The variables and data structures which manage the floating point execution are all declared in the file dlx.h as part of the basic DLX structure. The variables num_add_units, num_div_units, and num_mul_units specify how many of each type of floating point execution unit are available on the machine. The variables fp_add_latency, fp_div_latency, and fp_mul_latency specify the corresponding latencies (in clock cycles) of each of the execution units. All six of these variables have default values which may be overridden via command line parameters when DLXsim is invoked.
The variable FPstatusReg is the status register which is examined on a BFPT or BFPF instruction. The various floating point set instructions (EQF, NED, etc.) write to this register.
The array fp_add_units contains the status of all the floating point adders during execution. If fp_add_units[i] is zero, adder i is available. A non-zero value means that the unit is currently performing an operation - the value specifies the clock cycle when the operation will complete. The array fp_div_units and fp_mul_units contain analogous information for the floating point dividers and multipliers. All three structures can be accessed through the array fp_units which is an array of pointers to the three execution unit status arrays.
The array waiting_FPRs contains 32 elements, corresponding to the 32 floating point registers in DLX. A zero in waiting_FPRs[i] means floating point register Fi can be read from; it contains its most current value. A non-zero value means register Fi is the destination register of a pending floating point operation (one which has issued but not yet completed). Attempting to read or write to such a register means a hazard condition exists. The non-zero value indicates the cycle at which the writeback to the register will occur.
The variable FPopsList points to the chain of pending floating point operations. Each item in this chain is of type FPop, a structure with the following fields:
The variable checkFP is a copy of the ready field of the first floating point operation on the pending operation list. If its value is zero, no floating point operations are pending. Otherwise checkFP indicates when the next (soonest) floating point operation will complete. This provides for very quick checking in the fast-path of the simulator. Only one value needs to be checked in a cycle when no writebacks should occur.
Many of the previously discussed structures refer to a clock cycle count when a particular operation will complete. The current clock cycle is kept in the variable cycleCount. This variable is incremented each time the simulator executes its main loop. It is also incremented an extra time when a load stall is detected since the floating point units are still executing during a load stall. When the cycle count reaches a large value specified by the constant CYC_CNT_RESET, cycleCount is ``reset'' back to a small number (5), and all references to clock cycles in the floating point data structures are adjusted accordingly. This operation is necessary to prevent cycleCount from overflowing, becoming negative, and thereby wreaking havoc on the sorted list of pending operations. Making cycleCount an unsigned integer does not work, since there are still problems with sorting the pending operations when cycle counts ``wrap around'' to zero.