From b2282508d347bdb7bb5cd21c482de13c4cf90f0f Mon Sep 17 00:00:00 2001
From: Chad Rosier <mcrosier@codeaurora.org>
Date: Tue, 10 Dec 2013 21:33:59 +0000
Subject: [PATCH] [AArch64] Refactor the NEON floating-point absolute
 difference LLVM AArch64 intrinsic to use f32/f64 types, rather than their
 vector equivalents.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@196965 91177308-0d34-0410-b5e6-96231b3b80d8
---
 include/llvm/IR/IntrinsicsAArch64.td     |  4 +++-
 lib/Target/AArch64/AArch64InstrNEON.td   | 13 +++++++++++--
 test/CodeGen/AArch64/neon-scalar-fabd.ll | 14 ++++----------
 3 files changed, 18 insertions(+), 13 deletions(-)

diff --git a/include/llvm/IR/IntrinsicsAArch64.td b/include/llvm/IR/IntrinsicsAArch64.td
index ddbbc7a2e7d..37abbe7db8a 100644
--- a/include/llvm/IR/IntrinsicsAArch64.td
+++ b/include/llvm/IR/IntrinsicsAArch64.td
@@ -304,7 +304,9 @@ def int_aarch64_neon_vabs :
   Intrinsic<[llvm_v1i64_ty], [llvm_v1i64_ty], [IntrNoMem]>;
 
 // Scalar Absolute Difference
-def int_aarch64_neon_vabd : Neon_2Arg_Intrinsic;
+def int_aarch64_neon_vabd :
+  Intrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>, LLVMMatchType<0>],
+            [IntrNoMem]>;
 
 // Scalar Negate Value
 def int_aarch64_neon_vneg :
diff --git a/lib/Target/AArch64/AArch64InstrNEON.td b/lib/Target/AArch64/AArch64InstrNEON.td
index a07bec015d2..1f862d4cff6 100644
--- a/lib/Target/AArch64/AArch64InstrNEON.td
+++ b/lib/Target/AArch64/AArch64InstrNEON.td
@@ -4176,6 +4176,15 @@ multiclass Neon_Scalar3Same_HS_size_patterns<SDPatternOperator opnode,
             (INSTS FPR32:$Rn, FPR32:$Rm)>;
 }
 
+multiclass Neon_Scalar3Same_fabd_SD_size_patterns<SDPatternOperator opnode,
+                                                  Instruction INSTS,
+                                                  Instruction INSTD> {
+  def : Pat<(f32 (opnode (f32 FPR32:$Rn), (f32 FPR32:$Rm))),
+            (INSTS FPR32:$Rn, FPR32:$Rm)>;
+  def : Pat<(f64 (opnode (f64 FPR64:$Rn), (f64 FPR64:$Rm))),
+            (INSTD FPR64:$Rn, FPR64:$Rm)>;
+}
+
 multiclass Neon_Scalar3Same_SD_size_patterns<SDPatternOperator opnode,
                                              Instruction INSTS,
                                              Instruction INSTD> {
@@ -5199,8 +5208,8 @@ defm : Neon_Scalar3Same_cmp_SD_size_patterns<int_aarch64_neon_vcagt,
 
 // Scakar Floating-point Absolute Difference
 defm FABD: NeonI_Scalar3Same_SD_sizes<0b1, 0b1, 0b11010, "fabd">;
-defm : Neon_Scalar3Same_SD_size_patterns<int_aarch64_neon_vabd,
-                                         FABDsss, FABDddd>;
+defm : Neon_Scalar3Same_fabd_SD_size_patterns<int_aarch64_neon_vabd,
+                                              FABDsss, FABDddd>;
 
 // Scalar Absolute Value
 defm ABS : NeonI_Scalar2SameMisc_D_size<0b0, 0b01011, "abs">;
diff --git a/test/CodeGen/AArch64/neon-scalar-fabd.ll b/test/CodeGen/AArch64/neon-scalar-fabd.ll
index 75686d32064..6343310a3c0 100644
--- a/test/CodeGen/AArch64/neon-scalar-fabd.ll
+++ b/test/CodeGen/AArch64/neon-scalar-fabd.ll
@@ -4,10 +4,7 @@ define float @test_vabds_f32(float %a, float %b) {
 ; CHECK-LABEL: test_vabds_f32
 ; CHECK: fabd {{s[0-9]+}}, {{s[0-9]+}}, {{s[0-9]+}}
 entry:
-  %vabd.i = insertelement <1 x float> undef, float %a, i32 0
-  %vabd1.i = insertelement <1 x float> undef, float %b, i32 0
-  %vabd2.i = call <1 x float> @llvm.aarch64.neon.vabd.v1f32(<1 x float> %vabd.i, <1 x float> %vabd1.i)
-  %0 = extractelement <1 x float> %vabd2.i, i32 0
+  %0 = call float @llvm.aarch64.neon.vabd.f32(float %a, float %a)
   ret float %0
 }
 
@@ -15,12 +12,9 @@ define double @test_vabdd_f64(double %a, double %b) {
 ; CHECK-LABEL: test_vabdd_f64
 ; CHECK: fabd {{d[0-9]+}}, {{d[0-9]+}}, {{d[0-9]+}}
 entry:
-  %vabd.i = insertelement <1 x double> undef, double %a, i32 0
-  %vabd1.i = insertelement <1 x double> undef, double %b, i32 0
-  %vabd2.i = call <1 x double> @llvm.aarch64.neon.vabd.v1f64(<1 x double> %vabd.i, <1 x double> %vabd1.i)
-  %0 = extractelement <1 x double> %vabd2.i, i32 0
+  %0 = call double @llvm.aarch64.neon.vabd.f64(double %a, double %b)
   ret double %0
 }
 
-declare <1 x double> @llvm.aarch64.neon.vabd.v1f64(<1 x double>, <1 x double>)
-declare <1 x float> @llvm.aarch64.neon.vabd.v1f32(<1 x float>, <1 x float>)
+declare double @llvm.aarch64.neon.vabd.f64(double, double)
+declare float @llvm.aarch64.neon.vabd.f32(float, float)
-- 
2.34.1