Leaves (Jul 16 2024 at 00:42): Leaves (Jul 16 2024 at 00:42):Regalloc2 is great, but I have a question. Is this a failure of the heuristics? Or have I set up my machine instruction structure incorrectly? The test machine instruction that I pass to regalloc2 has no Defs or Clobbers so there is no reason it couldn't just TEST EAX,EAX.
Chris Fallin (Jul 16 2024 at 02:13):You’re right in this particular case that move seems extraneous; it’s hard to say more without actually running your environment and getting trace logs from RA2 to see why it’s doing that. (And to be clear I unfortunately don’t have the spare cycles for that, much as I would like to!)
Chris Fallin (Jul 16 2024 at 02:15):I would verify that the vreg placed in rax is not clobbered elsewhere and that that there aren’t any register constraints forcing this
Amanieu (Jul 16 2024 at 14:45):Could the clobber on call be interfering with regalloc somehow?
Leaves (Jul 16 2024 at 19:28):I figured out why it is happening.
The test instruction is testing 32 bit EAX, prior to it, a truncate takes the RAX from the function call and converts it to a 32bit. Although the truncate instruction isn't actually emitted, it still creates a Def, see the below description:
10: call_rel32 (Def: v22i fixed(p0i)), ...
11: truncate (Use: v22i any),(Def: v29i reuse(0)),Clobbers {}
12: test_rs_r (Use: v29i any),(Use: v29i reg),Clobbers {}
So regalloc thinks that the value in is being mutated(because I'm telling it it is) and creates a mov into a different register to avoid mutating RAX which is used later after the current function returns.
So, not a problem with regalloc2, a problem with my own design decisions.
Chris Fallin (Jul 16 2024 at 20:16):Glad to see the mystery solved!
Last updated: Dec 23 2024 at 13:07 UTC