[llvm] r257794 - [CodeGen] Don't assume fp_to_fp16 produces i16 when legalizing it.

[Ahmed Bougacha via llvm-commits]

Author: ab
Date: Thu Jan 14 13:45:36 2016
New Revision: 257794

URL: http://llvm.org/viewvc/llvm-project?rev=257794&view=rev
Log:
[CodeGen] Don't assume fp_to_fp16 produces i16 when legalizing it.

Since r230276, we support an improved legalization for f64->f16,
which goes through a temporary f32, improving codegen when
f32->f16 is legal but not f64->f16. This requires unsafe-fp-math.

However, that legalization assumed that the second step, producing
a pseudo-softened f16, had type i16. That's not true on targets
with illegal i16, such as ARM.

Use the initial f64->f16 result type instead.

Modified:
    llvm/trunk/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
    llvm/trunk/test/CodeGen/ARM/fp16.ll

Modified: llvm/trunk/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp?rev=257794&r1=257793&r2=257794&view=diff
==============================================================================
--- llvm/trunk/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp (original)
+++ llvm/trunk/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp Thu Jan 14 13:45:36 2016
@@ -3416,7 +3416,7 @@ bool SelectionDAGLegalize::ExpandNode(SD
         SDValue FloatVal = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Op,
                                        DAG.getIntPtrConstant(0, dl));
         Results.push_back(
-            DAG.getNode(ISD::FP_TO_FP16, dl, MVT::i16, FloatVal));
+            DAG.getNode(ISD::FP_TO_FP16, dl, Node->getValueType(0), FloatVal));
       }
     }
     break;

Modified: llvm/trunk/test/CodeGen/ARM/fp16.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/ARM/fp16.ll?rev=257794&r1=257793&r2=257794&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/ARM/fp16.ll (original)
+++ llvm/trunk/test/CodeGen/ARM/fp16.ll Thu Jan 14 13:45:36 2016
@@ -1,10 +1,16 @@
 ; RUN: llc -mtriple=armv7a--none-eabi < %s | FileCheck --check-prefix=CHECK --check-prefix=CHECK-HARDFLOAT-EABI %s
 ; RUN: llc -mtriple=armv7a--none-gnueabi < %s | FileCheck --check-prefix=CHECK --check-prefix=CHECK-HARDFLOAT-GNU %s
-; RUN: llc -mattr=+vfp3,+fp16 < %s | FileCheck --check-prefix=CHECK --check-prefix=CHECK-FP16 %s
 ; RUN: llc -mtriple=armv8-eabihf < %s | FileCheck --check-prefix=CHECK --check-prefix=CHECK-ARMV8 %s
 ; RUN: llc -mtriple=thumbv7m-eabi < %s | FileCheck --check-prefix=CHECK --check-prefix=CHECK-SOFTFLOAT-EABI %s
 ; RUN: llc -mtriple=thumbv7m-gnueabi < %s | FileCheck --check-prefix=CHECK --check-prefix=CHECK-SOFTFLOAT-GNU %s
 
+;; +fp16 is special: it has f32->f16 (unlike v7), but not f64->f16 (unlike v8).
+;; This exposes unsafe-fp-math optimization opportunities; test that.
+; RUN: llc -mattr=+vfp3,+fp16 < %s |\
+; RUN:   FileCheck --check-prefix=CHECK --check-prefix=CHECK-FP16 --check-prefix=CHECK-FP16-SAFE %s
+; RUN: llc -mattr=+vfp3,+fp16 < %s -enable-unsafe-fp-math |\
+; RUN:   FileCheck --check-prefix=CHECK --check-prefix=CHECK-FP16 --check-prefix=CHECK-FP16-UNSAFE %s
+
 target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-n32"
 target triple = "armv7---eabihf"
 
@@ -76,7 +82,10 @@ define i16 @test_to_fp16(double %in) {
 
 ; CHECK-HARDFLOAT-GNU: bl __aeabi_d2h
 
-; CHECK-FP16: bl __aeabi_d2h
+; CHECK-FP16-SAFE: bl __aeabi_d2h
+
+; CHECK-FP16-UNSAFE:      vcvt.f32.f64 s0, d0
+; CHECK-FP16-UNSAFE-NEXT: vcvtb.f16.f32 s0, s0
 
 ; CHECK-ARMV8: vcvtb.f16.f64 [[TMP:s[0-9]+]], d0
 ; CHECK-ARMV8: vmov r0, [[TMP]]