llvm dev - Sep 2016 - LLVM v3.9.0 and math built-ins

A little while ago I asked a question on CFE-Dev about a change in the
behaviour of programs using the ISO C math functions, although that question
should have been put to LLVM-Dev.  But I got excellent clarification of the
problem anyway.  However, since then I have been trying to adapt our
out-of-tree implementation to get the previous behaviour.  The problem is
that something like:

 

#include <math.h>

 

extern double foo(double);

 

int useMathName() {

  if ((exp(1.0) < 2.71) || (exp(1.0) > 2.72))

    return -1;

  return 0;

}

 

int useOtherName() {

  if ((foo(1.0) < 2.71) || (foo(1.0) > 2.72))

    return -1;

  return 0;

}

 

With v3.8.0 the compiler elided both the calls to 'exp' and the tests,
so
the function 'useMathName' reduced to simply:

 

return 0;

 

But this was not correct as it ignored the possibility that the math
functions could have the side-effect of changing 'errno', and this was
fixed
in v3.9.0, and the calls are no longer elided, though using the
"as-if"
rule, the tests are still eliminated and 'useMathName' becomes:

 

(void)exp(1.0);

(void)exp(1.0);

return 0;

 

So I changed our implementation so that '-fno-math-errno' is the default
for
the tool-chain, and the '-fmath-errno' option is not passed on to the
CC1
phase.  I expected that this would allow the compiler to elide the calls,
but it is not doing so.  I don't want to use '-ffast-math' as this
has lots
of FP issues, and I don't want to make it the default.

 

Any idea why the math functions are not elided?  I am using '-O3' and
implying '-fno-math-errno'.  I have verified that our Toolchain
implements
'IsMathErrnoDefault' and returns 'false', and that the option
'-fmath-errno'
is not being passed to the CC1 stage.  Since our math function
implementation does not in fact change 'errno', it is very desirable
that
this elision occurs in an FP safe way.

 

Thanks,

 

            MartinO

 

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20160916/c1685c19/attachment.html>

> On Sep 16, 2016, at 1:30 AM, Martin J. O'Riordan via llvm-dev
<llvm-dev at lists.llvm.org> wrote:
> 
> A little while ago I asked a question on CFE-Dev about a change in the
behaviour of programs using the ISO C math functions, although that question
should have been put to LLVM-Dev.  But I got excellent clarification of the
problem anyway.  However, since then I have been trying to adapt our out-of-tree
implementation to get the previous behaviour.  The problem is that something
like:
>  
> #include <math.h>
>  
> extern double foo(double);
>  
> int useMathName() {
>   if ((exp(1.0) < 2.71) || (exp(1.0) > 2.72))
>     return -1;
>   return 0;
> }
>  
> int useOtherName() {
>   if ((foo(1.0) < 2.71) || (foo(1.0) > 2.72))
>     return -1;
>   return 0;
> }
>  
> With v3.8.0 the compiler elided both the calls to ‘exp’ and the tests, so
the function ‘useMathName’ reduced to simply:
>  
> return 0;
>  
> But this was not correct as it ignored the possibility that the math
functions could have the side-effect of changing ‘errno’, and this was fixed in
v3.9.0, and the calls are no longer elided, though using the “as-if” rule, the
tests are still eliminated and ‘useMathName’ becomes:
>  
> (void)exp(1.0);
> (void)exp(1.0);

It is not clear to me: isn’t the spec saying that errno shall be set on
underflow and overflow. The constant folding should be able to detect this and
fold if it does not happen?

Also, because we don’t really model errno, we don’t constant fold these call
anymore :(

> return 0;
>  
> So I changed our implementation so that ‘-fno-math-errno’ is the default
for the tool-chain, and the ‘-fmath-errno’ option is not passed on to the CC1
phase.  I expected that this would allow the compiler to elide the calls, but it
is not doing so.  I don’t want to use ‘-ffast-math’ as this has lots of FP
issues, and I don’t want to make it the default.
>  
> Any idea why the math functions are not elided?  I am using ‘-O3’ and
implying ‘-fno-math-errno’.  I have verified that our Toolchain implements
‘IsMathErrnoDefault’ and returns ‘false’, and that the option ‘-fmath-errno’ is
not being passed to the CC1 stage.  Since our math function implementation does
not in fact change ‘errno’, it is very desirable that this elision occurs in an
FP safe way.
You can check that clang when emitting the IR is adding the attribute
"readnone” on the declaration of exp(). If it does not then it assumes
errno.

