Back to PBT run
test.mlir
// RUN: loom-raise-opt --loom-lower-scf-to-dfg --mlir-print-op-generic %s | FileCheck %s
llvm.func @imported_callable(i64) -> i64
func.func private @native_callable(%arg_a: i32) -> i32 {
return %arg_a : i32
}
dataflow.thread private @thread_0 domain(#dataflow.thread_domain<dense>)(
%limit: index, %memory: memref<?xindex>) ctrl (%ctrl: none) {
"loom.spatial_region"(%limit, %memory)
<{operandSegmentSizes = array<i32: 1, 0, 1, 0>,
resultSegmentSizes = array<i32: 0, 0>}> ({
^bb0(%lim: index, %target: memref<?xindex>):
%zero = arith.constant 0 : index
%one = arith.constant 1 : index
%two = arith.constant 2 : index
%flag = arith.cmpi ult, %zero, %two : index
scf.parallel (%iv0) = (%zero) to (%two) step (%one) {
%mx0 = arith.muli %zero, %two : index
%ix0 = arith.addi %mx0, %iv0 : index
scf.parallel (%iv1) = (%zero) to (%two) step (%one) {
%mx1 = arith.muli %ix0, %two : index
%ix1 = arith.addi %mx1, %iv1 : index
%ld2 = memref.load %target[%ix1] : memref<?xindex>
%ad2 = arith.addi %ld2, %one : index
memref.store %ad2, %target[%ix1] : memref<?xindex>
scf.reduce
}
scf.reduce
}
"loom.spatial_yield"()
<{operandSegmentSizes = array<i32: 0, 0>}> : () -> ()
}) {graph_name = "g_thread_0", source_maps = []} :
(index, memref<?xindex>) -> ()
dataflow.thread.yield
}
dataflow.thread private @thread_1 domain(#dataflow.thread_domain<dense>)(
%limit: index, %memory: memref<?xindex>) ctrl (%ctrl: none) {
"loom.spatial_region"(%limit, %memory)
<{operandSegmentSizes = array<i32: 1, 0, 1, 0>,
resultSegmentSizes = array<i32: 0, 0>}> ({
^bb0(%lim: index, %target: memref<?xindex>):
%zero = arith.constant 0 : index
%one = arith.constant 1 : index
%two = arith.constant 2 : index
%flag = arith.cmpi ult, %zero, %two : index
%wr3 = scf.while (%wa3 = %zero) : (index) -> index {
%wc3 = arith.cmpi ult, %wa3, %two : index
scf.condition(%wc3) %wa3 : index
} do {
^bb0(%wb3: index):
%wr4 = scf.while (%wa4 = %zero) : (index) -> index {
%wc4 = arith.cmpi ult, %wa4, %two : index
scf.condition(%wc4) %wa4 : index
} do {
^bb0(%wb4: index):
memref.store %one, %target[%zero] : memref<?xindex>
%ld6 = memref.load %target[%zero] : memref<?xindex>
%ad6 = arith.addi %ld6, %one : index
memref.store %ad6, %target[%zero] : memref<?xindex>
%wn4 = arith.addi %wb4, %one : index
scf.yield %wn4 : index
}
memref.store %one, %target[%zero] : memref<?xindex>
%wn3 = arith.addi %wb3, %one : index
scf.yield %wn3 : index
}
%wr8 = scf.while (%wa8 = %zero) : (index) -> index {
%wc8 = arith.cmpi ult, %wa8, %two : index
scf.condition(%wc8) %wa8 : index
} do {
^bb0(%wb8: index):
%wr9 = scf.while (%wa9 = %zero) : (index) -> index {
%wc9 = arith.cmpi ult, %wa9, %two : index
scf.condition(%wc9) %wa9 : index
} do {
^bb0(%wb9: index):
memref.store %one, %target[%zero] : memref<?xindex>
