Hi! The MIN_EXPR/MAX_EXPR handling in *build_binary_op is minimal (especially for C FE), because min/max aren't expressions the languages contain directly. I'm using those for the #pragma omp atomic x = x < y ? y : x; forms, but e.g. for the attached testcase we normally reject _Complex int vs. int comparisons, in C++ due to MIN/MAX_EXPR we were diagnosing it as invalid types for <unknown> while in C we accept it and ICEd later on.
The following patch will try build_binary_op with LT_EXPR on the operands first to get needed diagnostics and fail if it returns error_mark_node. Bootstrapped/regtested on x86_64-linux and i686-linux, committed to trunk. 2022-02-16 Jakub Jelinek <ja...@redhat.com> PR c/104531 * c-omp.cc (c_finish_omp_atomic): For MIN_EXPR/MAX_EXPR, try first build_binary_op with LT_EXPR and only if that doesn't return error_mark_node call build_modify_expr. * c-c++-common/gomp/atomic-31.c: New test. --- gcc/c-family/c-omp.cc.jj 2022-02-11 00:19:22.107067674 +0100 +++ gcc/c-family/c-omp.cc 2022-02-15 16:06:24.173311609 +0100 @@ -353,8 +353,13 @@ c_finish_omp_atomic (location_t loc, enu } bool save = in_late_binary_op; in_late_binary_op = true; - x = build_modify_expr (loc, blhs ? blhs : lhs, NULL_TREE, opcode, - loc, rhs, NULL_TREE); + if ((opcode == MIN_EXPR || opcode == MAX_EXPR) + && build_binary_op (loc, LT_EXPR, blhs ? blhs : lhs, rhs, + true) == error_mark_node) + x = error_mark_node; + else + x = build_modify_expr (loc, blhs ? blhs : lhs, NULL_TREE, opcode, + loc, rhs, NULL_TREE); in_late_binary_op = save; if (x == error_mark_node) return error_mark_node; --- gcc/testsuite/c-c++-common/gomp/atomic-31.c.jj 2022-02-15 17:02:17.938486108 +0100 +++ gcc/testsuite/c-c++-common/gomp/atomic-31.c 2022-02-15 17:04:23.811729201 +0100 @@ -0,0 +1,11 @@ +/* c/104531 */ +/* { dg-do compile } */ + +int x; + +void +foo (_Complex int y) +{ + #pragma omp atomic compare /* { dg-error "invalid operands" } */ + x = x > y ? y : x; +} Jakub