From ff4027d152d0d928540dbc19a41840a3a59ff23d Mon Sep 17 00:00:00 2001
From: Archibald Elliott <archibald.elliott@arm.com>
Date: Fri, 23 Sep 2022 18:41:58 +0100
Subject: [PATCH] [ARM] Support fp16/bf16 using t constraint

fp16 and bf16 values can be used in GCC's inline assembly using the "t"
constraint, which means "VFP floating-point registers s0-s31" - fp16 and
bf16 values are stored in S registers too.

This change ensures that LLVM is compatible with GCC for programs that
use fp16 and the 't' constraint.

Fixes #57753

Differential Revision: https://reviews.llvm.org/D134553
---
 llvm/lib/Target/ARM/ARMISelLowering.cpp    |  2 ++
 llvm/test/CodeGen/ARM/inlineasm-fp-half.ll | 32 ++++++++++++++++++++++
 2 files changed, 34 insertions(+)
 create mode 100644 llvm/test/CodeGen/ARM/inlineasm-fp-half.ll

diff --git a/llvm/lib/Target/ARM/ARMISelLowering.cpp b/llvm/lib/Target/ARM/ARMISelLowering.cpp
index 544fd0efd05b..68bc71b28585 100644
--- a/llvm/lib/Target/ARM/ARMISelLowering.cpp
+++ b/llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -20163,6 +20163,8 @@ RCPair ARMTargetLowering::getRegForInlineAsmConstraint(
     case 't':
       if (VT == MVT::Other)
         break;
+      if (VT == MVT::f16 || VT == MVT::bf16)
+        return RCPair(0U, &ARM::HPRRegClass);
       if (VT == MVT::f32 || VT == MVT::i32)
         return RCPair(0U, &ARM::SPRRegClass);
       if (VT.getSizeInBits() == 64)
diff --git a/llvm/test/CodeGen/ARM/inlineasm-fp-half.ll b/llvm/test/CodeGen/ARM/inlineasm-fp-half.ll
new file mode 100644
index 000000000000..b2a894f49c52
--- /dev/null
+++ b/llvm/test/CodeGen/ARM/inlineasm-fp-half.ll
@@ -0,0 +1,32 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=arm -mattr=+armv8.2-a,+fp-armv8,+fullfp16,+bf16,-neon %s -o - | FileCheck %s
+; RUN: llc -mtriple=thumb -mattr=+armv8.2-a,+fp-armv8,+fullfp16,+bf16,-neon %s -o - | FileCheck %s
+
+
+define arm_aapcscc half @f(half %x) nounwind {
+; CHECK-LABEL: f:
+; CHECK:       @ %bb.0: @ %entry
+; CHECK-NEXT:    vmov.f16 s0, r0
+; CHECK-NEXT:    @APP
+; CHECK-NEXT:    vsqrt.f16 s0, s0
+; CHECK-NEXT:    @NO_APP
+; CHECK-NEXT:    vmov r0, s0
+; CHECK-NEXT:    bx lr
+entry:
+  %0 = tail call half asm "vsqrt.f16 $0, $1", "=t,t"(half %x)
+  ret half %0
+}
+
+define arm_aapcscc bfloat @h(bfloat %x) nounwind {
+; CHECK-LABEL: h:
+; CHECK:       @ %bb.0: @ %entry
+; CHECK-NEXT:    vmov.f16 s0, r0
+; CHECK-NEXT:    @APP
+; CHECK-NEXT:    vmov.f32 s0, s0
+; CHECK-NEXT:    @NO_APP
+; CHECK-NEXT:    vmov.f16 r0, s0
+; CHECK-NEXT:    bx lr
+entry:
+  %0 = tail call bfloat asm "vmov.f32 $0, $1", "=t,t"(bfloat %x)
+  ret bfloat %0
+}