An Issue with REP blocks, Single step interrupts and HUB execution

ozpropdev · 2017-01-23 10:45

Hi Chip

I have been playing with the single step interrupt mechanism in P2 and it works great!
However I am experiencing one little issue when stepping into a REP block in HUB exec mode.
Here is the demo REP block that I used for COG,LUT and HUB exec testing.

rep_block		mov	delta,#0
			rep	@.rb_end,#32
			add	alpha,#1
			sub	bravo,#2
			mul	charlie,#2
.rb_end
			bmask	delta,#19
			ret

Here is the results from my P2 debugger (work in progres) for COG exec mode.
COG address advances from $007 to $00B as expected.

-----------------------------------------------------------------------------------------(P2 Debugger)
STACK : --_00000 --_00000 --_00000 --_00000 --_00000 --_00000 --_00000 --_00000
COG | $00040: "Delta"       $00000000 %00000000_00000000_00000000_00000000 #0
00007: FCDC0420              REP     #$002,#$020
00008: F1047A01              ADD     $03D,#$001
00009: F1847C02              SUB     $03E,#$002
0000A: FA047E02              MUL     $03F,#$002
Flags (CZ) = _Z
>*
-----------------------------------------------------------------------------------------(P2 Debugger)
STACK : --_00000 --_00000 --_00000 --_00000 --_00000 --_00000 --_00000 --_00000
COG | $00040: "Delta"       $00000000 %00000000_00000000_00000000_00000000 #0
0000B: F99C8013              BMASK   $040,#$013
Flags (CZ) = _Z

Here is the results for LUT exec mode.
Address advances from $283 to $287 as expected.

-----------------------------------------------------------------------------------------(P2 Debugger)
STACK : -Z_0000E --_00000 --_00000 --_00000 --_00000 --_00000 --_00000 --_00000
COG | $00040: "Delta"       $00000000 %00000000_00000000_00000000_00000000 #0
00283: FCDC0420              REP     #$002,#$020
00284: F1047A01              ADD     $03D,#$001
00285: F1847C02              SUB     $03E,#$002
00286: FA047E02              MUL     $03F,#$002
Flags (CZ) = _Z
>*
-----------------------------------------------------------------------------------------(P2 Debugger)
STACK : -Z_0000E --_00000 --_00000 --_00000 --_00000 --_00000 --_00000 --_00000
COG | $00040: "Delta"       $00000000 %00000000_00000000_00000000_00000000 #0
00287: F99C8013              BMASK   $040,#$013
Flags (CZ) = _Z

And finally here is the HUb exec mode results.
Note that the address advances past the expected address of $7014 to $7018.
The expected "BMASK" instruction has executed as proven by the cog register "delta" contents.

-----------------------------------------------------------------------------------------(P2 Debugger)
STACK : -Z_00289 -Z_0000E --_00000 --_00000 --_00000 --_00000 --_00000 --_00000
COG | $00040: "Delta"       $00000000 %00000000_00000000_00000000_00000000 #0
07004: FCDC0420              REP     #$002,#$020
07008: F1047A01              ADD     $03D,#$001
0700C: F1847C02              SUB     $03E,#$002
07010: FA047E02              MUL     $03F,#$002
Flags (CZ) = _Z
>*
-----------------------------------------------------------------------------------------(P2 Debugger)
STACK : -Z_00289 -Z_0000E --_00000 --_00000 --_00000 --_00000 --_00000 --_00000
COG | $00040: "Delta"       $000FFFFF %00000000_00001111_11111111_11111111 #1048575
07018: FD600031              RET
Flags (CZ) = _Z

If the instruction following the REP block is a 2 long variant (i.e AUGx,ALTx etc) it is executed
and skipped in the same manner.

Hope I made sense...
Cheers
Brian

Rayman · 2017-01-23 11:41

I vaguely recall that there are some things you can't do in hubexec mode.

Isn't REP one of them?

ozpropdev · 2017-01-23 11:58

The REP block works fine in HUB exec mode.
It's just the single step that has an issue.

