[DAG] Avoid strictfp nodes in lowering of llvm.fcanonicalize
The default lowering of llvm.fcanonicalize, implemented in #106370 and
#142105, follows the suggestions in
https://llvm.org/docs/LangRef.html#llvm-canonicalize-intrinsic and
implements canonicalization as a multiplication by 1.0. The problem with
this implementation is that the compiler tries to "optimize" the
multiplication by removing it. To prevent from this, the existing
implementation uses STRICT_FMUL for the multiplication, even in the
default mode.
This solution has some drawbacks. First, STRICT_* nodes are designed to
represent the side effects of floating-point operations. In the default
Mode, these side effects are ignored for all operation, so the
occurrence of strict node in this context is a misuse, it impedes the
development of strictfp support.
Second, the solution relies on the absence of optimizations for
STRICT_FMUL. However, STRICT_* nodes should also be optimized, and
future development may invalidate this assumption.
This patch modifies the default expansion of llvm.fcanonicalize. Instead
of using STRICT_FMUL it represents the multiplication by a special node,
which is changed to FMUL immediately before the instruction selection.
llvm-project
d3acd55d - [DAG] Avoid strictfp nodes in lowering of llvm.fcanonicalize
