This is Richi's prototype patch in https://gcc.gnu.org/bugzilla/show_bug.cgi?id=36043#c23 with fixes for blocks larger than one reg, big-endian, and BLOCK_REG_PADDING. I also removed the operand_subword_force since we may as well let narrow_bit_field_mem in extract_bit_field do that for us. It is necessary to do the BLOCK_REG_PADDING shift after we've loaded the block or else repeat the bit-field extraction in that case.
Bootstrapped and regression tested (-m32 and -m64) x86_64-linux and powerpc64-linux. OK to apply? I'll also throw together a testcase or three. For execute tests I'm thinking of using sbrk to locate an odd sized struct such that access past the end segfaults, rather than mmap/munmap as was done in the pr36043 testcase. Does that sound reasonable? PR target/65408 PR target/58744 PR middle-end/36043 * calls.c (load_register_parameters): Don't load past end of mem unless suitably aligned. Index: gcc/calls.c =================================================================== --- gcc/calls.c (revision 221435) +++ gcc/calls.c (working copy) @@ -2090,6 +2090,26 @@ load_register_parameters (struct arg_data *args, i (XEXP (args[i].value, 0), size))) *sibcall_failure = 1; + if (size % UNITS_PER_WORD == 0 + || MEM_ALIGN (mem) % BITS_PER_WORD == 0) + move_block_to_reg (REGNO (reg), mem, nregs, args[i].mode); + else + { + if (nregs > 1) + move_block_to_reg (REGNO (reg), mem, nregs - 1, + args[i].mode); + rtx dest = gen_rtx_REG (word_mode, REGNO (reg) + nregs - 1); + unsigned int bitoff = (nregs - 1) * BITS_PER_WORD; + unsigned int bitsize = size * BITS_PER_UNIT - bitoff; + rtx x = extract_bit_field (mem, bitsize, bitoff, 1, + dest, word_mode, word_mode); + if (BYTES_BIG_ENDIAN) + x = expand_shift (LSHIFT_EXPR, word_mode, x, + BITS_PER_WORD - bitsize, dest, 1); + if (x != dest) + emit_move_insn (dest, x); + } + /* Handle a BLKmode that needs shifting. */ if (nregs == 1 && size < UNITS_PER_WORD #ifdef BLOCK_REG_PADDING @@ -2097,22 +2117,18 @@ load_register_parameters (struct arg_data *args, i #else && BYTES_BIG_ENDIAN #endif - ) + ) { - rtx tem = operand_subword_force (mem, 0, args[i].mode); - rtx ri = gen_rtx_REG (word_mode, REGNO (reg)); - rtx x = gen_reg_rtx (word_mode); + rtx dest = gen_rtx_REG (word_mode, REGNO (reg)); int shift = (UNITS_PER_WORD - size) * BITS_PER_UNIT; - enum tree_code dir = BYTES_BIG_ENDIAN ? RSHIFT_EXPR - : LSHIFT_EXPR; + enum tree_code dir = (BYTES_BIG_ENDIAN + ? RSHIFT_EXPR : LSHIFT_EXPR); + rtx x; - emit_move_insn (x, tem); - x = expand_shift (dir, word_mode, x, shift, ri, 1); - if (x != ri) - emit_move_insn (ri, x); + x = expand_shift (dir, word_mode, dest, shift, dest, 1); + if (x != dest) + emit_move_insn (dest, x); } - else - move_block_to_reg (REGNO (reg), mem, nregs, args[i].mode); } /* When a parameter is a block, and perhaps in other cases, it is -- Alan Modra Australia Development Lab, IBM