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

1 files changed, 13 insertions, 0 deletions

diff --git a/llvm/lib/CodeGen/TargetLoweringBase.cpp b/llvm/lib/CodeGen/TargetLoweringBase.cpp
index c21ffaf..ad0f95b 100644
--- a/llvm/lib/CodeGen/TargetLoweringBase.cpp
+++ b/llvm/lib/CodeGen/TargetLoweringBase.cpp
@@ -274,6 +274,9 @@ RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) {
       return FPROUND_F128_F16;
     if (OpVT == MVT::ppcf128)
       return FPROUND_PPCF128_F16;
+  } else if (RetVT == MVT::bf16) {
+    if (OpVT == MVT::f32)
+      return FPROUND_F32_BF16;
   } else if (RetVT == MVT::f32) {
     if (OpVT == MVT::f64)
       return FPROUND_F64_F32;
@@ -1373,6 +1376,16 @@ void TargetLoweringBase::computeRegisterProperties(
     }
   }
 
+  // Decide how to handle bf16. If the target does not have native bf16 support,
+  // promote it to f32, because there are no bf16 library calls (except for
+  // converting from f32 to bf16).
+  if (!isTypeLegal(MVT::bf16)) {
+    NumRegistersForVT[MVT::bf16] = NumRegistersForVT[MVT::f32];
+    RegisterTypeForVT[MVT::bf16] = RegisterTypeForVT[MVT::f32];
+    TransformToType[MVT::bf16] = MVT::f32;
+    ValueTypeActions.setTypeAction(MVT::bf16, TypePromoteFloat);
+  }
+
   // Loop over all of the vector value types to see which need transformations.
   for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
        i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {