This fixes the handling of subregister extract copies. This
will allow AMDGPU to remove its implementation of
shouldRewriteCopySrc, which exists as a 10 year old workaround
to this bug. peephole-opt-fold-reg-sequence-subreg.mir will
show the expected improvement once the custom implementation
is removed.
The copy coalescing processing here is overly abstracted
from what's actually happening. Previously when visiting
coalescable copy-like instructions, we would parse the
sources one at a time and then pass the def of the root
instruction into findNextSource. This means that the
first thing the new ValueTracker constructed would do
is getVRegDef to find the instruction we are currently
processing. This adds an unnecessary step, placing
a useless entry in the RewriteMap, and required skipping
the no-op case where getNewSource would return the original
source operand. This was a problem since in the case
of a subregister extract, shouldRewriteCopySource would always
say that it is useful to rewrite and the use-def chain walk
would abort, returning the original operand. Move the process
to start looking at the source operand to begin with.
This does not fix the confused handling in the uncoalescable
copy case which is proving to be more difficult. Some currently
handled cases have multiple defs from a single source, and other
handled cases have 0 input operands. It would be simpler if
this was implemented with isCopyLikeInstr, rather than guessing
at the operand structure as it does now.
There are some improvements and some regressions. The
regressions appear to be downstream issues for the most part. One
of the uglier regressions is in PPC, where a sequence of insert_subrgs
is used to build registers. I opened #125502 to use reg_sequence instead,
which may help.
The worst regression is an absurd SPARC testcase using a <251 x fp128>,
which uses a very long chain of insert_subregs.
We need improved subregister handling locally in PeepholeOptimizer,
and other pasess like MachineCSE to fix some of the other regressions.
We should handle subregister composes and folding more indexes
into insert_subreg and reg_sequence.
As of rev ea222be0d, LLVMs assembler will actually try to honour the
"fill value" part of p2align directives. X86 printed these as 0x90, which
isn't actually what it wanted: we want multi-byte nops for .text
padding. Compiling via a textual assembly file produces single-byte
nop padding since ea222be0d but the built-in assembler will produce
multi-byte nops. This divergent behaviour is undesirable.
To fix: don't set the byte padding field for x86, which allows the
assembler to pick multi-byte nops. Test that we get the same multi-byte
padding when compiled via textual assembly or directly to object file.
Added same-align-bytes-with-llasm-llobj.ll to that effect, updated
numerous other tests to not contain check-lines for the explicit padding.
This stops reporting CostPerUse 1 for `R8`-`R15` and `XMM8`-`XMM31`.
This was previously done because instruction encoding require a REX
prefix when using them resulting in longer instruction encodings. I
found that this regresses the quality of the register allocation as the
costs impose an ordering on eviction candidates. I also feel that there
is a bit of an impedance mismatch as the actual costs occure when
encoding instructions using those registers, but the order of VReg
assignments is not primarily ordered by number of Defs+Uses.
I did extensive measurements with the llvm-test-suite wiht SPEC2006 +
SPEC2017 included, internal services showed similar patterns. Generally
there are a log of improvements but also a lot of regression. But on
average the allocation quality seems to improve at a small code size
regression.
Results for measuring static and dynamic instruction counts:
Dynamic Counts (scaled by execution frequency) / Optimization Remarks:
Spills+FoldedSpills -5.6%
Reloads+FoldedReloads -4.2%
Copies -0.1%
Static / LLVM Statistics:
regalloc.NumSpills mean -1.6%, geomean -2.8%
regalloc.NumReloads mean -1.7%, geomean -3.1%
size..text mean +0.4%, geomean +0.4%
Static / LLVM Statistics:
mean -2.2%, geomean -3.1%) regalloc.NumSpills
mean -2.6%, geomean -3.9%) regalloc.NumReloads
mean +0.6%, geomean +0.6%) size..text
Static / LLVM Statistics:
regalloc.NumSpills mean -3.0%
regalloc.NumReloads mean -3.3%
size..text mean +0.3%, geomean +0.3%
Differential Revision: https://reviews.llvm.org/D133902
Test updates were performed using:
https://gist.github.com/nikic/98357b71fd67756b0f064c9517b62a34
These are only the test updates where the test passed without
further modification (which is almost all of them, as the backend
is largely pointer-type agnostic).
To generate zero value, the PXOR instruction need 3 operands that is
tied to the same vreg. If is not good in SSA form and with undef value
two address instruction pass may convert
`%0:vr128 = PXORrr undef %0, undef %0`
to `%1:vr128 = PXORrr undef %1:vr128(tied-def 0), undef %0:vr128`.
It is not expected.
It can be simplified to SET0 instruction which only take 1 destination
operand. It should be more friendly to two address instruction pass and
register allocation pass.
`%0:vr128 = V_SET0`
Also add AVX1 code path so that it is consistant to other code.
Differential Revision: https://reviews.llvm.org/D124903
This is a follow up of D99010. We didn't consider the live range of shape registers when hoist ldtilecfg. There maybe risks, e.g. we happen to insert it to an invalid range of some registers and get unexpected error.
This patch fixes this problem by storing the value to corresponding stack place of ldtilecfg after all its definition immediately.
This patch also fix a problem in previous code: If we don't have a ldtilecfg which dominates all AMX instructions, we cannot initialize shapes for other ldtilecfg.
There're still some optimization points left. E.g. eliminate unused mov instructions, break the def-use dependency before RA etc.
Reviewed By: LuoYuanke, xiangzhangllvm
Differential Revision: https://reviews.llvm.org/D99966
The previous code calculated the first ldtilecfg by dominating all AMX registers' def. This may result in the ldtilecfg being inserted into a loop.
This patch try to calculate the nearest point where all shapes of AMX registers are reachable.
Reviewed By: LuoYuanke
Differential Revision: https://reviews.llvm.org/D99010
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
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. We
analyze the regmask of call instruction and insert ldtilecfg if there is
any tile data register live across the call. Inserting the sttilecfg
before the call is unneccessary, because the tile config doesn't change
and we can just reload the config.
Besides we also need check tile config register interference. Since we
don't model the config register we should check interference from the
ldtilecfg to each tile data register def.
ldtilecfg
/ \
BB1 BB2
/ \
call BB3
/ \
%1=tileload %2=tilezero
We can start from the instruction of each tile def, and backward to
ldtilecfg. If there is any call instruction, and tile data register is
not preserved, we should insert ldtilecfg after the call instruction.
Differential Revision: https://reviews.llvm.org/D94155
