hdl.ast.Value.word_select() works incorrectly on actual platform (ECP5 Versa)

[ghost]

Issue description

According to its docstring, the function word_select(offset, width) of the class nmigen.hdl.ast.Value should return the slice [offset*width:(offset+1):width] of the Value subject. It returns such an expected result in simulation (using PySim), but does not when used on an actual platform, such as the ECP5 Versa evaluation board.

Steps to reproduce the issue

1. Prepare the following nMigen script:

   import argparse
   import os
   
   from nmigen import *
   from nmigen.back import pysim
   from nmigen_boards.versa_ecp5 import VersaECP5Platform
   
   
   class Debug(Elaboratable):
       def __init__(self, simulate):
           self.simulate = simulate
   
       def elaborate(self, platform):
           m = Module()
   
           if self.simulate:
               leds = Signal(8)
           else:
               cd_sync = ClockDomain(reset_less=True)
               m.domains += cd_sync
               m.d.comb += cd_sync.clk.eq(platform.request("clk100").i)
               leds = [platform.request("led", i).o for i in range(8)]
   
           dummy = Signal(8)
           # Note: on the board I'm using, the LED outputs seem to be inverted
           #       e.g. leds[0].eq(1) --> LED#0 turns OFF
           #       e.g. leds[0].eq(0) --> LED#0 turns ON
   
           # Using word_select
           m.d.comb += [
               dummy.word_select(0, 2).eq(~0b00),
               dummy.word_select(1, 2).eq(~0b10),
               dummy.word_select(2, 2).eq(~0b11),
               dummy.word_select(3, 2).eq(~0b01)
           ]
   
           # Using slices
           m.d.comb += [
               dummy[0:2].eq(~0b00),
               dummy[2:4].eq(~0b10),
               dummy[4:6].eq(~0b11),
               dummy[6:8].eq(~0b01)
           ]
   
           m.d.sync += [
               leds[i].eq(dummy[i]) for i in range(0, 8)
           ]
   
           return m
   
   
   if __name__ == "__main__":
       parser = argparse.ArgumentParser()
       parser.add_argument("--simulate", action="store_true")
       build_dir = "./result"
       args = parser.parse_args()
   
       platform = VersaECP5Platform()
       top = Debug(args.simulate)
   
       if args.simulate:
           os.makedirs(build_dir, exist_ok=True)
           with pysim.Simulator(top,
                   vcd_file=open(os.path.join(build_dir, "debug.vcd"), "w"),
                   gtkw_file=open(os.path.join(build_dir, "debug.gtkw"), "w")) as sim:
               sim.add_clock(1e-6)
               sim.run_until(1000e-6, run_passive=True)
       else:
           platform.build(top, build_dir=build_dir)
   

2. Comment out the block starting with the comment line "Using slices". Now we are testing with word_select.

3. Under an environment with nMigen 0.1rc1 and other necessary tools (such as nextpnr, Trellis and OpenOCD) properly installed, run the script with Python. Add --simulate as the argument to see the VCD from simulation.

4. Use OpenOCD to program the ECP5 board, given that a working Config file is ready (see the example versa.cfg from HeavyX's README). For example, use the following command after building the bitstream:

openocd -f versa.cfg -c "transport select jtag; init; svf result/top.svf; exit"

5. The LEDs lights up like below (inverted bit pattern == 11101100):

6. Now, repeat Step 2 but instead of commenting out the "Using slices" block, comment out the "Using word_select" this time. Now we are testing with slicing.

7. Repeat Steps 3 and 4.

8. The LEDs lights up like below (inverted bit pattern == 01111000):

Comparison between Actual and Expected Results

If we compare the simulation result (i.e. VCDs) from the two scenarios, they are the same and as expected (bit pattern == 10000111 == 01111000 inverted (due to LED polarity issue)):

However, the actual LED result when using word_select() is inconsistent with the simulation result, as revealed in the picture of Step 5.

Using LED is just an example to quickly show this issue. I have actually encountered the issue when using word_select() to grab slices over in nmigen-soc (link). So maybe it is a potential bug in nMigen?

[whitequark]

Interesting. It could be a bit width issue. Could you try to reproduce this by simulating in iverilog?

[ld-cd]

The balls are attached to the low side of the LEDs (writing 1 would turn them off unless the board layer inverts), so perhaps the issue with nmigen soc is different?

[ghost]

@ld-cd Hi! Could you please re-phrase your question? I see that the ECP5 Versa board defines that the LEDs light up when the balls are driven low, but I don't understand its correlation to nmigen-soc. Thanks.

[ld-cd]

Oh, I was just mentioning that the inverted bit pattern you are seeing could be because of the way they are wired, leaving anything funky you mentioned you had happen with nmigen soc as a result of something else.

[ld-cd]

Sorry, I believe I misread the issue