LSR has some logic that tries to aggressively reuse registers in
formula. This can lead to sub-optimal decision in complex loops where
the backend it trying to use shouldFavorPostInc. This disables the
re-use in those situations.
Differential Revision: https://reviews.llvm.org/D79301
VMEM soft clauses only contain VMEM and FLAT instructions. Teaching
GCNHazardRecognizer::checkSoftClauseHazards that other kinds of
instructions will naturally break the clause means there are far fewer
cases where it has to insert an s_nop instruction to forcibly break the
clause.
Differential Revision: https://reviews.llvm.org/D79353
Marking a section as ALLOC tells the ELF loader to load the section into memory.
As we do not want to load the notes into VRAM, the flag should not be there.
On AMDHSA, .note is still marked as ALLOC, apparently this is currently
needed for OpenCL (see https://reviews.llvm.org/D74995).
Differential Revision: https://reviews.llvm.org/D76278
A PREDICATE_CAST(PREDICATE_CAST(X)) can be converted to a
PREDICATE_CAST(X) as the operation can convert between any forms of
predicates (v4i1/v8i1/v16i1/i32). Unfortunately I got the type wrong on
one of the rarer converts, which would lead to invalid nodes during
isel. This fixes it up to use the correct type.
Differential Revision: https://reviews.llvm.org/D79402
Unless we're truncating an 'all-bits' result, using PACKSS for vXi64->vXi32 truncation causes problems with later combines as ComputeNumSignBits struggles to see through BITCASTs to smaller types. If we don't use PACKSS in these cases then we fallback to shuffles which are usually just as good.
Summary:
This fixes a few things that are connected. It is very hard to provide
an independent test case for each of those fixes, because they are
interconnected and sometimes one masks another. The provided test case
triggers some of those bugs below but not all.
---
1. Background:
`placeBlockMarker` takes a BB, and if the BB is a destination of some
branch, it places `end_block` marker there, and computes the nearest
common dominator of all predecessors (what we call 'header') and places
a `block` marker there.
When we first place markers, we traverse BBs from top to bottom. For
example, when there are 5 BBs A, B, C, D, and E and B, D, and E are
branch destinations, if mark the BB given to `placeBlockMarker` with `*`
and draw a rectangle representing the border of `block` and `end_block`
markers, the process is going to look like
```
-------
----- |-----|
--- |---| ||---||
|A| ||A|| |||A|||
--- --> |---| --> ||---||
*B | B | || B ||
C | C | || C ||
D ----- |-----|
E *D | D |
E -------
*E
```
which means when we first place markers, we go from inner to outer
scopes. So when we place a `block` marker, if the header already
contains other `block` or `try` marker, it has to belong to an inner
scope, so the existing `block`/`try` markers should go _after_ the new
marker. This was the assumption we had.
But after placing all markers we run `fixUnwindMismatches` function.
There we do some control flow transformation and create some branches,
and we call `placeBlockMarker` again to place `block`/`end_block`
markers for those newly created branches. We can't assume that we are
traversing branch destination BBs from top to bottom now because we are
basically inserting some new markers in the middle of existing markers.
Fix:
In `placeBlockMarker`, we don't have the assumption that the BB given is
in the order of top to bottom, and when placing `block` markers,
calculates whether existing `block` or `try` markers are inner or
outer scopes with respect to the current scope.
---
2. Background:
In `fixUnwindMismatches`, when there is a call whose correct unwind
destination mismatches the current destination after initially placing
`try` markers, we wrap that with a new nested `try`/`catch`/`end` and
jump to the correct handler within the new `catch`. The correct handler
code is split as a separate BB from its original EH pad so it can be
branched to. Here's an example:
- Before
```
mbb:
call @foo <- Unwind destination mismatch!
wrong-ehpad:
catch
...
cont:
end_try
...
correct-ehpad:
catch
[handler code]
```
- After
```
mbb:
try (new)
call @foo
nested-ehpad: (new)
catch (new)
local.set n / drop (new)
br %handleri (new)
nested-end: (new)
end_try (new)
wrong-ehpad:
catch
...
cont:
end_try
...
correct-ehpad:
catch
local.set n / drop (new)
handler: (new)
end_try
[handler code]
```
Note that after this transformation, it is possible there are no calls
to actually unwind to `correct-ehpad` here. `call @foo` now
branches to `handler`, and there can be no other calls to unwind to
`correct-ehpad`. In this case `correct-ehpad` does not have any
predecessors anymore.
