jitlayers: Use GlobalISel on AArch64 at -O0/-O1 by IanButterworth · Pull Request #60339 · JuliaLang/julia

IanButterworth · 2025-12-07T04:08:51Z

On-top of #59950
Required llvm/llvm-project#171499 which is backported in #59950

Alternative to #60338

GlobalISel is LLVM's modern instruction selector that is designed to replace
both FastISel and SelectionDAG. On AArch64, it is mature and enabled by default
at -O0 in upstream LLVM.

This enables GlobalISel with fallback mode on AArch64, which provides faster
instruction selection than SelectionDAG while maintaining correctness by
falling back to SelectionDAG for unsupported patterns.

Note: This requires RemoveJuliaAddrspacesPass to run before codegen, which is
already the case in the current pipeline (see pipeline.cpp comment about
GlobalISel not liking Julia's address spaces).

Co-authored-by: Claude

gbaraldi · 2025-12-08T14:51:11Z

I'm more comfortable with this than fastisel

giordano · 2025-12-08T14:52:33Z

Does this need a pkgeval?

IanButterworth · 2025-12-08T16:20:47Z

Yeah a pkgeval on aarch64 would be good. @maleadt is that still possible?

xal-0 · 2025-12-08T20:41:29Z

I believe this was enabled at one point for ahead-of-time compilation but disabled because of miscompiles? I should probably report/fix the easy ones upstream...
#54140 (comment)

The atomic Float16 miscompile still happens on this version:

$ ./usr/bin/julia --banner=short -O1
  o  | Version 1.14.0-DEV.1348 (2025-12-07)
 o o | HEAD/dc0539badaa* (fork: 26 commits, 7 days)
julia> code_native(setindex!, (Threads.Atomic{Float16}, Float16); debuginfo=:none)
[...]
"_julia_setindex!_600":                 ; @"julia_setindex!_600"
	stp	x29, x30, [sp, #-16]!           ; 16-byte Folded Spill
	mov	x29, sp
	mrs	w8, NZCV
	;DEBUG_VALUE: setindex!:v <- $w8
	stlrh	w8, [x0]
	ldp	x29, x30, [sp], #16             ; 16-byte Folded Reload
	ret

Compare with -O1 on master:

"_julia_setindex!_615":                 ; @"julia_setindex!_615"
	;DEBUG_VALUE: setindex!:x <- [$x0+0]
	;DEBUG_VALUE: setindex!:v <- $h0
	stp	x29, x30, [sp, #-16]!           ; 16-byte Folded Spill
	mov	x29, sp
	;DEBUG_VALUE: setindex!:x <- [$x0+0]
	;DEBUG_VALUE: setindex!:v <- $h0
	fmov	w8, s0
	stlrh	w8, [x0]
	ldp	x29, x30, [sp], #16             ; 16-byte Folded Reload
	ret

IIRC you can get this to trigger during bootstrapping with -O1.

oscardssmith · 2025-12-08T20:50:01Z

yeah. we should definitely be reporting any bugs in global isel we're finding. IIUC Global Isel is intended to be the isel future (completely replacing SelectionDAG) so we want it to be as solid as possible.

maleadt · 2025-12-09T13:08:45Z

Yeah a pkgeval on aarch64 would be good. @maleadt is that still possible?

Not easily, no; it's a very manual process.
The ARM machine in question is now also used for Base CI, so I can't take it over for a full-scale PkgEval run without compromising CI resources.

oscardssmith · 2025-12-09T13:45:25Z

should we also use it at O2/O3? iiuc, for aarch64 it's generally expected to produce equally good code

IanButterworth · 2025-12-09T14:41:14Z

@xal-0 I added a test based on your example. It fails locally.

gbaraldi · 2025-12-09T19:37:30Z

You can probably ask Claude to turn this into an llc test case for an LLVM issue.

IanButterworth · 2025-12-09T19:47:00Z

llvm/llvm-project#171494

IanButterworth · 2025-12-13T12:46:13Z

llvm/llvm-project#171499 has been merged.

