struct Node {
size_t _left;
Node* left() {
return cast(Node*) (_left & 1);
}
bool useLeft() {
return left is null;
}
}
Result:
Internal error: ..\ztc\cgelem.c 3387
This one is really hard to reproduce. If you change this test program in even very small ways, it's no longer reproduced.
Comment #1 by clugdbug — 2009-11-12T07:17:10Z
I can't reproduce this (on Windows) with either D2 or D1. I tried several compiler versions, it compiled correctly in all cases.
Comment #2 by dsimcha — 2009-11-13T08:02:26Z
Didn't realize it at the time, but I've looked into this further. It only happens when you compile w/ -O -inline.
Comment #3 by clugdbug — 2009-11-13T12:42:02Z
Reduced test case. Compile with -O (-inline not required). Not a regression; fails on DMD0.165.
bool foo() {
int x;
return cast(void*) (x & 1) == null;
}
Internal error: ztc\cgelem.c 3387
Really peculiar thing is that replacing & with | or ^ makes the bug disappear.
Comment #4 by clugdbug — 2009-11-13T13:17:00Z
This assert is actually completely harmless. It's part of a check for if the equality comparison can be reduced to a single byte comparison. This can happen if it's a pointer with all the high bits clear, which includes null. Here are 3 cases which trigger the same bug.
bool foo()
{
int x;
// return cast(void*) (x & 1) == null; // Case 1
// return cast(void*) (x & 1) == cast(void *)(2); // Case 2
return cast(bool function()) (x & 1) == null; // Case 3
}
================
PATCH:
cgelem.c, line 3387.
/* Try to convert to byte/word comparison for ((x & c)==d)
when mask c essentially casts x to a smaller type
*/
if (OPTIMIZER &&
e1->Eoper == OPand &&
e1->E2->Eoper == OPconst &&
(sz = tysize(e2->Ety)) > CHARSIZE)
{ int op;
- assert(tyintegral(e2->Ety));
+ assert(tyintegral(e2->Ety) || (e2->Ety == TYnptr));
#if TX86 /* ending up with byte ops in A regs */
if (!(el_tolong(e2) & ~CHARMASK) &&
!(el_tolong(e1->E2) & ~CHARMASK)
)
{
Comment #5 by leandro.lucarella — 2009-11-23T06:48:08Z