[llvmgcc] Passing large structures by value to functions results in large code.

[llvmbot]

Bugzilla Link	745
Resolution	FIXED
Resolved on	Mar 06, 2010 14:00
Version	1.0
OS	All
Depends On	llvm/llvm-bugzilla-archive#1521
Reporter	LLVM Bugzilla Contributor

Extended Description

#include <stdlib.h>

struct RT {
char a;
char b;
int c;
int d[50][50];
} glob;

int main(int argc, char **argv) {
glob.a = 1;
glob.b = 2;
glob.c = 3;
printf("glob.1 = %d\n", *(&glob + 1));
}

compiles 62 seconds, spits out 18kb bc file with 5063 tmp vars. For
larger arrays (say 1000 X 2000) it never returns. GCC has no problems
compiling such examples.

A 10-liner shouldn't kill the compiler :-)

Domagoj

[lattner]

It would be nice to fix this, but it's totally unrealistic code (i.e. performance of any code that uses this sort
of thing would be completely horrible).

Also, most of the compile time is probably spent on the same reason as Bug 473.

-Chris

[lattner]

fwiw., this code should use memcpy instead of individual field copies.

[lattner]

*** Bug llvm/llvm-bugzilla-archive#1104 has been marked as a duplicate of this bug. ***

[lattner]

*** Bug llvm/llvm-bugzilla-archive#1498 has been marked as a duplicate of this bug. ***

[lattner]

This has been fixed in llvm and llvm-gcc, and awaits specific targets to pick it up. You should be good on x86 and x86-64 now. Thanks to Rafael and Evan for making this happen for 2.2