From 948ac96cd8281b49f10f41224f6bc622480a0ed0 Mon Sep 17 00:00:00 2001
From: Ludovic Raess <ludovic.rass@gmail.com>
Date: Tue, 6 Oct 2026 13:29:37 +0200
Subject: [PATCH] GCN: spell out the host's i128 alignment in the datalayout

Since AMDGPU_LLVM_Backend_jll 23.1.1+3, the external back-end aligns i128
to 16 bytes, while Julia before 1.12 aligns Int128 to 8 bytes on the host,
so kernel arguments after an Int128 were read at the wrong offset
(JuliaGPU/AMDGPU.jl#1002). The back-end replaces the module's datalayout
with its own, but since 23.1.1+4 keeps a 64-bit i128 alignment when the
module asks for one. Append the host's alignment for the external back-end;
on 1.12+ this also makes the middle end agree with the back-end's i128:128.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Tim Besard <tim.besard@gmail.com>
---
 src/gcn.jl  | 12 ++++++++++++
 test/gcn.jl | 29 +++++++++++++++++++++++++++++
 2 files changed, 41 insertions(+)

diff --git a/src/gcn.jl b/src/gcn.jl
index cb4146f5..7e4a1cfa 100644
--- a/src/gcn.jl
+++ b/src/gcn.jl
@@ -49,6 +49,18 @@ function llvm_machine(target::GCNCompilerTarget)
     return tm
 end
 
+# the external back-end replaces the module's datalayout with its own, which aligns i128 to
+# 16 bytes. it does keep a 64-bit alignment if the module asks for one (since
+# AMDGPU_LLVM_Backend_jll 23.1.1+4), which Julia needs before 1.12 to lay out kernel
+# arguments like the host does.
+function llvm_datalayout(target::GCNCompilerTarget)
+    dl = @invoke llvm_datalayout(target::AbstractCompilerTarget)
+    (dl === nothing || target.backend !== :external) && return dl
+    @dispose dl=dl begin
+        string(dl) * "-i128:$(8 * Base.datatype_alignment(Int128))"
+    end
+end
+
 
 # AMDGPU calls the device scope `agent`, and the subgroup one `wavefront`
 llvm_syncscope(::GCNCompilerTarget, name::String) =
diff --git a/test/gcn.jl b/test/gcn.jl
index 90ce90a7..243cdc60 100644
--- a/test/gcn.jl
+++ b/test/gcn.jl
@@ -338,6 +338,35 @@ end
     end
 end
 
+@testset "Int128 kernel arguments" begin
+    # the back-end must lay out kernel arguments like the host does, which before Julia 1.12
+    # aligns Int128 to 8 bytes (JuliaGPU/AMDGPU.jl#1002)
+    mod = @eval module $(gensym())
+        kernel(a::Int32, b::Int128, c::Tuple{Int32,Int128}, d::Int64) = return
+    end
+    types = Tuple{Int32, Int128, Tuple{Int32,Int128}, Int64}
+    arg_offset(i) = fieldoffset(types, i)
+    arg_size(i) = sizeof(fieldtype(types, i))
+
+    @test @filecheck begin
+        @check "amdhsa.kernels:"
+        @check_next "- .args:"
+        @check ".offset: $(arg_offset(1)){{\$}}"
+        @check_next ".size: $(arg_size(1)){{\$}}"
+        @check_next ".value_kind: by_value"
+        @check ".offset: $(arg_offset(2)){{\$}}"
+        @check_next ".size: $(arg_size(2)){{\$}}"
+        @check_next ".value_kind: by_value"
+        @check ".offset: $(arg_offset(3)){{\$}}"
+        @check_next ".size: $(arg_size(3)){{\$}}"
+        @check_next ".value_kind: by_value"
+        @check ".offset: $(arg_offset(4)){{\$}}"
+        @check_next ".size: $(arg_size(4)){{\$}}"
+        @check_next ".value_kind: by_value"
+        GCN.code_native(mod.kernel, types; dump_module=true, kernel=true)
+    end
+end
+
 @testset "skip scalar trap" begin
     mod = @eval module $(gensym())
         workitem_idx_x() = ccall("llvm.amdgcn.workitem.id.x", llvmcall, Int32, ())