LDC1.32 generates wrong binary when using extern (C)

[hsadeghi]

Hi all,

We have a rather big project including c/c++ and d code, when I updated our d compiler from ldc_1.30 to ldc_1.32 one of our tests failed,
it is rather a long code but I bring the part for which wrong binary (target arch: RISCV64) has been generated:

void* wrap(alias fun)()
{
  import std.traits;
  import std.meta;

  alias Args = Parameters!fun;

  extern (C)
  static void bariKernelFun(TaskParams!Args* taskParams)
  {
    version (D_BetterC)
    {
      version (sst)
      {
        // Reconstruct arguments
        import std.traits;

        // Arguments without const (such that we can edit them)
        alias UnconstArgs = staticMap!(Unconst, Args);
        UnconstArgs argsRecon;

        static foreach (i, Target; UnconstArgs)
        {
          bariPrintF(LogLevel.DEBUG, "---- Reconstructing ulong %d to a %.*s\n\0",
            taskParams[0][i],
            Target.stringof.length, Target.stringof.ptr);
          static if (hasMember!(Target, "bariObject"))
          {
            // Target is something like Vector!ulong
            argsRecon[i] = Target("reconstructed", taskParams[0][i]);
          }
          else
          {
            // Target is something like double
            argsRecon[i] = (cast(Target*)(&(taskParams[0][i])))[0];
          }
        }

        // call function and put the const-s back in
        fun(cast(Args) argsRecon);

        static foreach (i, Target; Args)
        {
          static if (hasMember!(Target, "bariObject") && !is(Target == const))
          {
            pragma(msg, mangledName!fun, " flushes a ", Target.stringof, " on task completion (arg ", i, ")");
            bariPrintF(LogLevel.DEBUG, "---- Flushing output object %.*s with OID %lu\n\0",
              Target.stringof.length, Target.stringof.ptr, taskParams[0][i]);
            argsRecon[i].bariObject.flush();
          }
        }

      }
      else
      {
        fun(taskParams.args);
      }
    }

the issue appears to be connected to fun(args), I compared (RISCV)assembly code and the only difference between correct and wrong binary is something like "ld a0, 24(a0)" which in the wrong binary, becomes "ld a0, 0(a0)" (basically and index becomes zero !)
as far as I tested later versions (1.34 and 1.33) are ok and it just appears in v1.32

I was wondering if anyone has encountered the same issue and knows why is it happening in just one specific version ?

Many thanks in advance for your remarks,
Hadi

[kinke]

Hi! This sounds similar to #4362, a v1.31 regression that was fixed in v1.32.1.

[hsadeghi]

I just tested v1.31 and got the same issue, so I can verify that that the regression you mentioned is the cause.
Many thanks @kinke