[RISCV][LibCall] Add libcall for i64 -> bf16 (#130024)

Add support for lowering i64 -> bf16 with libcall.

Author: tclin914

llvm/include/llvm/IR/RuntimeLibcalls.def

@@ -444,6 +444,7 @@ HANDLE_LIBCALL(SINTTOFP_I32_F64, "__floatsidf")
 HANDLE_LIBCALL(SINTTOFP_I32_F80, "__floatsixf")
 HANDLE_LIBCALL(SINTTOFP_I32_F128, "__floatsitf")
 HANDLE_LIBCALL(SINTTOFP_I32_PPCF128, "__gcc_itoq")
+HANDLE_LIBCALL(SINTTOFP_I64_BF16, "__floatdibf")
 HANDLE_LIBCALL(SINTTOFP_I64_F16, "__floatdihf")
 HANDLE_LIBCALL(SINTTOFP_I64_F32, "__floatdisf")
 HANDLE_LIBCALL(SINTTOFP_I64_F64, "__floatdidf")
@@ -462,6 +463,7 @@ HANDLE_LIBCALL(UINTTOFP_I32_F64, "__floatunsidf")
 HANDLE_LIBCALL(UINTTOFP_I32_F80, "__floatunsixf")
 HANDLE_LIBCALL(UINTTOFP_I32_F128, "__floatunsitf")
 HANDLE_LIBCALL(UINTTOFP_I32_PPCF128, "__gcc_utoq")
+HANDLE_LIBCALL(UINTTOFP_I64_BF16, "__floatundibf")
 HANDLE_LIBCALL(UINTTOFP_I64_F16, "__floatundihf")
 HANDLE_LIBCALL(UINTTOFP_I64_F32, "__floatundisf")
 HANDLE_LIBCALL(UINTTOFP_I64_F64, "__floatundidf")

llvm/lib/CodeGen/TargetLoweringBase.cpp

@@ -312,6 +312,8 @@ RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) {
     if (RetVT == MVT::ppcf128)
       return SINTTOFP_I32_PPCF128;
   } else if (OpVT == MVT::i64) {
+    if (RetVT == MVT::bf16)
+      return SINTTOFP_I64_BF16;
     if (RetVT == MVT::f16)
       return SINTTOFP_I64_F16;
     if (RetVT == MVT::f32)
@@ -358,6 +360,8 @@ RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) {
     if (RetVT == MVT::ppcf128)
       return UINTTOFP_I32_PPCF128;
   } else if (OpVT == MVT::i64) {
+    if (RetVT == MVT::bf16)
+      return UINTTOFP_I64_BF16;
     if (RetVT == MVT::f16)
       return UINTTOFP_I64_F16;
     if (RetVT == MVT::f32)

llvm/test/CodeGen/RISCV/bfloat-convert.ll

@@ -1098,17 +1098,99 @@ define bfloat @fcvt_bf16_wu_load(ptr %p) nounwind {
   ret bfloat %1
 }
 
-; TODO: The following tests error on rv32 with zfbfmin enabled.
-
-; define bfloat @fcvt_bf16_l(i64 %a) nounwind {
-;   %1 = sitofp i64 %a to bfloat
-;   ret bfloat %1
-; }
+define bfloat @fcvt_bf16_l(i64 %a) nounwind {
+; CHECK32ZFBFMIN-LABEL: fcvt_bf16_l:
+; CHECK32ZFBFMIN:       # %bb.0:
+; CHECK32ZFBFMIN-NEXT:    addi sp, sp, -16
+; CHECK32ZFBFMIN-NEXT:    sw ra, 12(sp) # 4-byte Folded Spill
+; CHECK32ZFBFMIN-NEXT:    call __floatdibf
+; CHECK32ZFBFMIN-NEXT:    lw ra, 12(sp) # 4-byte Folded Reload
+; CHECK32ZFBFMIN-NEXT:    addi sp, sp, 16
+; CHECK32ZFBFMIN-NEXT:    ret
+;
+; RV32ID-LABEL: fcvt_bf16_l:
+; RV32ID:       # %bb.0:
+; RV32ID-NEXT:    addi sp, sp, -16
+; RV32ID-NEXT:    sw ra, 12(sp) # 4-byte Folded Spill
+; RV32ID-NEXT:    call __floatdisf
+; RV32ID-NEXT:    call __truncsfbf2
+; RV32ID-NEXT:    fmv.x.w a0, fa0
+; RV32ID-NEXT:    lui a1, 1048560
+; RV32ID-NEXT:    or a0, a0, a1
+; RV32ID-NEXT:    fmv.w.x fa0, a0
+; RV32ID-NEXT:    lw ra, 12(sp) # 4-byte Folded Reload
+; RV32ID-NEXT:    addi sp, sp, 16
+; RV32ID-NEXT:    ret
+;
+; CHECK64ZFBFMIN-LABEL: fcvt_bf16_l:
+; CHECK64ZFBFMIN:       # %bb.0:
+; CHECK64ZFBFMIN-NEXT:    fcvt.s.l fa5, a0
+; CHECK64ZFBFMIN-NEXT:    fcvt.bf16.s fa0, fa5
+; CHECK64ZFBFMIN-NEXT:    ret
+;
+; RV64ID-LABEL: fcvt_bf16_l:
+; RV64ID:       # %bb.0:
+; RV64ID-NEXT:    addi sp, sp, -16
+; RV64ID-NEXT:    sd ra, 8(sp) # 8-byte Folded Spill
+; RV64ID-NEXT:    fcvt.s.l fa0, a0
+; RV64ID-NEXT:    call __truncsfbf2
+; RV64ID-NEXT:    fmv.x.w a0, fa0
+; RV64ID-NEXT:    lui a1, 1048560
+; RV64ID-NEXT:    or a0, a0, a1
+; RV64ID-NEXT:    fmv.w.x fa0, a0
+; RV64ID-NEXT:    ld ra, 8(sp) # 8-byte Folded Reload
+; RV64ID-NEXT:    addi sp, sp, 16
+; RV64ID-NEXT:    ret
+  %1 = sitofp i64 %a to bfloat
+  ret bfloat %1
+}
 
