error: ran out of registers during register allocation

[pageexec]

Bugzilla Link	11078
Resolution	FIXED
Resolved on	Oct 12, 2011 18:38
Version	trunk
OS	Linux
Blocks	#4440
Attachments	preprocessed source
CC	@efriedma-quic,@stoklund

Extended Description

when compiling linux-3.0.4/arch/x86/kernel/machine_kexec_32.c i get the following error on i386:

linux-3.0.4-pax-clang/arch/x86/include/asm/cmpxchg_32.h:79:15: error: ran out of registers during register allocation
asm volatile("\n1:\t"
^
with the following inline asm:

79 »·······asm volatile("\n1:\t"
80 »·······»······· LOCK_PREFIX "cmpxchg8b %0\n\t"
81 »·······»······· "jnz 1b"
82 »·······»······· : "=m" (*ptr), "+A" (prev)
83 »·······»······· : "b" (low), "c" (high)
84 »·······»······· : "memory");

this code used to compile fine some months ago. preprocessed source attached, the command line needed to compile it:

clang -Wp,-MD,arch/x86/kernel/.machine_kexec_32.o.d -nostdinc -Iarch/x86/include/generated -Iinclude -include include/generated/autoconf.h -D__KERNEL__ -Wall -Wundef -Wstrict-prototypes -Wno-trigraphs -Qunused-arguments -W -Wno-unused-parameter -Wno-missing-field-initializers -Wno-self-assign -Wno-unused-value -Wno-format -mno-sse -Wno-unknown-warning-option -fcatch-undefined-behavior -fno-strict-aliasing -fno-common -Werror-implicit-function-declaration -Wno-format-security -fno-delete-null-pointer-checks -Wno-empty-body -O2 -m32 -msoft-float -mregparm=3 -freg-struct-return -mpreferred-stack-boundary=2 -march=i686 -mtune=core2 -maccumulate-outgoing-args -ffreestanding -DCONFIG_AS_CFI=1 -DCONFIG_AS_CFI_SIGNAL_FRAME=1 -DCONFIG_AS_CFI_SECTIONS=1 -pipe -Wno-sign-compare -fno-asynchronous-unwind-tables -mno-sse -mno-mmx -mno-sse2 -mno-3dnow -Wframe-larger-than=1024 -fno-stack-protector -Wno-unused-but-set-variable -fno-omit-frame-pointer -fno-optimize-sibling-calls -g -femit-struct-debug-baseonly -pg -Wdeclaration-after-statement -Wno-pointer-sign -fno-strict-overflow -fconserve-stack -D"KBUILD_STR(s)=#s" -D"KBUILD_BASENAME=KBUILD_STR(machine_kexec_32)" -D"KBUILD_MODNAME=KBUILD_STR(machine_kexec_32)" -c -o machine_kexec_32.o machine_kexec_32.i

[efriedma-quic]

Reduced testcase
Attached fails with "llc b.ll -disable-fp-elim=1". Not entirely sure why; there are clearly enough registers available here.

[stoklund]

The problem is %vreg2:

0L BB#0: derived from LLVM BB %entry
16L %vreg1 = MOV32rm <fi#-1>, 1, %noreg, 0, %noreg; mem:LD4FixedStack-1 GR32:%vreg1
32L %vreg2 = MOV32rm %vreg1, 1, %noreg, 0, %noreg; mem:LD4[%pgd] GR32_ABCD:%vreg2 GR32:%vreg1
80L TEST8ri %vreg2:sub_8bit, 1, %EFLAGS; GR32_ABCD:%vreg2
96L JNE_4 <BB#2>, %EFLAGS<imp-use,kill>
112L JMP_4 <BB#1>
Successors according to CFG: BB#1 BB#2

128L BB#1: derived from LLVM BB %if.then
Predecessors according to CFG: BB#0
144L INLINEASM es:bar [sideeffect], $0:[regdef], %vreg4<def,dead>, $1:[mem], %noreg, 1, %noreg, <ga:@__forc
e_order>, %noreg, $2:[clobber], %EFLAGS<earlyclobber,imp-def,dead>; GR32:%vreg4
Successors according to CFG: BB#2

160L BB#2: derived from LLVM BB %if.end
Predecessors according to CFG: BB#0 BB#1
176L %vreg5 = MOV32r0 %EFLAGS<imp-def,dead>; GR32:%vreg5
192L %EBX = MOV32r0 %EFLAGS<imp-def,dead>
208L %ECX = MOV32r0 %EFLAGS<imp-def,dead>
240L %EAX = COPY %vreg5; GR32:%vreg5
256L %EDX = MOV32r0 %EFLAGS<imp-def,dead>
272L INLINEASM es:foo [sideeffect], $0:[mem], %vreg2, 1, %noreg, 0, %noreg, $1:[regdef], %EAX<imp-def
,dead>, %EDX<imp-def,dead>, $2:[reguse], %EBX, $3:[reguse], %ECX, $4:[reguse] [tiedto:$1], %EAX, %EDX, $5:[c
lobber], %EFLAGS<earlyclobber,imp-def,dead>; GR32_ABCD:%vreg2
288L RET

The register gets constrained to GR32_ABCD because of 'TEST8ri %vreg2:sub_8bit', but the INLINEASM already has fixed arguments for all of the GR32_ABCD registers.

Because of that, the live range is split in three:

Inflated %vreg9 to GR32
queuing new interval: %vreg9,3.506945e-03 = [88d,264d:0) 0@88d
queuing new interval: %vreg10,6.644737e-03 = [32d,88d:0) 0@32d
queuing new interval: %vreg11,inf = [264d,272d:0) 0@264d

The %vreg11 register isolates the INLINEASM, and it should be inflated to GR32 like %vreg9, but since we don't capture the inline asm constraints in the MI operands, the register class inflater backs off and leaves it as the safe GR32.

[stoklund]

r141836

[edwintorok]

mentioned in issue #4440