r280832 added 32-bit support for emitting conditional tail-calls, but
dropped imp-used parameter registers. This went unnoticed until
r281113, which added 64-bit support, as this is only exposed with
parameter passing via registers.
Don't drop the imp-used parameters.
llvm-svn: 281223
Trying to infer the 'returned' attribute if an argument is already
'returned' can lead to verification failure: inference might determine
that a different argument is passed through which would result in two
different arguments marked as 'returned'.
This fixes PR30350.
llvm-svn: 281221
Summary: This removes disabled instructions from match tables so we will not match them at all.
Reviewers: tstellarAMD, vpykhtin, artem.tamazov
Subscribers: wdng, nhaehnle, arsenm
Differential Revision: https://reviews.llvm.org/D24452
llvm-svn: 281216
For the common pattern (CMPZ (AND x, #bitmask), #0), we can do some more efficient instruction selection if the bitmask is one consecutive sequence of set bits (32 - clz(bm) - ctz(bm) == popcount(bm)).
1) If the bitmask touches the LSB, then we can remove all the upper bits and set the flags by doing one LSLS.
2) If the bitmask touches the MSB, then we can remove all the lower bits and set the flags with one LSRS.
3) If the bitmask has popcount == 1 (only one set bit), we can shift that bit into the sign bit with one LSLS and change the condition query from NE/EQ to MI/PL (we could also implement this by shifting into the carry bit and branching on BCC/BCS).
4) Otherwise, we can emit a sequence of LSLS+LSRS to remove the upper and lower zero bits of the mask.
1-3 require only one 16-bit instruction and can elide the CMP. 4 requires two 16-bit instructions but can elide the CMP and doesn't require materializing a complex immediate, so is also a win.
llvm-svn: 281215
If a constant is unamed_addr and is only used within one function, we can save
on the code size and runtime cost of an indirection by changing the global's storage
to inside the constant pool. For example, instead of:
ldr r0, .CPI0
bl printf
bx lr
.CPI0: &format_string
format_string: .asciz "hello, world!\n"
We can emit:
adr r0, .CPI0
bl printf
bx lr
.CPI0: .asciz "hello, world!\n"
This can cause significant code size savings when many small strings are used in one
function (4 bytes per string).
llvm-svn: 281213
Summary:
This test was not testing the intrinsics. A function like this:
define %v4f32 @test_v4f32.floor(%v4f32 %a){
...
%1 = call %v4f32 @llvm.floor.v4f32(%v4f32 %a)
...
}
is transformed into the following assembly:
_test_v4f32.floor: @ @test_v4f32.floor
...
bl _floorf
...
In each function tested, there are two CHECK: one that checked
for the label and another one for the intrinsic that should be used
inside the function (in our case, "floor"). However, although the
first CHECK was matching the label, the second was not matching the
intrinsic, but the second "floor" in the same line as the label.
This is fixed by making the first CHECK match the entire line.
Reviewers: jmolloy, rengolin
Subscribers: rengolin, llvm-commits
Differential Revision: https://reviews.llvm.org/D24398
llvm-svn: 281211
Unlike SDag, we use a separate G_GEP instruction (much simplified, only taking
a single byte offset) to preserve the pointer type information through
selection.
llvm-svn: 281205
Some generic instructions have multiple types. While in theory these always be
discovered by inspecting the single definition of each generic vreg, in
practice those definitions won't always be local and traipsing through a big
function to find them will not be fun.
So this changes MIRPrinter to print out the type of uses as well as defs, if
they're known to be different or not known to be the same.
On the parsing side, we're a little more flexible: provided each register is
given a type in at least one place it's mentioned (and all types are
consistent) we accept the MIR. This doesn't introduce ambiguity but makes
writing tests manually a bit less painful.
llvm-svn: 281204
Remove createNode() and any API that depending on it, and add
HasCreateNode to the list of checks for HasObsoleteCustomizations. Now
an ilist *never* allocates (this was already true for iplist).
This factors out all the differences between iplist and ilist. I'll aim
to rename both to "owning_ilist" eventually, to call out the interesting
(not exactly intrusive) ownership semantics. In the meantime, I've left
both names around to reduce code churn.
