Add DAG optimisation for FP16_TO_FP

The FP16_TO_FP node only uses the bottom 16 bits of its input, so the
following pattern can be optimised by removing the AND:

  (FP16_TO_FP (AND op, 0xffff)) -> (FP16_TO_FP op)

This is a common pattern for ARM targets when functions have __fp16
arguments, as they are passed as floats (so that they get passed in the
correct registers), but then bitcast and truncated to ignore the top 16
bits.

llvm-svn: 245832

2 files changed, 57 insertions, 0 deletions

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 607c3bf2336..bd02b1e156b 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -313,6 +313,7 @@ namespace {
     SDValue visitMGATHER(SDNode *N);
     SDValue visitMSCATTER(SDNode *N);
     SDValue visitFP_TO_FP16(SDNode *N);
+    SDValue visitFP16_TO_FP(SDNode *N);
 
     SDValue visitFADDForFMACombine(SDNode *N);
     SDValue visitFSUBForFMACombine(SDNode *N);
@@ -1411,6 +1412,7 @@ SDValue DAGCombiner::visit(SDNode *N) {
   case ISD::MSCATTER:           return visitMSCATTER(N);
   case ISD::MSTORE:             return visitMSTORE(N);
   case ISD::FP_TO_FP16:         return visitFP_TO_FP16(N);
+  case ISD::FP16_TO_FP:         return visitFP16_TO_FP(N);
   }
   return SDValue();
 }
@@ -13122,6 +13124,21 @@ SDValue DAGCombiner::visitFP_TO_FP16(SDNode *N) {
   return SDValue();
 }
 
+SDValue DAGCombiner::visitFP16_TO_FP(SDNode *N) {
+  SDValue N0 = N->getOperand(0);
+
+  // fold fp16_to_fp(op & 0xffff) -> fp16_to_fp(op)
+  if (N0->getOpcode() == ISD::AND) {
+    ConstantSDNode *AndConst = getAsNonOpaqueConstant(N0.getOperand(1));
+    if (AndConst && AndConst->getAPIntValue() == 0xffff) {
+      return DAG.getNode(ISD::FP16_TO_FP, SDLoc(N), N->getValueType(0),
+                         N0.getOperand(0));
+    }
+  }
+
+  return SDValue();
+}
+
 /// Returns a vector_shuffle if it able to transform an AND to a vector_shuffle
 /// with the destination vector and a zero vector.
 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
diff --git a/llvm/test/CodeGen/ARM/fp16-args.ll b/llvm/test/CodeGen/ARM/fp16-args.ll
new file mode 100644
index 00000000000..31a20f85483
--- /dev/null
+++ b/llvm/test/CodeGen/ARM/fp16-args.ll
@@ -0,0 +1,40 @@
+; RUN: llc -float-abi soft -mattr=+fp16 < %s | FileCheck %s --check-prefix=CHECK --check-prefix=SOFT
+; RUN: llc -float-abi hard -mattr=+fp16 < %s | FileCheck %s --check-prefix=CHECK --check-prefix=HARD
+
+target datalayout = "e-m:e-p:32:32-i64:64-v128:64:128-a:0:32-n32-S64"
+target triple = "armv7a--none-eabi"
+
+define float @foo(float %a.coerce, float %b.coerce) {
+entry:
+  %0 = bitcast float %a.coerce to i32
+  %tmp.0.extract.trunc = trunc i32 %0 to i16
+  %1 = bitcast i16 %tmp.0.extract.trunc to half
+  %2 = bitcast float %b.coerce to i32
+  %tmp1.0.extract.trunc = trunc i32 %2 to i16
+  %3 = bitcast i16 %tmp1.0.extract.trunc to half
+  %4 = fadd half %1, %3
+  %5 = bitcast half %4 to i16
+  %tmp5.0.insert.ext = zext i16 %5 to i32
+  %6 = bitcast i32 %tmp5.0.insert.ext to float
+  ret float %6
+; CHECK: foo:
+
+; SOFT: vmov    {{s[0-9]+}}, r1
+; SOFT: vmov    {{s[0-9]+}}, r0
+; SOFT: vcvtb.f32.f16   {{s[0-9]+}}, {{s[0-9]+}}
+; SOFT: vcvtb.f32.f16   {{s[0-9]+}}, {{s[0-9]+}}
+; SOFT: vadd.f32        {{s[0-9]+}}, {{s[0-9]+}}, {{s[0-9]+}}
+; SOFT: vcvtb.f16.f32   {{s[0-9]+}}, {{s[0-9]+}}
+; SOFT: vmov    r0, {{s[0-9]+}}
+
+; HARD-NOT: vmov
+; HARD-NOT: uxth
+; HARD: vcvtb.f32.f16   {{s[0-9]+}}, s1
+; HARD: vcvtb.f32.f16   {{s[0-9]+}}, s0
+; HARD: vadd.f32        {{s[0-9]+}}, {{s[0-9]+}}, {{s[0-9]+}}
+; HARD: vcvtb.f16.f32   s0, {{s[0-9]+}}
+; HARD-NOT: vmov
+; HARD-NOT: uxth
+
+; CHECK: bx lr
+}