https://issues.dlang.org/show_bug.cgi?id=16189

Kirill Kryukov <kkryukov@gmail.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |kkryukov@gmail.com

--

https://issues.dlang.org/show_bug.cgi?id=16189

ag0aep6g@gmail.com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |wrong-code
                 CC|                            |ag0aep6g@gmail.com
                 OS|Windows                     |All

--- Comment #1 from ag0aep6g@gmail.com ---
Slightly more reduced:

----
void main()
{
    ubyte[9][1] data;
    size_t a = 0;
    loop:
    data[0] = data[a];
    a--;
    bool b = false;
    if (b) goto loop;
    assert(a == -1); // Fails with -O
}
----

Also happens on Linux.

--

https://issues.dlang.org/show_bug.cgi?id=16189

--- Comment #2 from Kirill Kryukov <kkryukov@gmail.com> ---
A possible workaround: change "a--;" into "{ auto a2 = a - 1; a = a2; }".

(This is NOT to suggest that the bug does not need fixing, as it's annoying as hell that even simplest C-like code does not work correctly.)

As for previous reduction - it hurts my eyes to see size_t (unsigned type) compared for equality with -1, so I suggest to at least use ptrdiff_t.

--

https://issues.dlang.org/show_bug.cgi?id=16189

Ketmar Dark <ketmar@ketmar.no-ip.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |ketmar@ketmar.no-ip.org

--- Comment #3 from Ketmar Dark <ketmar@ketmar.no-ip.org> ---
funny thing: adding `asm { nop; }` anywhere in `main` seems to turn the optimizer off, and the code works again. a hackish way to control optimizer on per-function basis. ;-)

--

https://issues.dlang.org/show_bug.cgi?id=16189

--- Comment #4 from Kirill Kryukov <kkryukov@gmail.com> ---
I thought this might be a regression, so I tested dmd versions down to 2.050. But no, the bug is there in each and every one of them.

--

https://issues.dlang.org/show_bug.cgi?id=16189

Seb <greensunny12@gmail.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |greensunny12@gmail.com

--- Comment #5 from Seb <greensunny12@gmail.com> ---
See also the discussion in the forum: https://forum.dlang.org/post/jguphfhffgtdtfittkyw@forum.dlang.org

My summary:

The optimizer seems to be confused and wrongly precomputes a.
Note that if you remove e.g. "if (b) goto loop;", a will be correctly
statically set in the printf too:

---
                mov     RSI,0FFFFFFFFh                ; look ma - a is now -1
                lea     RDI,FLAT:.rodata[00h][RIP]
                xor     EAX,EAX                       ; set eax to 0
                call      printf@PLT32
---

For comparison the assembly of -O.
Here with printf and a different value to be compared with a because it's
easier to read:

---
main:
                push    RBP
                mov     RBP,RSP
                sub     RSP,010h
                lea     RAX,-010h[RBP]
                xor     ECX,ECX
                mov     [RAX],RCX
                mov     8[RAX],CL
                lea     RSI,-010h[RBP]
                lea     RDI,-010h[RBP]
                movsd
                movsb
                mov     RSI,01C71C71C71C71C70h          ; 2 function argument
(value of a)
                lea     RDI,FLAT:.rodata[00h][RIP]      ; "%d" (1st function
argument)
                xor     EAX,EAX                         ; set eax to 0
                call      printf@PLT32                  ; printf("%d", a)
                mov     EDX,0Ch
                lea     RSI,_TMP0@PC32[RIP]             ; load function
arguments (in reverse order)
                lea     RDI,_TMP0@PC32[RIP]
                call      __assert@PLT32                ; values load, let's
call assert
                add     [RAX],AL
.text.main      ends
        end
----

--

https://issues.dlang.org/show_bug.cgi?id=16189

bitter.taste@gmx.com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |bitter.taste@gmx.com

--- Comment #6 from bitter.taste@gmx.com ---
The LICM pass is too eager and given this small snippet
```
void main()
{
    ubyte[9][1] data;
    size_t a = 0;
    loop:
    data[0] = data[a];
    a--;
    bool b = false;
    if (b) goto loop;
    assert(a == -1); // Fails with -O
}
```
`a' is considered to be as an "index" variable into `data' with starting value
0 and that's decremented until `a < 1', this condition is most likely derived
by the bound-checking condition.
As a result we get

```
Adding (a(1) =  ((_TMP1(3) -  #data(0)) /  9LL )) to exit block B4
```

that doesn't play well with the constant folding pass

```
const prop (_TMP1(3) replaced by #data(0).xfffffffffffffff7)
```

as shown by

```
const prop (a(1) replaced by 2049638230412172400LL )
```

--

https://issues.dlang.org/show_bug.cgi?id=16189

Walter Bright <bugzilla@digitalmars.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |industry
                 CC|                            |bugzilla@digitalmars.com

--

https://issues.dlang.org/show_bug.cgi?id=16189

--- Comment #7 from Walter Bright <bugzilla@digitalmars.com> ---
https://github.com/dlang/dmd/pull/8074

--

https://issues.dlang.org/show_bug.cgi?id=16189

Walter Bright <bugzilla@digitalmars.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Hardware|x86_64                      |All

--