Prolog/Epilog Insertion going crazy on PPC

[Quuxplusone]

Bugzilla Link	PR2964
Status	RESOLVED FIXED
Importance	P normal
Reported by	Chris Lattner (clattner@nondot.org)
Reported on	2008-10-28 01:53:39 -0700
Last modified on	2008-10-29 13:42:41 -0700
Version	1.0
Hardware	PC All
CC	dalej@apple.com, evan.cheng@apple.com, llvm-bugs@lists.llvm.org
Fixed by commit(s)
Attachments
Blocks
Blocked by
See also

go into test/CodeGen/PowerPC and do this:

$ llvm-as < compare-fcmp-ord.ll | llvm-extract -func test | llc -march=ppc32 -
print-machineinstrs -debug-pass=Executions | & less

Before PEI, the code looks like this:

entry: 0x1812ab4, LLVM BB @0x14021a0, ID#0:
Live Ins: %F1 %F2
        %CR7<def> = FCMPUD %F1<kill>, %F2<kill>
        %R2<def> = MFCR %CR0<imp-use,kill>, %CR1<imp-use,kill>, %CR2<imp-use,kill>, %CR3<imp-use,kill>,
%CR4<imp-use,kill>, %CR5<imp-use,kill>, %CR6<imp-use,kill>, %CR7<imp-use,kill>
        %R3<def> = RLWINM %R2, 29, 31, 31
        %R2<def> = RLWINM %R2<kill>, 31, 31, 31
        %R3<def> = OR %R2<kill>, %R3<kill>
        BLR 20, %reg0, %R3<imp-use,kill>, %LR<imp-use>

Afterwards, it looks like this:

entry: 0x1812ab4, LLVM BB @0x14021a0, ID#0:
Live Ins: %F1 %F2 %CR2 %CR3 %CR4
        %R0<def> = MFCR %CR0<imp-use>, %CR1<imp-use>, %CR2<imp-use>, %CR3<imp-use>, %CR4<imp-use>, %CR5<
imp-use>, %CR6<imp-use>, %CR7<imp-use>
        %R0<def> = RLWINM %R0, 8, 0, 31
        STW %R0<kill>, -4, %R1
        %R0<def> = MFCR %CR0<imp-use>, %CR1<imp-use>, %CR2<imp-use>, %CR3<imp-use>, %CR4<imp-use>, %CR5<
imp-use>, %CR6<imp-use>, %CR7<imp-use>
        %R0<def> = RLWINM %R0, 12, 0, 31
        STW %R0<kill>, -8, %R1
        %R0<def> = MFCR %CR0<imp-use>, %CR1<imp-use>, %CR2<imp-use>, %CR3<imp-use>, %CR4<imp-use>, %CR5<
imp-use>, %CR6<imp-use>, %CR7<imp-use>
        %R0<def> = RLWINM %R0, 16, 0, 31
        STW %R0<kill>, -12, %R1
        %CR7<def> = FCMPUD %F1<kill>, %F2<kill>
        %R2<def> = MFCR %CR0<imp-use,kill>, %CR1<imp-use,kill>, %CR2<imp-use,kill>, %CR3<imp-use,kill>,
%CR4<imp-use,kill>, %CR5<imp-use,kill>, %CR6<imp-use,kill>, %CR7<imp-use,kill>
        %R3<def> = RLWINM %R2, 29, 31, 31
        %R2<def> = RLWINM %R2<kill>, 31, 31, 31
        %R3<def> = OR %R2<kill>, %R3<kill>
        %R0<def> = LWZ -12, %R1
        %R0<def> = RLWINM %R0, 16, 0, 31
        %CR4<def> = MTCRF %R0
        %R0<def> = LWZ -8, %R1
        %R0<def> = RLWINM %R0, 20, 0, 31
        %CR3<def> = MTCRF %R0
        %R0<def> = LWZ -4, %R1
        %R0<def> = RLWINM %R0, 24, 0, 31
        %CR2<def> = MTCRF %R0
        BLR 20, %reg0, %R3<imp-use,kill>, %LR<imp-use>

This is a huge code quality regression from the past.  It looks like
individuals CR is being saved and restored.  CR7 (which is def'd by this code)
is a call clobber register according to this:

PPCInstrInfo.td:422.

[Quuxplusone]

I precipitated this by marking MFCR as using all 7 CR subregisters, which it does. This exposed two underlying problems:

It saves callee-saved CR subregisters (and, I think, other registers, the code appears generic) if they are used; it needs to do this only if they're stored into. (I already fixed the same problem for LR.)

It saves CR subregisters individually (it is saving CR2,3 and 4 which are callee-saved) instead of saving CR as a unit. It always did this and nobody noticed before.

I'll take care of this.

[Quuxplusone]

Thanks Dale!

[Quuxplusone]

I've reverted the change that exposed this for now, thus "fixing" the problem.

[Quuxplusone]

Awesome, thanks Dale.  I added a testcase here:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20081027/069103.html