Fix crash in Beats on x86-64, minor other fixes #18131
Conversation
hrydgard
force-pushed
the
assorted-fixes-4
branch
from
de368ea to
accd9b1
Compare
| @@ -429,6 +429,7 @@ namespace MIPSComp { | |||
| skipStore = J_CC(CC_NE); | |||
|
|
|||
| CompITypeMemWrite(op, 32, safeMemFuncs.writeU32); | |||
| gpr.MapReg(rt, true, true); | |||
There was a problem hiding this comment.
Wait, this looks wrong. You can't map in a conditional, it could corrupt other regs from the perspective of the other condition. This needs to happen before the J_CC?
-[Unknown]
There was a problem hiding this comment.
Since the reg is only written to in both branches, I actually think it should be fine here?
There was a problem hiding this comment.
But hm, I'll clean it up.
There was a problem hiding this comment.
No, because another reg might spill. If it only spills in one conditional, then that other reg is corrupted.
-[Unknown]
There was a problem hiding this comment.
A problem is that CompITypeMemWrite itself locks and unlocks regs, so I can't reliably do it that way, can I?
There was a problem hiding this comment.
Hm, I think the actual bug is in CompITypeMemWrite, in the gpr.R(rt).IsImm() case, since normally we'd end up with rt being mapped coming out of it.
I'll fix that.
There was a problem hiding this comment.
Well, not so much a bug as CompITypeMemWrite just not being designed to leave rt writable, fixable with a flag.
Fixes #18130 , the sc instruction failed to map the destination register.
Only use #18126 in soft-gpu mode (for now)
Add a couple of asserts.