register spilled to wrong location

[llvmbot]

Bugzilla Link	9134
Resolution	WONTFIX
Resolved on	Nov 25, 2011 14:14
Version	trunk
OS	Linux
Attachments	test case
Reporter	LLVM Bugzilla Contributor
CC	@asl,@d0k

Extended Description

When registers are spilled around call sites, the first spill is to the wrong stack offset.

The result of the test code produced by "llc -march=systemz" is:

zalloc: # @​zalloc
stmg %r14, %r15, 112(%r15)
stg %r2, 168(%r15) # <--exceeds frame size
# spills to old %r15
aghi %r15, -184 # expand frame 160+24
brasl %r14, malloc
stg %r2, 160(%r15) # superfluous but OK
lg %r4, 168(%r15) # <--reload from new r15
lghi %r3, 0
brasl %r14, memset
lg %r1, 160(%r15)
stg %r1, 176(%r15)
lg %r2, 176(%r15)
lmg %r14, %r15, 296(%r15)
br %r14

[asl]

It just might be a bug in systemz's frameindex elimination

[llvmbot]

I think the problem stems from the test case calling a function in the initial BB. This causes a spill of R2. The code in SystemZFrameLowering::emitPrologue thinks this spill is part of the callee-save area. This simple patch makes things better, but is still most probably wrong:

Index: lib/Target/SystemZ/SystemZFrameLowering.cpp

--- lib/Target/SystemZ/SystemZFrameLowering.cpp (revision 124724)
+++ lib/Target/SystemZ/SystemZFrameLowering.cpp (working copy)
@@ -110,7 +110,7 @@
uint64_t NumBytes = StackSize - getOffsetOfLocalArea();

// Skip the callee-saved push instructions.

• while (MBBI != MBB.end() &&

• if (MBBI != MBB.end() &&
  (MBBI->getOpcode() == SystemZ::MOV64mr ||
  MBBI->getOpcode() == SystemZ::MOV64mrm))
  ++MBBI;

This will skip only one push to the stack, which should be the one created by SystemZFrameLowering::spillCalleeSavedRegisters, but not the later spill of R2.

[asl]

The patch is definitely invalid. There might be multiple callee-saved reg spill instructions depending on the allocation pattern.

[llvmbot]

Well I did say it was "most probably wrong" :-)

It just helps me get further running test cases.

[d0k]

The SystemZ backend was removed.