On Thu, Mar 16, 2023 at 9:25 AM Qing Zhao via Gcc-patches <gcc-patches@gcc.gnu.org> wrote: > > Hi, > > Recently, we discovered some floating point precision diffs when using GCC8 > to build our > application on arm64: After some investigation, it turns out that this is due > to the > -ffp-contract=fast option that is on by default. Therefore, we have to > explicitly add > -ffp-contract=off and do a full-rebuild. > > GCC by default turns -ffp-contract=fast on. > https://gcc.gnu.org/onlinedocs/gcc-8.5.0/gcc/Optimize-Options.html#Optimize-Options > https://gcc.gnu.org/onlinedocs/gcc-12.2.0/gcc/Optimize-Options.html#Optimize-Options > > " > -ffp-contract=style > -ffp-contract=off disables floating-point expression contraction. > -ffp-contract=fast enables > floating-point expression contraction such as forming of fused multiply-add > operations if > the target has native support for them. -ffp-contract=on enables > floating-point expression > contraction if allowed by the language standard. This is currently not > implemented and > treated equal to -ffp-contract=off. > > The default is -ffp-contract=fast. > " > > This can be shown by a small example for arm64 with gcc8.5 in > https://godbolt.org/z/MxYfnG8TE. > Only when adding -std=c89 explicitly, this transformaton is off. > > another exmaple also shows that Clang and MSVC only allow this transformation > when speifiying > ffast-math and fp:fast: https://godbolt.org/z/o54bYfPbP > > When searching online, we found that there were similar discussions recently > on the exact same issue: > https://github.com/dotnet/runtime/issues/64604 > https://github.com/dotnet/runtime/issues/64591 > > a summary of these discussions is: > > 1. "fmadd" is a fused operation and will return a different result for many > inputs; > 2. therefore, -ffp-contract=fast is not a safe optimization to be on by > default; > 3. Clang and MSVC only allow this when specifying ffast-math and fp:fast > since this is not an > IEEE754 compliant optimization; > 4. The reasons why GCC turns on this option by default are: > A. GNU C language spec allows such transformation. > B. this did not expose real problem for most X86/X64 apps previously since > FMA instructions > didn't exist until 2013 when the FMA3 instruction set was added, and > also these instructions > were not always available.. > 5. Arm64 has fused multiply-add instructions as "baseline" and are always > available. therefore > -ffp-contract=fast exposed more serious problems on Arm64 platforms.
This summary ignores x87 and even ignores PowerPC in GCC having FMA even before clang/LLVM was around. > > our major question: > > Should GCC turn off -ffp-contract=fast by default since it's not IEEE754 > compliant and more > modern processors have the FMA instructions available by default? NO. We have this debate every few years and such. Thanks, Andrew Pinski > > Thanks. > > Qing