One of the deleted APIs is the ilist copy constructor. I've lifted up
and tested iplist::cloneFrom (ala simple_ilist::cloneFrom) as a
replacement.
Users of ilist<> and iplist<> that want the list to allocate nodes have
a few options:
- use std::list;
- use AllocatorList or BumpPtrList (or build a similarly trivial list);
- use cloneFrom (which is explicit at the call site); or
- allocate at the call site.
See r280573, r281177, r281181, and r281182 for examples of what to do if
you're updating out-of-tree code.
llvm-svn: 281184
- Add AllocatorList, a non-intrusive list that owns an LLVM-style
allocator and provides a std::list-like interface (trivially built on
top of simple_ilist),
- add a typedef (and unit tests) for BumpPtrList, and
- use BumpPtrList for the list of llvm::yaml::Token (i.e., TokenQueueT).
TokenQueueT has no need for the complexity of an intrusive list. The
only reason to inherit from ilist was to customize the allocator.
TokenQueueT was the only example in-tree of using ilist<> in a truly
non-intrusive way.
Moreover, this removes the final use of the non-intrusive
ilist_traits<>::createNode (after r280573, r281177, and r281181). I
have a WIP patch that removes this customization point (and the API that
relies on it) that I plan to commit soon.
Note: AllocatorList owns the allocator, which limits the viable API
(e.g., splicing must be on the same list). For now I've left out
any problematic API. It wouldn't be hard to split AllocatorList into
two layers: an Impl class that calls DerivedT::getAlloc (via CRTP), and
derived classes that handle Allocator ownership/reference/etc semantics;
and then implement splice with appropriate assertions; but TBH we should
probably just customize the std::list allocators at that point.
llvm-svn: 281182
Force IVUsers to be moved instead of copied, properly update Parent
pointers in IVStrideUse when IVUsers is moved, and make sure we have
move constructors available in iplist and ilist.
I came across this in a WIP patch that deleted the copy constructors
from ilist. I was surprised to find that IVUsersAnalysis couldn't be
registered in the new pass manager.
It's not clear to me whether IVUsers was getting moved only when empty,
but if it was being moved when it was non-empty then this fixes a
pointer invalidation bug and should give some sort of speedup. Note
that the bugfix would be necessary even for a copy constructor.
llvm-svn: 281181
ilist_iterator::reset was unnecessary API, and wasn't any clearer (or
safer) at the call site than constructing a temporary and assigning it
to the iterator.
llvm-svn: 281175
Now that MachineBasicBlock::reverse_instr_iterator knows when it's at
the end (since r281168 and r281170), implement
MachineBasicBlock::reverse_iterator directly on top of an
ilist::reverse_iterator by adding an IsReverse template parameter to
MachineInstrBundleIterator. This replaces another hard-to-reason-about
use of std::reverse_iterator on list iterators, matching the changes for
ilist::reverse_iterator from r280032 (see the "out of scope" section at
the end of that commit message). MachineBasicBlock::reverse_iterator
now has a handle to the current node and has obvious invalidation
semantics.
r280032 has a more detailed explanation of how list-style reverse
iterators (invalidated when the pointed-at node is deleted) are
different from vector-style reverse iterators like std::reverse_iterator
(invalidated on every operation). A great motivating example is this
commit's changes to lib/CodeGen/DeadMachineInstructionElim.cpp.
Note: If your out-of-tree backend deletes instructions while iterating
on a MachineBasicBlock::reverse_iterator or converts between
MachineBasicBlock::iterator and MachineBasicBlock::reverse_iterator,
you'll need to update your code in similar ways to r280032. The
following table might help:
[Old] ==> [New]
delete &*RI, RE = end() delete &*RI++
RI->erase(), RE = end() RI++->erase()
reverse_iterator(I) std::prev(I).getReverse()
reverse_iterator(I) ++I.getReverse()
--reverse_iterator(I) I.getReverse()
reverse_iterator(std::next(I)) I.getReverse()
RI.base() std::prev(RI).getReverse()
RI.base() ++RI.getReverse()
--RI.base() RI.getReverse()
std::next(RI).base() RI.getReverse()
(For more details, have a look at r280032.)
llvm-svn: 281172
Add an assertion to the MachineInstrBundleIterator from instr_iterator
that the underlying iterator is valid. This is possible know that we
can check ilist_node::isSentinel (since r281168), and is consistent with
the constructors from MachineInstr* and MachineInstr&.
