6733 – Regression(2.054) ICE(cod2.c) pure nothrow func with side-effect parameters

I have no idea what the heck is going on, but this crashes on DMD32 D Compiler v2.055:
struct Zero { }
void test(T1, T2)(T1 a, T2 b) { }
void main() { test(Zero(), Zero()); }

Oh, and the error is:
Internal error: ..\ztc\cod2.c 4624

This worked in 2.053 and earlier.

I just bisected it down to:
commit 4c9661fa9fbd427909a334133dfc7f3869e47c31
Author: Walter Bright <walter@walterbright.com>
Date: Thu Jun 23 00:50:46 2011 -0700
 nothrow inference
Reverting it from tip of master yields a successful build of the code above.

(In reply to comment #3)
> I just bisected it down to:
> 
> commit 4c9661fa9fbd427909a334133dfc7f3869e47c31
> Author: Walter Bright <walter@walterbright.com>
> Date: Thu Jun 23 00:50:46 2011 -0700
> 
> nothrow inference
> 
> Reverting it from tip of master yields a successful build of the code above.
Thanks. The backend failure is occuring inside a comma expression (x, y). Fails because x isn't an expression, it's just a parameter (presumably the struct literal).

Actually this is more a pseudo-regresssion, it's just an expansion of an existing bug into a few more cases. Walter's commit has nothing to do with the root cause. The test case below fails in exactly the same way on 2.025 (but passed on 2.023).
struct Zero { int x; }
void test(T)(T a, T b) pure nothrow { } 
void main() { test(Zero(7), Zero(4)); }
The bug is triggered when test() gets completely optimized away because it's pure nothrow.

Reduced test case:
void bug6733(int a, int b) pure nothrow { } 
void main() {
 int z;
 bug6733(z++, z++);
}
This is definitely a backend bug. What happens is, that since it's pure nothrow and the result is unused, it's a no-side-effect call (OPcallns).
In the first optimisation step (optelem in cgelem.c), the call gets discarded, and simply replaced with the parameter list (wrapped in an OPparam). If the parameters had no side-effects, the whole thing would be discarded.
If there's only one with a side-effect, it's the only thing that's left. But if there are TWO with side-effects, the OPparam remains.
The rest of the backend can't cope with a naked OPparam. Boom.
Solution would be to replace the OPparam with comma expressions.

https://github.com/D-Programming-Language/dmd/commit/6536012f7e23d5dc5928567224d47cc7d9997ceb
https://github.com/D-Programming-Language/dmd/commit/ab50833d7ca7fde0380525ed11e66908c440667a 

*** Issue 6890 has been marked as a duplicate of this issue. ***