As a starting point, have you tried to pass -fno-math-errno?

— 
Mehdi

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20160916/3e4524c9/attachment.html>

Hi Mehdi,

 

The ISO C specification does permit the math functions to modify ‘errno’, but I
thought that the ‘-fno-math-errno’  option was to tell the optimiser to assume
that ‘errno’ is not modified by the math functions.  Explicitly providing
‘-fno-math-errno’ is not restoring the elision optimisation that was performed
by LLVM v3.8, and this is really only a driver option, with ‘-fmath-errno’
passed onto CC1 if ‘MathErrno’ is true, and omitted if false.

 

I certainly agree that if ‘MathErrno’ is true, then the compiler should not
elide the call as it has to assume that there are memory write side-effects.  It
already models the math functions as built-ins and determine the expected answer
when the arguments are constants, and elides the tests in the simple example
below - all this is good, and when I use ‘-fno-builtin’ it stops correctly
disabled this optimisation.

 

I modified my sources to force ‘readnone’ and even tried
‘setDoesNotAccessMemory()’ on the math functions, but this still results in one
call occurring and the code generated is effectively:

 

(void)exp(1.0);

return 0;

 

This is better, but still not as optimal as it was in v3.8.  With this change,
compiling with ‘-fmath-errno’ reverts to inserting two calls.

 

While calling these functions with constants is not terribly common in C code
(mostly contrived test cases), it is a lot more common in real C++ code via
template expansions and inlining.  With ‘-O3 -S -emit-llvm -fno-math-errno’ plus
my change to call ‘setDoesNotAccessMemory()’, the IR generated is as follows:

 

; ModuleID = 'src/mathElisions.c'

source_filename = "src/mathElisions.c"

target datalayout =
"e-m:e-p:32:32-f64:64-i64:64-v128:64-v64:64-v32:32-v16:16-n8:16:32-S64"

target triple = "shave"

 

; Function Attrs: nounwind readnone

define i32 @useMathName() local_unnamed_addr #0 {

  %1 = tail call float @exp(float 1.000000e+00)

  ret i32 0

}

 

; Function Attrs: readnone

declare float @exp(float) local_unnamed_addr #1

 

; Function Attrs: nounwind

define i32 @useOtherName() local_unnamed_addr #2 {

  %1 = tail call float @foo(float 1.000000e+00) #4

  %2 = fcmp olt float %1, 0x4005AE1480000000

  br i1 %2, label %7, label %3

 

; <label>:3:                                      ; preds = %0

  %4 = tail call float @foo(float 1.000000e+00) #4

  %5 = fcmp ogt float %4, 0x4005C28F60000000

  br i1 %5, label %7, label %6

 

; <label>:6:                                      ; preds = %3

  br label %7

 

; <label>:7:                                      ; preds = %0, %3, %6

  %8 = phi i32 [ 0, %6 ], [ -1, %3 ], [ -1, %0 ]

  ret i32 %8

}

 

declare float @foo(float) local_unnamed_addr #3

 

attributes #0 = { nounwind readnone
"disable-tail-calls"="false"
"less-precise-fpmad"="false"
"no-frame-pointer-elim"="true"
"no-frame-pointer-elim-non-leaf"
"no-infs-fp-math"="false"
"no-jump-tables"="false"
"no-nans-fp-math"="false"
"no-signed-zeros-fp-math"="false"
"stack-protector-buffer-size"="8"
"target-cpu"="myriad2.2"
"unsafe-fp-math"="false"
"use-soft-float"="false" }

attributes #1 = { readnone "disable-tail-calls"="false"
"less-precise-fpmad"="false"
"no-frame-pointer-elim"="true"
"no-frame-pointer-elim-non-leaf"
"no-infs-fp-math"="false"
"no-nans-fp-math"="false"
"no-signed-zeros-fp-math"="false"
"stack-protector-buffer-size"="8"
"target-cpu"="myriad2.2"
"unsafe-fp-math"="false"
"use-soft-float"="false" }

