Hi! When looking at PR114175 (although that bug seems to be now a riscv backend bug), I've noticed that for the TYPE_NO_NAMED_ARGS_STDARG_P functions which return value through hidden reference, like #include <stdarg.h>
struct S { char a[64]; }; int n; struct S foo (...) { struct S s = {}; va_list ap; va_start (ap); for (int i = 0; i < n; ++i) if ((i & 1)) s.a[0] += va_arg (ap, double); else s.a[0] += va_arg (ap, int); va_end (ap); return s; } we were incorrectly calling assign_parms_setup_varargs twice, once at the start of the function and once in if (cfun->stdarg && !DECL_CHAIN (parm)) assign_parms_setup_varargs (&all, &data, false); where parm is the last and only "named" parameter. The first call, guarded with TYPE_NO_NAMED_ARGS_STDARG_P, was added in r13-3549 and is needed for int bar (...) etc. functions using va_start/va_arg/va_end, otherwise the FOR_EACH_VEC_ELT (fnargs, i, parm) in which the other call is will not iterate at all. But we shouldn't be doing that if we have the hidden return pointer. With the following patch on the above testcase with -O0 -std=c23 the assembly difference is: pushq %rbp .cfi_def_cfa_offset 16 .cfi_offset 6, -16 movq %rsp, %rbp .cfi_def_cfa_register 6 pushq %rbx subq $192, %rsp .cfi_offset 3, -24 - movq %rdi, -192(%rbp) - movq %rsi, -184(%rbp) - movq %rdx, -176(%rbp) - movq %rcx, -168(%rbp) - movq %r8, -160(%rbp) - movq %r9, -152(%rbp) - testb %al, %al - je .L2 - movaps %xmm0, -144(%rbp) - movaps %xmm1, -128(%rbp) - movaps %xmm2, -112(%rbp) - movaps %xmm3, -96(%rbp) - movaps %xmm4, -80(%rbp) - movaps %xmm5, -64(%rbp) - movaps %xmm6, -48(%rbp) - movaps %xmm7, -32(%rbp) -.L2: movq %rdi, -312(%rbp) movq %rdi, -192(%rbp) movq %rsi, -184(%rbp) movq %rdx, -176(%rbp) movq %rcx, -168(%rbp) movq %r8, -160(%rbp) movq %r9, -152(%rbp) testb %al, %al - je .L13 + je .L12 movaps %xmm0, -144(%rbp) movaps %xmm1, -128(%rbp) movaps %xmm2, -112(%rbp) movaps %xmm3, -96(%rbp) movaps %xmm4, -80(%rbp) movaps %xmm5, -64(%rbp) movaps %xmm6, -48(%rbp) movaps %xmm7, -32(%rbp) -.L13: +.L12: plus some renumbering of labels later on which clearly shows that because of this bug, we were saving all the registers twice rather then once. With -O2 -std=c23 some of it is DCEd, but we still get subq $160, %rsp .cfi_def_cfa_offset 168 - testb %al, %al - je .L2 - movaps %xmm0, 24(%rsp) - movaps %xmm1, 40(%rsp) - movaps %xmm2, 56(%rsp) - movaps %xmm3, 72(%rsp) - movaps %xmm4, 88(%rsp) - movaps %xmm5, 104(%rsp) - movaps %xmm6, 120(%rsp) - movaps %xmm7, 136(%rsp) -.L2: movq %rdi, -24(%rsp) movq %rsi, -16(%rsp) movq %rdx, -8(%rsp) movq %rcx, (%rsp) movq %r8, 8(%rsp) movq %r9, 16(%rsp) testb %al, %al - je .L13 + je .L12 movaps %xmm0, 24(%rsp) movaps %xmm1, 40(%rsp) movaps %xmm2, 56(%rsp) movaps %xmm3, 72(%rsp) movaps %xmm4, 88(%rsp) movaps %xmm5, 104(%rsp) movaps %xmm6, 120(%rsp) movaps %xmm7, 136(%rsp) -.L13: +.L12: difference, i.e. this time not all, but the floating point args were conditionally all saved twice. Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? 2024-03-01 Jakub Jelinek <ja...@redhat.com> * function.cc (assign_parms): Only call assign_parms_setup_varargs early for TYPE_NO_NAMED_ARGS_STDARG_P functions if fnargs is empty. --- gcc/function.cc.jj 2024-01-12 13:47:20.834428745 +0100 +++ gcc/function.cc 2024-02-29 21:14:35.275889093 +0100 @@ -3650,7 +3650,8 @@ assign_parms (tree fndecl) assign_parms_initialize_all (&all); fnargs = assign_parms_augmented_arg_list (&all); - if (TYPE_NO_NAMED_ARGS_STDARG_P (TREE_TYPE (fndecl))) + if (TYPE_NO_NAMED_ARGS_STDARG_P (TREE_TYPE (fndecl)) + && fnargs.is_empty ()) { struct assign_parm_data_one data = {}; assign_parms_setup_varargs (&all, &data, false); Jakub