https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117182
--- Comment #2 from Kazumoto Kojima <kkojima at gcc dot gnu.org> --- Created attachment 59366 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=59366&action=edit a reduced test case (-m4 -mlra -O2 -ftrace) qemu-sh4-static failed with Unhandled trap: 0x1a0 pc=0x004004f4 sr=0x00000001 pr=0x004004f0 fpscr=0x00080000 spc=0x00000000 ssr=0x00000000 gbr=0x3ffbd1e0 vbr=0x00000000 sgr=0x00000000 dbr=0x00000000 delayed_pc=0x0040055e fpul=0x00000000 r0=0x00000000 r1=0x00000001 r2=0x00000002 r3=0x00000000 r4=0x407ffda0 r5=0x407ffdb4 r6=0x407ffebc r7=0x00000000 r8=0x407ffdec r9=0x407ffdf0 r10=0x407ffde0 r11=0x407ffddc r12=0x00400ca4 r13=0x3ffffa40 r14=0x00000000 r15=0x407ffdd4 r16=0x00000000 r17=0x00000000 r18=0x00000000 r19=0x00000000 r20=0x00000000 r21=0x00000000 r22=0x00000000 r23=0x00000000 in conditional delay slot (delayed_pc=0x0040055e) It turned out that fldi1 was executed when fpscr == 0x80001 unexpectedly. It seems qemu fails with the unhandled trap for that situation. The fldi1 insn in problem is generated with lra-constraints.cc:curr_insn_transform. The LRA dump says Changing pseudo 408 in operand 1 of insn 406 on equiv 1.0e+0 Considering alt=0 of insn 406: (0) =f (1) f 1 Non-pseudo reload: reject+=2 1 Non input pseudo reload: reject++ overall=9,losers=1,rld_nregs=1 Considering alt=1 of insn 406: (0) r (1) r 0 Non input pseudo reload: reject++ 1 Non-pseudo reload: reject+=2 1 Non input pseudo reload: reject++ overall=16,losers=2 -- refuse Considering alt=2 of insn 406: (0) f (1) G Bad operand -- refuse Considering alt=3 of insn 406: (0) f (1) H overall=0,losers=0,rld_nregs=0 Choosing alt 3 in insn 406: (0) f (1) H {movsf_ie_ra} Looks LRA tries to substitute a reg-reg move into a constant load when the source pseudo reg is equivalent to a constant. Unfortunately this is unsafe on SH. If the constant is the SFmode 0.0 or 1.0, then that constant load will end up with fldi0/1 which is usable only with single precision mode. We have already cannot_substitute_mem_equiv_p target hook to avoid similar transformations when the source pseudo reg is equivalent to a memory. I'm trying to add cannot_substitute_const_equiv_p target hook to avoid that substitution to fldi0/1.
