Implement {Integer|Long}.remainderUnsigned intrinsic

Implement intrinsic for RISC-V target

Integer.remainderUnsigned: +13.71%
Long.remainderUnsigned:    +1623.19%

Test: testrunner.py --target --64 --ndebug --optimizing
Change-Id: I4c3c47bad8fcbdde723e68d19aad45be0d4ff8bb

5 files changed, 38 insertions, 6 deletions

diff --git a/compiler/optimizing/code_generator_arm64.h b/compiler/optimizing/code_generator_arm64.h
index 78049c5675..04694ece3e 100644
--- a/compiler/optimizing/code_generator_arm64.h
+++ b/compiler/optimizing/code_generator_arm64.h
@@ -124,6 +124,8 @@ const vixl::aarch64::CPURegList callee_saved_fp_registers(vixl::aarch64::CPURegi
 Location ARM64ReturnLocation(DataType::Type return_type);
 
 #define UNIMPLEMENTED_INTRINSIC_LIST_ARM64(V) \
+  V(IntegerRemainderUnsigned)                 \
+  V(LongRemainderUnsigned)                    \
   V(StringStringIndexOf)                      \
   V(StringStringIndexOfAfter)                 \
   V(StringBufferAppend)                       \
diff --git a/compiler/optimizing/code_generator_arm_vixl.h b/compiler/optimizing/code_generator_arm_vixl.h
index 51bee1cd77..05e5bdf2b7 100644
--- a/compiler/optimizing/code_generator_arm_vixl.h
+++ b/compiler/optimizing/code_generator_arm_vixl.h
@@ -124,6 +124,8 @@ using VIXLUInt32Literal = vixl::aarch32::Literal<uint32_t>;
   V(UnsafeCASLong)   /* High register pressure */                          \
   V(SystemArrayCopyChar)                                                   \
   V(LongDivideUnsigned)                                                    \
+  V(IntegerRemainderUnsigned)                                              \
+  V(LongRemainderUnsigned)                                                 \
   V(CRC32Update)                                                           \
   V(CRC32UpdateBytes)                                                      \
   V(CRC32UpdateByteBuffer)                                                 \
diff --git a/compiler/optimizing/code_generator_x86.h b/compiler/optimizing/code_generator_x86.h
index 93f8e6ed9b..6ce0c506a0 100644
--- a/compiler/optimizing/code_generator_x86.h
+++ b/compiler/optimizing/code_generator_x86.h
@@ -55,6 +55,8 @@ static constexpr size_t kRuntimeParameterFpuRegistersLength =
   V(IntegerHighestOneBit)                   \
   V(LongHighestOneBit)                      \
   V(LongDivideUnsigned)                     \
+  V(IntegerRemainderUnsigned)               \
+  V(LongRemainderUnsigned)                  \
   V(CRC32Update)                            \
   V(CRC32UpdateBytes)                       \
   V(CRC32UpdateByteBuffer)                  \
diff --git a/compiler/optimizing/code_generator_x86_64.h b/compiler/optimizing/code_generator_x86_64.h
index 7a5e4aa894..81c8ead32e 100644
--- a/compiler/optimizing/code_generator_x86_64.h
+++ b/compiler/optimizing/code_generator_x86_64.h
@@ -69,6 +69,8 @@ static constexpr FloatRegister non_volatile_xmm_regs[] = { XMM12, XMM13, XMM14,
   V(FP16Compare)                               \
   V(FP16Min)                                   \
   V(FP16Max)                                   \
+  V(IntegerRemainderUnsigned)                  \
+  V(LongRemainderUnsigned)                     \
   V(StringStringIndexOf)                       \
   V(StringStringIndexOfAfter)                  \
   V(StringBufferAppend)                        \
diff --git a/compiler/optimizing/intrinsics_riscv64.cc b/compiler/optimizing/intrinsics_riscv64.cc
index 0a9ac872db..b71232b4e4 100644
--- a/compiler/optimizing/intrinsics_riscv64.cc
+++ b/compiler/optimizing/intrinsics_riscv64.cc
@@ -595,7 +595,7 @@ void IntrinsicCodeGeneratorRISCV64::VisitLongNumberOfTrailingZeros(HInvoke* invo
   EmitIntegralUnOp(invoke, [&](XRegister rd, XRegister rs1) { __ Ctz(rd, rs1); });
 }
 
-static void GenerateDivideUnsigned(HInvoke* invoke, CodeGeneratorRISCV64* codegen) {
+static void GenerateDivRemUnsigned(HInvoke* invoke, bool is_div, CodeGeneratorRISCV64* codegen) {
   LocationSummary* locations = invoke->GetLocations();
   Riscv64Assembler* assembler = codegen->GetAssembler();
   DataType::Type type = invoke->GetType();
@@ -611,10 +611,18 @@ static void GenerateDivideUnsigned(HInvoke* invoke, CodeGeneratorRISCV64* codege
   codegen->AddSlowPath(slow_path);
   __ Beqz(divisor, slow_path->GetEntryLabel());
 
-  if (type == DataType::Type::kInt32) {
-    __ Divuw(out, dividend, divisor);
+  if (is_div) {
+    if (type == DataType::Type::kInt32) {
+      __ Divuw(out, dividend, divisor);
+    } else {
+      __ Divu(out, dividend, divisor);
+    }
   } else {
-    __ Divu(out, dividend, divisor);
+    if (type == DataType::Type::kInt32) {
+      __ Remuw(out, dividend, divisor);
+    } else {
+      __ Remu(out, dividend, divisor);
+    }
   }
 
   __ Bind(slow_path->GetExitLabel());
@@ -625,7 +633,7 @@ void IntrinsicLocationsBuilderRISCV64::VisitIntegerDivideUnsigned(HInvoke* invok
 }
 
 void IntrinsicCodeGeneratorRISCV64::VisitIntegerDivideUnsigned(HInvoke* invoke) {
-  GenerateDivideUnsigned(invoke, codegen_);
+  GenerateDivRemUnsigned(invoke, /*is_div=*/true, codegen_);
 }
 
 void IntrinsicLocationsBuilderRISCV64::VisitLongDivideUnsigned(HInvoke* invoke) {
@@ -633,7 +641,23 @@ void IntrinsicLocationsBuilderRISCV64::VisitLongDivideUnsigned(HInvoke* invoke)
 }
 
 void IntrinsicCodeGeneratorRISCV64::VisitLongDivideUnsigned(HInvoke* invoke) {
-  GenerateDivideUnsigned(invoke, codegen_);
+  GenerateDivRemUnsigned(invoke, /*is_div=*/true, codegen_);
+}
+
+void IntrinsicLocationsBuilderRISCV64::VisitIntegerRemainderUnsigned(HInvoke* invoke) {
+  CreateIntIntToIntSlowPathCallLocations(allocator_, invoke);
+}
+
+void IntrinsicCodeGeneratorRISCV64::VisitIntegerRemainderUnsigned(HInvoke* invoke) {
+  GenerateDivRemUnsigned(invoke, /*is_div=*/false, codegen_);
+}
+
+void IntrinsicLocationsBuilderRISCV64::VisitLongRemainderUnsigned(HInvoke* invoke) {
+  CreateIntIntToIntSlowPathCallLocations(allocator_, invoke);
+}
+
+void IntrinsicCodeGeneratorRISCV64::VisitLongRemainderUnsigned(HInvoke* invoke) {
+  GenerateDivRemUnsigned(invoke, /*is_div=*/false, codegen_);
 }
 
 #define VISIT_INTRINSIC(name, low, high, type, start_index)                              \