%wn9 = arith.addi %wb9, %one : index
scf.yield %wn9 : index
}
%ld11 = memref.load %target[%zero] : memref<?xindex>
%ad11 = arith.addi %ld11, %one : index
memref.store %ad11, %target[%zero] : memref<?xindex>
%wn8 = arith.addi %wb8, %one : index
scf.yield %wn8 : index
}
"loom.spatial_yield"()
<{operandSegmentSizes = array<i32: 0, 0>}> : () -> ()
}) {graph_name = "g_thread_1", source_maps = []} :
(index, memref<?xindex>) -> ()
dataflow.thread.yield
}
// CHECK: "builtin.module"() ({
// CHECK-NEXT: "llvm.func"() <{CConv = #llvm.cconv<ccc>, function_type = !llvm.func<i64 (i64)>, linkage = #llvm.linkage<external>, sym_name = "imported_callable", unnamed_addr = 0 : i64, visibility_ = 0 : i64}> ({
// CHECK-NEXT: }) : () -> ()
// CHECK-NEXT: "func.func"() <{function_type = (i32) -> i32, sym_name = "native_callable", sym_visibility = "private"}> ({
// CHECK-NEXT: ^bb0(%arg12: i32):
// CHECK-NEXT: "func.return"(%arg12) : (i32) -> ()
// CHECK-NEXT: }) : () -> ()
// CHECK-NEXT: "dataflow.thread"() <{domain = #dataflow.thread_domain<dense>, function_type = (index, memref<?xindex>) -> (), sym_name = "thread_0", sym_visibility = "private"}> ({
// CHECK-NEXT: ^bb0(%arg9: index, %arg10: memref<?xindex>, %arg11: none):
// CHECK-NEXT: %152 = "dataflow.graph.launch"(%arg11, %arg9, %arg10) <{callee = @g_thread_0, operandSegmentSizes = array<i32: 1, 1, 0, 1, 0>, resultSegmentSizes = array<i32: 0, 0, 1>, source_maps = []}> : (none, index, memref<?xindex>) -> none
// CHECK-NEXT: "dataflow.thread.yield"(%152) : (none) -> ()
// CHECK-NEXT: }) : () -> ()
// CHECK-NEXT: "dataflow.thread"() <{domain = #dataflow.thread_domain<dense>, function_type = (index, memref<?xindex>) -> (), sym_name = "thread_1", sym_visibility = "private"}> ({
// CHECK-NEXT: ^bb0(%arg6: index, %arg7: memref<?xindex>, %arg8: none):
// CHECK-NEXT: %151 = "dataflow.graph.launch"(%arg8, %arg6, %arg7) <{callee = @g_thread_1, operandSegmentSizes = array<i32: 1, 1, 0, 1, 0>, resultSegmentSizes = array<i32: 0, 0, 1>, source_maps = []}> : (none, index, memref<?xindex>) -> none
// CHECK-NEXT: "dataflow.thread.yield"(%151) : (none) -> ()
// CHECK-NEXT: }) : () -> ()
// CHECK-NEXT: "dataflow.graph"() <{arg_attrs = [{}, {llvm.noalias}], function_type = (index, memref<?xindex>) -> (), input_segments = array<i32: 1, 0, 1>, result_segments = array<i32: 0, 0, 0>, sym_name = "g_thread_0", sym_visibility = "private"}> ({
// CHECK-NEXT: ^bb0(%arg3: none, %arg4: index, %arg5: memref<?xindex>):
// CHECK-NEXT: %127 = "dataflow.constant"(%arg3) <{const_value = 1 : index}> : (none) -> index
// CHECK-NEXT: %128 = "dataflow.constant"(%arg3) <{const_value = 2 : index}> : (none) -> index
// CHECK-NEXT: %129 = "dataflow.constant"(%arg3) <{const_value = 0 : index}> : (none) -> index
