fabs not marked readnone

[nlewycky]

Bugzilla Link	2520
Resolution	FIXED
Resolved on	Jul 08, 2008 05:13
Version	trunk
OS	Linux

Extended Description

According to fabs(3), none of fabs fabsf or fabsl can have any errors. As such, they should be marked readnone since all they ever do is return their own input.

$ cat fabs.c
#include <math.h>

double f(double *x, double *y) { return fabs(*x + *y); }

$ llvm-gcc -O2 fabs.c -S -o - -emit-llvm
; ModuleID = 'fabs.c'
target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:32:32"
target triple = "i386-pc-linux-gnu"

define double @​f(double* %x, double* %y) nounwind {
entry:
%tmp2 = load double* %x, align 8 ; [#uses=1]
%tmp4 = load double* %y, align 8 ; [#uses=1]
%tmp5 = add double %tmp2, %tmp4 ; [#uses=1]
%tmp6 = tail call double @​fabs( double %tmp5 ) ; [#uses=1]
ret double %tmp6
}

declare double @​fabs(double)

This stops some optimizations from working (GVN in particular) because it doesn't know that fabs() won't be writing to one of the double* arguments.

[llvmbot]

This occurs because llvm-convert generates a call to fabs
directly itself (as opposed to converting gcc's fabs builtin)
and forgets to generate attributes. It's not clear to me
what the correct attributes are: "const" (readnone) is the
obvious choice, but, since the result of fabs presumably depends
on the rounding mode, "pure" (readonly) is probably more correct
unless -ffast-math is specified. gcc has a whole infrastructure
for determining the right attributes given parameter flags; it
would be nice to make use of it here.

[lattner]

fabs is actually really easy. It never depends on the rounding mode and doesn't change behavior based on -ffast-math etc. It should always be readnone.

[llvmbot]

Fixed here:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20080707/064680.html