This is a more general version of D109273. Though it doesn't
peek through bitcasts or rearange broadcasts.
Reviewed By: LuoYuanke
Differential Revision: https://reviews.llvm.org/D109295
d8faf03807 implemented general-regs-only for X86 by disabling all features
with vector instructions. But the CRC32 instruction in SSE4.2 ISA, which uses
only GPRs, also becomes unavailable. This patch adds a CRC32 feature for this
instruction and allows it to be used with general-regs-only.
Reviewed By: pengfei
Differential Revision: https://reviews.llvm.org/D105462
The xmm variant have half the throughput (and +1cy latency) of the mmx variants, but are still 1uop.
I still need to do more thorough testing of SLM on test-suite before fixing the obvious bad numbers for WritePMULLD.
But this helps the D103695 helper script get to more accurate numbers for vXi32 multiplies of extended operands (i.e. we can use PMADDWD, PMULLW/PMULHW etc). Matches what Intel AoM / Agner / llvm-exegesis reports.
The xmm variant have half the throughput (and +1cy latency) of the mmx variants, but are still 1uop.
I still need to do more thorough testing of SLM on test-suite before fixing the obvious bad numbers for WritePMULLD.
But this helps the D103695 helper script get to more accurate numbers for vXi32 multiplies of extended operands (i.e. we can use PMADDWD, PMULLW/PMULHW etc). Matches what Intel AoM / Agner / llvm-exegesis reports.
For RMW instructions, the load and store hold the MEC for an extra cycle, but within the same single uop. This is alluded to in the Intel AOM:
"The MEC also owns the MEC RSV, which is responsible for scheduling of all loads and stores. Load and
store instructions go through addresses generation phase in program order to avoid on-the-fly memory
ordering later in the pipeline. Therefore, an unknown address will stall younger memory instructions."
Noticed while trying to get a cheap SLM test box up and running with llvm-exegesis - RMW arithmetic is always 1uop - and matches what Agner / InstLatX64 report as well.
These were all set to the same best case mul i32 values (which seems to be the only version of MUL that SLM actually performs well with).
Noticed while trying to improve multiplication costs for vectorization via the D103695 helper script. Confirmed with Intel AoM / Agner / InstLatX64.
SLM PBLENDVB is just as bad as BLENDVPD/PS - so model it as such, fixing the rr vs rm uops diff as well. The Intel AoM appears to have a copy+paste typo with PBLENDW, it doesn't match Agner or InstLatX64.
Noticed while investigating some of the weird discrepancies reported by the D103695 helper script (SLM had much better vector shift throughputs than it should).
Fixing this link error:
ld: error: relocation R_X86_64_PC32 cannot be used against symbol __asan_report_load...; recompile with -fPIC
Reviewed By: vitalybuka
Differential Revision: https://reviews.llvm.org/D109183
https://reviews.llvm.org/D56686 was supposed to allow these to
work on Windows without needing to enable the xsave feature to
match MSVC. It seems this didn't work because the backend isel
patterns would still block it.
This patch removes the predicates from the isel patterns.
Fixes PR51706.
Reviewed By: pengfei
Differential Revision: https://reviews.llvm.org/D109097
PMADDWD(v8i16 x, v8i16 y) == (v4i32) { (int)x[0]*y[0] + (int)x[1]*y[1], ..., (int)x[6]*y[6] + (int)x[7]*y[7] }
Currently combineMulToPMADDWD only folds cases where the upper 17 bits of both vXi32 inputs are known zero (i.e. the first half is positive and the second half of the pair is zero in each 2xi16 pair), this can be relaxed to only require one zero-extended input if the other input has at least 17 sign bits.
That way the sign of the result is still preserved, and the second half is still zero.
Noticed while investigating PR47437.
Differential Revision: https://reviews.llvm.org/D108522
Please refer to
https://lists.llvm.org/pipermail/llvm-dev/2021-September/152440.html
(and that whole thread.)
TLDR: the original patch had no prior RFC, yet it had some changes that
really need a proper RFC discussion. It won't be productive to discuss
such an RFC, once it's actually posted, while said patch is already
committed, because that introduces bias towards already-committed stuff,
and the tree is potentially in broken state meanwhile.
While the end result of discussion may lead back to the current design,
it may also not lead to the current design.
Therefore i take it upon myself
to revert the tree back to last known good state.
This reverts commit 4c4093e6e3.
This reverts commit 0a2b1ba33a.
This reverts commit d9873711cb.
This reverts commit 791006fb8c.
This reverts commit c22b64ef66.
This reverts commit 72ebcd3198.
This reverts commit 5fa6039a5f.
This reverts commit 9efda541bf.
This reverts commit 94d3ff09cf.
