https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93376
--- Comment #8 from CVS Commits <cvs-commit at gcc dot gnu.org> --- The master branch has been updated by Jakub Jelinek <ja...@gcc.gnu.org>: https://gcc.gnu.org/g:c124b345e460780e3c3d4a6e58d0e6e03da982a1 commit r10-6178-gc124b345e460780e3c3d4a6e58d0e6e03da982a1 Author: Jakub Jelinek <ja...@redhat.com> Date: Thu Jan 23 16:17:56 2020 +0100 i386: Fix ICEs on TImode signed overflow add/sub patterns [PR93376] The following testcase ICEs, because during try_combine of i3: (insn 18 17 19 2 (parallel [ (set (reg:CCO 17 flags) (eq:CCO (plus:OI (sign_extend:OI (reg:TI 96)) (const_int 1 [0x1])) (sign_extend:OI (plus:TI (reg:TI 96) (const_int 1 [0x1]))))) (set (reg:TI 98) (plus:TI (reg:TI 96) (const_int 1 [0x1]))) ]) "pr93376.c":8:10 223 {*addvti4_doubleword_1} (expr_list:REG_UNUSED (reg:TI 98) (expr_list:REG_DEAD (reg:TI 96) (nil)))) and i2: (insn 17 37 18 2 (set (reg:TI 96) (const_wide_int 0x7fffffffffffffffffffffffffffffff)) "pr93376.c":8:10 65 {*movti_internal} (nil)) the eq in there gets simplified into: (eq:CCO (const_wide_int 0x080000000000000000000000000000000) (const_wide_int 0x80000000000000000000000000000000)) and simplify-rtx.c tries to simplify it by simplifying MINUS of the two operands. Now, i386 defines MAX_BITSIZE_MODE_ANY_INT to 128, because OImode and XImode are used mainly as a placeholder for the vector modes; these new signed overflow patterns are an exception to that, but what they really need is just TImode precision + 1 (maybe 2 worst case) bits at any time. wide-int.h defines WIDE_INT_MAX_ELTS in a way that it contains one more HWI above number of HWIs to cover WIDE_INT_MAX_ELTS, so on i386 that is 3 HWIs, meaning that TImode precision + 1/2 bits is still representable in there. Unfortunately, the way wi::sub_large is implemented, it needs not just those 3 HWIs, but one HWI above the maximum of the lengths of both operands, which means it buffer overflows, overwrites the following precision in wide_int_storage and ICEs later on. The need for 4 HWIs is only temporary, because canonize immediately after it canonicalizes it back to 3 HWIs only. The patch is something suggested by Richard S., avoid using OImode for this and instead use a partial int mode that is smaller. 2020-01-23 Jakub Jelinek <ja...@redhat.com> PR target/93376 * config/i386/i386-modes.def (POImode): New mode. (MAX_BITSIZE_MODE_ANY_INT): Change from 128 to 160. * config/i386/i386.md (DPWI): New mode attribute. (addv<mode>4, subv<mode>4): Use <DPWI> instead of <DWI>. (QWI): Rename to... (QPWI): ... this. Use POI instead of OI for TImode. (*addv<dwi>4_doubleword, *addv<dwi>4_doubleword_1, *subv<dwi>4_doubleword, *subv<dwi>4_doubleword_1): Use <QPWI> instead of <QWI>. * gcc.dg/pr93376.c: New test.
