Has something changed in variadic function call ABI...?


  the trunk version fails in executing code that used to work a few weeks ago for code that constructs variadic function calls in x86_64. Here an example of our code (which is generated by a function call stub generator for our interpreter):

   // put arguments in dval, or lval or u[]
   long fptr = (long)&Printf; // <--- our variadic function
   __asm__ __volatile__("movlpd %0, %%xmm0" :: "m" (dval[0]) : "%xmm0");
   __asm__ __volatile__("movlpd %0, %%xmm1" :: "m" (dval[1]) : "%xmm1");
   __asm__ __volatile__("movlpd %0, %%xmm2" :: "m" (dval[2]) : "%xmm2");
   __asm__ __volatile__("movlpd %0, %%xmm3" :: "m" (dval[3]) : "%xmm3");
   __asm__ __volatile__("movlpd %0, %%xmm4" :: "m" (dval[4]) : "%xmm4");
   __asm__ __volatile__("movlpd %0, %%xmm5" :: "m" (dval[5]) : "%xmm5");
   __asm__ __volatile__("movlpd %0, %%xmm6" :: "m" (dval[6]) : "%xmm6");
   __asm__ __volatile__("movlpd %0, %%xmm7" :: "m" (dval[7]) : "%xmm7");
   __asm__ __volatile__("movq %0, %%rdi" :: "m" (lval[0]) : "%rdi");
   __asm__ __volatile__("movq %0, %%rsi" :: "m" (lval[1]) : "%rsi");
   __asm__ __volatile__("movq %0, %%rdx" :: "m" (lval[2]) : "%rdx");
   __asm__ __volatile__("movq %0, %%rcx" :: "m" (lval[3]) : "%rcx");
   __asm__ __volatile__("movq %0, %%r8" :: "m" (lval[4]) : "%r8");
   __asm__ __volatile__("movq %0, %%r9" :: "m" (lval[5]) : "%r9");
   __asm__ __volatile__("movq %0, %%r10" :: "m" (fptr) : "%r10");
   // (umax+2)*8 = 176
   __asm__ __volatile__("subq $176, %rsp");
   __asm__ __volatile__("movq %0, %%rax \n\t"
                        "movq %%rax, 0(%%rsp)" :: "m" (u[0].lval) : "%rax");
   __asm__ __volatile__("movq %0, %%rax \n\t"
                        "movq %%rax, 8(%%rsp)" :: "m" (u[1].lval) : "%rax");
   __asm__ __volatile__("movq %0, %%rax \n\t"
                        "movq %%rax, 16(%%rsp)" :: "m" (u[2].lval) : "%rax");
   __asm__ __volatile__("movq %0, %%rax \n\t"
                        "movq %%rax, 24(%%rsp)" :: "m" (u[3].lval) : "%rax");
   __asm__ __volatile__("movq %0, %%rax \n\t"
                        "movq %%rax, 32(%%rsp)" :: "m" (u[4].lval) : "%rax");
   __asm__ __volatile__("movq %0, %%rax \n\t"
                        "movq %%rax, 40(%%rsp)" :: "m" (u[5].lval) : "%rax");
   __asm__ __volatile__("movq %0, %%rax \n\t"
                        "movq %%rax, 48(%%rsp)" :: "m" (u[6].lval) : "%rax");
   __asm__ __volatile__("movq %0, %%rax \n\t"
                        "movq %%rax, 56(%%rsp)" :: "m" (u[7].lval) : "%rax");
   __asm__ __volatile__("movq %0, %%rax \n\t"
                        "movq %%rax, 64(%%rsp)" :: "m" (u[8].lval) : "%rax");
   __asm__ __volatile__("movq %0, %%rax \n\t"
                        "movq %%rax, 72(%%rsp)" :: "m" (u[9].lval) : "%rax");
   __asm__ __volatile__("movq %0, %%rax \n\t"
                        "movq %%rax, 80(%%rsp)" :: "m" (u[10].lval) : "%rax");
   __asm__ __volatile__("movq %0, %%rax \n\t"
                        "movq %%rax, 88(%%rsp)" :: "m" (u[11].lval) : "%rax");
   __asm__ __volatile__("movq %0, %%rax \n\t"
                        "movq %%rax, 96(%%rsp)" :: "m" (u[12].lval) : "%rax");
   __asm__ __volatile__("movq %0, %%rax \n\t"
                        "movq %%rax, 104(%%rsp)" :: "m" (u[13].lval) : "%rax");
   __asm__ __volatile__("movq %0, %%rax \n\t"
                        "movq %%rax, 112(%%rsp)" :: "m" (u[14].lval) : "%rax");
   __asm__ __volatile__("movq %0, %%rax \n\t"
                        "movq %%rax, 120(%%rsp)" :: "m" (u[15].lval) : "%rax");
   __asm__ __volatile__("movq %0, %%rax \n\t"
                        "movq %%rax, 128(%%rsp)" :: "m" (u[16].lval) : "%rax");
   __asm__ __volatile__("movq %0, %%rax \n\t"
                        "movq %%rax, 136(%%rsp)" :: "m" (u[17].lval) : "%rax");
   __asm__ __volatile__("movq %0, %%rax \n\t"
                        "movq %%rax, 144(%%rsp)" :: "m" (u[18].lval) : "%rax");
   __asm__ __volatile__("movq %0, %%rax \n\t"
                        "movq %%rax, 152(%%rsp)" :: "m" (u[19].lval) : "%rax");
   __asm__ __volatile__("movl $8, %eax"); // number of used xmm registers
   __asm__ __volatile__("call *%r10");
   __asm__ __volatile__("addq $176, %rsp");
   __asm__ __volatile__("movq %%rax, %0" : "=m" (u[0].lval) :: "memory"); // get return value
   __asm__ __volatile__("movq %%rdi, %0" : "=m" (u[1].lval) :: "memory"); // get return value (icc C++ object)
       G__letint(result7, 67, (long) u[0].lval);
    return(1 || funcname || hash || result7 || libp) ;

