Hubert Tong via llvm-dev
2020-Sep-13 01:37 UTC
[llvm-dev] Invalid transformation in LibCallSimplifier::replacePowWithSqrt?
The transformation in LibCallSimplifier::replacePowWithSqrt with respect to -Inf uses a select instruction, which based on the observed behaviour, incorporates the side effects of the unchosen branch. This means that (for pow) a call to sqrt(-Inf) is materialized. Such a call is specified as having a domain error (C17 subclause 7.12.7.5) since the operand is less than zero. Contrast this with pow(-Inf, 0.5), which is specified by C17 subclause F.10.4.4 as having a result of +Inf (indicating an exact result for the operation and, since IEEE Std 754-2008 subclause 9.1.1 states that domain errors are to be indicated by a NaN result, a lack of a domain error). It is noted that the above statements were made notwithstanding the ERRORS section of pow() in POSIX.1-2017 XSH Chapter 3, which specifies a domain error except perhaps by deference to the C standard due to a conflict between the POSIX and the C wording. The transformation in question causes EDOM for pow(-Inf, 0.5) even on platforms where the system library implementation of pow does not cause this situation to arise. A sample program that (on some platforms, such as Linux on x86-64) completes successfully with optimizations off, and aborts with LLVM's optimization follows; -fmath-errno does not help, and it is not expected to, because it is designed to retain setting errno to non-zero (not to prevent spuriously setting errno). #include <errno.h> volatile double inf = -__builtin_inf(); double pow(double, double); void abort(void); int main(void) { errno = 0; pow(inf, 0.5); if (errno != 0) abort(); } Compiler Explorer link: https://godbolt.org/z/5Wr66M Is the transformation actually valid in some way? If not, I see no instances where the LibCallSimplier generates conditional branches. Retaining the transformation in some way without generating conditional branches would probably involve bailing out if infinities are still in play. -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20200912/311cac5f/attachment.html>
Sanjay Patel via llvm-dev
2020-Sep-14 16:45 UTC
[llvm-dev] Invalid transformation in LibCallSimplifier::replacePowWithSqrt?
Yes, that looks like a bug. The transform is ok in general for negative numbers, but -Inf is a special-case for pow(), right? If so, we probably need an extra check of the input with "isKnownNeverInfinity()". If there are other errno divergences for edge case values, we may need to check other conditions. On Sat, Sep 12, 2020 at 9:37 PM Hubert Tong via llvm-dev < llvm-dev at lists.llvm.org> wrote:> The transformation in LibCallSimplifier::replacePowWithSqrt with respect > to -Inf uses a select instruction, which based on the observed behaviour, > incorporates the side effects of the unchosen branch. This means that (for > pow) a call to sqrt(-Inf) is materialized. Such a call is specified as > having a domain error (C17 subclause 7.12.7.5) since the operand is less > than zero. Contrast this with pow(-Inf, 0.5), which is specified by C17 > subclause F.10.4.4 as having a result of +Inf (indicating an exact result > for the operation and, since IEEE Std 754-2008 subclause 9.1.1 states that > domain errors are to be indicated by a NaN result, a lack of a domain > error). > > It is noted that the above statements were made notwithstanding the ERRORS > section of pow() in POSIX.1-2017 XSH Chapter 3, which specifies a domain > error except perhaps by deference to the C standard due to a conflict > between the POSIX and the C wording. > > The transformation in question causes EDOM for pow(-Inf, 0.5) even on > platforms where the system library implementation of pow does not cause > this situation to arise. > > A sample program that (on some platforms, such as Linux on x86-64) > completes successfully with optimizations off, and aborts with LLVM's > optimization follows; -fmath-errno does not help, and it is not expected > to, because it is designed to retain setting errno to non-zero (not to > prevent spuriously setting errno). > > #include <errno.h> > volatile double inf = -__builtin_inf(); > > double pow(double, double); > void abort(void); > int main(void) { > errno = 0; > pow(inf, 0.5); > if (errno != 0) abort(); > } > > Compiler Explorer link: https://godbolt.org/z/5Wr66M > > Is the transformation actually valid in some way? If not, I see no > instances where the LibCallSimplier generates conditional branches. > Retaining the transformation in some way without generating conditional > branches would probably involve bailing out if infinities are still in play. > > _______________________________________________ > LLVM Developers mailing list > llvm-dev at lists.llvm.org > https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20200914/09e98d11/attachment.html>
Hubert Tong via llvm-dev
2020-Sep-14 16:58 UTC
[llvm-dev] Invalid transformation in LibCallSimplifier::replacePowWithSqrt?
On Mon, Sep 14, 2020 at 12:45 PM Sanjay Patel <spatel at rotateright.com> wrote:> Yes, that looks like a bug. The transform is ok in general for negative > numbers, but -Inf is a special-case for pow(), right? > If so, we probably need an extra check of the input with > "isKnownNeverInfinity()". >There is an extra check there already, but it uses "select" instead of branching. One question is if branching is okay to use instead. Or perhaps you mean the transform should not be done unless "isKnownNeverInfinity()" returns true.> If there are other errno divergences for edge case values, we may need to > check other conditions. > > On Sat, Sep 12, 2020 at 9:37 PM Hubert Tong via llvm-dev < > llvm-dev at lists.llvm.org> wrote: > >> The transformation in LibCallSimplifier::replacePowWithSqrt with respect >> to -Inf uses a select instruction, which based on the observed behaviour, >> incorporates the side effects of the unchosen branch. This means that (for >> pow) a call to sqrt(-Inf) is materialized. Such a call is specified as >> having a domain error (C17 subclause 7.12.7.5) since the operand is less >> than zero. Contrast this with pow(-Inf, 0.5), which is specified by C17 >> subclause F.10.4.4 as having a result of +Inf (indicating an exact result >> for the operation and, since IEEE Std 754-2008 subclause 9.1.1 states that >> domain errors are to be indicated by a NaN result, a lack of a domain >> error). >> >> It is noted that the above statements were made notwithstanding the >> ERRORS section of pow() in POSIX.1-2017 XSH Chapter 3, which specifies a >> domain error except perhaps by deference to the C standard due to a >> conflict between the POSIX and the C wording. >> >> The transformation in question causes EDOM for pow(-Inf, 0.5) even on >> platforms where the system library implementation of pow does not cause >> this situation to arise. >> >> A sample program that (on some platforms, such as Linux on x86-64) >> completes successfully with optimizations off, and aborts with LLVM's >> optimization follows; -fmath-errno does not help, and it is not expected >> to, because it is designed to retain setting errno to non-zero (not to >> prevent spuriously setting errno). >> >> #include <errno.h> >> volatile double inf = -__builtin_inf(); >> >> double pow(double, double); >> void abort(void); >> int main(void) { >> errno = 0; >> pow(inf, 0.5); >> if (errno != 0) abort(); >> } >> >> Compiler Explorer link: https://godbolt.org/z/5Wr66M >> >> Is the transformation actually valid in some way? If not, I see no >> instances where the LibCallSimplier generates conditional branches. >> Retaining the transformation in some way without generating conditional >> branches would probably involve bailing out if infinities are still in play. >> >> _______________________________________________ >> LLVM Developers mailing list >> llvm-dev at lists.llvm.org >> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev >> >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20200914/6e857a8b/attachment.html>
Seemingly Similar Threads
- Invalid transformation in LibCallSimplifier::replacePowWithSqrt?
- Invalid transformation in LibCallSimplifier::replacePowWithSqrt?
- Undef and Poison round table follow-up & a plan
- [LLVMdev] a question about constant fold for fdiv
- [LLVMdev] a question about constant fold for fdiv