[X86][BF16] Add libcall for F80 -> BF16 (#109116)

This fixes #108936, but the calling convention doesn't match with GCC. I
doubt we have such a lib function for now, so leave the calling
convention as is.

Author: phoebewang

compiler-rt/lib/builtins/CMakeLists.txt

@@ -192,6 +192,7 @@ set(GENERIC_SOURCES
 set(BF16_SOURCES
     extendbfsf2.c
     truncdfbf2.c
+    truncxfbf2.c
     truncsfbf2.c
 )
 

compiler-rt/lib/builtins/truncxfbf2.c

@@ -0,0 +1,13 @@
+//===-- lib/truncxfbf2.c - long double -> bfloat conversion -------*- C -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#define SRC_80
+#define DST_BFLOAT
+#include "fp_trunc_impl.inc"
+
+COMPILER_RT_ABI dst_t __truncxfbf2(long double a) { return __truncXfYf2__(a); }

llvm/include/llvm/IR/RuntimeLibcalls.def

@@ -367,6 +367,7 @@ HANDLE_LIBCALL(FPROUND_F128_F16, "__trunctfhf2")
 HANDLE_LIBCALL(FPROUND_PPCF128_F16, "__trunctfhf2")
 HANDLE_LIBCALL(FPROUND_F32_BF16, "__truncsfbf2")
 HANDLE_LIBCALL(FPROUND_F64_BF16, "__truncdfbf2")
+HANDLE_LIBCALL(FPROUND_F80_BF16, "__truncxfbf2")
 HANDLE_LIBCALL(FPROUND_F64_F32, "__truncdfsf2")
 HANDLE_LIBCALL(FPROUND_F80_F32, "__truncxfsf2")
 HANDLE_LIBCALL(FPROUND_F128_F32, "__trunctfsf2")

llvm/lib/CodeGen/TargetLoweringBase.cpp

@@ -169,6 +169,8 @@ RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) {
       return FPROUND_F32_BF16;
     if (OpVT == MVT::f64)
       return FPROUND_F64_BF16;
+    if (OpVT == MVT::f80)
+      return FPROUND_F80_BF16;
   } else if (RetVT == MVT::f32) {
     if (OpVT == MVT::f64)
       return FPROUND_F64_F32;

llvm/test/CodeGen/X86/bfloat.ll

@@ -1953,3 +1953,25 @@ define void @PR92471(ptr %0, ptr %1) nounwind {
   store <7 x float> %4, ptr %1, align 4
   ret void
 }
+
+define bfloat @PR108936(x86_fp80 %0) nounwind {
+; X86-LABEL: PR108936:
+; X86:       # %bb.0:
+; X86-NEXT:    subl $12, %esp
+; X86-NEXT:    fldt {{[0-9]+}}(%esp)
+; X86-NEXT:    fstpt (%esp)
+; X86-NEXT:    calll __truncxfbf2
+; X86-NEXT:    addl $12, %esp
+; X86-NEXT:    retl
+;
+; CHECK-LABEL: PR108936:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    subq $24, %rsp
+; CHECK-NEXT:    fldt {{[0-9]+}}(%rsp)
+; CHECK-NEXT:    fstpt (%rsp)
+; CHECK-NEXT:    callq __truncxfbf2@PLT
+; CHECK-NEXT:    addq $24, %rsp
+; CHECK-NEXT:    retq
+  %2 = fptrunc x86_fp80 %0 to bfloat
+  ret bfloat %2
+}