This can cause a bug in `placeBlockMarker`, because we may need to place
`end_block` marker in `handler`, and `placeBlockMarker` computes the
nearest common dominator of all predecessors. If one of `handler`'s
predecessor (here `correct-ehpad`) does not have any predecessors, i.e.,
no way of reaching it, we cannot correctly compute the common dominator
of predecessors of `handler`, and end up placing no `block`/`end`
markers. This bug actually sometimes masks the bug 1.
Fix:
When we have an EH pad that does not have any predecessors after this
transformation, deletes all its successors, so that its successors don't
have any dangling predecessors.
---
3. Background:
Actually the `handler` BB in the example shown in bug 2 doesn't need
`end_block` marker, despite it being a new branch destination, because
it already has `end_try` marker which can serve the same purpose. I just
put that example there for an illustration purpose. There is a case we
actually need to place `end_block` marker: when the branch dest is the
appendix BB. The appendix BB is created when there is a call that is
supposed to unwind to the caller ends up unwinding to a wrong EH pad. In
this case we also wrap the call with a nested `try`/`catch`/`end`,
create an 'appendix' BB at the very end of the function, and branch to
that BB, where we rethrow the exception to the caller.
Fix:
When we don't actually need to place block markers, we don't.
---
4. In case we fall through to the continuation BB after the catch block,
after extracting handler code in `fixUnwindMismatches` (refer to bug 2
for an example), we now have to add a branch to it to bypass the
handler.
- Before
```
try
...
(falls through to 'cont')
catch
handler body
end
<-- cont
```
- After
```
try
...
br %cont (new)
catch
end
handler body
<-- cont
```
The problem is, we haven't been placing a new `end_block` marker in the
`cont` BB in this case. We should, and this fixes it. But it is hard to
provide a test case that triggers this bug, because the current
compilation pipeline from .ll to .s does not generate this kind of code;
we always have a `br` after `invoke`. But code without `br` is still
valid, and we can have that kind of code if we have some pipeline
changes or optimizations later. Even mir test cases cannot trigger this
part for now, because we don't encode auxiliary EH-related data
structures (such as `WasmEHFuncInfo`) in mir now. Those functionalities
can be added later, but I don't think we should block this fix on that.
Reviewers: dschuff
Subscribers: sbc100, jgravelle-google, hiraditya, sunfish, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D79324
This patch makes the folding of or(A, B) into not(and(not(A), not(B)))
more agressive for I1 vector. This only affects Thumb2 MVE and improves
codegen, because it removes a lot of msr/mrs instructions on VPR.P0.
This patch also adds a xor(vcmp) -> !vcmp fold for MVE.
Differential Revision: https://reviews.llvm.org/D77202
This patch adds an implementation of PerformVSELECTCombine in the
ARM DAG Combiner that transforms vselect(not(cond), lhs, rhs) into
vselect(cond, rhs, lhs).
Normally, this should be done by the target-independent DAG Combiner,
but it doesn't handle the kind of constants that we generate, so we
have to reimplement it here.
Differential Revision: https://reviews.llvm.org/D77712
Register live ranges may have had gaps that after coalescing should be
removed. This is done by adding a new segment to the range, and merging
it with neighboring segments. When doing so, do not assume that each
subrange of the register ended at the same index. If a subrange ended
earlier, adding this segment could make the live range invalid.
Instead, if the subrange is not live at the start of the segment,
extend it first.
Today symbol names generated for machine basic block sections use a
unary encoding to reduce bloat. This is essential when every basic block
in the binary is assigned a symbol however with basic block clusters
(rG05192e585ce175b55f2a26b83b4ed7882785c8e6) when we only need to
generate a few non-temporary symbols we can assign more descriptive
names making them more user friendly. With this change -
Cold cluster section for function foo is named "foo.cold"
Exception cluster section for function foo is named "foo.eh"
Other cluster sections identified by their ids are named "foo.ID"
Using this format works well with existing tools. It will demangle as
expected and works with existing symbolizers, profilers and debuggers
out of the box.
