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

[Phoebe Wang via llvm-commits]

https://github.com/phoebewang created https://github.com/llvm/llvm-project/pull/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.

>From cf39af6bab4bd5c2b0579adea610aba0d3e8271b Mon Sep 17 00:00:00 2001
From: "Wang, Phoebe" <phoebe.wang at intel.com>
Date: Wed, 18 Sep 2024 17:39:59 +0800
Subject: [PATCH] [X86][BF16] Add libcall for F80 -> BF16

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.
---
 llvm/include/llvm/IR/RuntimeLibcalls.def |  1 +
 llvm/lib/CodeGen/TargetLoweringBase.cpp  |  2 ++
 llvm/test/CodeGen/X86/bfloat.ll          | 22 ++++++++++++++++++++++
 3 files changed, 25 insertions(+)

diff --git a/llvm/include/llvm/IR/RuntimeLibcalls.def b/llvm/include/llvm/IR/RuntimeLibcalls.def
index c3d5ef9f4e4f82..69cf43140ad4bd 100644
--- a/llvm/include/llvm/IR/RuntimeLibcalls.def
+++ b/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")
diff --git a/llvm/lib/CodeGen/TargetLoweringBase.cpp b/llvm/lib/CodeGen/TargetLoweringBase.cpp
index eb3190c7cd247a..9fdde454559171 100644
--- a/llvm/lib/CodeGen/TargetLoweringBase.cpp
+++ b/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;
diff --git a/llvm/test/CodeGen/X86/bfloat.ll b/llvm/test/CodeGen/X86/bfloat.ll
index 3759909a2ccc8e..3144fd56d9ccf3 100644
--- a/llvm/test/CodeGen/X86/bfloat.ll
+++ b/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 at PLT
+; CHECK-NEXT:    addq $24, %rsp
+; CHECK-NEXT:    retq
+  %2 = fptrunc x86_fp80 %0 to bfloat
+  ret bfloat %2
+}