This suppresses warning `unused variable 'MRI' [-Wunused-variable]`
for those fusions that don't need `MRI`.

Add rounding mode attribute to `arith`. This attribute can be used in
different FP `arith` operations to control rounding mode. Rounding modes
correspond to IEEE 754-specified rounding modes. Use in `arith.truncf` folding.

As this is not supported in dialects other than LLVM, conversion should fail for
now in case this attribute is present.

---------

Signed-off-by: Victor Perez <victor.perez@codeplay.com>

1. Remove 'AllModes' and 'DefaultMode' suffixes for DecoderTables under
default HwMode.
2. Introduce a less aggressive suppression for HwMode DecoderTable, only
reduce necessary tables duplications. This allows encodings under
different HwModes to retain the original DecoderNamespace.
3. Change 'suppress-per-hwmode-duplicates' command option from bool type
to enum type, allowing users to choose what level of suppression to use.

This patch reorder the lexical block for the translation unit, visible update block for the TU and
the viisble upaete block for the extern C context after the type decl
offsets block.

This should be a NFC patch.

This is helpful for later optimizations for eliding unreachable
declarations in the global module fragment. See the comments in
https://github.com/llvm/llvm-project/pull/76930.

Simply, if we want to get the reachable sets of declaratins during the
writing process, we need to write the file-level context later than the
process of writing declarations (which is the main process to determine
the reachable set).

`RegisterClassInfo::getRegPressureSetLimit` has been changed to return a
smaller value than before so the limit may become negative in later
calculations. As a workaround, change to use
`TargetRegisterInfo::getRegPressureSetLimit`.
Also improve tests.

Fixes: #85243.

-- Convert `math.fpowi` to a series of `arith.mulf` operations.
-- If the power is negative, we divide the result by 1.

Intrinsic declared to have sideeffects, but it's done only to prevent
moving it. Removing unused ones is OK.

Exacted from #84850 for easier review.

The exposition-only type trait `pair-like` includes `ranges::subrange`,
but in every single case excludes `ranges::subrange` from the list. This
patch introduces two new traits `__tuple_like_no_subrange` and
`__pair_like_no_subrange`, which exclude `ranges::subrange` from the
possible matches. `__pair_like` is no longer required, and thus removed.
`__tuple_like` is implemented as `__tuple_like_no_subrange` or a
`ranges::subrange` specialization.

This is recommitted as the test and fix for
llvm.vector.reduce.fmaximum/fminimum are covered in #80553 and #80697

Follow up to #86049.
clang-armv8-quick build bot can trigger this branch.

RFC:
https://discourse.llvm.org/t/rfc-add-llvm-experimental-hot-intrinsic-or-llvm-hot/77641

RFC:
https://discourse.llvm.org/t/rfc-add-llvm-experimental-hot-intrinsic-or-llvm-hot/77641

These intrinsics will not be lowered to code.

RFC:
https://discourse.llvm.org/t/rfc-add-llvm-experimental-hot-intrinsic-or-llvm-hot/77641

The earlier implementation on AMDGPU used explicit token operands at
SI_CALL and SI_CALL_ISEL. This is now replaced with CONVERGENCECTRL_GLUE
operands, with the following effects:

- The treatment of tokens at call-like operations is now consistent with
the treatment at intrinsics.
- Support for tail calls using implicit tokens at SI_TCRETURN "just
works".
- The extra parameter at call-like instructions is eliminated, thus
restoring those instructions and their handling to the original state.

The new glue node is placed after the existing glue node for the
outgoing call parameters, which seems to not interfere with selection of
the call-like nodes.

RFC:
https://discourse.llvm.org/t/rfc-add-llvm-experimental-hot-intrinsic-or-llvm-hot/77641

`RHS` can be shorter than `LHS`.

Reported by asan after #83440.

The goal is to have ability to change logic compile time based on PGO
data.

Our primary application is removing UBSAN checks from hot code.
Then we'd like to use this for libc++ hardening and regular debug
asserts.
Previous attempt is #84214.

Benefits from special intrinsic vs #84214:
1. Resulting binary is 3% faster than removing traps (on
"test-suite/MultiSource/Benchmarks" with PGO+ThinLTO)
2. Intrinsic can be used from source code to change behavior from C/C++
program. E.g. enabling asserts in cold code.
3. Easier to match basic blocks.

RFC:
https://discourse.llvm.org/t/rfc-add-llvm-experimental-hot-intrinsic-or-llvm-hot/77641



---------

Co-authored-by: Nikita Popov <npopov@redhat.com>

Fixes: https://github.com/llvm/llvm-project/issues/62252

Depends on: #73667

Reverts llvm/llvm-project#79489

We know that the issues was with asan/annotations. We can revert it.

[msan] Nit: use init_origins parameter instead of __msan_get_track_origins() in InitShadowWithReExec (#86994)

This fixes a nit I had accidentally introduced in
https://github.com/llvm/llvm-project/pull/85142

I don't think the value of __msan_get_track_origins() will change
between the start and end of InitShadowWithReExec, but it's cleaner to
use the parameter.

- The raw profile format change is in https://github.com/llvm/llvm-project/pull/82711

This reverts commit d2b63ed0. CI again.
Leading theory is bad interaction between cmake and buildbot, some guesses at PR86737

Both `Transpose` and `Broadcast` specify the `SameVariadicOperandSize`
trait. However neither has a variadic operand let alone more than one.

This is likely a relic from copying the boilerplate of the `Reduce`
definition.

Reverts llvm/llvm-project#86991

Some bots are broken with a leak being detected now.

This speeds up registered op creation by 10-11% by allowing lookup by
TypeID instead of StringRef.

This can break your build/tests at runtime with an error that you're creating
an unregistered operation that you have registered. If so you are likely using
a class inheriting from the "real" operation. See for example in this patch the
case of:

  class ConstantIndexOp : public arith::ConstantOp {


If one is using `builder.create<ConstantIndexOp>()` they actually create an
`arith.constant` operation, but the builder will fetch the TypeID for 
the `ConstantIndexOp` class which does not correspond to any registered
operation. To fix it the `ConstantIndexOp` class got this addition:

  static ::mlir::TypeID resolveTypeID() { return TypeID::get<ConstantOp>(); }

We have the ability to load .dwp files with a .debug_info.dwo section
that exceeds 4GB. There were 4 locations that were using 32 bit offsets
and lengths to extract variable locations, and if a DIE was over the 4GB
barrier, we would truncate the block offset for the variable locations
and the variable expression would be garbage. This fixes the issues. It
isn't possible to add a test for this as we don't want to create a 4GB
.dwp file on test machines.

For more context on isa predicates, see:
https://github.com/llvm/llvm-project/pull/83753.

This reverts commit a2982a29.
Buildbot failure that motivated the revert looks unrelated.

It was shown experimentally that this may have some benefit on newer HW.

Fixes BUILD_SHARED_LIBS builds after #87191 made helpers non-inline.

An inverted condition causes `SymbolFileDWARFDebugMap::FindTypes` to
bail out after inspecting the first .o file in each module.

The same kind of bug is found in
`SymbolFileDWARFDebugMap::ParseDeclsForContext`.

Correct both early exit conditions and add a regression test for lookup
of up a type defined in a secondary compilation unit.

Fixes #87176