// CHECK-NEXT: %130 = "arith.muli"(%129, %128) <{overflowFlags = #arith.overflow<none>}> : (index, index) -> index
// CHECK-NEXT: %131 = "arith.addi"(%130, %129) <{overflowFlags = #arith.overflow<none>}> : (index, index) -> index
// CHECK-NEXT: %132 = "arith.addi"(%142#0, %127) <{overflowFlags = #arith.overflow<none>}> : (index, index) -> index
// CHECK-NEXT: %133 = "arith.muli"(%129, %128) <{overflowFlags = #arith.overflow<none>}> : (index, index) -> index
// CHECK-NEXT: %134 = "arith.addi"(%133, %127) <{overflowFlags = #arith.overflow<none>}> : (index, index) -> index
// CHECK-NEXT: %135 = "arith.addi"(%144#0, %127) <{overflowFlags = #arith.overflow<none>}> : (index, index) -> index
// CHECK-NEXT: %136 = "arith.muli"(%127, %128) <{overflowFlags = #arith.overflow<none>}> : (index, index) -> index
// CHECK-NEXT: %137 = "arith.addi"(%136, %129) <{overflowFlags = #arith.overflow<none>}> : (index, index) -> index
// CHECK-NEXT: %138 = "arith.addi"(%146#0, %127) <{overflowFlags = #arith.overflow<none>}> : (index, index) -> index
// CHECK-NEXT: %139 = "arith.muli"(%127, %128) <{overflowFlags = #arith.overflow<none>}> : (index, index) -> index
// CHECK-NEXT: %140 = "arith.addi"(%139, %127) <{overflowFlags = #arith.overflow<none>}> : (index, index) -> index
// CHECK-NEXT: %141 = "arith.addi"(%148#0, %127) <{overflowFlags = #arith.overflow<none>}> : (index, index) -> index
// CHECK-NEXT: %142:2 = "dataflow.load"(%arg5, %131, %arg3) : (memref<?xindex>, index, none) -> (index, none)
// CHECK-NEXT: %143 = "dataflow.store"(%arg5, %131, %132, %142#1) : (memref<?xindex>, index, index, none) -> none
// CHECK-NEXT: %144:2 = "dataflow.load"(%arg5, %134, %arg3) : (memref<?xindex>, index, none) -> (index, none)
// CHECK-NEXT: %145 = "dataflow.store"(%arg5, %134, %135, %144#1) : (memref<?xindex>, index, index, none) -> none
// CHECK-NEXT: %146:2 = "dataflow.load"(%arg5, %137, %arg3) : (memref<?xindex>, index, none) -> (index, none)
// CHECK-NEXT: %147 = "dataflow.store"(%arg5, %137, %138, %146#1) : (memref<?xindex>, index, index, none) -> none
// CHECK-NEXT: %148:2 = "dataflow.load"(%arg5, %140, %arg3) : (memref<?xindex>, index, none) -> (index, none)
// CHECK-NEXT: %149 = "dataflow.store"(%arg5, %140, %141, %148#1) : (memref<?xindex>, index, index, none) -> none
// CHECK-NEXT: %150:4 = "dataflow.sync"(%143, %145, %147, %149) : (none, none, none, none) -> (none, none, none, none)
// CHECK-NEXT: "dataflow.graph.return"(%150#0) <{operandSegmentSizes = array<i32: 0, 0, 0, 1>}> : (none) -> ()
// CHECK-NEXT: }) : () -> ()
// CHECK-NEXT: "dataflow.graph"() <{arg_attrs = [{}, {llvm.noalias}], function_type = (index, memref<?xindex>) -> (), input_segments = array<i32: 1, 0, 1>, result_segments = array<i32: 0, 0, 0>, sym_name = "g_thread_1", sym_visibility = "private"}> ({
// CHECK-NEXT: ^bb0(%arg0: none, %arg1: index, %arg2: memref<?xindex>):
