[PATCH] Restore the sqrt -> llvm.sqrt mapping in fast-math mode

[Hal Finkel]

----- Original Message -----
> 
> On Thu, Sep 12, 2013 at 2:44 PM, Hal Finkel < hfinkel at anl.gov >
> wrote:
> 
> 
> 
> 
> 
> ----- Original Message -----
> > 
> > On Thu, Sep 12, 2013 at 2:03 PM, Hal Finkel < hfinkel at anl.gov >
> > wrote:
> > 
> > 
> > 
> > 
> > Hello,
> > 
> > Please review the attached patch which restores the libm sqrt* ->
> > @llvm.sqrt* mapping, but only in fast-math mode (specifically, when
> > the UnsafeFPMath or NoNaNsFPMath CodeGen options are enabled). The
> > @llvm.sqrt* intrinsics have slightly different semantics from the
> > libm call, specifically, they are undefined when given a non-zero
> > negative number (the libm calls will always return NaN for any
> > negative number).
> > 
> > This mapping was removed in r100613, and replaced with a TODO, but
> > at
> > that time the fast-math flags were not yet implemented. Now that we
> > have these, restoring this mapping is important because it will
> > enable autovectorization of sqrt calls in loops (at least in
> > fast-math mode).
> > 
> > 
> > 
> > 
> > This is dangerous, if LangRef is actually correct. People don't
> > associate -ffast-math with "my program will crash at random". :) Of
> > course, LangRef is probably overstating the issue.
> 
> I agree, and the LangRef does indeed say "undefined behavior", but I
> assume that should really mean, "returns an undefined value." Do you
> agree?
> 
> 
> 
> Well, if we map llvm.sqrt to sqrt and sqrt sets errno, we really do
> mean "undefined behavior"... or at least something more that
> "returns an undefined value".

Agreed, but this possibility-of-setting-errno problem exists for all LLVM libm-style intrinsics, and so also exists for pow() [for which we currently do this exact kind of replacement whenever -fmath-errno=0]. So, FWIW, this is not without precedent.

> 
> 
> 
> 
> > 
> > 
> > That said, there's actually a general issue here: if we map the
> > LLVM
> > intrinsics to libc functions, and the libc functions set errno, we
> > could break code that depends on errno for non-math calls (e.g.
> > fopen().)
> 
> Perhaps, but I'm not changing that here. For one thing, if the
> mapping does, in effect, sqrt -> llvm.sqrt -> sqrt, and only when
> -fmath-errno=0. Are you worried about cases where the libm functions
> actually do set errno (even though we have -fmath-errno=0)?
> 
> 
> 
> 
> I'm not sure our implementation of -fno-math-errno is safe: according
> to the gcc manual, it isn't equivalent to marking the math functions
> with attribute((const)). (For example, the gcc manual's definition
> allows transforming a call to sqrt() into the SSE sqrt instruction,
> but it doesn't allow hoisting a call to sqrt out of arbitrary loops
> on a machine where the sqrt() call could set errno.)

I'm sure it is not safe, for the very reason that you highlight. Nevertheless, this is a long-standing problem, affecting the implementation of -fmath-errno=0 on all systems for which libm math functions actually do set errno, and will require a general solution (fairly orthogonal to this patch).

I recommend that we:

 1. Commit this change (so that we can autovectorize calls to sqrt()).
 2. Have a discussion about how to actually solve this problem: I think that it involves making a specific function attribute for setting errno, and teaching the alias analysis infrastructure to do something sensible with it.

 -Hal

> 
> 
> -Eli

-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory