The following fails (reduced from arm bootstrap): $ cat t.cc struct gcc_options { int x_flag_tree_phiprop; int x_flag_tree_pre; int x_flag_tree_pta; int x_flag_tree_reassoc; int x_flag_tree_scev_cprop; int x_flag_tree_sink; int x_flag_tree_slp_vectorize; int x_flag_web; } cl_optimization_restore_opts; signed char cl_optimization_restore_ptr_1, cl_optimization_restore_ptr_0; void cl_optimization_restore() { cl_optimization_restore_opts .x_flag_tree_phiprop = cl_optimization_restore_opts.x_flag_tree_pre = cl_optimization_restore_opts.x_flag_tree_pta = cl_optimization_restore_opts.x_flag_tree_reassoc = cl_optimization_restore_opts.x_flag_tree_scev_cprop = cl_optimization_restore_opts.x_flag_tree_sink = cl_optimization_restore_opts.x_flag_tree_slp_vectorize = cl_optimization_restore_ptr_0; cl_optimization_restore_opts.x_flag_web = cl_optimization_restore_ptr_1; } $ gcc/xgcc -B gcc -c t.cc -O2 -march=armv8-a+crypto during GIMPLE pass: slp t.cc: In function ‘void cl_optimization_restore()’: t.cc:12:6: internal compiler error: in supportable_widening_operation, at tree-vect-stmts.cc:12199 12 | void cl_optimization_restore() { | ^~~~~~~~~~~~~~~~~~~~~~~ 0x1dd0b13 supportable_widening_operation(vec_info*, tree_code, _stmt_vec_info*, tree_node*, tree_node*, tree_code*, tree_code*, int*, vec<tree_node*, va_heap, vl_ptr>*) /home/alecop01/toolchain/src/gcc/gcc/tree-vect-stmts.cc:12199 0x1dd5ef5 vectorizable_conversion /home/alecop01/toolchain/src/gcc/gcc/tree-vect-stmts.cc:5064 0x1df36c6 vect_analyze_stmt(vec_info*, _stmt_vec_info*, bool*, _slp_tree*, _slp_instance*, vec<stmt_info_for_cost, va_heap, vl_ptr>*) /home/alecop01/toolchain/src/gcc/gcc/tree-vect-stmts.cc:11256 0x13a164d vect_slp_analyze_node_operations_1 /home/alecop01/toolchain/src/gcc/gcc/tree-vect-slp.cc:5957 0x13a164d vect_slp_analyze_node_operations /home/alecop01/toolchain/src/gcc/gcc/tree-vect-slp.cc:6147 0x13a1556 vect_slp_analyze_node_operations /home/alecop01/toolchain/src/gcc/gcc/tree-vect-slp.cc:6126 0x13a3b37 vect_slp_analyze_operations(vec_info*) /home/alecop01/toolchain/src/gcc/gcc/tree-vect-slp.cc:6387 0x13a8ce8 vect_slp_analyze_bb_1 /home/alecop01/toolchain/src/gcc/gcc/tree-vect-slp.cc:7372 0x13a8ce8 vect_slp_region /home/alecop01/toolchain/src/gcc/gcc/tree-vect-slp.cc:7419 0x13a9b9d vect_slp_bbs /home/alecop01/toolchain/src/gcc/gcc/tree-vect-slp.cc:7610 0x13aa1c1 vect_slp_function(function*) /home/alecop01/toolchain/src/gcc/gcc/tree-vect-slp.cc:7698 0x13b3ec6 execute /home/alecop01/toolchain/src/gcc/gcc/tree-vectorizer.cc:1532 Please submit a full bug report, with preprocessed source (by using -freport-bug). Please include the complete backtrace with any bug report. See <https://gcc.gnu.org/bugs/> for instructions. $ gcc/xgcc -B gcc -v Reading specs from gcc/specs COLLECT_GCC=gcc/xgcc COLLECT_LTO_WRAPPER=gcc/lto-wrapper Target: arm-linux-gnueabihf Configured with: /home/alecop01/toolchain/src/gcc/configure --prefix=/data_sdb/toolchain/cc1s/arm-lin --enable-languages=c,c++ --disable-bootstrap --target=arm-linux-gnueabihf Thread model: posix Supported LTO compression algorithms: zlib gcc version 13.0.0 20221128 (experimental) (GCC)
Going backwards: gcc_assert (VECTOR_MODE_P (intermediate_mode)); .. intermediate_mode = insn_data[icode1].operand[0].mode; ... if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) return false;
OK, so we have vec_unpacks_lo_v2si unpacking V2Si to DImode. I can fix that easily.
The master branch has been updated by Richard Biener <rguenth@gcc.gnu.org>: https://gcc.gnu.org/g:2200b7054638bbf157192a4dca23886dd420becb commit r13-4374-g2200b7054638bbf157192a4dca23886dd420becb Author: Richard Biener <rguenther@suse.de> Date: Mon Nov 28 17:26:15 2022 +0100 tree-optimization/107896 - allow v2si to dimode unpacks The following avoids ICEing for V2SI -> DImode vec_unpacks_lo. PR tree-optimization/107896 * tree-vect-stmts.cc (supportable_widening_operation): Handle non-vector mode intermediate mode.
Should be fixed.