// CHECK-NEXT: %0 = "dataflow.constant"(%arg0) <{const_value = 0 : index}> : (none) -> index
// CHECK-NEXT: %1 = "dataflow.constant"(%arg0) <{const_value = 1 : index}> : (none) -> index
// CHECK-NEXT: %2 = "dataflow.constant"(%arg0) <{const_value = 2 : index}> : (none) -> index
// CHECK-NEXT: %3 = "dataflow.carry"(%8, %arg0, %51#0) : (i1, none, none) -> none
// CHECK-NEXT: %4 = "dataflow.carry"(%8, %0, %54) : (i1, index, index) -> index
// CHECK-NEXT: %5 = "dataflow.invariant"(%8, %2) : (i1, index) -> index
// CHECK-NEXT: %6 = "dataflow.carry"(%8, %arg0, %53) : (i1, none, none) -> none
// CHECK-NEXT: %7 = "dataflow.carry"(%8, %arg0, %53) : (i1, none, none) -> none
// CHECK-NEXT: %8 = "arith.cmpi"(%4, %5) <{predicate = 6 : i64}> : (index, index) -> i1
// CHECK-NEXT: %9:2 = "dataflow.demux"(%8, %3) : (i1, none) -> (none, none)
// CHECK-NEXT: %10:2 = "dataflow.gate"(%8, %3) : (i1, none) -> (i1, none)
// CHECK-NEXT: %11:2 = "dataflow.demux"(%10#0, %10#1) : (i1, none) -> (none, none)
// CHECK-NEXT: %12:2 = "dataflow.demux"(%8, %6) : (i1, none) -> (none, none)
// CHECK-NEXT: %13:2 = "dataflow.demux"(%8, %7) : (i1, none) -> (none, none)
// CHECK-NEXT: %14:2 = "dataflow.demux"(%8, %4) : (i1, index) -> (index, index)
// CHECK-NEXT: %15 = "dataflow.invariant"(%8, %2) : (i1, index) -> index
// CHECK-NEXT: %16:2 = "dataflow.gate"(%8, %15) : (i1, index) -> (i1, index)
// CHECK-NEXT: %17:2 = "dataflow.demux"(%16#0, %16#1) : (i1, index) -> (index, index)
// CHECK-NEXT: %18 = "dataflow.invariant"(%8, %1) : (i1, index) -> index
// CHECK-NEXT: %19:2 = "dataflow.gate"(%8, %18) : (i1, index) -> (i1, index)
// CHECK-NEXT: %20:2 = "dataflow.demux"(%19#0, %19#1) : (i1, index) -> (index, index)
// CHECK-NEXT: %21 = "dataflow.invariant"(%8, %0) : (i1, index) -> index
// CHECK-NEXT: %22:2 = "dataflow.gate"(%8, %21) : (i1, index) -> (i1, index)
// CHECK-NEXT: %23:2 = "dataflow.demux"(%22#0, %22#1) : (i1, index) -> (index, index)
// CHECK-NEXT: %24 = "dataflow.carry"(%28, %10#1, %30#1) : (i1, none, none) -> none
// CHECK-NEXT: %25 = "dataflow.carry"(%28, %22#1, %45) : (i1, index, index) -> index
// CHECK-NEXT: %26 = "dataflow.invariant"(%28, %16#1) : (i1, index) -> index
// CHECK-NEXT: %27 = "dataflow.carry"(%28, %13#1, %44) : (i1, none, none) -> none
// CHECK-NEXT: %28 = "arith.cmpi"(%25, %26) <{predicate = 6 : i64}> : (index, index) -> i1
// CHECK-NEXT: %29:2 = "dataflow.demux"(%28, %24) : (i1, none) -> (none, none)
// CHECK-NEXT: %30:2 = "dataflow.gate"(%28, %24) : (i1, none) -> (i1, none)
// CHECK-NEXT: %31:2 = "dataflow.demux"(%30#0, %30#1) : (i1, none) -> (none, none)
// CHECK-NEXT: %32:2 = "dataflow.demux"(%28, %27) : (i1, none) -> (none, none)
// CHECK-NEXT: %33:2 = "dataflow.demux"(%28, %25) : (i1, index) -> (index, index)
