[RISCV] Support Inline ASM for the bf16 type.

This patch makes the RISCV-V asm constraint `f` recognize the bfloat
type.

Author: circYuan

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

@@ -20155,7 +20155,8 @@ RISCVTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
         return std::make_pair(0U, &RISCV::GPRPairRegClass);
       return std::make_pair(0U, &RISCV::GPRNoX0RegClass);
     case 'f':
-      if (Subtarget.hasStdExtZfhmin() && VT == MVT::f16)
+      if ((Subtarget.hasStdExtZfhmin() && VT == MVT::f16) ||
+          (Subtarget.hasStdExtZfbfmin() && VT == MVT::bf16))
         return std::make_pair(0U, &RISCV::FPR16RegClass);
       if (Subtarget.hasStdExtF() && VT == MVT::f32)
         return std::make_pair(0U, &RISCV::FPR32RegClass);
@@ -20273,6 +20274,11 @@ RISCVTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
         unsigned HReg = RISCV::F0_H + RegNo;
         return std::make_pair(HReg, &RISCV::FPR16RegClass);
       }
+      if (Subtarget.hasStdExtZfbfmin() && VT == MVT::bf16){
+        unsigned RegNo = FReg - RISCV::F0_F;
+        unsigned HReg = RISCV::F0_H + RegNo;
+        return std::make_pair(HReg, &RISCV::FPR16RegClass);
+      }
     }
   }
 
@@ -20949,6 +20955,14 @@ bool RISCVTargetLowering::splitValueIntoRegisterParts(
     return true;
   }
 
+  // Since the inline asm only use the first type in the RegisterClass, the bf16
+  // inline asm would choose the f16 from FPR16RegClass for doing the copy, and
+  // we correct the behavior here to avoid generating wrong SelectionDAG.
+  if (ValueVT == MVT::bf16 && PartVT == MVT::f16){
+    Parts[0] = Val;
+    return true;
+  }
+
   if (ValueVT.isScalableVector() && PartVT.isScalableVector()) {
     LLVMContext &Context = *DAG.getContext();
     EVT ValueEltVT = ValueVT.getVectorElementType();

llvm/test/CodeGen/RISCV/inline-asm-f-constraint-bf16.ll

@@ -0,0 +1,77 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=riscv32 -mattr=+f,+experimental-zfbfmin -target-abi=ilp32 -verify-machineinstrs < %s \
+; RUN:   | FileCheck -check-prefix=RV32F %s
+; RUN: llc -mtriple=riscv64 -mattr=+f,+experimental-zfbfmin -target-abi=lp64 -verify-machineinstrs < %s \
+; RUN:   | FileCheck -check-prefix=RV64F %s
+; RUN: llc -mtriple=riscv32 -mattr=+d,+experimental-zfbfmin -target-abi=ilp32 -verify-machineinstrs < %s \
+; RUN:   | FileCheck -check-prefix=RV32F %s
+; RUN: llc -mtriple=riscv64 -mattr=+d,+experimental-zfbfmin -target-abi=lp64 -verify-machineinstrs < %s \
+; RUN:   | FileCheck -check-prefix=RV64F %s
+
+@gf = external global float
+
+define float @constraint_f_float(bfloat %a) nounwind {
+; RV32F-LABEL: constraint_f_float:
+; RV32F:       # %bb.0:
+; RV32F-NEXT:    fmv.h.x fa5, a0
+; RV32F-NEXT:    #APP
+; RV32F-NEXT:    fcvt.s.bf16 fa5, fa5
+; RV32F-NEXT:    #NO_APP
+; RV32F-NEXT:    fmv.x.w a0, fa5
+; RV32F-NEXT:    ret
+;
+; RV64F-LABEL: constraint_f_float:
+; RV64F:       # %bb.0:
+; RV64F-NEXT:    fmv.h.x fa5, a0
+; RV64F-NEXT:    #APP
+; RV64F-NEXT:    fcvt.s.bf16 fa5, fa5
+; RV64F-NEXT:    #NO_APP
+; RV64F-NEXT:    fmv.x.w a0, fa5
+; RV64F-NEXT:    ret
+  %1 = load float, float* @gf
+  %2 = tail call float asm "fcvt.s.bf16 $0, $1", "=f,f"(bfloat %a)
+  ret float %2
+}
+
+define float @constraint_f_float_abi_name(bfloat %a) nounwind {
+; RV32F-LABEL: constraint_f_float_abi_name:
+; RV32F:       # %bb.0:
+; RV32F-NEXT:    fmv.h.x fa0, a0
+; RV32F-NEXT:    #APP
+; RV32F-NEXT:    fcvt.s.bf16 ft0, fa0
+; RV32F-NEXT:    #NO_APP
+; RV32F-NEXT:    fmv.x.w a0, ft0
+; RV32F-NEXT:    ret
+;
+; RV64F-LABEL: constraint_f_float_abi_name:
+; RV64F:       # %bb.0:
+; RV64F-NEXT:    fmv.h.x fa0, a0
+; RV64F-NEXT:    #APP
+; RV64F-NEXT:    fcvt.s.bf16 ft0, fa0
+; RV64F-NEXT:    #NO_APP
+; RV64F-NEXT:    fmv.x.w a0, ft0
+; RV64F-NEXT:    ret
+  %1 = load float, float* @gf
+  %2 = tail call float asm "fcvt.s.bf16 $0, $1", "={ft0},{fa0}"(bfloat %a)
+  ret float %2
+}
+
+define bfloat @constraint_gpr(bfloat %x) {
+; RV32F-LABEL: constraint_gpr:
+; RV32F:       # %bb.0:
+; RV32F-NEXT:    .cfi_def_cfa_offset 0
+; RV32F-NEXT:    #APP
+; RV32F-NEXT:    mv a0, a0
+; RV32F-NEXT:    #NO_APP
+; RV32F-NEXT:    ret
+;
+; RV64F-LABEL: constraint_gpr:
+; RV64F:       # %bb.0:
+; RV64F-NEXT:    .cfi_def_cfa_offset 0
+; RV64F-NEXT:    #APP
+; RV64F-NEXT:    mv a0, a0
+; RV64F-NEXT:    #NO_APP
+; RV64F-NEXT:    ret
+  %1 = tail call bfloat asm sideeffect alignstack "mv $0, $1", "={x10},{x10}"(bfloat %x)
+  ret bfloat %1
+}