[HEXAGON] Add support to lower "FREEZE a half(f16)" instruction on Hexagon and fix the isel-buildvector-v2f16.ll assertion

[quic-asaravan]

No description provided.

[llvmbot]

@llvm/pr-subscribers-backend-hexagon

Author: Abinaya Saravanan (quic-asaravan)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/130977.diff

3 Files Affected:

• (modified) llvm/lib/Target/Hexagon/HexagonISelLowering.h (+1)
• (modified) llvm/lib/Target/Hexagon/HexagonISelLoweringHVX.cpp (+11-11)
• (added) llvm/test/CodeGen/Hexagon/fp16-promote.ll (+44)

diff --git a/llvm/lib/Target/Hexagon/HexagonISelLowering.h b/llvm/lib/Target/Hexagon/HexagonISelLowering.h
index aaa9c65c1e07e..4df88b3a8abd7 100644
--- a/llvm/lib/Target/Hexagon/HexagonISelLowering.h
+++ b/llvm/lib/Target/Hexagon/HexagonISelLowering.h
@@ -362,6 +362,7 @@ class HexagonTargetLowering : public TargetLowering {
   shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override {
     return AtomicExpansionKind::LLSC;
   }
+  bool softPromoteHalfType() const override { return true; }
 
 private:
   void initializeHVXLowering();
diff --git a/llvm/lib/Target/Hexagon/HexagonISelLoweringHVX.cpp b/llvm/lib/Target/Hexagon/HexagonISelLoweringHVX.cpp
index f77bda3138354..de2bac6905530 100644
--- a/llvm/lib/Target/Hexagon/HexagonISelLoweringHVX.cpp
+++ b/llvm/lib/Target/Hexagon/HexagonISelLoweringHVX.cpp
@@ -1618,17 +1618,6 @@ HexagonTargetLowering::LowerHvxBuildVector(SDValue Op, SelectionDAG &DAG)
   for (unsigned i = 0; i != Size; ++i)
     Ops.push_back(Op.getOperand(i));
 
-  // First, split the BUILD_VECTOR for vector pairs. We could generate
-  // some pairs directly (via splat), but splats should be generated
-  // by the combiner prior to getting here.
-  if (VecTy.getSizeInBits() == 16*Subtarget.getVectorLength()) {
-    ArrayRef<SDValue> A(Ops);
-    MVT SingleTy = typeSplit(VecTy).first;
-    SDValue V0 = buildHvxVectorReg(A.take_front(Size/2), dl, SingleTy, DAG);
-    SDValue V1 = buildHvxVectorReg(A.drop_front(Size/2), dl, SingleTy, DAG);
-    return DAG.getNode(ISD::CONCAT_VECTORS, dl, VecTy, V0, V1);
-  }
-
   if (VecTy.getVectorElementType() == MVT::i1)
     return buildHvxVectorPred(Ops, dl, VecTy, DAG);
 
@@ -1645,6 +1634,17 @@ HexagonTargetLowering::LowerHvxBuildVector(SDValue Op, SelectionDAG &DAG)
     return DAG.getBitcast(tyVector(VecTy, MVT::f16), T0);
   }
 
+  // First, split the BUILD_VECTOR for vector pairs. We could generate
+  // some pairs directly (via splat), but splats should be generated
+  // by the combiner prior to getting here.
+  if (VecTy.getSizeInBits() == 16 * Subtarget.getVectorLength()) {
+    ArrayRef<SDValue> A(Ops);
+    MVT SingleTy = typeSplit(VecTy).first;
+    SDValue V0 = buildHvxVectorReg(A.take_front(Size / 2), dl, SingleTy, DAG);
+    SDValue V1 = buildHvxVectorReg(A.drop_front(Size / 2), dl, SingleTy, DAG);
+    return DAG.getNode(ISD::CONCAT_VECTORS, dl, VecTy, V0, V1);
+  }
+
   return buildHvxVectorReg(Ops, dl, VecTy, DAG);
 }
 
diff --git a/llvm/test/CodeGen/Hexagon/fp16-promote.ll b/llvm/test/CodeGen/Hexagon/fp16-promote.ll
new file mode 100644
index 0000000000000..1ef0a133ce30a
--- /dev/null
+++ b/llvm/test/CodeGen/Hexagon/fp16-promote.ll
@@ -0,0 +1,44 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -march=hexagon  < %s | FileCheck %s
+
+define half @freeze_half_undef() nounwind {
+; CHECK-LABEL: freeze_half_undef:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    {
+; CHECK-NEXT:     call __truncsfhf2
+; CHECK-NEXT:     r0 = #0
+; CHECK-NEXT:     allocframe(#0)
+; CHECK-NEXT:    }
+; CHECK-NEXT:    {
+; CHECK-NEXT:     call __extendhfsf2
+; CHECK-NEXT:    }
+; CHECK-NEXT:    {
+; CHECK-NEXT:     call __truncsfhf2
+; CHECK-NEXT:     r0 = sfadd(r0,r0)
+; CHECK-NEXT:    }
+; CHECK-NEXT:    {
+; CHECK-NEXT:     r31:30 = dealloc_return(r30):raw
+; CHECK-NEXT:    }
+  %y1 = freeze half undef
+  %t1 = fadd half %y1, %y1
+  ret half %t1
+}
+
+define half @freeze_half_poison(half %maybe.poison) {
+; CHECK-LABEL: freeze_half_poison:
+; CHECK:  // %bb.0:
+; CHECK:    {
+; CHECK-NEXT:     call __extendhfsf2
+; CHECK-NEXT:     allocframe(r29,#0):raw
+; CHECK-NEXT:    }
+; CHECK-NEXT:    {
+; CHECK-NEXT:     call __truncsfhf2
+; CHECK-NEXT:     r0 = sfadd(r0,r0)
+; CHECK-NEXT:    }
+; CHECK-NEXT:    {
+; CHECK-NEXT:     r31:30 = dealloc_return(r30):raw
+; CHECK-NEXT:    }
+  %y1 = freeze half %maybe.poison
+  %t1 = fadd half %y1, %y1
+  ret half %t1
+}

[quic-asaravan]

This fixes #117337.

[quic-asaravan]

@iajbar @androm3da Can you please review this ?

[iajbar]

Fixes #107791

[alexrp]

Fixes #107791

You mean #117337 right? LGTM; thanks!

[alexrp]

I suppose this is technically an ABI change, thus precluding a backport to 20.1.1, right?

[androm3da]

@tstellar is this kind of change one that prevents backporting? Or requires a 20.2.x?