// CHECK-NEXT: %34 = "dataflow.invariant"(%28, %19#1) : (i1, index) -> index
// CHECK-NEXT: %35:2 = "dataflow.gate"(%28, %34) : (i1, index) -> (i1, index)
// CHECK-NEXT: %36:2 = "dataflow.demux"(%35#0, %35#1) : (i1, index) -> (index, index)
// CHECK-NEXT: %37 = "dataflow.invariant"(%28, %22#1) : (i1, index) -> index
// CHECK-NEXT: %38:2 = "dataflow.gate"(%28, %37) : (i1, index) -> (i1, index)
// CHECK-NEXT: %39:2 = "dataflow.demux"(%38#0, %38#1) : (i1, index) -> (index, index)
// CHECK-NEXT: %40:2 = "dataflow.sync"(%30#1, %32#1) : (none, none) -> (none, none)
// CHECK-NEXT: %41 = "dataflow.store"(%arg2, %38#1, %35#1, %40#0) : (memref<?xindex>, index, index, none) -> none
// CHECK-NEXT: %42:2 = "dataflow.load"(%arg2, %38#1, %41) : (memref<?xindex>, index, none) -> (index, none)
// CHECK-NEXT: %43 = "arith.addi"(%42#0, %35#1) <{overflowFlags = #arith.overflow<none>}> : (index, index) -> index
// CHECK-NEXT: %44 = "dataflow.store"(%arg2, %38#1, %43, %42#1) : (memref<?xindex>, index, index, none) -> none
// CHECK-NEXT: %45 = "arith.addi"(%33#1, %35#1) <{overflowFlags = #arith.overflow<none>}> : (index, index) -> index
// CHECK-NEXT: %46:2 = "dataflow.demux"(%30#0, %30#1) : (i1, none) -> (none, none)
// CHECK-NEXT: %47:4 = "dataflow.sync"(%46#0, %31#0, %36#0, %39#0) : (none, none, index, index) -> (none, none, index, index)
// CHECK-NEXT: %48 = "dataflow.mux"(%30#0, %47#0, %46#1) : (i1, none, none) -> none
// CHECK-NEXT: %49 = "dataflow.carry"(%28, %10#1, %48) : (i1, none, none) -> none
// CHECK-NEXT: %50:2 = "dataflow.demux"(%28, %49) : (i1, none) -> (none, none)
// CHECK-NEXT: %51:2 = "dataflow.sync"(%29#0, %50#0) : (none, none) -> (none, none)
// CHECK-NEXT: %52:2 = "dataflow.sync"(%51#0, %32#0) : (none, none) -> (none, none)
// CHECK-NEXT: %53 = "dataflow.store"(%arg2, %22#1, %19#1, %52#0) : (memref<?xindex>, index, index, none) -> none
// CHECK-NEXT: %54 = "arith.addi"(%14#1, %19#1) <{overflowFlags = #arith.overflow<none>}> : (index, index) -> index
// CHECK-NEXT: %55:2 = "dataflow.demux"(%10#0, %51#0) : (i1, none) -> (none, none)
// CHECK-NEXT: %56:2 = "dataflow.sync"(%55#0, %11#0) : (none, none) -> (none, none)
// CHECK-NEXT: %57:2 = "dataflow.sync"(%56#0, %17#0) : (none, index) -> (none, index)
// CHECK-NEXT: %58:2 = "dataflow.sync"(%20#0, %23#0) : (index, index) -> (index, index)
// CHECK-NEXT: %59:2 = "dataflow.sync"(%57#0, %58#0) : (none, index) -> (none, index)
// CHECK-NEXT: %60 = "dataflow.mux"(%10#0, %59#0, %55#1) : (i1, none, none) -> none
// CHECK-NEXT: %61 = "dataflow.carry"(%8, %arg0, %60) : (i1, none, none) -> none
// CHECK-NEXT: %62:2 = "dataflow.demux"(%8, %61) : (i1, none) -> (none, none)
// CHECK-NEXT: %63:2 = "dataflow.sync"(%9#0, %62#0) : (none, none) -> (none, none)
// CHECK-NEXT: %64 = "dataflow.carry"(%69, %63#0, %111#0) : (i1, none, none) -> none
