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D95858
With predicate masks, AVX512 can efficiently perform variable-index vector insertion with 2 broadcasts + 1 comparison, avoiding a lot of aliased memory traffic.
Differential Revision: https://reviews.llvm.org/D95779
This is directly analogous to the existing no_caller_saved_registers, but with the opposite intention. A function or call so marked shifts the responsibility of spilling the usual CSRs to it's caller.
An indirect call site and callee which don't agree on the attribute is ill defined.
The motivation for this change is that being able to prune callee saves (without modifying other details of the calling convention) is sometimes useful when generating stubs and adapters. There's no intention to expose this as a source language feature; this is expected to be used by frontends to implement adapters where warranted.
Some specific examples of use cases:
* GC compatible compiled code wants to call an externally defined library function without needing to track pointer values through CSRs.
* debug enabled code wants to call precompiled library which doesn't provide enough information to track CSRs while preserving debug quality in caller.
* adapter stub entering hand written assembler which doesn't follow normal calling conventions.
We can only legally extract from the lowest 128-bit subvector, so extract the correct subvector to allow us to handle 256/512-bit vector element extracts.
This is an optimized approach for D94155.
Previous code build the model that tile config register is the user of
each AMX instruction. There is a problem for the tile config register
spill. When across function, the ldtilecfg instruction may be inserted
on each AMX instruction which use tile config register. This cause all
tile data register clobber.
To fix this issue, we remove the model of tile config register. Instead,
we analyze the AMX instructions between one call to another. We will
insert ldtilecfg after the first call if we find any AMX instructions.
Reviewed By: LuoYuanke
Differential Revision: https://reviews.llvm.org/D95136
Source Drift happens when the sources are updated after profiling the binary
but before building the final optimized binary. If the source has changed since
the profiles were obtained, optimizing basic blocks might be sub-optimal. This
only applies to BasicBlockSection::List as it creates clusters of basic blocks
using basic block ids. Source drift can invalidate these groupings leading to
sub-optimal code generation with regards to performance.
PGO source drift for a particular function can be detected using function
metadata added in D95495.
When source drift is deected, disable basic block clusters by default
which can be re-enabled with -mllvm option
bbsections-detect-source-drift=false.
Differential Revision: https://reviews.llvm.org/D95593
If the shuffle mask can't be widened to match the original extracted element width, see if the upper bits are zeroable - which allows us to extract+zero-extend the smaller extraction.
This de-pessimizes the arguably more usual case of no masked mem intrinsics,
and gets rid of one more Dominator Tree recalculation.
As per llvm/test/CodeGen/X86/opt-pipeline.ll,
there's one more Dominator Tree recalculation left, we could get rid of.
Unlike VPERMILPS, VPERMILPD can have non-repeating masks in each 128-bit subvector, we weren't accounting for this when folding vperm2f128(vpermilpd(x,c),vpermilpd(y,c)) -> vpermilpd(vperm2f128(x,y),c).
I'm intending to add support for this but wanted to get a minimal fix in first for merging into 12.xx.
Fixes PR48908
This fully de-pessimizes the common case of no indirectbr's,
(where we don't actually need to do anything to preserve domtree)
and avoids domtree recomputation in the case there were indirectbr's.
Note that two indirectbr's could have a common successor, and not all
successors of an indirectbr's are meant to survive the expansion.
Though, the code assumes that an indirectbr's doesn't have
duplicate successors, those *should* have been deduplicated
by simplifycfg or something already.
This moves SinkIntoLoop from MachineLICM to MachineSink. The motivation for
this work is that hoisting is a canonicalisation transformation, but we do not
really have a good story to sink instructions back if that is better, e.g. to
reduce live-ranges, register pressure and spilling. This has been discussed a
few times on the list, the latest thread is:
https://lists.llvm.org/pipermail/llvm-dev/2020-December/147184.html
There it was pointed out that we have the LoopSink IR pass, but that works on
IR, lacks register pressure informatiom, and is focused on profile guided
optimisations, and then we have MachineLICM and MachineSink that both perform
sinking. MachineLICM is more about hoisting and CSE'ing of hoisted
instructions. It also contained a very incomplete and disabled-by-default
SinkIntoLoop feature, which we now move to MachineSink.
Getting loop-sinking to do something useful is going to be at least a 3-step
approach:
1) This is just moving the code and is almost a NFC, but contains a bug fix.
