https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112758
Jakub Jelinek <jakub at gcc dot gnu.org> changed: What |Removed |Added ---------------------------------------------------------------------------- CC| |ebotcazou at gcc dot gnu.org, | |jakub at gcc dot gnu.org, | |law at gcc dot gnu.org, | |segher at gcc dot gnu.org --- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> --- Seems expand_compound_operation is called on (sign_extend:DI (mem/c:SI (lo_sum:DI (reg/f:DI 144) (symbol_ref:DI ("globalVar") [flags 0x86] <var_decl 0x7fffea1f9b40 globalVar>)) [1 globalVar+0 S4 A32])) and takes the 7336 tem = gen_lowpart (mode, XEXP (x, 0)); 7337 if (!tem || GET_CODE (tem) == CLOBBER) 7338 return x; 7339 tem = simplify_shift_const (NULL_RTX, ASHIFT, mode, 7340 tem, modewidth - pos - len); 7341 tem = simplify_shift_const (NULL_RTX, unsignedp ? LSHIFTRT : ASHIFTRT, 7342 mode, tem, modewidth - len); path on it, mode being DImode, modewidth 64, pos 0, len 32. The second simplify_shift_const is then called with (ashift:DI (subreg:DI (mem/c:SI (lo_sum:DI (reg/f:DI 144) (symbol_ref:DI ("globalVar") [flags 0x86] <var_decl 0x7fffea1f9b40 globalVar>)) [1 globalVar+0 S4 A32]) 0) (const_int 32 [0x20])) and triggers the 10810 /* If this was (ashiftrt (ashift foo C1) C2) and FOO has more 10811 than C1 high-order bits equal to the sign bit, we can convert 10812 this to either an ASHIFT or an ASHIFTRT depending on the 10813 two counts. 10814 10815 We cannot do this if VAROP's mode is not SHIFT_UNIT_MODE. */ 10816 10817 if (code == ASHIFTRT && first_code == ASHIFT 10818 && int_varop_mode == shift_unit_mode 10819 && (num_sign_bit_copies (XEXP (varop, 0), shift_unit_mode) 10820 > first_count)) 10821 { 10822 varop = XEXP (varop, 0); 10823 count -= first_count; 10824 if (count < 0) 10825 { 10826 count = -count; 10827 code = ASHIFT; 10828 } 10829 10830 continue; 10831 } optimization in there. As RISC V is WORD_REGISTER_OPERATIONS target with load_extend_op (E_SImode) == SIGN_EXTEND, it triggers the: 5444 /* For paradoxical SUBREGs on machines where all register operations 5445 affect the entire register, just look inside. Note that we are 5446 passing MODE to the recursive call, so the number of sign bit 5447 copies will remain relative to that mode, not the inner mode. 5448 5449 This works only if loads sign extend. Otherwise, if we get a 5450 reload for the inner part, it may be loaded from the stack, and 5451 then we lose all sign bit copies that existed before the store 5452 to the stack. */ 5453 if (WORD_REGISTER_OPERATIONS 5454 && load_extend_op (inner_mode) == SIGN_EXTEND 5455 && paradoxical_subreg_p (x) 5456 && MEM_P (SUBREG_REG (x))) 5457 return cached_num_sign_bit_copies (SUBREG_REG (x), mode, 5458 known_x, known_mode, known_ret); path and so the sign-extension in the end folds into just a paradoxical subreg of the MEM. But probably something in the combiner then just sees a paradoxical SUBREG and thinks that all the bits above the SUBREG_REG are undefined and picks ZERO_EXTEND. I'm afraid I don't know enough about WORD_REGISTER_OPERATIONS to know what is right and what is not.