Regalloc should deal with running out of registers more gracefully

[efriedma-quic]

Bugzilla Link	2084
Resolution	WORKSFORME
Resolved on	Oct 14, 2011 21:26
Version	unspecified
OS	Linux

Extended Description

Testcase (note that it is impossible to allocate 8 general regs to an asm on x86):

int a() {int a,b,c,d,di,si,bp,sp;; asm volatile ("%0 %1 %2 %3 %4 %5 %6 %7" :
"=r"(a), "=r"(b), "=r"(c), "=r"(d), "=r"(di), "=r"(si), "=r"(bp), "=r"(sp) : );
return a;}

gcc gives the following error:
: In function ‘a’:
:1: error: can't find a register in class ‘GENERAL_REGS’ while
reloading ‘asm’

llc -regalloc=linearscan hangs.

llc -regalloc=local silently puts multiple outputs into the same register.

llc -regalloc=simple gives the following assertion:
llc: /home/eli/llvm/lib/CodeGen/RegAllocSimple.cpp:128: unsigned
int::RegAllocSimple::getFreeReg(unsigned int): Assertion `RI+regIdx !=
RE && "Not enough registers!"' failed.

llc -regalloc=bigblock silently puts multiple outputs into the same register.

[llvmbot]

I think all it needs to do is to scan the current MI to see if another operand
has already allocated a register that overlaps with the register that's chosen.
I would only put the check in DEBUG though because compile time is very
important when it comes to local allocator.

[efriedma-quic]

The situation here is improved a lot on trunk:

:1:50: error: ran out of registers during register allocation
int a() {int a,b,c,d,di,si,bp,sp;; asm volatile ("%0 %1 %2 %3 %4 %5 %6 %7" :
^