pabs3 has quit [Read error: Connection reset by peer]
pabs3 has joined #riscv
pabs3 has quit [Quit: Don't rest until all the world is paved in moss and greenery.]
pabs3 has joined #riscv
mlw has joined #riscv
fuwei has joined #riscv
andyc has joined #riscv
fuwei has quit [Ping timeout: 246 seconds]
fuwei has joined #riscv
mark4o has joined #riscv
DrachenMaus has joined #riscv
lagash has quit [Ping timeout: 246 seconds]
markh has quit [Read error: Connection reset by peer]
DragonMaus has quit [Quit: No Ping reply in 180 seconds.]
motherfsck has quit [Ping timeout: 246 seconds]
mark4o is now known as markh
prabhakalad has quit [Ping timeout: 246 seconds]
Starfoxxes has quit [Ping timeout: 246 seconds]
motherfsck has joined #riscv
Starfoxxes has joined #riscv
motherfsck has quit [Ping timeout: 246 seconds]
atka has quit [Ping timeout: 255 seconds]
atka has joined #riscv
danilogondolfo has joined #riscv
cleger has joined #riscv
heat has joined #riscv
prabhakalad has joined #riscv
Andre_Z has joined #riscv
rattlesnake has quit [Ping timeout: 252 seconds]
rattlesn- has joined #riscv
inad922 has joined #riscv
<inad922>
In case anyone needs a gdb with python support for Risc-V the way to build it is to clone https://github.com/riscv-collab/riscv-gnu-toolchain and then `./configure --prefix=/opt/riscv && GDB_TARGET_FLAGS_EXTRA="--with-python=$(which python)" make` . Obviously replace the prefix and the python path as needed.
inad922 has quit [Client Quit]
lemoniter has joined #riscv
X-Scale has quit [Quit: Client closed]
andyc has quit [Quit: Connection closed for inactivity]
lagash has joined #riscv
jedesa has quit [Ping timeout: 244 seconds]
jacklsw has quit [Ping timeout: 255 seconds]
andyc has joined #riscv
levitating has joined #riscv
jobol has joined #riscv
X-Scale has joined #riscv
clever has quit [Ping timeout: 276 seconds]
clever has joined #riscv
jedesa has joined #riscv
aethlas has quit [Quit: bye]
aethlas has joined #riscv
jedesa has quit [Ping timeout: 248 seconds]
jedesa has joined #riscv
X-Scale has quit [Quit: Client closed]
cousteau has joined #riscv
<cousteau>
Hi!
<cousteau>
In GCC, -O2 results in misaligned branch targets (if the C extension is enabled), despite of -falign-labels being enabled by -O2. Is that normal? This can cause extra delay on certain RISC-V implementations; I understood the point of enabling that option in -O2 was to avoid this delay.
aethlas has quit [Quit: bye]
<cousteau>
-falign-labels=4 works fine, but -falign-labels doesn't align the labels at all, so I don't get why that'd be added to -O2.
aethlas has joined #riscv
<palmer>
cousteau: can you file a GCC bug? It sounds like a bug, we probably just missed something
<cousteau>
(asking because maybe the default alignment for -falign-labels should be changed, which I suppose is something the RISC-V developers are the ones in charge of fixing)
<cousteau>
palmer: OK, but I first wanted to make sure this was indeed a bug
<cousteau>
The GCC documentation for -falign-labels says that with no default value it might just be 1 meaning no alignment.
<cousteau>
But maybe that's only for architectures where misalignment has no consequences (e.g. VLIW ones like Intel where instructions can typically be as short as one byte)
<palmer>
Ya, but if everyone else is aligning them we probably should too. It'll really depend on performance, though, so we'll have to go talk about it
<palmer>
So not 100% sure it's a bug, but it's at least worth discussing
<cousteau>
Also I'm not sure if misaligned branch targets are an issue on most RISC-V implementations. They are in CVE4
gioyik has joined #riscv
<cousteau>
I did notice an important performance penalty on a very tight loop with/without alignment. (10 clock cycles per iteration with alignment, 11 without, so like a 10% slower)
<palmer>
Ya, they're slow on some implemtations. It's pretty common
<palmer>
At a bare minimum we should get the tuning defaults right, I don't think we even have one for that uarch? So maybe that's the answer here
<cousteau>
I also couldn't manage to figure out where the default value of -falign-labels is changed in the GCC code.
<cousteau>
Yeah, there only seem to be -mtune parameters for SiFive stuff.
<cousteau>
Anyway I have no idea where those tuning parameters are all defined
<cousteau>
So where exactly do I report this? GCC mailing list? Some RISC-V specific one?
