[LLVMbugs] [Bug 6450] New: LiveIntervals code doesn't properly identify unspillable registers.

[bugzilla-daemon at llvm.org]

http://llvm.org/bugs/show_bug.cgi?id=6450

           Summary: LiveIntervals code doesn't properly identify
                    unspillable registers.
           Product: new-bugs
           Version: unspecified
          Platform: PC
        OS/Version: Linux
            Status: NEW
          Severity: enhancement
          Priority: P5
         Component: new bugs
        AssignedTo: unassignedbugs at nondot.org
        ReportedBy: mwb.cde at googlemail.com
                CC: llvmbugs at cs.uiuc.edu


Created an attachment (id=4342)
 --> (http://llvm.org/bugs/attachment.cgi?id=4342)
LiveInterval patch - add no_spill field

The spiller code in LiveIntervalAnalysis.cpp and Spiller.cpp asserts
when passed a register with infinite weight even if the register is a
valid choice for spilling.

Registers that can't be spilled, because they are generated by the
spiller for a register placed on the stack, are identified by setting
their interval weight to HUGE_VALF (== inf). This makes them
indistinguishable from normal registers that have infinite weight,
which can occur with deeply nested loops such as generated by the test
case for bug 5846.

The attached LiveIntervals patch adds boolean field no_spill to class
LiveInterval, to indicate whether the interval represents a spillable
register, and updates the spiller code to set and test the field. It
does not change the setting and testing of the interval weight except
to remove the assertions so this patch should not affect any existing
code. The patch has been tested for the CodeGen component.

The RegAllocLinScan patch is a partial, untested change to
RegAllocLinearScan to make use of the no_spill field when trying to
find a register to spill. It gets the test-case for bug 5846 to
compile but is only meant as an example of use and has not been
tested.

-- 
Configure bugmail: http://llvm.org/bugs/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are on the CC list for the bug.