6502 Timing of Interrupt Handling

From VisualChips

Jump to: navigation, search

This page contains some as-yet unpolished extracts from postings by user Hydrophilic on commodore128.org, material used by permission.

This page contains work in progress and unanswered questions, which should be answered by reference to visual6502 simulation URLs.


Interrupt handling sequence

The 6502 performs an interrupt as a 7-cycle instruction sequence which starts with an instruction fetch. (Is this true when the interrupted instruction is a branch?) The fetched instruction is substituted by a BRK in the IR.

  1. IRQ during INC showing the latest point at which an IRQ will affect the next instruction.
  2. IRQ during SEI which does take effect before the I bit is set
  3. IRQ one half-cycle later during SEI which does not take effect: I has been set and masks the interrupt

Interrupts colliding

  • this needs to be studied and verified. Observations by Hydrophilic follow

This simulation shows a lost NMI. NMI is brought low when doing an IRQ acknowledge. Specifically, 1/2 cycle before fetching the IRQ vector (cycle 13 phase 2). NMI remains low for 2.5 cycles. NMI returns high on cycle 16 phase 1.

The NMI is never serviced. This *might* be due #NMIP being automatically cleared after fetching PC high during any interrupt response...

Interrupts and changes to I mask bit

Instructions such as SEI and CLI affect the status register during the following instruction, due the the 6502 pipelining. Therefore the masking of the interrupt does not take place until the following instruction is already underway. However, RTI restores the status register early, and so the restored I mask bit already is in effect for the next instruction.

  1. IRQ during CLI (IRQ has no effect on following instruction, interrupt occurs on the one after that.) Described as "effect of CLI is delayed by one opcode"

Interrupts during branches

Some simulations for discussion

  1. late NMI (NMI during IRQ handling, causing NMI vector to be fetched and followed.)
  2. later NMI (NMI during IRQ handling, just prior to vector fetch, too late to usurp the IRQ, does not even interrupt the first instruction of the IRQ handler, but the second one.)
 Perhaps this is because 'doIRQ' (line 480) is not set during the last vector fetch; that is, during cycle 15 when fetching $FFFF, 'doIRQ' is still false.  I don't know why, considering both INTG and 'normal' (line 629) have 'standard' values and there is NMI pending.
 After starting to work on INX, 'doIRQ' finally gets set correctly so that 'normal' and INTG get triggered at the end of its execution (cycle 18).  And then finally (cycle 19) the NMI is processed
  1. lost NMI (NMI during IRQ handling, showing that 2.5 cycles is too short for an NMI to be serviced if it falls during this critical time of fetching the IRQ vector)
  2. IRQ delayed by branch. As can be seen from previous simulations, the CPU will normally examine 'do IRQ' on the last cycle of an instruction and if it is set, will clear 'normal' (line 629) and set INTG.

However, in the last cycle of BNE (branch taken, no page cross), although the IRQ has been asserted and 'do IRQ' has been set true (cycle 6), the 'normal' and INTG lines are not updated. So the CPU continues with the next instruction, also BNE. Again 'do IRQ' is examined and the 'normal' and INTG are updated, but not during the last cycle of the instruction (as per MOS Tech specs) but actually during the next-to-last cycle (see cycle 13). Note if the branch were not taken, it would be the last cycle of the instruction.

Perhaps the designers considered cycle 2 of any branch instruction to be the 'natural' end and check 'do IRQ' there... and only there... unless a page boundary is crossed.

Now that I think of it, the fact that INTG is set on the second cycle of any branch instruction (regardless if branch is taken or not), means this line is set 1 cycles earlier than normal if the branch does get taken, and 2 cycles earlier than normal if the branch crosses a page boundary.

(The above commentary, as a pastebomb, should be revisited and tidied up and reconciled with other sources. If we fail to explain, we can remove our explanations and leave only our evidence.)


Personal tools