Changelog:

Your task

  1. Start from the single-cycle processor you implemented in the previous lab. Alternately, we will supply a seqlab solution (in Collab under the resources folder; after the lab is due) in case you did not finish all of the lab. Make a copy of this file called seqhw.hcl.

  2. Implement all the remaining Y86 instructions in seqhw.hcl:
    • jXX (conditional jumps)
    • mrmovq
    • pushq
    • popq
    • call
    • ret
  3. Test your code with make test-seqhw.

  4. Submit to kytos

Advice/Hints

Stat

The Stat should be

Implementing jXX

To add conditional support to JXX, you should update the PC to the immediate value (valC) only if the conditions are met. You should already have something like wire conditionsMet:1 from implementing cmovXX in lab.

Testing jXX

If jXX is correctly implemented, the following (which is found in y86/jxx.yo) should run for 19 steps, visiting hex addresses 0, a, 14, 27, 29, 1d, 14, 27, 29, 1d, 14, 27, 29 1d, 14, 27, 29, 1d, and 26, then halting at address 26:

      	irmovq $3, %rax
	irmovq $-1, %rbx
a:
	jmp b
c:
	jge a
	halt
b:
	addq %rbx, %rax
	jmp c

    

You can test this by running the output of the simulator through the grep tool to select out just a subset of lines:

      linux> ./hclrs seqhw.hcl y86/jxx.yo | grep 'pc ='
pc = 0x0; loaded [30 f0 03 00 00 00 00 00 00 00 : irmovq $0x3, %rax]
pc = 0xa; loaded [30 f3 ff ff ff ff ff ff ff ff : irmovq $0xffffffffffff, %rbx]
pc = 0x14; loaded [70 27 00 00 00 00 00 00 00 : jmp 0x27]
pc = 0x27; loaded [60 30 : addq %rbx, %rax]
pc = 0x29; loaded [70 1d 00 00 00 00 00 00 00 : jmp 0x1d]
pc = 0x1d; loaded [75 14 00 00 00 00 00 00 00 : jge 0x14]
pc = 0x14; loaded [70 27 00 00 00 00 00 00 00 : jmp 0x27]
pc = 0x27; loaded [60 30 : addq %rbx, %rax]
pc = 0x29; loaded [70 1d 00 00 00 00 00 00 00 : jmp 0x1d]
pc = 0x1d; loaded [75 14 00 00 00 00 00 00 00 : jge 0x14]
pc = 0x14; loaded [70 27 00 00 00 00 00 00 00 : jmp 0x27]
pc = 0x27; loaded [60 30 : addq %rbx, %rax]
pc = 0x29; loaded [70 1d 00 00 00 00 00 00 00 : jmp 0x1d]
pc = 0x1d; loaded [75 14 00 00 00 00 00 00 00 : jge 0x14]
pc = 0x14; loaded [70 27 00 00 00 00 00 00 00 : jmp 0x27]
pc = 0x27; loaded [60 30 : addq %rbx, %rax]
pc = 0x29; loaded [70 1d 00 00 00 00 00 00 00 : jmp 0x1d]
pc = 0x1d; loaded [75 14 00 00 00 00 00 00 00 : jge 0x14]
pc = 0x26; loaded [00 : halt]

    

Implementing mrmovq

  1. Memory is accessed by setting mem_addr to the memory address in question and either
    • setting mem_readbit to 0, mem_writebit to 1, and mem_input to the value to write to memory, which will cause the memory system to write a 4-byte value to memory; or
    • setting mem_readbit to 1 and mem_writebit to 0, which will cause the memory system to read a 4-byte value from memory into mem_output.
  2. You will also need to compute the memory address as reg_outputB + valC (the book suggests you do this in the ALU, meaning the same mux you used for OPq’s adding and subtracting). This is the same calculation used for rmmovq.

Testing mrmovq

If both memory moves are implemented correctly, the following (y86/rrmr.yo) should result in %rdx containing 0x20000 and address 0xa2 containing byte 0x02.

      mrmovq 2(%rax), %rax
rmmovq %rax, 160(%rax)
mrmovq 158(%rax), %rdx

    

The first instruction relies on %rax initially being 0.

Implementing pushq

Decode: read rA and %rsp

Execute: add −8 to %rsp

Memory: write reg_outputA to the address computed by that subtraction

Writeback: write the result of the subtraction back into %rsp

Testing pushq

The following code (y86/push.yo)

      irmovq $3, %rax
irmovq $256, %rsp
pushq %rax

    

should leave a 0x03 in address 0xf8 and an 0xf8 in %rsp

Implementing popq

Decode: read %rsp

Execute: add +8 to %rsp

Memory: read from the pre-add %rsp address (the value read in decode, not the value computed during execute)

Writeback: write both (1) the result of the addition back into %rsp and (2) the results of the read into rA

Testing popq

The following code (y86/pop.yo)

      irmovq $4, %rsp
popq %rax

    

should leave a 0xc in %rsp and a 0xfb0000000000000 in %rax

Implementing call

call is like pushq and jmp in general form

Decode: read %rsp

Execute: add −8 to %rsp

Memory: write the next instruction address (valP) to the address computed by that subtraction

Writeback: write the result of the subtraction back into %rsp

PC Update: the new PC (p_pc) should be valC, not valP.

Testing call

The following code (y86/call.yo)

      	irmovq $256, %rsp
	call a
	addq %rsp, %rsp
a:
	halt

    

should leave 0xf8 in %rsp and a 0x13 in address 0xf8

Implementing ret

ret is like popq and jmp in general form

Decode: read %rsp

Execute: add +8 to %rsp

Memory: read from the pre-added %rsp address

Writeback: write the result of the addition back into %rsp

PC Update: the new PC (p_pc) should be the value read from memory (mem_output), not valP.

Testing ret

The following code (y86/ret.yo)

      	irmovq $256, %rsp
	irmovq a, %rbx
	rmmovq %rbx, (%rsp)
	ret
	halt
a:
	irmovq $258, %rax
	halt

    

should run 6 cycles, leave %rax as 0x102, and %rsp as 0x108

General Testing

You can run the command make test-seqhw to test your processor over almost all the .yo files in the y86/ folder, comparing the output to supplied outputs in testdata/seq-reference and testdata/seq-traces. If your processor disagrees, you may find the traces in testdata/seq-traces helpful for debugging.

In addition, your code should behave the same as tools/yis when run on anything in the y86/ folder except