From ed839120c4a9e268c52f14b6443589456ec2b008 Mon Sep 17 00:00:00 2001
From: Tim Northover <tnorthover@apple.com>
Date: Tue, 1 Apr 2014 12:22:37 +0000
Subject: [PATCH] ARM64: add intrinsic for pmull (p64 x p64 = p128) operations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205302 91177308-0d34-0410-b5e6-96231b3b80d8
---
 include/llvm/IR/IntrinsicsARM64.td |  8 ++++++--
 lib/Target/ARM64/ARM64InstrInfo.td |  7 +++++++
 test/CodeGen/ARM64/vmul.ll         | 18 ++++++++++++++++++
 3 files changed, 31 insertions(+), 2 deletions(-)

diff --git a/include/llvm/IR/IntrinsicsARM64.td b/include/llvm/IR/IntrinsicsARM64.td
index b280d005d0c..d7f307e9ff6 100644
--- a/include/llvm/IR/IntrinsicsARM64.td
+++ b/include/llvm/IR/IntrinsicsARM64.td
@@ -74,8 +74,7 @@ let TargetPrefix = "arm64" in {  // All intrinsics start with "llvm.arm64.".
                 [IntrNoMem]>;
   class AdvSIMD_2VectorArg_Long_Intrinsic
     : Intrinsic<[llvm_anyvector_ty],
-                [LLVMTruncatedType<0>,
-                 LLVMTruncatedType<0>],
+                [LLVMTruncatedType<0>, LLVMTruncatedType<0>],
                 [IntrNoMem]>;
   class AdvSIMD_2VectorArg_Wide_Intrinsic
     : Intrinsic<[llvm_anyvector_ty],
@@ -178,6 +177,11 @@ let Properties = [IntrNoMem] in {
   def int_arm64_neon_umull : AdvSIMD_2VectorArg_Long_Intrinsic;
   def int_arm64_neon_pmull : AdvSIMD_2VectorArg_Long_Intrinsic;
 
+  // 64-bit polynomial multiply really returns an i128, which is not legal. Fake
+  // it with a v16i8.
+  def int_arm64_neon_pmull64 :
+        Intrinsic<[llvm_v16i8_ty], [llvm_i64_ty, llvm_i64_ty], [IntrNoMem]>;
+
   // Vector Extending Multiply
   def int_arm64_neon_fmulx : AdvSIMD_2FloatArg_Intrinsic;
 
diff --git a/lib/Target/ARM64/ARM64InstrInfo.td b/lib/Target/ARM64/ARM64InstrInfo.td
index c9a714b8dda..ee066a38f4f 100644
--- a/lib/Target/ARM64/ARM64InstrInfo.td
+++ b/lib/Target/ARM64/ARM64InstrInfo.td
@@ -2726,6 +2726,13 @@ defm USUBL   : SIMDLongThreeVectorBHS<1, 0b0010, "usubl",
 defm USUBW   : SIMDWideThreeVectorBHS<   1, 0b0011, "usubw",
                  BinOpFrag<(sub node:$LHS, (zext node:$RHS))>>;
 
+// Patterns for 64-bit pmull
+def : Pat<(int_arm64_neon_pmull64 V64:$Rn, V64:$Rm),
+          (PMULLv1i64 V64:$Rn, V64:$Rm)>;
+def : Pat<(int_arm64_neon_pmull64 (vector_extract (v2i64 V128:$Rn), (i64 1)),
+                                  (vector_extract (v2i64 V128:$Rm), (i64 1))),
+          (PMULLv2i64 V128:$Rn, V128:$Rm)>;
+
 // CodeGen patterns for addhn and subhn instructions, which can actually be
 // written in LLVM IR without too much difficulty.
 
diff --git a/test/CodeGen/ARM64/vmul.ll b/test/CodeGen/ARM64/vmul.ll
index 433c09d37f8..3ef0a76e204 100644
--- a/test/CodeGen/ARM64/vmul.ll
+++ b/test/CodeGen/ARM64/vmul.ll
@@ -1983,3 +1983,21 @@ define i64 @sqdmlsl_d(i32 %A, i32 %B, i64 %C) nounwind {
   %tmp5 = call i64 @llvm.arm64.neon.sqsub.i64(i64 %C, i64 %tmp4)
   ret i64 %tmp5
 }
+
+define <16 x i8> @test_pmull_64(i64 %l, i64 %r) nounwind {
+; CHECK-LABEL: test_pmull_64:
+; CHECK: pmull.1q
+  %val = call <16 x i8> @llvm.arm64.neon.pmull64(i64 %l, i64 %r)
+  ret <16 x i8> %val
+}
+
+define <16 x i8> @test_pmull_high_64(<2 x i64> %l, <2 x i64> %r) nounwind {
+; CHECK-LABEL: test_pmull_high_64:
+; CHECK: pmull2.1q
+  %l_hi = extractelement <2 x i64> %l, i32 1
+  %r_hi = extractelement <2 x i64> %r, i32 1
+  %val = call <16 x i8> @llvm.arm64.neon.pmull64(i64 %l_hi, i64 %r_hi)
+  ret <16 x i8> %val
+}
+
+declare <16 x i8> @llvm.arm64.neon.pmull64(i64, i64)
-- 
2.34.1