[Libclc-dev] [PATCH] amdgcn-amdhsa: Add get_global_size() implementation

[Tom Stellard via Libclc-dev]

---
 amdgcn-amdhsa/lib/SOURCES                     |  1 +
 amdgcn-amdhsa/lib/workitem/get_global_size.ll | 36 +++++++++++++++++++++++++++
 2 files changed, 37 insertions(+)
 create mode 100644 amdgcn-amdhsa/lib/workitem/get_global_size.ll

diff --git a/amdgcn-amdhsa/lib/SOURCES b/amdgcn-amdhsa/lib/SOURCES
index b33368d..257942c 100644
--- a/amdgcn-amdhsa/lib/SOURCES
+++ b/amdgcn-amdhsa/lib/SOURCES
@@ -1 +1,2 @@
+workitem/get_global_size.ll
 workitem/get_local_size.ll
diff --git a/amdgcn-amdhsa/lib/workitem/get_global_size.ll b/amdgcn-amdhsa/lib/workitem/get_global_size.ll
new file mode 100644
index 0000000..91e419d
--- /dev/null
+++ b/amdgcn-amdhsa/lib/workitem/get_global_size.ll
@@ -0,0 +1,36 @@
+declare i8 addrspace(2)* @llvm.amdgcn.dispatch.ptr() #0
+
+define i32 @get_global_size(i32) #1 {
+  %dispatch_ptr = call noalias nonnull dereferenceable(64) i8 addrspace(2)* @llvm.amdgcn.dispatch.ptr()
+  switch i32 %0, label %14 [
+    i32 0, label %2
+    i32 1, label %6
+    i32 2, label %10
+  ]
+
+; <label>:2:                                      ; preds = %1
+  %3 = getelementptr inbounds i8, i8 addrspace(2)* %dispatch_ptr, i64 12
+  %4 = bitcast i8 addrspace(2)* %3 to i32 addrspace(2)*
+  %5 = load i32, i32 addrspace(2)* %4, align 4, !invariant.load !0
+  ret i32 %5
+
+; <label>:6:                                      ; preds = %1
+  %7 = getelementptr inbounds i8, i8 addrspace(2)* %dispatch_ptr, i64 16
+  %8 = bitcast i8 addrspace(2)* %7 to i32 addrspace(2)*
+  %9 = load i32, i32 addrspace(2)* %8, align 4, !invariant.load !0
+  ret i32 %9
+
+; <label>:10:                                     ; preds = %1
+  %11 = getelementptr inbounds i8, i8 addrspace(2)* %dispatch_ptr, i64 20
+  %12 = bitcast i8 addrspace(2)* %11 to i32 addrspace(2)*
+  %13 = load i32, i32 addrspace(2)* %12, align 4, !invariant.load !0
+  ret i32 %13
+
+; <label>:14:                                     ; preds = %1
+  ret i32 1
+}
+
+attributes #0 = { nounwind readnone }
+attributes #1 = { alwaysinline norecurse nounwind readonly }
+
+!0 = !{}
-- 
2.7.4