Are there any other known bugs?

oscardssmith · 2026-02-24T14:32:59Z

Is this ready to go?

IanButterworth · 2026-02-24T14:42:57Z

This needs to wait for that fix to land in LLVM. If we wanted it backported there was some guidance here llvm/llvm-project#171499 (comment)

oscardssmith · 2026-02-24T14:50:21Z

Can we just add it to our LLVM branch?

IanButterworth · 2026-03-15T21:07:42Z

Trying this ontop of #59950 cc @giordano

giordano · 2026-03-15T21:16:42Z

I don't think your patch in included in the build at the moment present in #59950.

LLVM 21 replaced the `Stmt*` parameter in `SValBuilder::conjureSymbolVal` with `ConstCFGElementRef` (llvm/llvm-project#128251). Update all four call sites in GCChecker.cpp to use `C.getCFGElementRef()` instead of passing `Expr*` pointers directly. Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

LLVM 21 (llvm/llvm-project#138092) aligns i128 to 16 bytes when passing on the x86-32 stack. However, GCC does not support __int128 on 32-bit x86 at all, so C libraries use struct { int64_t, int64_t } which only has 4-byte stack alignment. This mismatch caused segfaults on Linux and incorrect results on Windows for ccalls involving Int128. Fix by using [4 x i32] instead of i128 as the byval type for large integer arguments, which naturally has 4-byte alignment matching the C ABI. Also fix abi_win32.cpp to pass large primitive types like Int128 by reference (byval) instead of directly, avoiding the same alignment issue. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

FastISel was disabled on AArch64 in 2015 (PR JuliaLang#13393) to fix issue JuliaLang#13321, but that issue was specifically about 32-bit ARM (ARMv7) segfaults during bootstrap. The AArch64 exclusion was added conservatively alongside the ARM fix. AArch64 FastISel has been actively maintained upstream with recent bug fixes: - llvm/llvm-project#75993 (Jan 2024) - llvm/llvm-project#133987 (May 2025) This enables faster instruction selection for JIT compilation on AArch64 at lower optimization levels, reducing compilation latency.

GlobalISel is LLVM's modern instruction selector that is designed to replace both FastISel and SelectionDAG. On AArch64, it is mature and enabled by default at -O0 in upstream LLVM. This enables GlobalISel with fallback mode on AArch64, which provides faster instruction selection than SelectionDAG while maintaining correctness by falling back to SelectionDAG for unsupported patterns. Note: This requires RemoveJuliaAddrspacesPass to run before codegen, which is already the case in the current pipeline (see pipeline.cpp comment about GlobalISel not liking Julia's address spaces). Co-Authored-By: Claude <noreply@anthropic.com>

Add JL_GC_PROMISE_ROOTED annotations in egal_set::insert() for val, list, and keyset. These class members are rooted by the caller via JL_GC_PUSH3, but the GC checker cannot track rooting through class member indirection. The post-put_key annotation on list is needed because the returned pointer is a new value assigned to the already- pushed member address. Suppress optin.core.FixedAddressDereference false positives in llvm-alloc-opt.cpp (LLVM ilist_node_base.h internals) and gc-stock.c (gc_read_stack bounds-checked address) — this checker is new in LLVM 21. Co-authored-by: Claude <noreply@anthropic.com>