Avoiding the new assertion in operator== and operator!= requires four
(!!!!) new overloads each.
(As an aside, I'm strongly in favour of:
- making the conversion from instr_iterator explicit;
- making the conversion from pointer explicit;
- making the conversion from reference explicit; and
- removing all the extra overloads of operator== and operator!= except
const_instr_iterator.
I'm not signing up for that at this point, but being clear about when
something is an MachineInstr-iterator (possibly instr_end()) vs
MachineInstr-bundle-iterator (possibly end()) vs MachineInstr* (possibly
nullptr) vs MachineInstr& (known valid) would surely make code
cleaner... and it would remove a ton of boilerplate from
MachineInstrBundleIterator operators.)
llvm-svn: 281170
This is a prep commit before fixing MachineBasicBlock::reverse_iterator
invalidation semantics, ala r281167 for ilist::reverse_iterator. This
changes MachineBasicBlock::Instructions to track which node is the
sentinel regardless of LLVM_ENABLE_ABI_BREAKING_CHECKS.
There's almost no functionality change (aside from ABI). However, in
the rare configuration:
#if !defined(NDEBUG) && !defined(LLVM_ENABLE_ABI_BREAKING_CHECKS)
the isKnownSentinel() assertions in ilist_iterator<>::operator* suddenly
have teeth for MachineInstr. If these assertions start firing for your
out-of-tree backend, have a look at the suggestions in the commit
message for r279314, and at some of the commits leading up to it that
avoid dereferencing the end() iterator.
llvm-svn: 281168
This adds two declarative configuration options for intrusive lists
(available for simple_ilist, iplist, and ilist). Both of these options
affect ilist_node interoperability and need to be passed both to the
node and the list. Instead of adding a new traits class, they're
specified as optional template parameters (in any order).
The two options:
1. Pass ilist_sentinel_tracking<true> or ilist_sentinel_tracking<false>
to control whether there's a bit on ilist_node "prev" pointer
indicating whether it's the sentinel. The default behaviour is to
use a bit if and only if LLVM_ENABLE_ABI_BREAKING_CHECKS.
2. Pass ilist_tag<TagA> and ilist_tag<TagB> to allow insertion of a
single node into two different lists (simultaneously).
I have an immediate use-case for (1) ilist_sentinel_tracking: fixing the
validation semantics of MachineBasicBlock::reverse_iterator to match
ilist::reverse_iterator (ala r280032: see the comments at the end of the
commit message there). I'm adding (2) ilist_tag in the same commit to
validate that the options framework supports expansion. Justin Bogner
mentioned this might enable a possible cleanup in SelectionDAG, but I'll
leave this to others to explore. In the meantime, the unit tests and
the comments for simple_ilist and ilist_node have usage examples.
Note that there's a layer of indirection to support optional,
out-of-order, template paramaters. Internal classes are templated on an
instantiation of the non-variadic ilist_detail::node_options.
User-facing classes use ilist_detail::compute_node_options to compute
the correct instantiation of ilist_detail::node_options.
The comments for ilist_detail::is_valid_option describe how to add new
options (e.g., ilist_packed_int<int NumBits>).
llvm-svn: 281167
This should *actually* fix PR30244. This cranks up the workaround for PR30188 so that we never sink loads or stores of allocas.
The idea is that these should be removed by SROA/Mem2Reg, and any movement of them may well confuse SROA or just cause unwanted code churn. It's not ideal that the midend should be crippled like this, but that unwanted churn can really cause significant regressions in important workloads (tsan).
llvm-svn: 281162
How I missed this locally is beyond me. I suspect llc didn't recompile. This is just changing the CHECK line back to what it was before r280364.
llvm-svn: 281161
Exposed by PR30244, we will split a block currently if we think we can sink at least one instruction. However this isn't right - the reason we split predecessors is so that we can sink instructions that otherwise couldn't be sunk because it isn't safe to do so - stores, for example.
So, change the heuristic to only split if it thinks it can sink at least one non-speculatable instruction.
Should fix PR30244.
llvm-svn: 281160
