Building kernels with frame pointers results in performance loss of roughly
10%, and it was suggested that I try -ffixed-reg to see if loss tracks the cost
of losing a register.  Doing so for rbp led to the gripe below.  To reproduce,
simply compile any kernel source tree with a config having CONFIG_FRAME_POINTER
disabled, and add -ffixed-rbp to CFLAGS.

crypto/cbc.c: In function 'crypto_cbc_decrypt':
crypto/cbc.c:191:1: internal compiler error: in dwarf2out_frame_debug_expr, at
dwarf2cfi.c:1520

This was gcc-4.8.5, but 6.1.1 does the very same.

Additionally, starting with -ffixed-r8, and progressively stealing more
registers, when I stole r8-r11, gcc became upset in a different way.

drivers/isdn/hisax/tei.c: In function 'tei_id_denied':
drivers/isdn/hisax/tei.c:208:1: internal compiler error: in assign_by_spills,
at lra-assigns.c:1281