RPCS3/llvm-mirror
1
0
Fork 0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-11-23 03:02:36 +01:00
Code Issues Projects Releases Wiki Activity

ReMat: fix overly cavalier attitude to sub-register indices

Browse Source 
There are two attempted optimisations in reMaterializeTrivialDef, trying to
avoid promoting the size of a register too much when rematerializing.
Unfortunately, both appear to be flawed. First, we see if the original register
would have worked, but this is inadequate. Consider:

    v1 = SOMETHING (v1 is QQ)
    v2:Q0 = COPY v1:Q1 (v1, v2 are QQ)
    ...
    uses of v2

In this case even though v2 *could* be used directly as the output of
SOMETHING, this would set the wrong bits of the QQ register involved. The
correct rematerialization must be:

    v2:Q0_Q1 = SOMETHING (v2 promoted to QQQ)
    ...
    uses of v2:Q1_Q2

For the second optimisation, if the correct remat is "v2:idx = SOMETHING" then
we can't necessarily expect v2 itself to be valid for SOMETHING, but we do try
to hunt for a class between v1 and v2 that works. Unfortunately, this is also
wrong:

    v1 = SOMETHING (v1 is QQ)
    v2:Q0_Q1 = COPY v1 (v1 is QQ, v2 is QQQ)
    ...
    uses of v2 as a QQQ

The canonical rematerialization here is "v2:Q0_Q1 = SOMETHING". However current
logic would decide that v2 could be a QQ (no interest is taken in later uses).

This patch, therefore, always accepts the widened register class without trying
to be clever. Generally there is no penalty to this (e.g. in the common GR32 <
GR64 case, expanding the width doesn't matter because it's not like you were
going to do anything else with the high bits of a GR32 register). It can
increase register pressure in cases like the ARM VFP regs though (multiple
non-overlapping but equivalent subregisters). Hopefully this situation is rare
enough that it won't matter.

Unfortunately, no in-tree targets actually expose this as far as I can tell
(there are so few isAsCheapAsAMove instructions for it to trigger on) so I've
been unable to produce a test. It was exposed in our ARM64 SPEC tests though,
and I will be adding a test there that we should be able to contribute
soon(TM).

llvm-svn: 199091
This commit is contained in:
Tim Northover 2014-01-13 10:47:01 +00:00
parent 20415ba0f4
commit 2c63f62596
1 changed files with 4 additions and 24 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

28
lib/CodeGen/RegisterCoalescer.cpp
View File

@ -816,31 +816,11 @@ bool RegisterCoalescer::reMaterializeTrivialDef(CoalescerPair &CP,
}
if (TargetRegisterInfo::isVirtualRegister(DstReg)) {
unsigned NewIdx = NewMI->getOperand(0).getSubReg();
const TargetRegisterClass *RCForInst;
if (NewIdx)
RCForInst = TRI->getMatchingSuperRegClass(MRI->getRegClass(DstReg), DefRC,
NewIdx);
MRI->setRegClass(DstReg, CP.getNewRC());
if (MRI->constrainRegClass(DstReg, DefRC)) {
// The materialized instruction is quite capable of setting DstReg
// directly, but it may still have a now-trivial subregister index which
// we should clear.
NewMI->getOperand(0).setSubReg(0);
} else if (NewIdx && RCForInst) {
// The subreg index on NewMI is essential; we still have to make sure
// DstReg:idx is in a class that NewMI can use.
MRI->constrainRegClass(DstReg, RCForInst);
} else {
// DstReg is actually incompatible with NewMI, we have to move to a
// super-reg's class. This could come from a sequence like:
// GR32 = MOV32r0
// GR8 = COPY GR32:sub_8
MRI->setRegClass(DstReg, CP.getNewRC());
updateRegDefsUses(DstReg, DstReg, DstIdx);
NewMI->getOperand(0).setSubReg(
TRI->composeSubRegIndices(SrcIdx, DefMI->getOperand(0).getSubReg()));
}
unsigned NewIdx = NewMI->getOperand(0).getSubReg();
updateRegDefsUses(DstReg, DstReg, DstIdx);
NewMI->getOperand(0).setSubReg(NewIdx);
} else if (NewMI->getOperand(0).getReg() != CopyDstReg) {
// The New instruction may be defining a sub-register of what's actually
// been asked for. If so it must implicitly define the whole thing.