Bug 24181 – reading double parameter from RCX rather than XMM1

Created attachment 1893 single file, 99 lines tinylisp.c I'm playing with https://github.com/Robert-van-Engelen/tinylisp/blob/main/src/tinylisp.c single file, 99 lines with GCC: ``` $ make -B tinylisp cc -c -o tinylisp.o tinylisp.c cc tinylisp.o -o tinylisp $ ./tinylisp tinylisp 930>(+ 2 3) 5 ``` wich dmd (importC): ``` $ dmd tinylisp.c # build fine $ ./tinylisp tinylisp 930>(+ 2 3) ERR 930>^C ``` something goes wrong at runtime for such simple example.

BTW, $ dmd --version DMD64 D Compiler v2.105.0 Copyright (C) 1999-2023 by The D Language Foundation, All Rights Reserved written by Walter Bright And LDC works fine: $ ldc2 -version LDC - the LLVM D compiler (1.34.0): based on DMD v2.104.2 and LLVM 16.0.6 built with LDC - the LLVM D compiler (1.34.0) Default target: x86_64-unknown-linux-gnu Host CPU: znver1 http://dlang.org - http://wiki.dlang.org/LDC $ ./tinylisp tinylisp 930>(+ 2 3) 5

Unfortunately, it's not a very simple example. It's very dense, compressed code with no comments. So far, I haven't found where the dmd version diverges from the gcc version. I'll look at it some more tomorrow.

Hmm, it does work with -m32, implying a code gen problem.

I tried reducing it with dustmite, but it's still pretty large: ``` int printf(); int sscanf(); char *strcpy(); int strcmp(); long strlen(); unsigned hp, sp = 1024, ATOM, PRIM, CONS = 0x7ffa, NIL; double cell[1024], nil, tru, err, env; double box(unsigned t, unsigned i) { double x; *(long *)&x = (long)t << 48 | i; return x; } unsigned ord(double x) { return *(unsigned *)&x; } unsigned equ(double x, double y) { return *(unsigned *)&x == *(long *)&y; } double atom(char *s) { unsigned i = 0; while (i < hp && strcmp((char *)cell + i, s)) i += 1; hp += strlen(strcpy((char *)cell + i, s)) + 1; return box(ATOM, i); } double cons(double x, double y) { cell[--sp] = x; cell[--sp] = y; return box(CONS, sp); } double car(double p) { return p ? cell[ord(p) + 1] : err; } double cdr(double p) { return p ? cell[ord(p)] : err; } double pair(double v, double x, double e) { return cons(cons(v, x), e); } double assoc(double v, double e) { while (e && !equ(v, car(car(e)))) e = cdr(e); return e ? cdr(car(e)) : err; } unsigned not(double x) { return x == NIL; } double f_add(double t) { double n; while (!not(t = cdr(t))) n += car(t); return n; } struct { char *s; double (*f)(); } prim[] = { "-", 0, "+", f_add, }; double apply(double f, double t) { return f ? prim[ord(f)].f(t) : f; } double eval(double x, double e) { return *(long *)&x >> 48 == ATOM ? assoc(x, e) : x ? apply(eval(car(x), e), x) : x; } char buf[40], see; char* input = "(+ 4 4)"; int inputI = 0; void look() { see = input[inputI++]; } unsigned seeing(char c) { return c ? see <= c : see; } char get() { char c = see; look(); return c; } char scan() { int i = 0; while (seeing(' ')) look(); return buf[i++] = get(); } double parse(); double list() { double x; return scan() == ')' ? nil : (x = parse(), cons(x, list())); } double atomic() { double n; int i; return sscanf(buf, "%lg%n", &n, &i) ? n : atom(buf); } double parse() { return *buf == '(' ? list() : atomic(); } void print(double x) { // printf("%016lX\n", *(unsigned long*)&x); if (x == ATOM) printf("ERR"); else printf("%.10lg", x); } int main() { int i; for (i = 0; prim[i].s; ++i) env = pair(atom(prim[i].s), box(PRIM, i), env); double result = eval((scan(), parse()), env); print(result); } ``` This program prints 8 with gcc, but ERR with dmd on linux 64-bit.

Ah, thanks Dennis! This is very helpful.

In the same github repo, there is a equivalent commented .c file (dmd generated binary has the same runtime error): https://github.com/Robert-van-Engelen/tinylisp https://github.com/Robert-van-Engelen/tinylisp/blob/main/src/tinylisp-commented.c and the design doc: https://github.com/Robert-van-Engelen/tinylisp/blob/main/tinylisp.pdf gcc works fine: $ make tinylisp-commented cc tinylisp-commented.c -o tinylisp-commented (env) zhou@joort:~/project/unzip/tinylisp/src$ ./tinylisp-commented tinylisp 930>(+ 3 5) 8 $ dmd tinylisp-commented.c tinylisp-commented.c: Error: module `tinylisp-commented` has non-identifier characters in filename, use module declaration instead $ ln -s tinylisp-commented.c tinylisp_commented.c $ dmd tinylisp_commented.c $ ./tinylisp_commented tinylisp 930>(+ 3 5) ERR # -m32 works confirmed: $ dmd -m32 tinylisp_commented.c $ ./tinylisp_commented tinylisp 930>(+ 3 5) 8

BTW, dmd generate binary also have runtime error with this single file lisp: https://github.com/Robert-van-Engelen/lisp/blob/main/src/lisp.c src$ dmd lisp.c src$ ./lisp lisp unbound define ERR 3 unbound symbol 8020+2000> # dmd -m32 works fine: src$ dmd -m32 lisp.c $ ./lisp lisp null? number? ... defmacro defun 6322+1929> GCC, and LDC2 works fine.

For this: unsigned equ(double x, double y) { return *(long long *)&x == *(long long *)&y; } The generated code is: equ: sub RSP,018h movq [RSP],XMM0 mov RAX,[RSP] cmp RAX,RCX <== ARRRGHHH!! setz AL movzx RAX,AL add RSP,018h ret The `y` parameter is passed in XMM1, not RCX. The cast to long long has fooled the code generator. I suspected something like this was happening, but it took a while to find it.

@WalterBright created dlang/dmd pull request #15707 "fix Issue 24181 - reading double parameter from RCX rather than XMM1" fixing this issue: - fix Issue 24181 - reading double parameter from RCX rather than XMM1 https://github.com/dlang/dmd/pull/15707

dlang/dmd pull request #15707 "fix Issue 24181 - reading double parameter from RCX rather than XMM1" was merged into master: - 9aacadb8604d7a7ee9c5226fe463ac4fe3115245 by Walter Bright: fix Issue 24181 - reading double parameter from RCX rather than XMM1 https://github.com/dlang/dmd/pull/15707