Seairth · 2017-01-23 12:33

I'm surprised debugging REP blocks works at all! I thought REP did not work correctly if you jumped inside of it. Has this changed?

evanh · 2017-01-23 12:46

I believe all interrupt forms are completely masked during a REP, amongst a few other shorter conditions like instruction prefixes. Goes looking ... ah and the WAITs too. It's all on page 20 of PropellerIIDocumentationv13.

evanh · 2017-01-23 12:48

The single step goes on the pending IRQ list until the REP is finished.

PS: I haven't read the opening post.

evanh · 2017-01-23 13:01

Ah, I should read the OP instead of following the other comments ... I see now the problem is the exit from the REP loop, not the loop itself. In HubExec it's going one instruction further than in CogExec.

ozpropdev · 2017-01-23 13:12

Just for kicks I placed two REP blocks back to back in HUB exec mode and a single step executes both blocks.

ozpropdev · 2017-01-23 13:20

and same goes for three REP blocks back to back....

cgracey · 2017-01-23 17:13

I'm looking over the Verilog now.

For REP in hub-exec mode, things are handled a bit differently than they are in cog-exec mode. I have an idea of what it could be, but I need to relearn this chunk of Verilog. Working on it....

Rayman · 2017-01-23 18:03

I'm still surprised REP works in hubexec mode at all.
Thought it would go off the rails if rep block didn't happen to be all in the fifo...

cgracey · 2017-01-23 19:48

Rayman wrote: »

I'm still surprised REP works in hubexec mode at all.
Thought it would go off the rails if rep block didn't happen to be all in the fifo...

The FIFO is not used. On any branch, including a REP loop, the pipeline is reloaded from memory.

cgracey · 2017-01-23 20:58

Ozpropdev,

I'm looking this over, still. I'm wondering if there are any differences for when the body of the REP block is only only one instruction. Any chance you could try out one-instruction cases?

ozpropdev · 2017-01-23 21:19

Chip
Same goes for a single instruction REP block.
In tiis case the instruction at $700C is executed too.

-----------------------------------------------------------------------------------------(P2 Debugger)
COG | $00040: "Delta"       $00000000 %00000000_00000000_00000000_00000000 #0
07004: FCDC0020              REP     #$000,#$020
07008: F1048001              ADD     $040,#$001
Flags (CZ) = _Z
>*
-----------------------------------------------------------------------------------------(P2 Debugger)
COG | $00040: "Delta"       $000FFFFF %00000000_00001111_11111111_11111111 #1048575
07010: FD600031              RET
Flags (CZ) = _Z
>

cgracey · 2017-01-23 21:52

ozpropdev wrote: »

Chip
Same goes for a single instruction REP block.
In tiis case the instruction at $700C is executed too.

-----------------------------------------------------------------------------------------(P2 Debugger)
COG | $00040: "Delta"       $00000000 %00000000_00000000_00000000_00000000 #0
07004: FCDC0020              REP     #$000,#$020
07008: F1048001              ADD     $040,#$001
Flags (CZ) = _Z
>*
-----------------------------------------------------------------------------------------(P2 Debugger)
COG | $00040: "Delta"       $000FFFFF %00000000_00001111_11111111_11111111 #1048575
07010: FD600031              RET
Flags (CZ) = _Z
>

How about the original test, but with a rep count of 1? It should allow you to step through the block, as there is, actually, no REP taking place.

ozpropdev · 2017-01-23 22:09

Same for a REP count of 1

-----------------------------------------------------------------------------------------(P2 Debugger)
COG | $00040: "Delta"       $00000000 %00000000_00000000_00000000_00000000 #0
07004: FCDC0401              REP     #$002,#$001
07008: F1047A01              ADD     $03D,#$001
0700C: F1847C02              SUB     $03E,#$002
07010: FA047E02              MUL     $03F,#$002
Flags (CZ) = _Z
>*
-----------------------------------------------------------------------------------------(P2 Debugger)
COG | $00040: "Delta"       $000FFFFF %00000000_00001111_11111111_11111111 #1048575
07018: FD600031              RET
Flags (CZ) = _Z
>

cgracey · 2017-01-23 22:16

ozpropdev wrote: »