<palmer>
If you can post that loop, then bugzilla is probably the best bet. Otherwise mailing list is fine, just the standard GCC lists
jedesa has quit [Ping timeout: 252 seconds]
<palmer>
There's probably a TARGET_ macro hook for it, I'd have to read the docs ;)
X-Scale has joined #riscv
<cousteau>
One reason I want to know where this is declared is to confirm that this hasn't been fixed yet
naoki has quit [Quit: naoki]
gioyik has quit [Ping timeout: 260 seconds]
jedesa has joined #riscv
TaelTydes has joined #riscv
jedesa has quit [Quit: jedesa]
gioyik has joined #riscv
jedesa has joined #riscv
gioyik has quit [Ping timeout: 260 seconds]
levitating has quit [Ping timeout: 246 seconds]
gioyik has joined #riscv
gioyik has quit [Ping timeout: 260 seconds]
Stat_headcrabed has joined #riscv
levitating has joined #riscv
gioyik has joined #riscv
gioyik has quit [Ping timeout: 260 seconds]
Tenkawa has joined #riscv
mlw has quit [Ping timeout: 252 seconds]
___nick___ has joined #riscv
andyc has quit [Quit: Connection closed for inactivity]
gioyik has joined #riscv
levitating has quit [Ping timeout: 252 seconds]
gioyik has quit [Ping timeout: 260 seconds]
Stat_headcrabed has quit [Quit: Stat_headcrabed]
levitating has joined #riscv
craigo has joined #riscv
gioyik has joined #riscv
gioyik has quit [Ping timeout: 260 seconds]
gioyik has joined #riscv
levitating has quit [Ping timeout: 246 seconds]
gioyik has quit [Ping timeout: 260 seconds]
Maylay has quit [Quit: Pipe Terminated]
Maylay has joined #riscv
Andre_Z has quit [Ping timeout: 246 seconds]
gioyik has joined #riscv
gioyik has quit [Ping timeout: 260 seconds]
levitating has joined #riscv
gioyik has joined #riscv
damian101 has quit [Ping timeout: 255 seconds]
levitating has quit [Ping timeout: 246 seconds]
gioyik has quit [Ping timeout: 260 seconds]
levitating has joined #riscv
X-Scale has quit [Quit: Client closed]
X-Scale has joined #riscv
gioyik has joined #riscv
levitating has quit [Remote host closed the connection]
levitating has joined #riscv
<palmer>
cousteau: the docs have a section for `@defmac LABEL_ALIGN (@var{label})` (there's PDF/HTML, I just reade the source)
<palmer>
looks like we're not doing anything special, but most other ports aren't either. So maybe that default behavior is the right thing, or maybe there's another hook we should be using
<palmer>
`struct riscv_tune_param` doesn't have any label alignment fields, though. So there's no uarch-dependent label alignment hooks, which I'd argue is a bug if there's uarch-specific alignment performance quirks
BootLayer has joined #riscv
levitating has quit [Remote host closed the connection]
levitating has joined #riscv
levitating has quit [Ping timeout: 272 seconds]
sakman_ has quit [Remote host closed the connection]
sakman_ has joined #riscv
<cousteau>
Hm. -falign-labels=256 makes all labels aligned to 256+8 (hex ends in 08). Wonder why the +8
vagrantc has joined #riscv
<cousteau>
That's with -O2. With -O1 it's +10 (0xA).
<jrtc27>
post-linking?
<cousteau>
Uh, not sure. I used godbolt, with "Compile to binary object".
<jrtc27>
no then
<jrtc27>
that'll just be a relocatable .o
<jrtc27>
so presumably this is R_RISCV_ALIGN not yet having been relaxed
<jrtc27>
if -mno-relax "fixes" it that'll verify it
<cousteau>
There are lots of R_RISCV_ALIGN/BRANCH things
<cousteau>
That thwarts my plans to use godbolt for debugging.
<cousteau>
Anyway. Does any expert in this sort of thing think that -falign-labels=0 should default to 4 and not 1 (or 2), so that -O2 aligns the branches?
vagrantc has quit [Quit: leaving]
vagrantc has joined #riscv
gioyik has quit [Quit: WeeChat 4.4.1]
<palmer>
cousteau: there should be a button in Godbolt to link to an executable and then disassemble, not sure if it works though
<palmer>
IMO we should have -falign-labels do something reasonable for the tuning we're targeting. For the generic stuff we'd need to go sort that out, but for the uarch-specific tunings it seems pretty straight-forward to just set it to the alignment targets for the system. Probably best to just discuss it upstream, though, hence he bugzilla ;)
prabhakalad has quit [Ping timeout: 246 seconds]
prabhakalad has joined #riscv
<cousteau>
OK, will submit a bug :)
<palmer>
cool
<cousteau>
Problem is that there are very few supported uarch
<palmer>
not sure what'll happen, but I think at least getting the tunings to match the HW should be straight-forward
<cousteau>
Most of them (all?) from SiFive. Maybe this is a job for CORE-V/OpenHW?
<palmer>
there's more in the recent releases, it's just a matter of getting someone to care enough to write one
<palmer>
they're generally pretty easy, as long as the tuning struct supports what you're trying to do then it's just boilerplate
<palmer>
and ya, it's generally best to try and get the HW vendor to do this. It's really their thing ;)
prabhakalad has quit [Ping timeout: 252 seconds]
prabhakalad has joined #riscv
danilogondolfo has quit [Quit: Leaving]
jbowen has quit [Ping timeout: 248 seconds]
jbowen has joined #riscv
X-Scale has quit [Quit: Client closed]
mlw has joined #riscv
<cousteau>
palmer: thanks! :)
cousteau has quit [Quit: Quit]
sakman_ has quit [Remote host closed the connection]