shetaye/omega-cpu

A 32-bit RISC CPU written in VHDL

omega-cpu's Issues

Race condition in output to register 30.

In the operations that write to register 30 as a status register (carry bit), if register 30 is also designated as the operation's "output" register, it is unclear what value actually gets written.

This should be clarified, probably by removing the ability to write to register 30 as a general-purpose register.

Add support for different output targets in assembler.

Currently, the assembler produces its machine-code output in lines of binary originally intended for a GHDL test bench. A post-process script exists to convert it to VHDL array format. There is no provision for straight binary output.

The assembler itself should support these three output formats, as it was originally designed. This can be done by adding new visitor classes alongside TextOutputVisitor and adding a command-line switch to choose among them when the finalOutput function is called:
https://github.com/Hyperion302/omega-cpu/blob/e6e8a70d20045856365634b33294ff635fceb88d/Assembler/OmegaAssembler.py#L963

Make 110111 instructions invalid.

Any instruction whose upper 6 bits are 110111 is treated as a NOP. These should raise an "invalid instruction" trap instead.

What does opcode 111 do?

It is unclear what happens when the processor encounters an instruction with the "reserved" opcode 111. This should be clarified and an "invalid instruction" trap raised on such an instruction.

Register 29 is used in two incompatible ways.

Currently, whenever an interrupt is triggered and serviced, the return address is supposed to be written to register 29.

However, register 29 is also usable as a general-purpose register and is designated by the calling convention as holding the return address from ordinary functions. This will cause a problem whenever an interrupt is serviced in the middle of a function call, wiping out the return address.

Assembler '.data' directive should support address operand.

The .data directive in the assembler is designed to support an address operand indicating where to put the segment, the same as the .text directive does. However, the assembler raises an error when an address is provided.

DIVI incorrectly uses divide-mode bit.

The divide-immediate instruction (DIVI), like its register counterpart, uses the least significant bit of its instruction as the "divide mode" bit to select between signed and unsigned mode.

However, this bit is also part of the immediate value, so any division by an even number is done in unsigned mode and any division by an odd number in signed mode.

To resolve this, DIVI must ignore the DM bit and operate exclusively in one mode or the other.
The documentation currently designates it as operating in signed mode.

Add assembler alerts for segment overlap.

When segment addresses are explicitly specified in the assembler, there may potentially be two or more instructions assigned to the same memory address (if, for example, two segments are given the same address).

Currently, the assembler handles this by the deterministic, but potentially confusing route of silently disambiguating in favor of the last instruction or directive, in document order. It might improve usability if the assembler alerted the user to any overlap.

Word alignment is not enforced.

It is possible to assign a value to the program counter that is not a multiple of 4, whether by outputting directly to register 31, or by using a register-mode jump instruction. Since instructions are always at word-aligned addresses (multiples of 4), this would cause some highly unexpected behavior.

This should be prevented by fixing the lower two bits of the PC permanently at zero.

Upper word of multiply product is unused.

When performing a multiplication, the ALU outputs the upper 32 bits of the product to the "register D" output, but this value goes unused by the controller. The R[D] field should be added to the multiply instruction to provide an outlet for this value.

Eliminate duplicate opcodes for SLLV, SLL, and EQ.

The instructions SLLV, SLL, and EQ` are usable with two different opcodes, as shown in Table 6 of the documentation. The duplicate opcodes, which the assembler does not use, should raise an "invalid instruction" trap instead.

Mangle label names in library functions.

The label names in the functions in the lib directory are of a sort that could be easily used in a user program (e.g., end), which is a problem since these labels are in a global namespace and hence have to be globally unique.

This problem could be alleviated by mangling the names to something much less likely to be used in a user program (e.g., print_string__end).

Traps are ignored.

Currently, although trap events such as divide overflow are raised within the processor, their handlers are no-ops and the program has no way of knowing that an error occurred. Eventually, these should be handled as interrupts.