Same for a REP count of 1

-----------------------------------------------------------------------------------------(P2 Debugger)
COG | $00040: "Delta"       $00000000 %00000000_00000000_00000000_00000000 #0
07004: FCDC0401              REP     #$002,#$001
07008: F1047A01              ADD     $03D,#$001
0700C: F1847C02              SUB     $03E,#$002
07010: FA047E02              MUL     $03F,#$002
Flags (CZ) = _Z
>*
-----------------------------------------------------------------------------------------(P2 Debugger)
COG | $00040: "Delta"       $000FFFFF %00000000_00001111_11111111_11111111 #1048575
07018: FD600031              RET
Flags (CZ) = _Z
>

What about cog-exec? I assume that was hub-exec.

ozpropdev · 2017-01-23 22:24

For a REP count of 1 in CogExec it still executes the entire block in one step but doesn't skip the instruction following the block like HubExec does.

-----------------------------------------------------------------------------------------(P2 Debugger)
COG | $00040: "Delta"       $00000000 %00000000_00000000_00000000_00000000 #0
00007: FCDC0401              REP     #$002,#$001
00008: F1047A01              ADD     $03D,#$001
00009: F1847C02              SUB     $03E,#$002
0000A: FA047E02              MUL     $03F,#$002
Flags (CZ) = _Z
>*
-----------------------------------------------------------------------------------------(P2 Debugger)
COG | $00040: "Delta"       $00000000 %00000000_00000000_00000000_00000000 #0
0000B: F99C8013              BMASK   $040,#$013
Flags (CZ) = _Z
>

cgracey · 2017-01-23 22:38

ozpropdev wrote: »

Just for kicks I placed two REP blocks back to back in HUB exec mode and a single step executes both blocks.

What about in cog-exec mode? Will it stop on the next REP, or execute that whole block, too?

ozpropdev · 2017-01-23 22:46

Back to back REP blocks in CogExec stop at the next block

-----------------------------------------------------------------------------------------(P2 Debugger)
00007: FCDC0401              REP     #$002,#$001
00008: F1048201              ADD     $041,#$001
00009: F1848402              SUB     $042,#$002
0000A: FA048602              MUL     $043,#$002
Flags (CZ) = _Z
>*
-----------------------------------------------------------------------------------------(P2 Debugger)
0000B: FCDC0401              REP     #$002,#$001
0000C: F1048201              ADD     $041,#$001
0000D: F1848402              SUB     $042,#$002
0000E: FA048602              MUL     $043,#$002
Flags (CZ) = _Z
>*

cgracey · 2017-01-23 22:47

ozpropdev wrote: »

Back to back REP blocks in CogExec stop at the next block

-----------------------------------------------------------------------------------------(P2 Debugger)
00007: FCDC0401              REP     #$002,#$001
00008: F1048201              ADD     $041,#$001
00009: F1848402              SUB     $042,#$002
0000A: FA048602              MUL     $043,#$002
Flags (CZ) = _Z
>*
-----------------------------------------------------------------------------------------(P2 Debugger)
0000B: FCDC0401              REP     #$002,#$001
0000C: F1048201              ADD     $041,#$001
0000D: F1848402              SUB     $042,#$002
0000E: FA048602              MUL     $043,#$002
Flags (CZ) = _Z
>*

Thanks for checking that. I'm still looking the code over, reconstructing how this works in my head.

cgracey · 2017-01-23 23:02

Here's the code that makes REP:

///////////
//	 //
//  REP	 //
//	 //
///////////

reg [11:0]	repdc;
reg  [9:0]	repic, repi;
reg [32:0]	repb;
reg		repa;

wire rep_si	=	!hubx && ~|d[8:0];				// rep, single instruction if cog exec and d[8:0] = 0
wire rep_sb	=	s == 1'b1;					// rep, single block if s = 1

wire repiz	=	~|repi;						// instruction counter zero?
wire repmore	=	|repb[32:1];					// infinite or more blocks?