attributes #2 = { nounwind "disable-tail-calls"="false"
"less-precise-fpmad"="false"
"no-frame-pointer-elim"="true"
"no-frame-pointer-elim-non-leaf"
"no-infs-fp-math"="false"
"no-jump-tables"="false"
"no-nans-fp-math"="false"
"no-signed-zeros-fp-math"="false"
"stack-protector-buffer-size"="8"
"target-cpu"="myriad2.2"
"unsafe-fp-math"="false"
"use-soft-float"="false" }

attributes #3 = { "disable-tail-calls"="false"
"less-precise-fpmad"="false"
"no-frame-pointer-elim"="true"
"no-frame-pointer-elim-non-leaf"
"no-infs-fp-math"="false"
"no-nans-fp-math"="false"
"no-signed-zeros-fp-math"="false"
"stack-protector-buffer-size"="8"
"target-cpu"="myriad2.2"
"unsafe-fp-math"="false"
"use-soft-float"="false" }

attributes #4 = { nounwind }

 

Our implementation has C ‘float’ and ‘double’ as 32-bit IEEE Single-Precision
and ‘long double’ as 64-bit IEEE Double-Precision.

 

Thanks,

 

            MartinO

 

From: mehdi.amini at apple.com <mailto:mehdi.amini at apple.com> 
[mailto:mehdi.amini at apple.com]
Sent: 16 September 2016 21:16



 

On Sep 16, 2016, at 1:30 AM, Martin J. O'Riordan via llvm-dev <llvm-dev
at lists.llvm.org <mailto:llvm-dev at lists.llvm.org> > wrote:

 

A little while ago I asked a question on CFE-Dev about a change in the behaviour
of programs using the ISO C math functions, although that question should have
been put to LLVM-Dev.  But I got excellent clarification of the problem anyway. 
However, since then I have been trying to adapt our out-of-tree implementation
to get the previous behaviour.  The problem is that something like:

 

#include <math.h>

 

extern double foo(double);

 

int useMathName() {

  if ((exp(1.0) < 2.71) || (exp(1.0) > 2.72))

    return -1;

  return 0;

}

 

int useOtherName() {

  if ((foo(1.0) < 2.71) || (foo(1.0) > 2.72))

    return -1;

  return 0;

}

 

With v3.8.0 the compiler elided both the calls to ‘exp’ and the tests, so the
function ‘useMathName’ reduced to simply:

 

return 0;

 

But this was not correct as it ignored the possibility that the math functions
could have the side-effect of changing ‘errno’, and this was fixed in v3.9.0,
and the calls are no longer elided, though using the “as-if” rule, the tests are
still eliminated and ‘useMathName’ becomes:

 

(void)exp(1.0);

(void)exp(1.0);

 

o me: isn’t the spec saying that errno shall be set on underflow and overflow.
The constant folding should be able to detect this and fold if it does not
happen?

 

Also, because we don’t really model errno, we don’t constant fold these call
anymore :(

 

return 0;

 

So I changed our implementation so that ‘-fno-math-errno’ is the default for the
tool-chain, and the ‘-fmath-errno’ option is not passed on to the CC1 phase.  I
expected that this would allow the compiler to elide the calls, but it is not
doing so.  I don’t want to use ‘-ffast-math’ as this has lots of FP issues, and
I don’t want to make it the default.

 

Any idea why the math functions are not elided?  I am using ‘-O3’ and implying
‘-fno-math-errno’.  I have verified that our Toolchain implements
‘IsMathErrnoDefault’ and returns ‘false’, and that the option ‘-fmath-errno’ is
not being passed to the CC1 stage.  Since our math function implementation does
not in fact change ‘errno’, it is very desirable that this elision occurs in an
FP safe way.

 

You can check that clang when emitting the IR is adding the attribute
"readnone” on the declaration of exp(). If it does not then it assumes
errno.

 

As a starting point, have you tried to pass -fno-math-errno?

 

— 

Mehdi

 

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20160920/5a6eb973/attachment.html>