// CHECK-NEXT: %65 = "dataflow.carry"(%69, %0, %117) : (i1, index, index) -> index
// CHECK-NEXT: %66 = "dataflow.invariant"(%69, %2) : (i1, index) -> index
// CHECK-NEXT: %67 = "dataflow.carry"(%69, %12#0, %116) : (i1, none, none) -> none
// CHECK-NEXT: %68 = "dataflow.carry"(%69, %13#0, %116) : (i1, none, none) -> none
// CHECK-NEXT: %69 = "arith.cmpi"(%65, %66) <{predicate = 6 : i64}> : (index, index) -> i1
// CHECK-NEXT: %70:2 = "dataflow.demux"(%69, %64) : (i1, none) -> (none, none)
// CHECK-NEXT: %71:2 = "dataflow.gate"(%69, %64) : (i1, none) -> (i1, none)
// CHECK-NEXT: %72:2 = "dataflow.demux"(%71#0, %71#1) : (i1, none) -> (none, none)
// CHECK-NEXT: %73:2 = "dataflow.demux"(%69, %67) : (i1, none) -> (none, none)
// CHECK-NEXT: %74:2 = "dataflow.demux"(%69, %68) : (i1, none) -> (none, none)
// CHECK-NEXT: %75:2 = "dataflow.demux"(%69, %65) : (i1, index) -> (index, index)
// CHECK-NEXT: %76 = "dataflow.invariant"(%69, %2) : (i1, index) -> index
// CHECK-NEXT: %77:2 = "dataflow.gate"(%69, %76) : (i1, index) -> (i1, index)
// CHECK-NEXT: %78:2 = "dataflow.demux"(%77#0, %77#1) : (i1, index) -> (index, index)
// CHECK-NEXT: %79 = "dataflow.invariant"(%69, %1) : (i1, index) -> index
// CHECK-NEXT: %80:2 = "dataflow.gate"(%69, %79) : (i1, index) -> (i1, index)
// CHECK-NEXT: %81:2 = "dataflow.demux"(%80#0, %80#1) : (i1, index) -> (index, index)
// CHECK-NEXT: %82 = "dataflow.invariant"(%69, %0) : (i1, index) -> index
// CHECK-NEXT: %83:2 = "dataflow.gate"(%69, %82) : (i1, index) -> (i1, index)
// CHECK-NEXT: %84:2 = "dataflow.demux"(%83#0, %83#1) : (i1, index) -> (index, index)
// CHECK-NEXT: %85 = "dataflow.carry"(%90, %71#1, %92#1) : (i1, none, none) -> none
// CHECK-NEXT: %86 = "dataflow.carry"(%90, %83#1, %105) : (i1, index, index) -> index
// CHECK-NEXT: %87 = "dataflow.invariant"(%90, %77#1) : (i1, index) -> index
// CHECK-NEXT: %88 = "dataflow.carry"(%90, %73#1, %104) : (i1, none, none) -> none
// CHECK-NEXT: %89 = "dataflow.carry"(%90, %74#1, %104) : (i1, none, none) -> none
// CHECK-NEXT: %90 = "arith.cmpi"(%86, %87) <{predicate = 6 : i64}> : (index, index) -> i1
// CHECK-NEXT: %91:2 = "dataflow.demux"(%90, %85) : (i1, none) -> (none, none)
// CHECK-NEXT: %92:2 = "dataflow.gate"(%90, %85) : (i1, none) -> (i1, none)
// CHECK-NEXT: %93:2 = "dataflow.demux"(%92#0, %92#1) : (i1, none) -> (none, none)
// CHECK-NEXT: %94:2 = "dataflow.demux"(%90, %88) : (i1, none) -> (none, none)
// CHECK-NEXT: %95:2 = "dataflow.demux"(%90, %89) : (i1, none) -> (none, none)
// CHECK-NEXT: %96:2 = "dataflow.demux"(%90, %86) : (i1, index) -> (index, index)
// CHECK-NEXT: %97 = "dataflow.invariant"(%90, %80#1) : (i1, index) -> index
// CHECK-NEXT: %98:2 = "dataflow.gate"(%90, %97) : (i1, index) -> (i1, index)
