https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99920
--- Comment #15 from Richard Biener <rguenth at gcc dot gnu.org> --- Dump of assembler code from 0x1094366c to 0x109436e0: 0x000000001094366c <operand_compare::hash_operand(tree_node const*, inchash::hash&, unsigned int)+2432>: lhz r0,0(r9) 0x0000000010943670 <operand_compare::hash_operand(tree_node const*, inchash::hash&, unsigned int)+2436>: clrldi r0,r0,48 0x0000000010943674 <operand_compare::hash_operand(tree_node const*, inchash::hash&, unsigned int)+2440>: stw r0,200(r31) 0x0000000010943678 <operand_compare::hash_operand(tree_node const*, inchash::hash&, unsigned int)+2444>: lwz r0,200(r31) => 0x000000001094367c <operand_compare::hash_operand(tree_node const*, inchash::hash&, unsigned int)+2448>: clrldi r0,r0,32 0x0000000010943680 <operand_compare::hash_operand(tree_node const*, inchash::hash&, unsigned int)+2452>: extsw r0,r0 0x0000000010943684 <operand_compare::hash_operand(tree_node const*, inchash::hash&, unsigned int)+2456>: ld r9,-11776(r2) 0x0000000010943688 <operand_compare::hash_operand(tree_node const*, inchash::hash&, unsigned int)+2460>: rldicr r0,r0,2,61 0x000000001094368c <operand_compare::hash_operand(tree_node const*, inchash::hash&, unsigned int)+2464>: add r9,r9,r0 0x0000000010943690 <operand_compare::hash_operand(tree_node const*, inchash::hash&, unsigned int)+2468>: lwz r0,0(r9) 0x0000000010943694 <operand_compare::hash_operand(tree_node const*, inchash::hash&, unsigned int)+2472>: stw r0,152(r31) r0 is ADDR_EXPR (130) (gdb) p/x $r2 $21 = 0x12b81a48 (gdb) p/x tree_code_type $22 = 0x123216a0 <tree_code_type> after ld r9,-11776(r2): (gdb) p/x $r9 $25 = 0x12a9fc38 rldicr r0,r0,2,61 (gdb) p $r0 $26 = 520 add r9,r9,r0 (gdb) p/x $r9 $27 = 0x12a9fe40 and then we load 0 (tcc_exceptional) from $r9, but (gdb) p tree_code_type[130] $31 = tcc_expression so sth is obviously broken here, but I don't understand ppc assembly enough to spot what. I guess r2 is "off"? It has this same value at function entry already. Are we maybe running into some linker relaxation / TOC overflow issue (given that -O2 works)? We're using -mno-minimal-toc when building libgcc but the host compiler uses plain -g -fno-PIE and we don't supply any special linker flags for cc1 either. Note this build uses binutils 2.25.0 (Bill, is that what you use as well?)