giordano · 2026-04-13T10:30:58Z

$ julia +nightly -e 'using InteractiveUtils; code_native(setindex!, (Threads.Atomic{Float16}, Float16); debuginfo=:none)'
        .file   "setindex!"
        .text
        .globl  "julia_setindex!_0"             // -- Begin function julia_setindex!_0
        .p2align        4
        .type   "julia_setindex!_0",@function
"julia_setindex!_0":                    // @"julia_setindex!_0"
; Function Signature: setindex!(Base.Threads.Atomic{Float16}, Float16)
// %bb.0:                               // %top
        //DEBUG_VALUE: setindex!:x <- [$x0+0]
        //DEBUG_VALUE: setindex!:v <- $h0
        stp     x29, x30, [sp, #-16]!           // 16-byte Folded Spill
                                        // kill: def $h0 killed $h0 def $s0
        //DEBUG_VALUE: setindex!:x <- [$x0+0]
        //DEBUG_VALUE: setindex!:v <- $h0
        fmov    w8, s0
        mov     x29, sp
        stlrh   w8, [x0]
        ldp     x29, x30, [sp], #16             // 16-byte Folded Reload
        ret
.Lfunc_end0:
        .size   "julia_setindex!_0", .Lfunc_end0-"julia_setindex!_0"
                                        // -- End function
        .section        ".note.GNU-stack","",@progbits
$ julia +pr60339 -O1 -e 'using InteractiveUtils; code_native(setindex!, (Threads.Atomic{Float16}, Float16); debuginfo=:none)'
        .file   "setindex!"
        .text
        .globl  "julia_setindex!_0"             // -- Begin function julia_setindex!_0
        .p2align        4
        .type   "julia_setindex!_0",@function
"julia_setindex!_0":                    // @"julia_setindex!_0"
; Function Signature: setindex!(Base.Threads.Atomic{Float16}, Float16)
// %bb.0:                               // %top
        stp     x29, x30, [sp, #-16]!           // 16-byte Folded Spill
                                        // kill: def $h0 killed $h0 def $s0
        //DEBUG_VALUE: setindex!:x <- [$x0+0]
        //DEBUG_VALUE: setindex!:v <- $h0
        fmov    w8, s0
        mov     x29, sp
        stlrh   w8, [x0]
        ldp     x29, x30, [sp], #16             // 16-byte Folded Reload
        ret
.Lfunc_end0:
        .size   "julia_setindex!_0", .Lfunc_end0-"julia_setindex!_0"
                                        // -- End function
        .section        ".note.GNU-stack","",@progbits
$ julia +pr60339 -O0 -e 'using InteractiveUtils; code_native(setindex!, (Threads.Atomic{Float16}, Float16); debuginfo=:none)'
        .file   "setindex!"
        .text
        .globl  "julia_setindex!_0"             // -- Begin function julia_setindex!_0
        .p2align        4
        .type   "julia_setindex!_0",@function
"julia_setindex!_0":                    // @"julia_setindex!_0"
; Function Signature: setindex!(Base.Threads.Atomic{Float16}, Float16)
// %bb.0:                               // %top
        //DEBUG_VALUE: setindex!:x <- [$x0+0]
        //DEBUG_VALUE: setindex!:v <- $h0
        stp     x29, x30, [sp, #-16]!           // 16-byte Folded Spill
        mov     x29, sp
                                        // kill: def $s0 killed $h0
        fmov    w8, s0
        stlrh   w8, [x0]
        ldp     x29, x30, [sp], #16             // 16-byte Folded Reload
        ret
.Lfunc_end0:
        .size   "julia_setindex!_0", .Lfunc_end0-"julia_setindex!_0"
                                        // -- End function
        .section        ".note.GNU-stack","",@progbits

@xal-0 looks good?

#60339 was auto-closed when #59950 merged and can't be reopened. Co-authored-by: Claude <noreply@anthropic.com>

JuliaLang#60339 was auto-closed when JuliaLang#59950 merged and can't be reopened. Co-authored-by: Claude <noreply@anthropic.com>

IanButterworth changed the title ~~jitlayers: Enable FastISel on AArch64 at -O0/-O1~~ jitlayers: Use GlobalISel on AArch64 at -O0/-O1 Dec 7, 2025

IanButterworth added the compiler:codegen Generation of LLVM IR and native code label Dec 7, 2025

IanButterworth force-pushed the ib/GlobalISel branch from d89fedf to dc0539b Compare December 7, 2025 04:13

IanButterworth marked this pull request as ready for review December 7, 2025 14:49

IanButterworth mentioned this pull request Dec 7, 2025

jitlayers: Enable FastISel on AArch64 at -O0/-O1 #60338

Closed

IanButterworth added the needs pkgeval Tests for all registered packages should be run with this change label Dec 8, 2025

