Fix callKillSet for CORINFO_HELP_ASSIGN_BYREF on x64.

[sandreenko]

Our emitter doesn't check liveness information consistency, so when we generate a LIR node that expands to a several machine instructions there is no way to check that nobody corrupts the data.

In this case STORE_OBJ generated

IN0006:        mov      rdi, r15
IN0007:        mov      rsi, r12
							Byref regs: 00009000 {r12 r15} => 00009040 {rsi r12 r15}
							Byref regs: 00009040 {rsi r12 r15} => 000090C0 {rsi rdi r12 r15}
							Call: GCvars=0000000000000000 {}, gcrefRegs=00000000 {}, byrefRegs=00009000 {r12 r15}
IN0008:        call     CORINFO_HELP_ASSIGN_BYREF
IN0009:        mov      ecx, 7
IN000a:        rep movsq 

where byrefRegs after call did not include rsi, rdi, because we trashed RBM_CALLEE_TRASH_NOGC that on Linux x64 incudes rsi, rdi.

coreclr/src/jit/emitxarch.cpp

Lines 6880 to 6884 in 1c47703

	/* Trim out any callee-trashed registers from the live set */
	gcrefRegs &= savedSet;
	byrefRegs &= savedSet;

Fixes #19361.

[sandreenko]

PTAL @dotnet/jit-contrib

I am trying to modify the repro from @jkotas to expose this issue in ci, if I success I will add it later.

[CarolEidt]

LGTM except comment suggestions

[CarolEidt]

Typo: rci -> rsi
Also, you may want to just be quite explicit that "This uses and defines rdi and rsi, so they remain as live GC refs or byrefs, and are not killed." or something like that.

[sandreenko]

Thank you, fixed.

[CarolEidt]

Same typo (RCI -> RSI)

[sandreenko]

Hm, win_x64 corefx test failed, lets try to rerun. Doesn't look like infra.
@dotnet-bot test
Windows_NT x64 Checked CoreFX Tests

[AndyAyersMS]

Any idea why gc stress didn't catch this?

[sandreenko]

Have not succeed in adding a repro. I will merge the fix only for now.

Any idea why gc stress didn't catch this?

The issue only marks struct fields as dead and we usually have other parts of the class alive (like this or temp variables that saves &field etc.). So I think it is pretty rare case when the gc can collect the whole class right after the large struct assignment.

Thank for your help with repro.