`regscan (repdc, 12'b0,							// p delta capture
go && rep,
	hubx	?	{{1'b0, d[8:0]} + 1'b1, 2'b0}			// for hub exec, must jump back by {d[8:0] + 1} * 4
		:	d[8:0])						// for cog exec, must subtract d[8:0] from p

`regscan (repic, 10'b0,							// instruction count capture
go && rep,
	d[8:0] + hubx)							// cog exec uses d[8:0], hub exec uses d[8:0] + 1

`regscan (repi, 10'b0,							// instruction counter
go && (rep || repa),
	rep	?	rep_si ? 9'b0 : d[8:0] - !hubx			// rep, load zero if single instr, else load instr count - 1
:	repiz	?	repic						// if instruction counter empty, reload
		:	repi - 1'b1)					// else, decrement

`regscan (repb, 33'b0,							// block counter
go && (rep || repa && repiz),
	rep	?	{~|s, rep_si ? s - 1'b1 : s}			// rep, capture infinite flag and block count
		:	{repb[32], repb[31:0] - 1'b1})			// preserve infinite flag, decrement block count

`regscan (repa, 1'b0,							// rep active flag
!ena || go && (rep || repa && repiz),
	!ena	?	1'b0
:	rep	?	!rep_si || !rep_sb				// rep, remain active if not single instruction or not single block
		:	repmore)					// after rep, remain active if infinite or more blocks

wire rept	=	rep	?	rep_si && !rep_sb		// rept on rep if cog exec and single instruction and multiple blocks
				:	repa && repiz && repmore;	// rept after rep when active and instruction counter zero and more blocks

wire [11:0] repd=	rep	?	d[8:0]				// p delta on rep, cog exec only
				:	repdc;				// p delta after rep

It usually takes me a few hours to get back into something like this and fully remember how it works.

Rayman · 2017-01-23 23:16

Is the part that blocks the interrupt in there somewhere?

cgracey · 2017-01-23 23:21

Rayman wrote: »

Is the part that blocks the interrupt in there somewhere?

Yes. The interrupt is inhibited if a REP instruction is executing or the register 'repa' is high.

What happens in hub-exec seems normal. I'm trying to figure out why cog-exec seems to work as one would expect. I think in cog-exec, 'repa' goes low one instruction early. I don't understand why, yet.

cgracey · 2017-01-24 07:54

I've been looking at this ALL day and I still haven't even figured out why I did D+1 for the hub exec instruction count, as opposed to just D for cog-exec. I should have made better notes in the Verilog code. I remember this was necessary, but I can't figure out why. It had something to do with a real JMP for hub-exec, as opposed to just a PC adjustment for cog-exec. But, why?

evanh · 2017-01-24 08:01

Oops, didn't see the verilog there ...

ozpropdev · 2017-01-24 09:20

Re: A REP count of 1

cgracey wrote: »

It should allow you to step through the block, as there is, actually, no REP taking place.

I note in your Verilog code the facility to handle a single pass block, and you suggest I should be able to step through the block.

// rep, remain active if not single instruction or not single block
// after rep, remain active if infinite or more blocks

The REP block still executes as an entire chunk with no step through in both CogExec and HubExec.

cgracey · 2017-01-24 09:33

ozpropdev wrote: »

Re: A REP count of 1

cgracey wrote: »

It should allow you to step through the block, as there is, actually, no REP taking place.

I note in your Verilog code the facility to handle a single pass block, and you suggest I should be able to step through the block.

// rep, remain active if not single instruction or not single block
// after rep, remain active if infinite or more blocks

The REP block still executes as an entire chunk with no step through in both CogExec and HubExec.

Okay. I see why it does that now. The register 'repb' doesn't go down to one until after the last loop back. Meanwhile, 'repa' is high. Only at the end of the last block does 'repa' go low. I think it's happening one instruction too late in the case of hub-exec.

I'm compiling some test version of the prop2 now, so that I can look at the PC and JMP bits during REP, both in cog-exec and hub-exec. I need to relearn what it's doing. It's like my brain unspooled some details over the last two months.

ozpropdev · 2017-01-24 09:38

I can post my debugger code if that would help.

cgracey · 2017-01-24 09:45

ozpropdev wrote: »

I can post my debugger code if that would help.

I'm watching the signals on a logic analyzer, which shows me what I need to see, for now.

Does your debugger drive a monitor or serial terminal? I'd like to use it, but I don't know what I need to know, in order to. I assume it's pretty simple, right? If you post it, I will try it out.

ozpropdev · 2017-01-24 09:57

It's a serial terminal based interface, nice and simple.
I'm preparing it for posting now....

An Issue with REP blocks, Single step interrupts and HUB execution

Comments