Icelake, Rocketlake and Tigerlake targets currently use the SkylakeServer scheduler model, despite being a later microarchitecture, leading to both reported bugs (PR48110) and discrepancies when comparing llvm-mca reports to other profiling tools (OSACA, uops, uica, etc.). And tbh I'm getting sick of llvm-mca getting blamed for what are backend scheduler model issues :-(
This patch doesn't attempt to fix any of these discrepancies - there should be no changes in codegen - its a setup patch that copies the skx model, renames all the resources, adds the additional ports (but doesn't reference them yet) and updates the llvm-exegesis pfm counter mappings (based off https://sourceforge.net/p/perfmon2/libpfm4/ci/master/tree/lib/events/intel_icl_events.h).
This should make it trivial for anyone with hardware access to use llvm-exegesis reports to iteratively improve the model (my attempts to get hold of a cheap tiger lake box haven't been fruitful yet....).
I will copy the SkylakeServer llvm-mca resource tests as follow up commits - the diff should entirely be the resource renames.
Differential Revision: https://reviews.llvm.org/D108914
The backend generally uses 64-bit immediates (e.g. what
MachineOperand::getImm() returns), so use that for analyzeCompare()
and optimizeCompareInst() as well. This avoids truncation for
targets that support immediates larger 32-bit. In particular, we
can avoid the bugprone value normalization hack in the AArch64
target.
This is a followup to D108076.
Differential Revision: https://reviews.llvm.org/D108875
The GLM/SLM special cases still get tested first but after the the MUL/DIV/REM pattern detection - this will be necessary for when we make the SLM vXi32 MUL canonicalization generic to improve PMULLW/PMULHW/PMADDDW cost support etc.
Remove method X86LowerAMXType::getRowFromCol since it's not used, and
it's causing a warning.
Reviewed By: LuoYuanke
Differential Revision: https://reviews.llvm.org/D108862
This fixes another reproducer from https://bugs.llvm.org/show_bug.cgi?id=51615
And again, the fix lies not in the code added in D105390
In this case, we completely don't check that the "broadcast-from-mem" we create
can actually fold the load. In this case, it's operand was not a load at all:
```
Combining: t16: v8i32 = vector_shuffle<0,u,u,u,0,u,u,u> t14, undef:v8i32
Creating new node: t29: i32 = undef
RepeatLoad:
t8: i32 = truncate t7
t7: i64 = extract_vector_elt t5, Constant:i64<0>
t5: v2i64,ch = load<(load (s128) from %ir.arg)> t0, t2, undef:i64
t2: i64,ch = CopyFromReg t0, Register:i64 %0
t1: i64 = Register %0
t4: i64 = undef
t3: i64 = Constant<0>
Combining: t15: v8i32 = undef
```
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D108821
x86_fp80 format allows values that do not fit any of IEEE-754 category.
Previously they were recognized by intrinsic __builtin_isnan as NaNs.
Now this intrinsic is implemented using instruction FXAM, which
distinguish between NaNs and unsupported values. It can make some
programs behave differently.
As a solution, this fix changes lowering of the intrinsic. If floating
point exceptions are ignored, llvm.isnan is lowered into unordered
comparison, as __buildtin_isnan was implemented earlier. In strictfp
functions the intrinsic is lowered using FXAM, which does not raise
exceptions even for signaling NaN, as required by IEEE-754 and C
standards.
Differential Revision: https://reviews.llvm.org/D108037
There are no paths into LowerFormalParms that have already specified
the sret register. We always materialize a virtual and then assign it
to the physical reg at the point of the return.
Differential Revision: https://reviews.llvm.org/D108762
Exegesis is faulty and sometimes when measuring throughput^-1
produces snippets that have loop-carried dependencies,
which must be what caused me to incorrectly measure it originally.
After looking much more carefully, the inverse throughput should match
that of the MULX w/ reg op.
As per llvm-exegesis measurements.
Looks like the NoRegister has some effect on the final code that is generated. My guess is that some optimization kicks in at the end?
When I use -S to dump the assembly I get the correct version with 'shrq $3, %r8':
movq %r9, %r8
shrq $3, %r8
movsbl 2147450880(%r8), %r8d
But, when I disassemble the final binary I get RAX in stead of R8:
mov %r9,%r8
shr $0x3,%rax
movsbl 0x7fff8000(%r8),%r8d
Reviewed By: vitalybuka
Differential Revision: https://reviews.llvm.org/D108745
Even though https://bugs.llvm.org/show_bug.cgi?id=51615
appears to be introduced by D105390, the fix lies here.
We can not replace a wide volatile load with a broadcast-from-memory,
because that would narrow the load, which isn't legal for volatiles.
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D108757
It turns out that SchedWrite WriteIMulH was always assigned to the low half of
the result of a MULX (rather than to the high half).
