From c33f8bb6a84acd469bad1ac09ff80a94ac005d01 Mon Sep 17 00:00:00 2001
From: Valentin Churavy <v.churavy@gmail.com>
Date: Mon, 22 Jun 2026 17:28:40 +0200
Subject: [PATCH 1/2] remove StaticArrays

---
 Project.toml              | 2 --
 src/KernelAbstractions.jl | 1 -
 src/pocl/backend.jl       | 2 --
 3 files changed, 5 deletions(-)

diff --git a/Project.toml b/Project.toml
index 4fe4ac21b..356e85a38 100644
--- a/Project.toml
+++ b/Project.toml
@@ -18,7 +18,6 @@ RandomNumbers = "e6cf234a-135c-5ec9-84dd-332b85af5143"
 SPIRVIntrinsics = "71d1d633-e7e8-4a92-83a1-de8814b09ba8"
 SPIRV_LLVM_Backend_jll = "4376b9bf-cff8-51b6-bb48-39421dff0d0c"
 SPIRV_Tools_jll = "6ac6d60f-d740-5983-97d7-a4482c0689f4"
-StaticArrays = "90137ffa-7385-5640-81b9-e52037218182"
 UUIDs = "cf7118a7-6976-5b1a-9a39-7adc72f591a4"
 pocl_standalone_jll = "54f56a70-6062-5590-a942-1226658f6c83"
 
@@ -49,7 +48,6 @@ SPIRVIntrinsics = "1"
 SPIRV_LLVM_Backend_jll = "22"
 SPIRV_Tools_jll = "2024.4, 2025.1"
 SparseArrays = "<0.0.1, 1.6"
-StaticArrays = "0.12, 1.0"
 UUIDs = "<0.0.1, 1.6"
 julia = "1.10"
 pocl_standalone_jll = "7.1.2"
diff --git a/src/KernelAbstractions.jl b/src/KernelAbstractions.jl
index 3881da55c..d6414ee22 100644
--- a/src/KernelAbstractions.jl
+++ b/src/KernelAbstractions.jl
@@ -12,7 +12,6 @@ import PrecompileTools
 import Atomix: @atomic, @atomicswap, @atomicreplace
 
 using MacroTools
-using StaticArrays
 using Adapt
 
 """
diff --git a/src/pocl/backend.jl b/src/pocl/backend.jl
index f23e20b0f..247b3466c 100644
--- a/src/pocl/backend.jl
+++ b/src/pocl/backend.jl
@@ -7,8 +7,6 @@ using ..POCL: device, clconvert, clfunction
 import KernelAbstractions as KA
 import KernelAbstractions.KernelIntrinsics as KI
 
-import StaticArrays
-
 import Adapt
 
 

From 0b58f9ba7f9b5bb309d0288955698078e46301f6 Mon Sep 17 00:00:00 2001
From: Valentin Churavy <v.churavy@gmail.com>
Date: Mon, 22 Jun 2026 17:38:41 +0200
Subject: [PATCH 2/2] [pocl] Use GPUCompiler.alloca for Scratchpad instead of
 MArray

Back the POCL `Scratchpad` (`@private`) with `GPUCompiler.alloca`, a direct
per-workitem stack allocation, instead of a StaticArrays `MArray`. The returned
`Ptr` is wrapped in a `CLDeviceArray` over OpenCL "Function" storage (LLVM
addrspace 0), where the SPIR-V target places allocas. Its alignment
(`Base.datatype_alignment(T)`) matches `CLDeviceArray`'s element accesses.

Requires GPUCompiler 1.23 (JuliaGPU/GPUCompiler.jl#859), which adds the
`alloca` intrinsic. Drops the now-unused StaticArrays import from the POCL
back-end (StaticArrays is still used by the CPU back-end).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---
 Project.toml        | 4 +++-
 src/pocl/backend.jl | 6 +++++-
 2 files changed, 8 insertions(+), 2 deletions(-)

diff --git a/Project.toml b/Project.toml
index 356e85a38..199a0708d 100644
--- a/Project.toml
+++ b/Project.toml
@@ -18,6 +18,7 @@ RandomNumbers = "e6cf234a-135c-5ec9-84dd-332b85af5143"
 SPIRVIntrinsics = "71d1d633-e7e8-4a92-83a1-de8814b09ba8"
 SPIRV_LLVM_Backend_jll = "4376b9bf-cff8-51b6-bb48-39421dff0d0c"
 SPIRV_Tools_jll = "6ac6d60f-d740-5983-97d7-a4482c0689f4"
+StaticArrays = "90137ffa-7385-5640-81b9-e52037218182"
 UUIDs = "cf7118a7-6976-5b1a-9a39-7adc72f591a4"
 pocl_standalone_jll = "54f56a70-6062-5590-a942-1226658f6c83"
 
@@ -35,7 +36,7 @@ SparseArraysExt = "SparseArrays"
 Adapt = "0.4, 1.0, 2.0, 3.0, 4"
 Atomix = "0.1, 1"
 EnzymeCore = "0.7, 0.8.1"
-GPUCompiler = "1.13.3"
+GPUCompiler = "1.23"
 InteractiveUtils = "1.6"
 LLVM = "9.4.1"
 LinearAlgebra = "1.6"
@@ -48,6 +49,7 @@ SPIRVIntrinsics = "1"
 SPIRV_LLVM_Backend_jll = "22"
 SPIRV_Tools_jll = "2024.4, 2025.1"
 SparseArrays = "<0.0.1, 1.6"
+StaticArrays = "0.12, 1.0"
 UUIDs = "<0.0.1, 1.6"
 julia = "1.10"
 pocl_standalone_jll = "7.1.2"
diff --git a/src/pocl/backend.jl b/src/pocl/backend.jl
index 247b3466c..ffa385f2f 100644
--- a/src/pocl/backend.jl
+++ b/src/pocl/backend.jl
@@ -220,7 +220,11 @@ end
 end
 
 @device_override @inline function KA.Scratchpad(ctx, ::Type{T}, ::Val{Dims}) where {T, Dims}
-    StaticArrays.MArray{KA.__size(Dims), T}(undef)
+    # private per-workitem scratch: a stack `alloca` (lowered by GPUCompiler) wrapped in a
+    # device array. the slot lives in OpenCL "Function" storage (LLVM addrspace 0), which is
+    # where the SPIR-V target places allocas.
+    ptr = POCL.GPUCompiler.alloca(T, Val(prod(Dims)))
+    CLDeviceArray(Dims, reinterpret(POCL.LLVMPtr{T, POCL.AS.Function}, ptr))
 end