floating point exceptions flags are ignored in constant folding

[llvmbot]

Bugzilla Link	3930
Resolution	INVALID
Resolved on	Apr 02, 2009 22:42
Version	unspecified
OS	All
Reporter	LLVM Bugzilla Contributor

Extended Description

I toyed around with http://llvm.org/demo/index.cgi

and found that this code :
#include <stdio.h>
#include <fenv.h>

int main(int argc, char **argv) {
#pragma STDC FENV_ACCESS ON
feclearexcept(FE_DIVBYZERO);
double d = 1.0/0.0;
if (fetestexcept(FE_DIVBYZERO))
printf("error\n");
else
printf("%e\n", d);
return 0;
}

is compiled this way :

define i32 @​main(i32 %argc, i8** nocapture %argv) nounwind {
entry:
%0 = tail call i32 @​feclearexcept(i32 4) nounwind ; [#uses=0]
%1 = tail call i32 @​fetestexcept(i32 4) nounwind ; [#uses=1]
%2 = icmp eq i32 %1, 0 ; [#uses=1]
br i1 %2, label %bb1, label %bb

bb: ; preds = %entry
%3 = tail call i32 @​puts(i8* getelementptr ([6 x i8]* @.str, i32 0, i32 0)) nounwind ; [#uses=0]
ret i32 0

bb1: ; preds = %entry
%4 = tail call i32 (i8*, ...)* @​printf(i8* noalias getelementptr ([4 x i8]* @.str1, i32 0, i32 0), double 0x7FF0000000000000) nounwind ; [#uses=0]
ret i32 0
}

it as basically no chance to print "error" like it does under a normal C compiler (tested on normal gcc x86_32).

the pragma is here to tell the compiler maker to be careful in floating point handling because there will be some tampering with the FP subsystem.

exactly the same kind of bug as Bug 3929

[lattner]

GCC ignores #pragma STDC FENV_ACCESS. This is a known issue. Please search the gcc bugzilla for details.

[llvmbot]

GCC ignores #pragma STDC FENV_ACCESS. This is a known issue. Please search
the gcc bugzilla for details.

This code works under gcc (mac) anyways, I was careful to test it and say so in the report.
If they ignore the pragma, the use the real code to decide if they can optimise or not. Which clearly LLVM doesn't.

[lattner]

No, they just get lucky. Try compiling at -O3 or some other high optimization level. GCC does not respect this. If this one test case happens to work, it is by chance, not design.

[llvmbot]

No, they just get lucky. Try compiling at -O3 or some other high optimization
level. GCC does not respect this. If this one test case happens to work, it
is by chance, not design.

both works under -O5

I don't feel like speaking about luck in this case. I'll check in gcc source code.