To avoid confusion, this patch swaps the two MULX writes in the tablegen
definition of MULX32/64. That way, write names better describe what they
actually refer to; this also avoids further complications if in future we decide
to reuse the same MulH writes to also model other scalar integer multiply
instructions. I also had to swap the latency values for the two MULX writes to
make sure that the change is effectively an NFC. In fact, none of the existing
x86 tests were affected by this small refactoring.
This patch also fixes a bug in MCA: a wrong latency value was propagated for
instructions that perform multiple writes to a same register. This last issue
was found by Roman while testing MULX on targets that define a different latency
for the Low/High part of the result.
Differential Revision: https://reviews.llvm.org/D108727
In-register structure returns are not special, and handled by lowering
to multiple-value tuples. We can tail-call from non-sret fns to
structure-returning functions, except on i686 where the sret pointer
is callee-pop.
Differential Revision: https://reviews.llvm.org/D105807
The implementation uses the int_asan_check_memaccess intrinsic to instrument the code. The intrinsic is replaced by a call to a function which performs the access check. The generated function names encode the input register name as a number using Reg - X86::NoRegister formula.
Reviewed By: vitalybuka
Differential Revision: https://reviews.llvm.org/D107850
Before this patch, WriteIMulH reported a latency value which is correct for the
RR variant of MULX, but not for the RM variant.
This patch fixes the issue by introducing a new WriteIMulHLd, which is meant to
be used only by the RM variant of MULX.
Differential Revision: https://reviews.llvm.org/D108701
Patch 9588b685c6 introduced dependency on ASAN. But it didn't
explicitly put LLVMInstrumentation as one of the library dependencies
such that the build will fail if we're building LLVM as shared libraries
(i.e. -DBUILD_SHARED_LIBS=ON).
This patch explicitly links X86CodeGen against the Instrumentation
component.
Differential Revision: https://reviews.llvm.org/D108662
The implementation uses the int_asan_check_memaccess intrinsic to instrument the code. The intrinsic is replaced by a call to a function which performs the access check. The generated function names encode the input register name as a number using Reg - X86::NoRegister formula.
Reviewed By: vitalybuka
Differential Revision: https://reviews.llvm.org/D107850
We don't have any vXi8 shift instructions (other than on XOP which is handled separately), so replace the shl(x,1) -> add(x,x) fold with shl(x,1) -> add(freeze(x),freeze(x)) to avoid the undef issues identified in PR50468.
Split off from D106675 as I'm still looking at whether we can fix the vXi16/i32/i64 issues with the D106679 alternative.
Differential Revision: https://reviews.llvm.org/D108139
Over in D105657, we started dropping instruction numbers (that become
variable locations) from call instructions, as we can't correctly represent
the x87 FP stack. Unfortunately, it turns out that the "special FP
instructions" that this pass transforms includes "every call instruction"
[0]. Thus, we've ended up dropping all return values from all calls. Ouch.
This patch adds a filter: only drop instruction numbers from calls if they
return something on the FP stack. Seeing how LLVM only allows a single
return value, this should drop instruction numbers on anything that returns
a float, and nothing else.
Rather than writing a new test, I've modified the original one to have a
positive and negative case: drop instruction number on a call with an
FP-stack modification, keep it on a plain call.
Differential Revision: https://reviews.llvm.org/D108580
Intended to be NFC. ARM/AArch64 don't appear to need adjustment.
TargetMachine::shouldAssumeDSOLocal is expected to be very simple, ideally
matching isDSOLocal(). The IR producers are expected to set dso_local correctly.
(While some may think this function can make producers' work easier, the
function is really not in a good position to set dso_local. See the various
special cases we duplicate from clang CodeGenModule.cpp.)
Reviewed By: mstorsjo
Differential Revision: https://reviews.llvm.org/D108514
Extend matchShuffleAsBlend to not only match against known in-place elements for BLEND shuffles, but use isElementEquivalent to determine if the shuffle mask's referenced element is the same as the in-place element.
This allows us to replace a number of insertps instructions with more general blendps instructions (better opportunities for commutation, concatenation etc.).
Before lowering shuffles, see if we can merge horizontal ops or canonicalize the shuffle mask to point to the same LHS/RHS of the HOps when an HOp's args are repeated.
Broadwell is mainly a die shrink of Haswell, but the model had many of the scheduling classes in different orders, making side-by-side comparisons very difficult.
The InstRW overrides are still quite different, but at least that part of the side-by-side diff is now in the same position.
This was noticed while I was trying to investigate diffs between llvm-mca and other perf analyzers in https://uica.uops.info/ - we used to be able to do diffs between most of the models very easily, but we seem to have lost that simplicity as classes have been altered, models have been refined and other models have rotted.
Craig Topper