-; define bfloat @fcvt_bf16_lu(i64 %a) nounwind {
-;   %1 = uitofp i64 %a to bfloat
-;   ret bfloat %1
-; }
+define bfloat @fcvt_bf16_lu(i64 %a) nounwind {
+; CHECK32ZFBFMIN-LABEL: fcvt_bf16_lu:
+; CHECK32ZFBFMIN:       # %bb.0:
+; CHECK32ZFBFMIN-NEXT:    addi sp, sp, -16
+; CHECK32ZFBFMIN-NEXT:    sw ra, 12(sp) # 4-byte Folded Spill
+; CHECK32ZFBFMIN-NEXT:    call __floatundibf
+; CHECK32ZFBFMIN-NEXT:    lw ra, 12(sp) # 4-byte Folded Reload
+; CHECK32ZFBFMIN-NEXT:    addi sp, sp, 16
+; CHECK32ZFBFMIN-NEXT:    ret
+;
+; RV32ID-LABEL: fcvt_bf16_lu:
+; RV32ID:       # %bb.0:
+; RV32ID-NEXT:    addi sp, sp, -16
+; RV32ID-NEXT:    sw ra, 12(sp) # 4-byte Folded Spill
+; RV32ID-NEXT:    call __floatundisf
+; RV32ID-NEXT:    call __truncsfbf2
+; RV32ID-NEXT:    fmv.x.w a0, fa0
+; RV32ID-NEXT:    lui a1, 1048560
+; RV32ID-NEXT:    or a0, a0, a1
+; RV32ID-NEXT:    fmv.w.x fa0, a0
+; RV32ID-NEXT:    lw ra, 12(sp) # 4-byte Folded Reload
+; RV32ID-NEXT:    addi sp, sp, 16
+; RV32ID-NEXT:    ret
+;
+; CHECK64ZFBFMIN-LABEL: fcvt_bf16_lu:
+; CHECK64ZFBFMIN:       # %bb.0:
+; CHECK64ZFBFMIN-NEXT:    fcvt.s.lu fa5, a0
+; CHECK64ZFBFMIN-NEXT:    fcvt.bf16.s fa0, fa5
+; CHECK64ZFBFMIN-NEXT:    ret
+;
+; RV64ID-LABEL: fcvt_bf16_lu:
+; RV64ID:       # %bb.0:
+; RV64ID-NEXT:    addi sp, sp, -16
+; RV64ID-NEXT:    sd ra, 8(sp) # 8-byte Folded Spill
+; RV64ID-NEXT:    fcvt.s.lu fa0, a0
+; RV64ID-NEXT:    call __truncsfbf2
+; RV64ID-NEXT:    fmv.x.w a0, fa0
+; RV64ID-NEXT:    lui a1, 1048560
+; RV64ID-NEXT:    or a0, a0, a1
+; RV64ID-NEXT:    fmv.w.x fa0, a0
+; RV64ID-NEXT:    ld ra, 8(sp) # 8-byte Folded Reload
+; RV64ID-NEXT:    addi sp, sp, 16
+; RV64ID-NEXT:    ret
+  %1 = uitofp i64 %a to bfloat
+  ret bfloat %1
+}
 
 define bfloat @fcvt_bf16_s(float %a) nounwind {
 ; CHECK32ZFBFMIN-LABEL: fcvt_bf16_s: