Skip to content

jitlayers: Use GlobalISel on AArch64 at -O0/-O1 #60339

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
IanButterworth wants to merge 4 commits into
base: master
Choose a base branch
from
+23 −1
Open

jitlayers: Use GlobalISel on AArch64 at -O0/-O1 #60339

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
12fdcd5
jitlayers: Enable FastISel on AArch64 at -O0/-O1
IanButterworth Dec 7, 2025
dc0539b
jitlayers: Use GlobalISel on AArch64 at -O0/-O1
IanButterworth Dec 7, 2025
cc950ea
add test
IanButterworth Dec 9, 2025
e49df8d
make test clearer
IanButterworth Dec 9, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 11 additions & 1 deletion src/jitlayers.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1897,7 +1897,17 @@ void optimizeDLSyms(Module &M) JL_NOTSAFEPOINT_LEAVE JL_NOTSAFEPOINT_ENTER {
void fixupTM(TargetMachine &TM) {
auto TheTriple = TM.getTargetTriple();
if (jl_options.opt_level < 2) {
if (!TheTriple.isARM() && !TheTriple.isPPC64() && !TheTriple.isAArch64())
// Try GlobalISel on AArch64 - it's the default in LLVM at -O0 and
// is apparently generally faster than SelectionDAG while producing good code.
// Use fallback mode so unsupported patterns fall back to SelectionDAG.
// Note: Requires RemoveJuliaAddrspacesPass to run before codegen
// because GlobalISel doesn't handle Julia's custom address spaces.
if (TheTriple.isAArch64()) {
TM.setGlobalISel(true);
TM.setGlobalISelAbort(GlobalISelAbortMode::Disable);
TM.setFastISel(false);
}
else if (!TheTriple.isARM() && !TheTriple.isPPC64())
TM.setFastISel(true);
else // FastISel seems to be buggy Ref #13321
TM.setFastISel(false);
Expand Down
12 changes: 12 additions & 0 deletions test/atomics.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1115,3 +1115,15 @@ function add_one57190!()
end

@test add_one57190!() == 1

# Test atomic Float16 operations at all optimization levels (GlobalISel miscompile on AArch64)
# See https://github.com/JuliaLang/julia/pull/54140#issuecomment-2855794363
for opt in 0:3
@test success(run(```$(Base.julia_cmd()) --startup-file=no -O$opt -e '
a = Threads.Atomic{Float16}(Float16(0))
a[] = Float16(1.5)
@assert a[] === Float16(1.5) "atomic Float16 store failed: got \$(a[]) expected 1.5 (opt level = -O$(Base.JLOptions().opt_level))"
a[] = Float16(3.25)
@assert a[] === Float16(3.25) "atomic Float16 store failed: got \$(a[]) expected 3.25 (opt level = -O$(Base.JLOptions().opt_level))"
'```))
end