swap order of implicit kernel arguments for better alignment

Both global_id_offset and enqueued_local_size are <3 x i32> implicit
kernel arguments always added when kernel uses global_id. Both take 12
bytes, but if following argument requires 8-byte alignment, an
additional 4 bytes of padding might be added. If both global_id_offset
and enqueued_local_size are reordered to be next to each other, they
will take 24 bytes, which offers better alignment for following
arguments.

Author: pkwasnie-intel

IGC/Compiler/Optimizer/OpenCLPasses/KernelArgs/KernelArgs.cpp

@@ -810,6 +810,7 @@ KernelArgsOrder::KernelArgsOrder(InputType layout)
             KernelArg::ArgType::RUNTIME_VALUE,
             KernelArg::ArgType::IMPLICIT_PAYLOAD_HEADER,
             KernelArg::ArgType::IMPLICIT_GLOBAL_OFFSET,
+            KernelArg::ArgType::IMPLICIT_ENQUEUED_LOCAL_WORK_SIZE,
 
             KernelArg::ArgType::PTR_LOCAL,
             KernelArg::ArgType::PTR_GLOBAL,
@@ -831,7 +832,6 @@ KernelArgsOrder::KernelArgsOrder(InputType layout)
             KernelArg::ArgType::IMPLICIT_LOCAL_SIZE,
             KernelArg::ArgType::IMPLICIT_STAGE_IN_GRID_ORIGIN,
             KernelArg::ArgType::IMPLICIT_STAGE_IN_GRID_SIZE,
-            KernelArg::ArgType::IMPLICIT_ENQUEUED_LOCAL_WORK_SIZE,
             KernelArg::ArgType::IMPLICIT_BINDLESS_OFFSET,
 
             KernelArg::ArgType::IMPLICIT_IMAGE_HEIGHT,
@@ -937,6 +937,7 @@ KernelArgsOrder::KernelArgsOrder(InputType layout)
             KernelArg::ArgType::RUNTIME_VALUE,
             KernelArg::ArgType::IMPLICIT_PAYLOAD_HEADER,
             KernelArg::ArgType::IMPLICIT_GLOBAL_OFFSET,
+            KernelArg::ArgType::IMPLICIT_ENQUEUED_LOCAL_WORK_SIZE,
             KernelArg::ArgType::PTR_LOCAL,
             KernelArg::ArgType::PTR_GLOBAL,
             KernelArg::ArgType::PTR_CONSTANT,
@@ -956,7 +957,6 @@ KernelArgsOrder::KernelArgsOrder(InputType layout)
             KernelArg::ArgType::IMPLICIT_LOCAL_SIZE,
             KernelArg::ArgType::IMPLICIT_STAGE_IN_GRID_ORIGIN,
             KernelArg::ArgType::IMPLICIT_STAGE_IN_GRID_SIZE,
-            KernelArg::ArgType::IMPLICIT_ENQUEUED_LOCAL_WORK_SIZE,
             KernelArg::ArgType::IMPLICIT_BINDLESS_OFFSET,
 
             KernelArg::ArgType::IMPLICIT_ARG_BUFFER,

IGC/ocloc_tests/features/constant_buffer/noinline.cl

@@ -19,16 +19,16 @@ SPDX-License-Identifier: MIT
 // CHECK-SUBROUTINE-NEXT:     - arg_type:        global_id_offset
 // CHECK-SUBROUTINE-NEXT:       offset:          0
 // CHECK-SUBROUTINE-NEXT:       size:            12
+// CHECK-SUBROUTINE-NEXT:     - arg_type:        enqueued_local_size
+// CHECK-SUBROUTINE-NEXT:       offset:          12
+// CHECK-SUBROUTINE-NEXT:       size:            12
 // CHECK-SUBROUTINE-NEXT:     - arg_type:        arg_bypointer
-// CHECK-SUBROUTINE-NEXT:       offset:          16
+// CHECK-SUBROUTINE-NEXT:       offset:          24
 // CHECK-SUBROUTINE-NEXT:       size:            8
 // CHECK-SUBROUTINE-NEXT:       arg_index:       0
 // CHECK-SUBROUTINE-NEXT:       addrmode:        stateless
 // CHECK-SUBROUTINE-NEXT:       addrspace:       global
 // CHECK-SUBROUTINE-NEXT:       access_type:     readwrite
-// CHECK-SUBROUTINE-NEXT:     - arg_type:        enqueued_local_size
-// CHECK-SUBROUTINE-NEXT:       offset:          24
-// CHECK-SUBROUTINE-NEXT:       size:            12
 // CHECK-SUBROUTINE-NEXT:   per_thread_payload_arguments:
 // CHECK-SUBROUTINE-NEXT:     - arg_type:        local_id
 // CHECK-SUBROUTINE-NEXT:       offset:          0