Fix byref attribute loss: copy param attrs after clone_into!

CloneFunctionInto rebuilds the AttributeList from scratch using VMap.
For byref params, VMap maps old args to addrspacecast instructions
(not Arguments), so dyn_cast<Argument> fails and byref(T) is silently
dropped. The subsequent setAttributes() overwrites any attrs we set
before clone_into!.

Without byref, the backend emits global_buffer(8) metadata instead of
by_value(sizeof(T)), causing HIP to copy only 8 bytes of struct data
as a "pointer" — leading to illegal address errors at runtime.

Also remove InferAddressSpaces (rely on AMDGPULowerKernelArguments
in codegen to trace addrspacecast chains for s_load).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: gbaraldi

src/gcn.jl

@@ -113,8 +113,8 @@ function add_kernarg_address_spaces!(@nospecialize(job::CompilerJob), mod::LLVM.
     end
 
     # insert addrspacecasts from kernarg (4) back to flat (0) so that the cloned IR
-    # (which expects flat pointers) continues to work. InferAddressSpaces will then
-    # propagate addrspace(4) through GEPs and loads, eliminating the casts.
+    # (which expects flat pointers) continues to work. The AMDGPU backend's
+    # AMDGPULowerKernelArguments traces these casts and produces s_load.
     new_args = LLVM.Value[]
     @dispose builder=IRBuilder() begin
         entry_bb = BasicBlock(new_f, "conversion")
@@ -127,9 +127,6 @@ function add_kernarg_address_spaces!(@nospecialize(job::CompilerJob), mod::LLVM.
             else
                 push!(new_args, parameters(new_f)[i])
             end
-            for attr in collect(parameter_attributes(f, i))
-                push!(parameter_attributes(new_f, i), attr)
-            end
         end
 
         # clone the original function body
@@ -144,6 +141,16 @@ function add_kernarg_address_spaces!(@nospecialize(job::CompilerJob), mod::LLVM.
         br!(builder, blocks(new_f)[2])
     end
 
+    # copy parameter attributes AFTER clone_into!, because CloneFunctionInto
+    # overwrites all attributes via setAttributes. For byref params, the VMap
+    # maps old args to addrspacecast instructions (not Arguments), so LLVM's
+    # attribute remapping silently drops them. We must re-add them here.
+    for i in 1:length(parameters(ft))
+        for attr in collect(parameter_attributes(f, i))
+            push!(parameter_attributes(new_f, i), attr)
+        end
+    end
+
     # replace the old function
     fn = LLVM.name(f)
     prune_constexpr_uses!(f)
@@ -152,23 +159,18 @@ function add_kernarg_address_spaces!(@nospecialize(job::CompilerJob), mod::LLVM.
     erase!(f)
     LLVM.name!(new_f, fn)
 
-    # propagate addrspace(4) through GEPs and loads, then clean up.
-    # InferAddressSpaces needs TargetTransformInfo (via TargetMachine) to know
-    # that flat address space is 0 on AMDGPU.
-    tm = llvm_machine(job.config.target)
+    # clean up the extra conversion block.
+    # NOTE: we do NOT run InferAddressSpaces here — the AMDGPU backend's
+    # AMDGPULowerKernelArguments pass traces addrspacecast chains during codegen
+    # and correctly produces s_load for addrspace(4) provenance. Running
+    # InferAddressSpaces with a TargetMachine can over-propagate addrspace(4)
+    # into pointer values loaded from the struct (which should remain flat/global).
     @dispose pb=NewPMPassBuilder() begin
-        add!(pb, NewPMFunctionPassManager()) do fpm
-            add!(fpm, InferAddressSpacesPass())
-        end
         add!(pb, NewPMFunctionPassManager()) do fpm
             add!(fpm, SimplifyCFGPass())
-            add!(fpm, SROAPass())
-            add!(fpm, EarlyCSEPass())
-            add!(fpm, InstCombinePass())
         end
-        run!(pb, mod, tm)
+        run!(pb, mod)
     end
-    dispose(tm)
 
     return functions(mod)[fn]
 end