Carry output in BarrelShifter

[gixslayer]

For our course where student design a processor I'm considering letting students implement a logical shift left and logical shift right (they already implement a rotate left). This is easy enough to accomplish with the BarrelShifter component. The problem is I'd also like to emulate the behavior of ARM/x86 in terms of the carry flag (last bit rotated/shifted out, unless a 0 rotation/shift was given). Doing this in circuits is possible, but trickier/messier than I'd like for these shifts. I therefore see the following options:

1. Have a BarrelShifter (variant) that outputs the desired carry bit natively in Digital, see here for an example implementation in terms of desired logic.
2. Implement and distribute a custom BarrelShifter variant.
3. Provide them with helper circuits that implement this.

Option 1 would be the nicest from our perspective, though I don't know if there is any desire for such a component, or if it would just clutter the library. Option 2 is undesirable as we currently only use stock Digital features, and I'd rather not maintain and push extra jar components to students for just this feature. Option 3 would would be acceptable, but create a bit more clutter at the circuit level compared to option 1.

What do you think of such a carry flag in the BarrelShifter? A useful addition, or too specific to our case?

[TheWildJarvi]

I think you should implement the shifter using wire splitters. It's really straight forward. You should have a underflow and overflow flag as well. I can post a circuit if more help is needed. The built-in barrel shifter is not synthesizable to HDL. - TWJ

…

________________________________ From: gixslayer ***@***.***> Sent: Monday, January 2, 2023 12:18:46 PM To: hneemann/Digital ***@***.***> Cc: Subscribed ***@***.***> Subject: [hneemann/Digital] Carry output in BarrelShifter (Discussion #1089) For our course where student design a processor I'm considering letting students implement a logical shift left and logical shift right (they already implement a rotate left). This is easy enough to accomplish with the BarrelShifter component. The problem is I'd also like to emulate the behavior of ARM/x86 in terms of the carry flag (last bit rotated/shifted out, unless a 0 rotation/shift was given). Doing this in circuits is possible, but trickier/messier than I'd like for these shifts. I therefore see the following options: 1. Have a BarrelShifter (variant) that outputs the desired carry bit natively in Digital, see here<gixslayer@2f68fb0> for an example implementation in terms of desired logic. 2. Implement and distribute a custom BarrelShifter variant. 3. Provide them with helper circuits that implement this. Option 1 would be the nicest from our perspective, though I don't know if there is any desire for such a component, or if it would just clutter the library. Option 2 is undesirable as we currently only use stock Digital features, and I'd rather not maintain and push extra jar components to students for just this feature. Option 3 would would be acceptable, but create a bit more clutter at the circuit level compared to option 1. What do you think of such a carry flag in the BarrelShifter? A useful addition, or too specific to our case? — Reply to this email directly, view it on GitHub<#1089>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AKQPU32DT63MKXX5STYTPU3WQMEXNANCNFSM6AAAAAATO7JTQA>. You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[hneemann]

What do you think of such a carry flag in the BarrelShifter? A useful addition, or too specific to our case?

The idea of a barrel shifter is to shift a variable number of bits.
What should be output at the carry flag when shifting by more than one bit or by zero bits?

[gixslayer]

As said I'm basing this on ARM/x86 instructions, as defined here for ARM. Shifting more than one bit is not an issue, as the carry output is defined by the last bit that was shifted/rotated out (an n bit rotation/shift is essentially just a sequence of n 1-bit rotates/shifts, thus the carry flag is determined by the last operation in that sequence). The zero bits case is trickier, as in ARM/x86 the carry flag (CF) would not be updated. Unless you want to have the carry output be a tri-state value, I figured setting it to false/0 would be a good alternative.

In our use case we have 32-bit values that are shifted/rotated by a variable amount, thus implementing this via splitters is not a better option as far as I can see.

[hneemann]

I would build a classic muxer-based barrel shifter with the 1-bit shifter in last place to create the carry bit. And I would then include this circuit as a subcircuit in the ALU.

[hneemann]

While building such a shifter, I found that it is easiest to use the built-in barrel shifter and use a single multiplexer to select the carry bit. If the circuit is to be exported to Verilog, however, a classic barrel shifter is required in any case, as the built-in barrel shifter is not exportable, as @TheWildJarvi correctly pointed out.

[gixslayer]

I eventually went with the solution to add subcircuits to the ALU for these instructions as I needed to tweak them slightly anyway for our exact semantics (allow shifting with a full 32-bit unsigned distance, not just the bottom 6 bits).

Also learned that the bit shift operators in test cases have a limited shift distance as the underlying Java bit shifts used mask the distance with 0x1f for int, and 0x3f for long. Nothing a little if-then-else couldn't solve, but that was a head scratcher for a bit.