$ c++filt _Z3foov.cold
foo() [clone .cold]
$ c++filt _Z3foov.eh
foo() [clone .eh]
$c++filt _Z3foov.1234
foo() [clone 1234]
Tests for basicblock-sections are updated with some cleanup where
appropriate.
Differential Revision: https://reviews.llvm.org/D79221
This a hack to fix illegal 32 to 16 bit copies.
The problem is when we make 16 bit subregs legal it creates
a huge amount of failures which can only be resolved at once
without a temporary hack like this.
The next step is to change operands, instruction definitions
and patterns until this hack is not needed.
Differential Revision: https://reviews.llvm.org/D79119
We allocated a suitably aligned frame index so we know that all the values
have ABI alignment.
For MIPS this avoids using pair of lwl + lwr instructions instead of a
single lw. I found this when compiling CHERI pure capability code where
we can't use the lwl/lwr unaligned loads/stores and and were to falling
back to a byte load + shift + or sequence.
This should save a few instructions for MIPS and possibly other backends
that don't have fast unaligned loads/stores.
It also improves code generation for CodeGen/X86/pr34653.ll and
CodeGen/WebAssembly/offset.ll since they can now use aligned loads.
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D78999
SelectionDAGBuilder currently doesn't propagate the known alignment of
the sret parameter. This is inefficient for MIPS and highly inefficient for
our out-of-tree CHERI-extended MIPS since we don't have lwl/lwr so fall back
to byte loads for align == 1.
Summary: This change enables all kind of carry out ISD opcodes to be selected according to the node divergence.
Reviewers: rampitec, arsenm, vpykhtin
Reviewed By: rampitec
Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D78091
Summary:
This patch adds AArch64ISD nodes for [S|U]MIN_PRED
and [S|U]MAX_PRED, and lowers both SVE intrinsics and
IR operations for min and max to these nodes.
There are two forms of these instructions for SVE: a predicated
form and an immediate (unpredicated) form. The patterns
which existed for the latter have been updated to match a
predicated node with an immediate and map this
to the immediate instruction.
Reviewers: sdesmalen, efriedma, dancgr, rengolin
Reviewed By: efriedma
Subscribers: huihuiz, tschuett, kristof.beyls, hiraditya, rkruppe, psnobl, cfe-commits, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D79087
Default legalization will create two v8i64 truncs to v8i32, concat
them to v16i32, and then truncate the rest of the way to v16i8.
Instead we can truncate directly from v8i64 to v8i8 in the lower
half of an xmm. Then concat the two halves to use vpunpcklqdq.
This is the same number of uops, but the dependency chain through
the uops is better since the halves are merged at the end.
I had to had SimplifyDemandedBits support for VTRUNC to prevent
a regression on vector-trunc-math.ll. combineTruncatedArithmetic
no longer gets a chance to shrink vXi64 mul so we were producing
the v8i64 multiply sequence using multiple PMULUDQs. With the
demanded bits fix we are able to prune out the extra ops leaving
just two PMULUDQs, one for each v8i64 half. This is twice the
width of the 2 v8i32 PMULLDs we had before, but PMULUDQ is 1
uop and PMULLD is 2. We also save some truncates. It's probably
worth using PMULUDQ even when PMULLQ is available since the latter
is 3 uops, but that will require a different change.
Differential Revision: https://reviews.llvm.org/D79231
Don't use $noreg for instructions that take register inputs.
Only allow $noreg for parts of memory operands.
Don't use index register with $rip base.
Use RETQ instead of the RET pseudo. This pass is after the
ExpandPseudo pass that converts RET to RETQ.
The two code paths have the same goal, legalizing a load of a non-byte-sized vector by loading the "flattened" representation in memory, slicing off each single element and then building a vector out of those pieces.
