https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96998
--- Comment #12 from CVS Commits <cvs-commit at gcc dot gnu.org> --- The master branch has been updated by Alex Coplan <acop...@gcc.gnu.org>: https://gcc.gnu.org/g:7de23b8c536397117bbea04a722fa1b86564dd7c commit r11-4563-g7de23b8c536397117bbea04a722fa1b86564dd7c Author: Alex Coplan <alex.cop...@arm.com> Date: Fri Oct 30 09:21:31 2020 +0000 aarch64: Fix PR96998 and restore code quality in combine This change fixes a bug in the AArch64 backend. Currently, we accept an odd sign_extract representation of addresses, but don't accept that same odd form of address as an LEA. This is the cause of PR96998. In the testcase given in the PR, combine produces: (insn 9 8 10 3 (set (mem:SI (plus:DI (sign_extract:DI (mult:DI (subreg:DI (reg/v:SI 92 [ g ]) 0) (const_int 4 [0x4])) (const_int 34 [0x22]) (const_int 0 [0])) (reg/f:DI 96)) [3 *i_5+0 S4 A32]) (asm_operands:SI ("") ("=Q") 0 [] [] [] test.c:11)) "test.c":11:5 -1 (expr_list:REG_DEAD (reg/v:SI 92 [ g ]) (nil))) Then LRA reloads the address and we ICE because we fail to recognize the sign_extract outside the mem: (insn 33 8 34 3 (set (reg:DI 100) (sign_extract:DI (ashift:DI (subreg:DI (reg/v:SI 92 [ g ]) 0) (const_int 2 [0x2])) (const_int 34 [0x22]) (const_int 0 [0]))) "test.c":11:5 -1 (nil)) The aarch64 changes here remove the support for this sign_extract representation of addresses, fixing PR96998. Now this by itself would regress code quality, so this change is paired with an improvement to combine which prevents an extract rtx from being emitted in this case: we now write the rtx above as a shift of an extend, which allows the combination to go ahead. Prior to this, combine.c:make_extraction() identified where we can emit an ashift of an extend in place of an extraction, but failed to make the corresponding canonicalization/simplification when presented with a mult by a power of two. Such a representation is canonical when representing a left-shifted address inside a mem. This change remedies this situation. For rtxes such as: (mult:DI (subreg:DI (reg:SI r) 0) (const_int 2^n)) where the bottom 32 + n bits are valid (the higher-order bits are undefined) and make_extraction() is being asked to sign_extract the lower (valid) bits, after the patch, we rewrite this as: (mult:DI (sign_extend:DI (reg:SI r)) (const_int 2^n)) instead of using a sign_extract. gcc/ChangeLog: PR target/96998 * combine.c (make_extraction): Also handle shifts written as (mult x 2^n), avoid creating an extract rtx for these. * config/aarch64/aarch64.c (aarch64_is_extend_from_extract): Delete. (aarch64_classify_index): Remove extract-based address handling. (aarch64_strip_extend): Likewise. (aarch64_rtx_arith_op_extract_p): Likewise, remove now-unused parameter. Update callers... (aarch64_rtx_costs): ... here. gcc/testsuite/ChangeLog: PR target/96998 * gcc.c-torture/compile/pr96998.c: New test.
