Sanjin Sijaric via llvm-dev
2016-Nov-22 21:22 UTC
[llvm-dev] Assert in register allocation during last chance recoloring
Hi,
I opened the bug #31122 for this issue.
There are two parts to this problem. Compiled using "llc
-debug-only=regalloc test.ll", the test case attached to the bug fails with
VirtRegMap.h:119: void llvm::VirtRegMap::clearVirt(unsigned int): Assertion
`Virt2PhysMap[virtReg] != NO_PHYS_REG && "attempt to clear a not
assigned
virtual register"' failed.
The code looks like:
%vreg647<def> = COPY %vreg645;
QQQQPR_with_dsub_7_then_ssub_0:%vreg647 QQQQPR:%vreg645
%vreg657<def> = COPY %vreg655;
QQQQPR_with_dsub_7_then_ssub_0:%vreg657 QQQQPR:%vreg655
%vreg679<def> = COPY %vreg678; QPR_VFP2:%vreg679 QPR:%vreg678
%vreg657:dsub_7_then_ssub_0<def> = VDIVS %vreg679:ssub_2,
%vreg647:dsub_7_then_ssub_0, pred:14, pred:%noreg;
QQQQPR_with_dsub_7_then_ssub_0:%vreg657,%vreg647 QPR_VFP2:%vreg679
%vreg659<def> = COPY %vreg657; QQQQPR:%vreg659
QQQQPR_with_dsub_7_then_ssub_0:%vreg657
This problem occurs during last chance recoloring, where %vreg679 doesn't
get allocated (and is marked as not spillable), so last chance recoloring
kicks in. It tries to allocate Q0 to %vreg679, where Q0_Q1_Q2_Q3 is
assigned to %vreg647, which interferes with %vreg679. Upon trying to
allocate a register to %vreg647, %vreg679 gets evicted from Q0. But
%vreg679 is a fixed register, so it shouldn't get evicted. Recoloring is
thus deemed successful, and upon returning from tryRecoloringCandidates, we
assert when trying to unassign %vreg679.
I can get around this by moving the virtual register under consideration in
tryLastChanceRecoloring (%vreg679) to RS_Done stage right before the call to
tryRecolringCandidates, and then restoring the original stage afterward.
This prevents the eviction of %vreg679 during recoloring. Alternatively, I
can check FixedRegisters in canEvictInterference and return false.
After the assert is gone, I run into "LLVM ERROR: ran out of registers
during register allocation". Increasing lcr-max-depth doesn't help.
To get past this error, I am thinking about spilling an actual physical
register if last chance recoloring fails. For example,
P = ..
virtX = .. Last chance recoloring fails,
. = P
.. = virtX
will result in something along the lines of
P = ..
spill P to fi#0 (original P)
P = .. (virtX is now in P)
..
Spill P to fi#1 (this is virtX)
Reload P form from fi#0 (original P)
. = P
..
Reload P from fi#1 (this is virtX)
= P
Does anyone have any other suggestions about dealing with this problem?
Thank you,
Sanjin
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20161122/bea27168/attachment.html>
Quentin Colombet via llvm-dev
2016-Nov-28 21:22 UTC
[llvm-dev] Assert in register allocation during last chance recoloring
Hi Sanjin, The spilling/spliting strategy for big registers is known to be weak so I am not surprised you run into this problem. I’ll have a closer look (sometime this week or next week) and I’ll comment in the PR. Cheers, -Quentin> On Nov 22, 2016, at 1:22 PM, Sanjin Sijaric via llvm-dev <llvm-dev at lists.llvm.org> wrote: > > Hi, > > I opened the bug #31122 for this issue. > > There are two parts to this problem. Compiled using "llc -debug-only=regalloc test.ll", the test case attached to the bug fails with > > VirtRegMap.h:119: void llvm::VirtRegMap::clearVirt(unsigned int): Assertion `Virt2PhysMap[virtReg] != NO_PHYS_REG && "attempt to clear a not assigned virtual register"' failed. > > The code looks like: > > %vreg647<def> = COPY %vreg645; QQQQPR_with_dsub_7_then_ssub_0:%vreg647 QQQQPR:%vreg645 > %vreg657<def> = COPY %vreg655; QQQQPR_with_dsub_7_then_ssub_0:%vreg657 QQQQPR:%vreg655 > %vreg679<def> = COPY %vreg678; QPR_VFP2:%vreg679 QPR:%vreg678 > %vreg657:dsub_7_then_ssub_0<def> = VDIVS %vreg679:ssub_2, %vreg647:dsub_7_then_ssub_0, pred:14, pred:%noreg; QQQQPR_with_dsub_7_then_ssub_0:%vreg657,%vreg647 QPR_VFP2:%vreg679 > %vreg659<def> = COPY %vreg657; QQQQPR:%vreg659 QQQQPR_with_dsub_7_then_ssub_0:%vreg657 > > This problem occurs during last chance recoloring, where %vreg679 doesn’t get allocated (and is marked as not spillable), so last chance recoloring kicks in. It tries to allocate Q0 to %vreg679, where Q0_Q1_Q2_Q3 is assigned to %vreg647, which interferes with %vreg679. Upon trying to allocate a register to %vreg647, %vreg679 gets evicted from Q0. But %vreg679 is a fixed register, so it shouldn’t get evicted. Recoloring is thus deemed successful, and upon returning from tryRecoloringCandidates, we assert when trying to unassign %vreg679. > > I can get around this by moving the virtual register under consideration in tryLastChanceRecoloring (%vreg679) to RS_Done stage right before the call to tryRecolringCandidates, and then restoring the original stage afterward. This prevents the eviction of %vreg679 during recoloring. Alternatively, I can check FixedRegisters in canEvictInterference and return false. > > After the assert is gone, I run into “LLVM ERROR: ran out of registers during register allocation”. Increasing lcr-max-depth doesn’t help. > > To get past this error, I am thinking about spilling an actual physical register if last chance recoloring fails. For example, > > P = …. > > virtX = …. Last chance recoloring fails, > > … = P > > …. = virtX > > will result in something along the lines of > > P = …. > > spill P to fi#0 (original P) > P = …. (virtX is now in P) > …. > Spill P to fi#1 (this is virtX) > Reload P form from fi#0 (original P) > … = P > …. > Reload P from fi#1 (this is virtX) > = P > > Does anyone have any other suggestions about dealing with this problem? > > Thank you, > Sanjin > > _______________________________________________ > LLVM Developers mailing list > llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org> > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev <http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev>-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20161128/828ae603/attachment.html>