The technique employed by `ExpandLoad` is slightly more convoluted and produces slightly better codegen on ARM, AMDGPU and x86 but suffers from some bugs (D78480) and is wrong for BE machines.
Differential Revision: https://reviews.llvm.org/D79096
Summary:
The current lowering of `select` on RISC-V uses a branch instruction to load a
register with one or other value. This is inefficient, especially in the case of
small constants that can be computed easily.
By implementing the TargetLowering::convertSelectOfConstantsToMath hook, some of
the simpler cases are covered that let us avoid introducing a branch in these
cases.
Reviewed By: luismarques
Differential Revision: https://reviews.llvm.org/D79260
Summary:
This just adds some simple cases for testing select of constants. There will be
a follow-up patch that improves code generation in some of these cases.
Reviewed By: luismarques
Differential Revision: https://reviews.llvm.org/D79259
Summary:
This patch addresses some weird assembly sequences we were seeing during
comparing floats. In particular, comparing a float to itself tells you whether
it is NaN or not, which we were doing correctly, but with an extra unneeded
`and` instruction.
This patch specialises the existing patterns to remove the `and` instructions
when both their operands are the same.
Reviewed By: luismarques, asb
Differential Revision: https://reviews.llvm.org/D78908
Summary:
I worked on adding some SelectionDag patterns to address code generated by these
examples, which came out of some differential testing against GCC. The pattern
additions will be in a follow-up patch.
Reviewers: luismarques, asb
Reviewed By: luismarques, asb
Differential Revision: https://reviews.llvm.org/D78907
I think some copy/pasting was used to create loops of different
VFs. But the increment of the induction variable wasn't updated
to match the VF.
This has no effect on the pattern matching we're testing, it just
helps the test make sense to the reader.
Summary:
As described in https://github.com/WebAssembly/simd/pull/209. This is
the final reorganization of the SIMD opcode space before
standardization. It has been landed in concert with corresponding
changes in other projects in the WebAssembly SIMD ecosystem.
Reviewers: aheejin
Subscribers: dschuff, sbc100, jgravelle-google, hiraditya, sunfish, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D79224
Handle concat_vectors(extract_subvector(broadcast(x)), extract_subvector(broadcast(x))) -> broadcast(x)
To expose this we also need collectConcatOps to recognise the insert_subvector(x, extract_subvector(x, lo), hi) subvector splat pattern
VMEM loads of the same type (sampler vs no sampler) are guaranteed to
write their result registers in order, so there is no need for an
s_waitcnt even if they write to overlapping vgprs.
Differential Revision: https://reviews.llvm.org/D79176
While restoring latency, check if any of the registers of
source instruction is a subregister of the successor instructions
apart from being same register.
This patch adds the x, t and g modifiers for inline asm from GCC. These will print a vector register as xmm*, ymm* or zmm* respectively.
I also fixed register names with modifiers with inteldialect so they are no longer printed with a leading %.
Patch by Amanieu d'Antras
Differential Revision: https://reviews.llvm.org/D78977
Also fix some cost tables for vXi1 types to match the costs entries for the types they will be promoted to.
Differential Revision: https://reviews.llvm.org/D79045
This pushes the NOT pattern up the DAG to help expose it for further combines (AND->ANDN in particular).
The PSHUFD/MOVDDUP 'splat' cases are the only ones I've seen in the wild so far, we can further generalize if/when we need to.
X86 matches several 'shift+xor' funnel shift patterns:
fold (or (srl (srl x1, 1), (xor y, 31)), (shl x0, y)) -> (fshl x0, x1, y)
fold (or (shl (shl x0, 1), (xor y, 31)), (srl x1, y)) -> (fshr x0, x1, y)
fold (or (shl (add x0, x0), (xor y, 31)), (srl x1, y)) -> (fshr x0, x1, y)
These patterns are also what we end up with the proposed expansion changes in D77301.
This patch moves these to DAGCombine's generic MatchFunnelPosNeg.
All existing X86 test cases still pass, and we just have a small codegen change in pr32282.ll.
Reviewed By: @spatel
Differential Revision: https://reviews.llvm.org/D78935
