www.digitalmars.com         C & C++   DMDScript  

digitalmars.D - Frustrated with dmd codegen bug

reply "H. S. Teoh" <hsteoh quickfur.ath.cx> writes:
Yesterday afternoon I ran into a performance issue in one of my D
projects, and thought, "well, it's simple, just compile with -profile,
identify the hotspot, optimize".  Unfortunately, doing that triggered a
latent codegen bug in -O that randomly causes runtime segfaults,
basically halting all progress.  I spent all morning today to reduce the
code in order to identify the codegen bug:

	https://issues.dlang.org/show_bug.cgi?id=18794

The only workaround I know of currently is to compile without -O.  But
that also means it's meaningless to compile with -profile, since it
would give misleading information (e.g., identify hotspots that aren't
really hotspots once the compiler actually optimizes the code).

The next best thing is to use LDC for better (and more reliable!)
codegen.  However, since this project was developed using dmd git master
(I know, I know), some non-trivial changes are needed to make it
compilable by LDC.

At this point, I'm *seriously* tempted to stop following dmd development
completely and just stick with LDC releases for my D projects. :-(


T

-- 
"The number you have dialed is imaginary. Please rotate your phone 90 degrees
and try again."
Apr 24 2018
next sibling parent jmh530 <john.michael.hall gmail.com> writes:
On Tuesday, 24 April 2018 at 18:53:02 UTC, H. S. Teoh wrote:

 [snip]


That's definitely weird. Problem seems to go away with a static 
array. Seems somehow related to impl[0]. Re-writing that as 
*impl.ptr and breaking apart some of the logic might help narrow 
down the issue.


bool method(int v)
{
     int wordIdx = v >> 6;
     int bitIdx = v & 0b00111111;

     func();

     if (impl.length < wordIdx)
     {
         import std.stdio : writeln;
         auto temp1 = (1UL << bitIdx);
         writeln(1UL << bitIdx); //testing with v=200, prints 256
         writeln(*impl.ptr & 256); //prints 0
         auto temp2 = (*impl.ptr & temp1); //program killed
         //writeln(*impl.ptr & 256); //if uncommented, program not 
killed
         return temp2 != 0;
     }
     else
     {
         return false;
     }
}
Apr 24 2018
prev sibling next sibling parent reply Basile B. <b2temp gmx.com> writes:
On Tuesday, 24 April 2018 at 18:53:02 UTC, H. S. Teoh wrote:

 Yesterday afternoon I ran into a performance issue in one of my 
 D projects, and thought, "well, it's simple, just compile with 
 -profile, identify the hotspot, optimize".  Unfortunately, 
 doing that triggered a latent codegen bug in -O that randomly 
 causes runtime segfaults, basically halting all progress.  I 
 spent all morning today to reduce the code in order to identify 
 the codegen bug:

 	https://issues.dlang.org/show_bug.cgi?id=18794


In the report you forgot to mention that the bug is only visible 
with -O -profile.
With just -O the provided test case works fine.
Apr 24 2018
next sibling parent reply "H. S. Teoh" <hsteoh quickfur.ath.cx> writes:
On Tue, Apr 24, 2018 at 08:18:55PM +0000, Basile B. via Digitalmars-d wrote:

 On Tuesday, 24 April 2018 at 18:53:02 UTC, H. S. Teoh wrote:

 Yesterday afternoon I ran into a performance issue in one of my D
 projects, and thought, "well, it's simple, just compile with
 -profile, identify the hotspot, optimize".  Unfortunately, doing
 that triggered a latent codegen bug in -O that randomly causes
 runtime segfaults, basically halting all progress.  I spent all
 morning today to reduce the code in order to identify the codegen
 bug:
 
 	https://issues.dlang.org/show_bug.cgi?id=18794


 
 In the report you forgot to mention that the bug is only visible with
 -O -profile.
 With just -O the provided test case works fine.


Really?! That's weird...  Originally the problem appeared with -O
-profile, but after reduction, I could reproduce the problem even
without -profile. (Just tested it again: the segfault happens without
-profile.)  Possibly it's affected by the OS/architecture?  Or perhaps
it's just very sensitive to the exact combination of flags +
environmental factors.


T

-- 
I'm still trying to find a pun for "punishment"...
Apr 24 2018
parent Basile B. <b2.temp gmx.com> writes:
On Tuesday, 24 April 2018 at 20:44:12 UTC, H. S. Teoh wrote:

 On Tue, Apr 24, 2018 at 08:18:55PM +0000, Basile B. via 
 Digitalmars-d wrote:

 On Tuesday, 24 April 2018 at 18:53:02 UTC, H. S. Teoh wrote:

 Yesterday afternoon I ran into a performance issue in one of 
 my D projects, and thought, "well, it's simple, just compile 
 with -profile, identify the hotspot, optimize".  
 Unfortunately, doing that triggered a latent codegen bug in 
 -O that randomly causes runtime segfaults, basically halting 
 all progress.  I spent all morning today to reduce the code 
 in order to identify the codegen bug:
 
 	https://issues.dlang.org/show_bug.cgi?id=18794


 
 In the report you forgot to mention that the bug is only 
 visible with
 -O -profile.
 With just -O the provided test case works fine.


 Really?! That's weird...  Originally the problem appeared with 
 -O -profile, but after reduction, I could reproduce the problem 
 even without -profile. (Just tested it again: the segfault 
 happens without -profile.)  Possibly it's affected by the 
 OS/architecture?  Or perhaps it's just very sensitive to the 
 exact combination of flags + environmental factors.


 T


I'm sorry to bring confusion but i've tested on linux x86_64. 
With dmd 2.079.1 and ~master (just updated it right now). Here 
for some reason -profile is always required to get the error :/
Apr 24 2018
prev sibling parent jmh530 <john.michael.hall gmail.com> writes:
On Tuesday, 24 April 2018 at 20:18:55 UTC, Basile B. wrote:

 [snip]

 In the report you forgot to mention that the bug is only 
 visible with -O -profile.
 With just -O the provided test case works fine.


I ran the test case on run.dlang.org with -O and it happens.
Apr 24 2018
prev sibling parent ag0aep6g <anonymous example.com> writes:
On 04/24/2018 08:53 PM, H. S. Teoh wrote:

 Yesterday afternoon I ran into a performance issue in one of my D
 projects, and thought, "well, it's simple, just compile with -profile,
 identify the hotspot, optimize".  Unfortunately, doing that triggered a
 latent codegen bug in -O that randomly causes runtime segfaults,
 basically halting all progress.  I spent all morning today to reduce the
 code in order to identify the codegen bug:
 
 	https://issues.dlang.org/show_bug.cgi?id=18794


Another one for my collection of bugs related to the bt instruction [1]. 
I have an open PR to fix one of those [2], but maybe the whole 
optimization should just be disabled until the issues are fixed.

[...]

 At this point, I'm *seriously* tempted to stop following dmd development
 completely and just stick with LDC releases for my D projects. :-(


Very understandable. DMD has too damn many wrong-code bugs.


[1] https://issues.dlang.org/show_bug.cgi?id=18750
[2] https://github.com/dlang/dmd/pull/8142
Apr 24 2018