This uses helper function `isLoopInvariant` that was factored out in D94082 and
added to MachineLoop.
2) A first functional change to make loop-sink a little bit less restrictive,
which it really is at the moment, is the change in D94308. This lets it do
more (alias) analysis using functions in MachineSink, making it a bit more
powerful. Nothing changes much: still off by default. But it shows that
MachineSink is a better home for this, and it starts using its functionality
like `hasStoreBetween`, and in the next step we can use `isProfitableToSinkTo`.
3) This is the going to be he interesting step: decision making when and how
many instructions to sink. This will be driven by the register pressure, and
deciding if reducing live-ranges and loop sinking will help in better
performance.
4) Once we are happy with 3), this should be enabled by default, that should be
the end goal of this exercise.
Differential Revision: https://reviews.llvm.org/D93694
Revert the change to use APInt::isSignedIntN from
5ff5cf8e057782e3e648ecf5ccf1d9990b53ee90.
Its clear that the games we were playing to avoid the topological
sort aren't working. So just fix it once and for all.
Fixes PR48888.
There are two use cases.
Assembler
We have accrued some code gated on MCAsmInfo::useIntegratedAssembler(). Some
features are supported by latest GNU as, but we have to use
MCAsmInfo::useIntegratedAs() because the newer versions have not been widely
adopted (e.g. SHF_LINK_ORDER 'o' and 'unique' linkage in 2.35, --compress-debug-sections= in 2.26).
Linker
We want to use features supported only by LLD or very new GNU ld, or don't want
to work around older GNU ld. We currently can't represent that "we don't care
about old GNU ld". You can find such workarounds in a few other places, e.g.
Mips/MipsAsmprinter.cpp PowerPC/PPCTOCRegDeps.cpp X86/X86MCInstrLower.cpp
AArch64 TLS workaround for R_AARCH64_TLSLD_MOVW_DTPREL_* (PR ld/18276),
R_AARCH64_TLSLE_LDST8_TPREL_LO12 (https://bugs.llvm.org/show_bug.cgi?id=36727https://sourceware.org/bugzilla/show_bug.cgi?id=22969)
Mixed SHF_LINK_ORDER and non-SHF_LINK_ORDER components (supported by LLD in D84001;
GNU ld feature request https://sourceware.org/bugzilla/show_bug.cgi?id=16833 may take a while before available).
This feature allows to garbage collect some unused sections (e.g. fragmented .gcc_except_table).
This patch adds `-fbinutils-version=` to clang and `-binutils-version` to llc.
It changes one codegen place in SHF_MERGE to demonstrate its usage.
`-fbinutils-version=2.35` means the produced object file does not care about GNU
ld<2.35 compatibility. When `-fno-integrated-as` is specified, the produced
assembly can be consumed by GNU as>=2.35, but older versions may not work.
`-fbinutils-version=none` means that we can use all ELF features, regardless of
GNU as/ld support.
Both clang and llc need `parseBinutilsVersion`. Such command line parsing is
usually implemented in `llvm/lib/CodeGen/CommandFlags.cpp` (LLVMCodeGen),
however, ClangCodeGen does not depend on LLVMCodeGen. So I add
`parseBinutilsVersion` to `llvm/lib/Target/TargetMachine.cpp` (LLVMTarget).
Differential Revision: https://reviews.llvm.org/D85474
The unwind info format requires that all adjustments are 8 byte aligned,
and the bottom three bits are masked out. Most Win64 calling conventions
have 32 bytes of shadow stack space for spilling parameters, and I
believe that constructing these fixed stack objects had the side effect
of ensuring an alignment of 8. However, the Intel regcall convention
does not have this shadow space, so when using that convention, it was
possible to make a 4 byte stack frame, which was impossible to describe
with unwind info.
Fixes pr48867
combineX86ShufflesConstants/canonicalizeShuffleMaskWithHorizOp can both handle/earlyout shuffles with inputs of different widths, so delay widening as late as possible to make it easier to match constant folds etc.
The plan is to eventually move the widening inside combineX86ShuffleChain so that we don't create any new nodes unless we successfully combine the shuffles.
rGbe69e66b1cd8 added the fold, but DAGCombiner.visitVECTOR_SHUFFLE doesn't merge shuffles if the inner shuffle is a splat, so we need to bail.
The non-fast-horiz-ops paths see some minor regressions, we might be able to improve on this after lowering to target shuffles.
Fix PR48823