with the trunk version this call into Printf() fails. While it used to work and while it still works with gcc 4.1 - 4.5. Does anybody know if something in the area has changed on purpose or if this is a regression?

Cheers, Fons.

Why put everything in separate __asm__ statements? It might be less
fragile if you'd put everything in a single one with appropriate

For instance, even though your early __asm__ statements clobber the
parameter registers (and %r10), the compiler is under no obligation to
*preserve* their new values until your call statement. At the very
least, try to replace those (and setting %eax to 8) with proper
constraints on the __asm__ containing the "call".

I'd also say -fomit-frame-pointer might break your code since you're
manipulating %rsp manually so you might want to make sure to use
-fno-omit-frame-pointer so the "m" constraints hopefully won't use
%rsp-relative addresses. (assuming lval, dval and/or u[] are on the

If you want to know what exactly is going on, I'd suggest compiling
with -S (or using a disassembler) to see whether any instructions get
inserted between your __asm__ statements, and what kind of addresses
the memory constraints generate.

Alternatively, you could just replace all of that asm with something like
  struct retval { int64_t a; int64_t b; };
  retval ret = ((retval (*)(...))fptr)(dval[0], ..., dval[7], lval[0],
..., lval[5], u[0].lval, ... u[19].lval);
  u[0].lval = ret.a;
  u[1].lval = ret.b;
and let your compiler figure out the best way to do this instead of
trying to re-implement the ABI manually...

Hoi Frits,

    this might work using C++, I'll look into it, but I remembered I had problems with pointers to virtual functions, but again it was quite a while ago I wrote this stuff and since then it has been working, with all x86_64 compilers including clang++ till a few days ago, so something definitely changed in the clang abi.

Cheers, Fons.

One of my other points was that the problem wasn't necessarily the
ABI. Did you look at the asm clang emits for that code?
That might give some insight into what's happening.

You didn't mention virtual functions in your original message, by the
way. Are you saying "Printf" could be a member function? (and a
virtual one at that?)
That could definitely complicate things...

I am now checking what is changed in the emitted code to understand what is going on. This code is generated by our dictionary generator that generates stubs for all global, static, member and virtual member functions you may find in a C++ class.

However, your small example gave me some ideas how I might get rid of assembler for all the cases. Thanks.

Cheers, Fons.