IanButterworth removed the needs pkgeval Tests for all registered packages should be run with this change label Dec 9, 2025

IanButterworth force-pushed the ib/GlobalISel branch from 36855b7 to e49df8d Compare December 9, 2025 17:53

oscardssmith requested a review from gbaraldi February 24, 2026 14:33

IanButterworth force-pushed the ib/GlobalISel branch from e49df8d to 9cb28b3 Compare March 15, 2026 21:06

IanButterworth requested a review from giordano as a code owner March 15, 2026 21:06

IanButterworth changed the base branch from master to mg/llvm-21 March 15, 2026 21:07

giordano and others added 2 commits March 23, 2026 11:24

Bump LLVM to v21

157c8fd

Don't register two EHFrameRegistrationPlugins

2119ae2

Update to LLVM v21.1.8

3542185

giordano force-pushed the mg/llvm-21 branch from fdb3d8b to 3542185 Compare March 23, 2026 11:30

giordano and others added 8 commits March 23, 2026 18:24

Temporarily (?) disable warning on maybe uninitialized

0e2ea5a

Merge branch 'master' into mg/llvm-21

970e96f

add test

2d54fdb

make test clearer

2a1edc0

IanButterworth force-pushed the ib/GlobalISel branch 3 times, most recently from 360eafb to 4b9356c Compare April 3, 2026 03:20

IanButterworth force-pushed the ib/GlobalISel branch from 4b9356c to 50697f8 Compare April 3, 2026 15:55

IanButterworth requested a review from xal-0 April 17, 2026 01:23

gbaraldi force-pushed the mg/llvm-21 branch from d4d184e to e071e3a Compare April 17, 2026 15:39

oscardssmith deleted the branch JuliaLang:mg/llvm-21 May 8, 2026 17:57

oscardssmith closed this May 8, 2026

IanButterworth mentioned this pull request May 8, 2026

jitlayers: Use GlobalISel on AArch64 at -O0/-O1 #61758

Merged

IanButterworth added a commit that referenced this pull request May 11, 2026

jitlayers: Use GlobalISel on AArch64 at -O0/-O1 (#61758)

5dee674

#60339 was auto-closed when #59950 merged and can't be reopened. Co-authored-by: Claude <noreply@anthropic.com>

hardikxk pushed a commit to hardikxk/julia that referenced this pull request May 18, 2026

jitlayers: Use GlobalISel on AArch64 at -O0/-O1 (JuliaLang#61758)

9a23a07

JuliaLang#60339 was auto-closed when JuliaLang#59950 merged and can't be reopened. Co-authored-by: Claude <noreply@anthropic.com>

Uh oh!

Conversation

IanButterworth commented Dec 7, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

gbaraldi commented Dec 8, 2025

Uh oh!

giordano commented Dec 8, 2025

Uh oh!

IanButterworth commented Dec 8, 2025

Uh oh!

xal-0 commented Dec 8, 2025

Uh oh!

oscardssmith commented Dec 8, 2025

Uh oh!

maleadt commented Dec 9, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

oscardssmith commented Dec 9, 2025

Uh oh!

IanButterworth commented Dec 9, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

gbaraldi commented Dec 9, 2025

Uh oh!

IanButterworth commented Dec 9, 2025

Uh oh!

IanButterworth commented Dec 13, 2025

Uh oh!

oscardssmith commented Feb 24, 2026

Uh oh!

IanButterworth commented Feb 24, 2026

Uh oh!

oscardssmith commented Feb 24, 2026

Uh oh!

IanButterworth commented Mar 15, 2026

Uh oh!

giordano commented Mar 15, 2026

Uh oh!

giordano commented Apr 13, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants

IanButterworth commented Dec 7, 2025 •

edited

Loading

maleadt commented Dec 9, 2025 •

edited

Loading

IanButterworth commented Dec 9, 2025 •

edited

Loading