Created attachment 47945 [details] reduced testcase This affects several targets, but I failed to reproduce this on x86_64. Compiler output: $ riscv64-unknown-linux-gnu-gcc -O2 testcase.c during GIMPLE pass: tailr testcase.c: In function 'bar.isra': testcase.c:4:1: internal compiler error: in fold_convert_loc, at fold-const.c:2435 4 | bar (unsigned char e) | ^~~ 0x631c58 fold_convert_loc(unsigned int, tree_node*, tree_node*) /repo/gcc-trunk/gcc/fold-const.c:2435 0x10e6d1c update_accumulator_with_ops /repo/gcc-trunk/gcc/tree-tailcall.c:832 0x10e9261 adjust_accumulator_values /repo/gcc-trunk/gcc/tree-tailcall.c:877 0x10e9261 eliminate_tail_call /repo/gcc-trunk/gcc/tree-tailcall.c:1018 0x10e9261 optimize_tail_call /repo/gcc-trunk/gcc/tree-tailcall.c:1041 0x10e9261 tree_optimize_tail_calls_1 /repo/gcc-trunk/gcc/tree-tailcall.c:1178 Please submit a full bug report, with preprocessed source if appropriate. Please include the complete backtrace with any bug report. See <https://gcc.gnu.org/bugs/> for instructions. $ riscv64-unknown-linux-gnu-gcc -v Using built-in specs. COLLECT_GCC=/repo/gcc-trunk/binary-latest-riscv64/bin/riscv64-unknown-linux-gnu-gcc COLLECT_LTO_WRAPPER=/repo/gcc-trunk/binary-trunk-20200301175845-g649e174102a-checking-yes-rtl-df-extra-riscv64/bin/../libexec/gcc/riscv64-unknown-linux-gnu/10.0.1/lto-wrapper Target: riscv64-unknown-linux-gnu Configured with: /repo/gcc-trunk//configure --enable-languages=c,c++ --enable-valgrind-annotations --disable-nls --enable-checking=yes,rtl,df,extra --with-cloog --with-ppl --with-isl --with-sysroot=/usr/riscv64-unknown-linux-gnu --build=x86_64-pc-linux-gnu --host=x86_64-pc-linux-gnu --target=riscv64-unknown-linux-gnu --with-ld=/usr/bin/riscv64-unknown-linux-gnu-ld --with-as=/usr/bin/riscv64-unknown-linux-gnu-as --disable-multilib --disable-libstdcxx-pch --prefix=/repo/gcc-trunk//binary-trunk-20200301175845-g649e174102a-checking-yes-rtl-df-extra-riscv64 Thread model: posix Supported LTO compression algorithms: zlib zstd gcc version 10.0.1 20200302 (experimental) (GCC)
I'm bisecting that right now.
Started with r10-3542-g0b92cf305dcf34387a8e2564e55ca8948df3b47a, so it was probably a latent issue.
Even void bar (unsigned char e) { bar (3); unsigned char c; c = -e; } ICEs with -O2 -fno-tree-dce the same way. So, the question is: 1) why on the #c0 testcase doesn't DCE do its job 2) why does the tailr pass consider it something that needs to be accumulated (in a void temporary)
--- gcc/tree-tailcall.c.jj 2020-01-12 11:54:38.517381665 +0100 +++ gcc/tree-tailcall.c 2020-03-03 20:22:17.324133660 +0100 @@ -339,7 +339,8 @@ process_assignment (gassign *stmt, && (non_ass_var = independent_of_stmt_p (op1, stmt, call, to_move))) ; - else if (op1 == *ass_var + else if (*ass_var + && op1 == *ass_var && (non_ass_var = independent_of_stmt_p (op0, stmt, call, to_move))) ; For the void calls *ass_var is NULL and if we have an UNARY_RHS assignment, op1 is also NULL and so will appear to match, while it shouldn't.
The master branch has been updated by Jakub Jelinek <jakub@gcc.gnu.org>: https://gcc.gnu.org/g:8e480ec1ddb307150cb36cf610b9eeab591de216 commit r10-7009-g8e480ec1ddb307150cb36cf610b9eeab591de216 Author: Jakub Jelinek <jakub@redhat.com> Date: Wed Mar 4 09:01:59 2020 +0100 tailcall: Fix up process_assignment [PR94001] When a function returns void or the return value is ignored, ass_var is NULL_TREE. The tail recursion handling generally assumes DCE has been performed and so doesn't expect to encounter useless assignments after the call and expects them to be part of the return value adjustment that need to be changed into tail recursion additions/multiplications. process_assignment does some verification and has a way to tell the caller to try to move dead or whatever other stmts that don't participate in the return value modifications before it is returned. For binary rhs assignments it is just fine, neither op0 nor op1 will be NULL_TREE and thus if *ass_var is NULL_TREE, it will not match, but unary rhs is handled by only setting op0 to rhs1 and setting op1 to NULL_TREE. And at this point, NULL_TREE == NULL_TREE and thus we think e.g. the c_2 = -e_3(D); dead stmt is actually a return value modification, so we queue it as multiplication and then create a void type SSA_NAME accumulator for it and ICE shortly after. Fixed by making sure op1 == *ass_var comparison is done only if *ass_var. 2020-03-04 Jakub Jelinek <jakub@redhat.com> PR tree-optimization/94001 * tree-tailcall.c (process_assignment): Before comparing op1 to *ass_var, verify *ass_var is non-NULL. * gcc.dg/pr94001.c: New test.
Fixed.
The releases/gcc-9 branch has been updated by Jakub Jelinek <jakub@gcc.gnu.org>: https://gcc.gnu.org/g:5de0dc84c75a43f78a03c7cdb7e7c443c641a7fa commit r9-8385-g5de0dc84c75a43f78a03c7cdb7e7c443c641a7fa Author: Jakub Jelinek <jakub@redhat.com> Date: Wed Mar 4 09:01:59 2020 +0100 tailcall: Fix up process_assignment [PR94001] When a function returns void or the return value is ignored, ass_var is NULL_TREE. The tail recursion handling generally assumes DCE has been performed and so doesn't expect to encounter useless assignments after the call and expects them to be part of the return value adjustment that need to be changed into tail recursion additions/multiplications. process_assignment does some verification and has a way to tell the caller to try to move dead or whatever other stmts that don't participate in the return value modifications before it is returned. For binary rhs assignments it is just fine, neither op0 nor op1 will be NULL_TREE and thus if *ass_var is NULL_TREE, it will not match, but unary rhs is handled by only setting op0 to rhs1 and setting op1 to NULL_TREE. And at this point, NULL_TREE == NULL_TREE and thus we think e.g. the c_2 = -e_3(D); dead stmt is actually a return value modification, so we queue it as multiplication and then create a void type SSA_NAME accumulator for it and ICE shortly after. Fixed by making sure op1 == *ass_var comparison is done only if *ass_var. 2020-03-04 Jakub Jelinek <jakub@redhat.com> PR tree-optimization/94001 * tree-tailcall.c (process_assignment): Before comparing op1 to *ass_var, verify *ass_var is non-NULL. * gcc.dg/pr94001.c: New test.