Since the default logic was based on having fast denormal/fma
features, and the default target has no features, we assumed flushing
by default. This fixes incorrectly assuming flushing in builds for
"generic" IR libraries.
The handling for no specified --cuda-gpu-arch in HIP is kind of
broken. Somewhere else forces a default target of gfx803, which does
not enable denormal handling by default. We don't see this default
switching here, so you'll end up with a different denormal mode
depending on whether you explicitly requested gfx803, or used it by
default.
I didn't realize HIP was a distinct offloading kind, so the subtarget
was looking for -march, which isn't correct for HIP. We also have the
possibility of different denormal defaults in the case of multiple
offload targets, so we need to thread the JobAction through the target
hook.
Apparently HIPToolChain does not subclass from AMDGPUToolChain, so
this was not applying the new denormal attributes. I'm not sure why
this doesn't subclass. Just copy the implementation for now.
There was a misisng space between the -march and --cuda-gpu-arch
arguments, so --cuda-gpu-arch wasn't actually being parsed. I'm not
sure what the intent of the sm_10 run lines were, but they error as an
unsupported architecture. Switch these to something else.
This reverts commit 737394c490.
The fp-model test was failing on platforms that enable denormal flushing
based on -ffast-math. This needs to reset to IEEE, not the default in
these cases.
Change-Id: Ibbad32f66d0d0b89b9c1173a3a96fb1a570ddd89
The IR hasn't switched the default yet, so explicitly add the ieee
attributes.
I'm still not really sure how the target default denormal mode should
interact with -fno-unsafe-math-optimizations. The target may have
selected the default mode to be non-IEEE based on the flags or based
on its true behavior, but we don't know which is the case. Since the
only users of a non-IEEE mode without a flag still support IEEE mode,
just reset to IEEE.
AMDGPU and x86 at least both have separate controls for whether
denormal results are flushed on output, and for whether denormals are
implicitly treated as 0 as an input. The current DAGCombiner use only
really cares about the input treatment of denormals.
Currently there are 4 different mechanisms for controlling denormal
flushing behavior, and about as many equivalent frontend controls.
- AMDGPU uses the fp32-denormals and fp64-f16-denormals subtarget features
- NVPTX uses the nvptx-f32ftz attribute
- ARM directly uses the denormal-fp-math attribute
- Other targets indirectly use denormal-fp-math in one DAGCombine
- cl-denorms-are-zero has a corresponding denorms-are-zero attribute
AMDGPU wants a distinct control for f32 flushing from f16/f64, and as
far as I can tell the same is true for NVPTX (based on the attribute
name).
Work on consolidating these into the denormal-fp-math attribute, and a
new type specific denormal-fp-math-f32 variant. Only ARM seems to
support the two different flush modes, so this is overkill for the
other use cases. Ideally we would error on the unsupported
positive-zero mode on other targets from somewhere.
Move the logic for selecting the flush mode into the compiler driver,
instead of handling it in cc1. denormal-fp-math/denormal-fp-math-f32
are now both cc1 flags, but denormal-fp-math-f32 is not yet exposed as
a user flag.
-cl-denorms-are-zero, -fcuda-flush-denormals-to-zero and
-fno-cuda-flush-denormals-to-zero will be mapped to
-fp-denormal-math-f32=ieee or preserve-sign rather than the old
attributes.
Stop emitting the denorms-are-zero attribute for the OpenCL flag. It
has no in-tree users. The meaning would also be target dependent, such
as the AMDGPU choice to treat this as only meaning allow flushing of
f32 and not f16 or f64. The naming is also potentially confusing,
since DAZ in other contexts refers to instructions implicitly treating
input denormals as zero, not necessarily flushing output denormals to
zero.
This also does not attempt to change the behavior for the current
attribute. The LangRef now states that the default is ieee behavior,
but this is inaccurate for the current implementation. The clang
handling is slightly hacky to avoid touching the existing
denormal-fp-math uses. Fixing this will be left for a future patch.
AMDGPU is still using the subtarget feature to control the denormal
mode, but the new attribute are now emitted. A future change will
switch this and remove the subtarget features.
Artem Belevich