Forbid divergent execution of work-group barriers

Author: vchuravy

src/macros.jl

@@ -81,11 +81,8 @@ function transform_gpu!(def, constargs, force_inbounds)
 end
 
 struct WorkgroupLoop
-    indicies::Vector{Any}
     stmts::Vector{Any}
     allocations::Vector{Any}
-    private_allocations::Vector{Any}
-    private::Set{Symbol}
     terminated_in_sync::Bool
 end
 
@@ -106,26 +103,18 @@ function find_sync(stmt)
 end
 
 # TODO proper handling of LineInfo
-function split(
-        stmts,
-        indicies = Any[], private = Set{Symbol}(),
-    )
+function split(stmts)
     # 1. Split the code into blocks separated by `@synchronize`
-    # 2. Aggregate `@index` expressions
-    # 3. Hoist allocations
-    # 4. Hoist uniforms
 
     current = Any[]
     allocations = Any[]
-    private_allocations = Any[]
     new_stmts = Any[]
     for stmt in stmts
         has_sync = find_sync(stmt)
         if has_sync
-            loop = WorkgroupLoop(deepcopy(indicies), current, allocations, private_allocations, deepcopy(private), is_sync(stmt))
+            loop = WorkgroupLoop(current, allocations, is_sync(stmt))
             push!(new_stmts, emit(loop))
             allocations = Any[]
-            private_allocations = Any[]
             current = Any[]
             is_sync(stmt) && continue
 
@@ -137,7 +126,7 @@ function split(
             function recurse(expr::Expr)
                 expr = unblock(expr)
                 if is_scope_construct(expr) && any(find_sync, expr.args)
-                    new_args = unblock(split(expr.args, deepcopy(indicies), deepcopy(private)))
+                    new_args = unblock(split(expr.args))
                     return Expr(expr.head, new_args...)
                 else
                     return Expr(expr.head, map(recurse, expr.args)...)
@@ -151,14 +140,10 @@ function split(
             push!(allocations, stmt)
             continue
         elseif @capture(stmt, @private lhs_ = rhs_)
-            push!(private, lhs)
-            push!(private_allocations, :($lhs = $rhs))
+            push!(allocations, :($lhs = $rhs))
             continue
         elseif @capture(stmt, lhs_ = rhs_ | (vs__, lhs_ = rhs_))
-            if @capture(rhs, @index(args__))
-                push!(indicies, stmt)
-                continue
-            elseif @capture(rhs, @localmem(args__) | @uniform(args__))
+            if @capture(rhs, @localmem(args__) | @uniform(args__))
                 push!(allocations, stmt)
                 continue
             elseif @capture(rhs, @private(T_, dims_))
@@ -170,7 +155,6 @@ function split(
                 end
                 alloc = :($Scratchpad(__ctx__, $T, Val($dims)))
                 push!(allocations, :($lhs = $alloc))
-                push!(private, lhs)
                 continue
             end
         end
@@ -180,7 +164,7 @@ function split(
 
     # everything since the last `@synchronize`
     if !isempty(current)
-        loop = WorkgroupLoop(deepcopy(indicies), current, allocations, private_allocations, deepcopy(private), false)
+        loop = WorkgroupLoop(current, allocations, false)
         push!(new_stmts, emit(loop))
     end
     return new_stmts
@@ -192,9 +176,7 @@ function emit(loop)
     body = Expr(:block, loop.stmts...)
     loopexpr = quote
         $(loop.allocations...)
-        $(loop.private_allocations...)
         if __active_lane__
-            $(loop.indicies...)
             $(unblock(body))
         end
     end