PPC stack references broken

[llvmbot]

Bugzilla Link	8749
Resolution	FIXED
Resolved on	Nov 07, 2018 00:17
Version	trunk
OS	MacOS X
Reporter	LLVM Bugzilla Contributor
CC	@asl

Extended Description

Run the following through llc -disable-fp-elim

; ModuleID = 'main.c'
target datalayout = "E-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f128:64:128-n32"
target triple = "powerpc-apple-darwin9.8"

define i32 @​main() nounwind {
entry:
%retval = alloca i32
%0 = alloca i32
%"alloca point" = bitcast i32 0 to i32
store i32 0, i32* %0, align 4
%1 = load i32* %0, align 4
store i32 %1, i32* %retval, align 4
br label %return

return: ; preds = %entry
%retval1 = load i32* %retval
ret i32 %retval1
}

You will see

stw r3, -12(r31)

but R31 is not set anywhere.
I suspect this was introduced in Anton's rewrite of this area, but haven't analyzed further.

[asl]

Hrm, interesting.

This was not a rewrite, just move stuff here and there. I will take a look.

[llvmbot]

I checked that this worked in 119096, before your patches, and failed in 120230, after them. That's not proof but I'm afraid you are the obvious suspect.

[asl]

Yes, this might easily be. I'm already looking into the problem.

[llvmbot]

Is anything happening on this?

[asl]

Yes. The root problem of this is that I slightly changed the behavior of prologue / epilogue expansion. I'm trying to restore the previous behavior w/o reverting all the stuff I committed :)

The main question is the following: is there any particular reason why function on PPC might reserve r31 reg, but does not actually use it?

[llvmbot]

Not sure what you mean by reserved. If you mean will using R31 for some purpose other than FP break something, I don't think so. But with disable-fp-elim you have to start out assuming it's reserved, and you can't unreserve it until fairly late, as register spilling may introduce FP references.

[asl]

Not sure what you mean by reserved. If you mean will using R31 for some
purpose other than FP break something, I don't think so. But with
disable-fp-elim you have to start out assuming it's reserved, and you can't
unreserve it until fairly late, as register spilling may introduce FP
references.
Ah, ok. Then yes, this should be broken :)

[asl]

Ok, it seems that PPC stack references lowering is pretty much fragile, depending on e.g. order of calls, etc.
I have a patch which fixes this particular testcase, hopefully it will fix all the stuff.

[asl]

Should be fixed in r122155. Dale, could you please check that it fixes all the remaining ppc issues appeared after the refactoring?

[llvmbot]

I was able to bootstrap on ppc using a built llvm-gcc as the installed compiler. That's good enough for me.

[asl]

Ok, thanks for checking!