// CHECK-NEXT: %99:2 = "dataflow.demux"(%98#0, %98#1) : (i1, index) -> (index, index)
// CHECK-NEXT: %100 = "dataflow.invariant"(%90, %83#1) : (i1, index) -> index
// CHECK-NEXT: %101:2 = "dataflow.gate"(%90, %100) : (i1, index) -> (i1, index)
// CHECK-NEXT: %102:2 = "dataflow.demux"(%101#0, %101#1) : (i1, index) -> (index, index)
// CHECK-NEXT: %103:2 = "dataflow.sync"(%92#1, %95#1) : (none, none) -> (none, none)
// CHECK-NEXT: %104 = "dataflow.store"(%arg2, %101#1, %98#1, %103#0) : (memref<?xindex>, index, index, none) -> none
// CHECK-NEXT: %105 = "arith.addi"(%96#1, %98#1) <{overflowFlags = #arith.overflow<none>}> : (index, index) -> index
// CHECK-NEXT: %106:2 = "dataflow.demux"(%92#0, %92#1) : (i1, none) -> (none, none)
// CHECK-NEXT: %107:4 = "dataflow.sync"(%106#0, %93#0, %99#0, %102#0) : (none, none, index, index) -> (none, none, index, index)
// CHECK-NEXT: %108 = "dataflow.mux"(%92#0, %107#0, %106#1) : (i1, none, none) -> none
// CHECK-NEXT: %109 = "dataflow.carry"(%90, %71#1, %108) : (i1, none, none) -> none
// CHECK-NEXT: %110:2 = "dataflow.demux"(%90, %109) : (i1, none) -> (none, none)
// CHECK-NEXT: %111:2 = "dataflow.sync"(%91#0, %110#0) : (none, none) -> (none, none)
// CHECK-NEXT: %112:2 = "dataflow.sync"(%111#0, %94#0) : (none, none) -> (none, none)
// CHECK-NEXT: %113:2 = "dataflow.load"(%arg2, %83#1, %112#0) : (memref<?xindex>, index, none) -> (index, none)
// CHECK-NEXT: %114:2 = "dataflow.sync"(%95#0, %113#1) : (none, none) -> (none, none)
// CHECK-NEXT: %115 = "arith.addi"(%113#0, %80#1) <{overflowFlags = #arith.overflow<none>}> : (index, index) -> index
// CHECK-NEXT: %116 = "dataflow.store"(%arg2, %83#1, %115, %114#0) : (memref<?xindex>, index, index, none) -> none
// CHECK-NEXT: %117 = "arith.addi"(%75#1, %80#1) <{overflowFlags = #arith.overflow<none>}> : (index, index) -> index
// CHECK-NEXT: %118:2 = "dataflow.demux"(%71#0, %111#0) : (i1, none) -> (none, none)
// CHECK-NEXT: %119:2 = "dataflow.sync"(%118#0, %72#0) : (none, none) -> (none, none)
// CHECK-NEXT: %120:2 = "dataflow.sync"(%119#0, %78#0) : (none, index) -> (none, index)
// CHECK-NEXT: %121:2 = "dataflow.sync"(%81#0, %84#0) : (index, index) -> (index, index)
// CHECK-NEXT: %122:2 = "dataflow.sync"(%120#0, %121#0) : (none, index) -> (none, index)
// CHECK-NEXT: %123 = "dataflow.mux"(%71#0, %122#0, %118#1) : (i1, none, none) -> none
// CHECK-NEXT: %124 = "dataflow.carry"(%69, %63#0, %123) : (i1, none, none) -> none
// CHECK-NEXT: %125:2 = "dataflow.demux"(%69, %124) : (i1, none) -> (none, none)
// CHECK-NEXT: %126:2 = "dataflow.sync"(%70#0, %125#0) : (none, none) -> (none, none)
// CHECK-NEXT: "dataflow.graph.return"(%126#0, %74#0) <{operandSegmentSizes = array<i32: 0, 0, 0, 2>}> : (none, none) -> ()
// CHECK-NEXT: }) : () -> ()
// CHECK-NEXT: }) : () -> ()
// CHECK-EMPTY: