[all-commits] [llvm/llvm-project] fb34d5: Promote bf16 to f32 when the target doesn't suppor...

[Benjamin Kramer via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: fb34d531af953119593be74753b89baf99fbc194
      https://github.com/llvm/llvm-project/commit/fb34d531af953119593be74753b89baf99fbc194
  Author: Benjamin Kramer <benny.kra at googlemail.com>
  Date:   2022-06-15 (Wed, 15 Jun 2022)

  Changed paths:
    M compiler-rt/lib/builtins/CMakeLists.txt
    M compiler-rt/lib/builtins/fp_trunc.h
    A compiler-rt/lib/builtins/truncsfbf2.c
    M llvm/include/llvm/CodeGen/ISDOpcodes.h
    M llvm/include/llvm/IR/RuntimeLibcalls.def
    M llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
    M llvm/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp
    M llvm/lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp
    M llvm/lib/CodeGen/TargetLoweringBase.cpp
    M llvm/lib/Target/X86/X86ISelLowering.cpp
    A llvm/test/CodeGen/X86/bfloat.ll

  Log Message:
  -----------
  Promote bf16 to f32 when the target doesn't support it

This is modeled after the half-precision fp support. Two new nodes are
introduced for casting from and to bf16. Since casting from bf16 is a
simple operation I opted to always directly lower it to integer
arithmetic. The other way round is more complicated if you want to
preserve IEEE semantics, so it's handled by a new __truncsfbf2
compiler-rt builtin.

This is of course very bare bones, but sufficient to get a semi-softened
fadd on x86.

Possible future improvements:
 - Targets with bf16 conversion instructions can now make fp_to_bf16 legal
 - The software conversion to bf16 can be replaced by a trivial
   implementation under fast math.

Differential Revision: https://reviews.llvm.org/D126953


  Commit: 8bc0bb956421e02bc1a1797363e979655dd326d6
      https://github.com/llvm/llvm-project/commit/8bc0bb956421e02bc1a1797363e979655dd326d6
  Author: Benjamin Kramer <benny.kra at googlemail.com>
  Date:   2022-06-15 (Wed, 15 Jun 2022)

  Changed paths:
    M compiler-rt/lib/builtins/CMakeLists.txt
    A compiler-rt/lib/builtins/truncdfbf2.c
    M llvm/include/llvm/IR/RuntimeLibcalls.def
    M llvm/lib/CodeGen/TargetLoweringBase.cpp
    M llvm/test/CodeGen/X86/bfloat.ll

  Log Message:
  -----------
  Add a conversion from double to bf16

This introduces a new compiler-rt function `__truncdfbf2`.


Compare: https://github.com/llvm/llvm-project/compare/43e7ba649535...8bc0bb956421