www.digitalmars.com         C & C++   DMDScript  

digitalmars.D.ldc - Missed optimization?

reply safety0ff <safety0ff.dev gmail.com> writes:
Hello,
I was experimenting with pure functions, exceptions and inlining 
and I noticed that LDC seems to generate inferior code to GDC in 
the cases I was considering.

For example:

http://ldc.acomirei.ru/#compilers:!((compiler:ldc,options:'-release+-O3+-boundscheck%3Doff',sourcez:PTBuFMCcGcEsHsB2ACAvMgLgC0vA7gDbjTQDcAUAUgObICGAJgwIwAUVitdAlMgA4BXSOHIBvcsmSwAZslZ006DtQB0AWzoAPbhOTiIMBIlbZchYtB2TJwjEJTL1WiuALQR1zDnzJE4PMgAopoAxuB8GEasAETR3BSSAL66tvb0yADUyMwUyZQ0yNLw8KywamoCGHQARkTIygDaALr0OuI24HaQKHQNAAxNKowsmfT9g8M55HnKyNV0kKXllTV1jS08Yrqz1QA0IQmYkACeerqSGCdnnpLVaGMDQ0xTnonIIXQYIVh6c/cALAAmUjIZKeEL3XqPSaHMHvT7fX4Q9CAgDMILhqW6cwyB2m%2BU4czoAC8lhUqrVwPUaM1WltJDt9odLqd2tY7ugoRNnodJMiHtyWLDdB8vj9RH90ECQfy0RiUp00tVcbkCbQAI4CTRklaU6mcWmbNmMvEXK5s26Q8ZPIXnI6su18q3Qnl2t6ixESjnIaWg5AgZACRDCOjfVYeUEihHiyU%2B4Hve5yyMdLooZV45JAA%3D%3D)),filterAsm:(commentOnly:!t,directives:!t,labels:!t),version:3

Compared to:

https://godbolt.org/g/WI12Bk

Even in the simplest case (function foo) contains duplicated 
memory load and comparison from the add1 function.

Changing to the throwless version LDC generates similar code for 
all cases.

Motivation:

I was thinking about autodecoding and wondering whether it was 
possible for the following example:
auto decode(immutable char[] s) pure
{
struct Result { dchar codepoint; uint advance; }
// throw Exception if s is invalid / erroneous
// return Result (decoded codepoint, number of chars to advance);
}
dchar front(string s) { return s.decode.codepoint; }
void popFront(ref string s) { s = s[s.decode.advance .. $];

ulong loop(string s)
{
   ulong checksum;
   while (s.length)
   {
     checksum += s.front;
     s.popFront;
   }
   return checksum;
}

To optimize the loop function (using inlining and purity):

ulong loop(string s)
{
   ulong checksum;
   while (s.length)
   {
     auto tmp = s.decode
     checksum += tmp.codepoint;
     s = [tmp.advance .. $];
   }
   return checksum;
}
Oct 21 2016
next sibling parent reply David Nadlinger via digitalmars-d-ldc <digitalmars-d-ldc puremagic.com> writes:
On 21 Oct 2016, at 22:14, safety0ff via digitalmars-d-ldc wrote:

 I was experimenting with pure functions, exceptions and inlining and I 
 noticed that LDC seems to generate inferior code to GDC in the cases I 
 was considering.
 […]
 https://godbolt.org/g/WI12Bk


That's indeed a bit surprising. The generated IR looks like this:

---
define i64  _D4test3fooFyAlZl({ i64, i64* } %a_arg) local_unnamed_addr 

   %.getAddressOf_dump.i2 = alloca { i64, i8* }, align 8 ; [#uses = 4, 
size/byte = 16]
   %.getAddressOf_dump.i = alloca { i64, i8* }, align 8 ; [#uses = 4, 
size/byte = 16]
   %a_arg.fca.1.extract = extractvalue { i64, i64* } %a_arg, 1 ; [#uses 
= 2]
   %1 = load i64, i64* %a_arg.fca.1.extract, align 8 ; [#uses = 2]
   %2 = bitcast { i64, i8* }* %.getAddressOf_dump.i to i8* ; [#uses = 2]
   call void  llvm.lifetime.start(i64 16, i8* %2)
   %3 = icmp eq i64 %1, 9223372036854775807        ; [#uses = 1]
   br i1 %3, label %if.i, label %_D4test4add1FNalZl.exit

[…]

_D4test4add1FNalZl.exit:                          ; preds = %0
   call void  llvm.lifetime.end(i64 16, i8* %2)
   %9 = load i64, i64* %a_arg.fca.1.extract, align 8 ; [#uses = 2]
   %10 = bitcast { i64, i8* }* %.getAddressOf_dump.i2 to i8* ; [#uses = 
2]
   call void  llvm.lifetime.start(i64 16, i8* %10)
   %11 = icmp eq i64 %9, 9223372036854775807       ; [#uses = 1]
   br i1 %11, label %if.i8, label %_D4test4add1FNalZl.exit9
---

There doesn't seem to be any stores which would potentially make %9 not 
foldable with %1. Tracking this down might be an interesting trip down 
the LLVM pass pipeline.

  — David
Oct 21 2016
parent reply David Nadlinger <code klickverbot.at> writes:
On Friday, 21 October 2016 at 22:19:26 UTC, David Nadlinger wrote:

 Tracking this down might be an interesting trip down the LLVM 
 pass pipeline.


I forgot to mention: Could you please create an entry on 
https://github.com/ldc-developers/ldc/issues for this?

  — David
Oct 21 2016
parent safety0ff <safety0ff.dev gmail.com> writes:
On Saturday, 22 October 2016 at 01:07:02 UTC, David Nadlinger 
wrote:

 I forgot to mention: Could you please create an entry on 
 https://github.com/ldc-developers/ldc/issues for this?

  — David


Done: https://github.com/ldc-developers/ldc/issues/1851
Oct 21 2016
prev sibling parent Daniel9 <andreikovalsa mail.ru> writes:
On Friday, 21 October 2016 at 21:14:56 UTC, safety0ff wrote:

 Hello,
 I was experimenting with pure functions, exceptions and 
 inlining and I noticed that LDC seems to generate inferior code 
 to GDC in the cases I was considering.

 [...]


what about?
Nov 01 2016