https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68416
--- Comment #1 from Ilya Enkovich <ienkovich at gcc dot gnu.org> ---
Seems the problem is in register allocation. Before allocation we have:
1: NOTE_INSN_DELETED
6: NOTE_INSN_BASIC_BLOCK 2
2: r95:DI=di:DI
REG_DEAD di:DI
3: r96:SI=si:SI
REG_DEAD si:SI
4: r97:BND64=bnd0:BND64
REG_DEAD bnd0:BND64
5: NOTE_INSN_FUNCTION_BEG
8: {unspec[r97:BND64,r95:DI] 92;[r95:DI]=unspec[[r95:DI]] 95;}
9: NOTE_INSN_DELETED
10: {unspec[r97:BND64,r95:DI+0x7] 93;[r95:DI+0x7]=unspec[[r95:DI+0x7]] 95;}
REG_DEAD r97:BND64
11: r87:DI=[r95:DI]
12: {r92:BND64=unspec[[unspec[r95:DI,r87:DI] 91]] 90;use [r95:DI];}
REG_DEAD r95:DI
13: r99:DI=sign_extend(r96:SI)
REG_DEAD r96:SI
14: NOTE_INSN_DELETED
15: r90:DI=r99:DI<<0x2+r87:DI
REG_DEAD r99:DI
REG_DEAD r87:DI
16: {unspec[r92:BND64,r90:DI] 92;[r90:DI]=unspec[[r90:DI]] 95;}
17: NOTE_INSN_DELETED
18: {unspec[r92:BND64,r90:DI+0x3] 93;[r90:DI+0x3]=unspec[[r90:DI+0x3]] 95;}
REG_DEAD r92:BND64
19: r103:SI=[r90:DI]
REG_DEAD r90:DI
24: ax:SI=r103:SI
REG_DEAD r103:SI
25: use ax:SI
After allocation:
1: NOTE_INSN_DELETED
6: NOTE_INSN_BASIC_BLOCK 2
4: [sp:DI-0x18]=bnd0:BND64
5: NOTE_INSN_FUNCTION_BEG
28: bnd1:BND64=[sp:DI-0x18]
8: {unspec[bnd1:BND64,di:DI] 92;[di:DI]=unspec[[di:DI]] 95;}
9: NOTE_INSN_DELETED
29: bnd2:BND64=[sp:DI-0x18]
10: {unspec[bnd2:BND64,di:DI+0x7] 93;[di:DI+0x7]=unspec[[di:DI+0x7]] 95;}
11: ax:DI=[di:DI]
12: {bnd3:BND64=unspec[[unspec[di:DI,ax:DI] 91]] 90;use [di:DI];}
30: [sp:DI-0x18]=bnd3:BND64
13: si:DI=sign_extend(si:SI)
14: NOTE_INSN_DELETED
15: ax:DI=si:DI<<0x2+ax:DI
31: bnd0:BND64=[sp:DI-0x18]
16: {unspec[bnd0:BND64,ax:DI] 92;[ax:DI]=unspec[[ax:DI]] 95;}
17: NOTE_INSN_DELETED
32: bnd1:BND64=[sp:DI-0x18]
18: {unspec[bnd1:BND64,ax:DI+0x3] 93;[ax:DI+0x3]=unspec[[ax:DI+0x3]] 95;}
19: ax:SI=[ax:DI]
24: ax:SI=ax:SI
25: use ax:SI
27: NOTE_INSN_DELETED
Here even simple r97:BND64=bnd0:BND64 goes through stack with a load at each
use like if bounds register usage kills it.
