Merge pull request eclipse-omr#7449 from knn-k/aarch64trn

Add vector trn instructions
dmitripivkine · Aug 27, 2024 · 35b34b4 · 35b34b4
2 parents dc1ec39 + 20448ec
commit 35b34b4
Show file tree

Hide file tree

Showing 4 changed files with 42 additions and 0 deletions.
diff --git a/compiler/aarch64/codegen/ARM64Debug.cpp b/compiler/aarch64/codegen/ARM64Debug.cpp
@@ -921,6 +921,10 @@ static const char *opCodeToNameMap[] =
    "vuzp2_8h",
    "vuzp2_4s",
    "vuzp2_2d",
+   "vtrn1_8b",
+   "vtrn1_16b",
+   "vtrn2_8b",
+   "vtrn2_16b",
    "vext16b",
    "vneg16b",
    "vneg8h",

diff --git a/compiler/aarch64/codegen/OMRInstOpCode.enum b/compiler/aarch64/codegen/OMRInstOpCode.enum
@@ -906,6 +906,10 @@
 		vuzp2_8h,                                                 	/* 0x4E405800	UZP2      	 */
 		vuzp2_4s,                                                 	/* 0x4E805800	UZP2      	 */
 		vuzp2_2d,                                                 	/* 0x4EC05800	UZP2      	 */
+		vtrn1_8b,                                               	/* 0x0E002800	TRN1      	 */
+		vtrn1_16b,                                              	/* 0x4E002800	TRN1      	 */
+		vtrn2_8b,                                               	/* 0x0E006800	TRN2      	 */
+		vtrn2_16b,                                              	/* 0x4E006800	TRN2      	 */
 	/* Vector extract */
 		vext16b,                                                  	/* 0x6E000000	EXT      	 */
 	/* Vector Data-processing (1 source) */

diff --git a/compiler/aarch64/codegen/OpBinary.cpp b/compiler/aarch64/codegen/OpBinary.cpp
@@ -907,6 +907,10 @@ const OMR::ARM64::InstOpCode::OpCodeBinaryEntry OMR::ARM64::InstOpCode::binaryEn
 		0x4E405800,	/* UZP2      	vuzp2_8h */
 		0x4E805800,	/* UZP2      	vuzp2_4s */
 		0x4EC05800,	/* UZP2      	vuzp2_2d */
+		0x0E002800,	/* TRN1      	vtrn1_8b */
+		0x4E002800,	/* TRN1      	vtrn1_16b */
+		0x0E006800,	/* TRN2      	vtrn2_8b */
+		0x4E006800,	/* TRN2      	vtrn2_16b */
 	/* Vector extract */
 		0x6E000000,	/* EXT       	vext16b	 */
 	/* Vector Data-processing (1 source) */

diff --git a/fvtest/compilerunittest/aarch64/BinaryEncoder.cpp b/fvtest/compilerunittest/aarch64/BinaryEncoder.cpp
@@ -1636,6 +1636,36 @@ INSTANTIATE_TEST_CASE_P(VectorUnzip2, ARM64Trg1Src2EncodingTest, ::testing::Valu
     std::make_tuple(TR::InstOpCode::vuzp2_2d, TR::RealRegister::v0, TR::RealRegister::v0, TR::RealRegister::v31, "4edf5800")
 ));
 
+INSTANTIATE_TEST_CASE_P(VectorTrn1, ARM64Trg1Src2EncodingTest, ::testing::Values(
+    std::make_tuple(TR::InstOpCode::vtrn1_8b, TR::RealRegister::v15, TR::RealRegister::v0, TR::RealRegister::v0, "0e00280f"),
+    std::make_tuple(TR::InstOpCode::vtrn1_8b, TR::RealRegister::v31, TR::RealRegister::v0, TR::RealRegister::v0, "0e00281f"),
+    std::make_tuple(TR::InstOpCode::vtrn1_8b, TR::RealRegister::v0, TR::RealRegister::v15, TR::RealRegister::v0, "0e0029e0"),
+    std::make_tuple(TR::InstOpCode::vtrn1_8b, TR::RealRegister::v0, TR::RealRegister::v31, TR::RealRegister::v0, "0e002be0"),
+    std::make_tuple(TR::InstOpCode::vtrn1_8b, TR::RealRegister::v0, TR::RealRegister::v0, TR::RealRegister::v15, "0e0f2800"),
+    std::make_tuple(TR::InstOpCode::vtrn1_8b, TR::RealRegister::v0, TR::RealRegister::v0, TR::RealRegister::v31, "0e1f2800"),
+    std::make_tuple(TR::InstOpCode::vtrn1_16b, TR::RealRegister::v15, TR::RealRegister::v0, TR::RealRegister::v0, "4e00280f"),
+    std::make_tuple(TR::InstOpCode::vtrn1_16b, TR::RealRegister::v31, TR::RealRegister::v0, TR::RealRegister::v0, "4e00281f"),
+    std::make_tuple(TR::InstOpCode::vtrn1_16b, TR::RealRegister::v0, TR::RealRegister::v15, TR::RealRegister::v0, "4e0029e0"),
+    std::make_tuple(TR::InstOpCode::vtrn1_16b, TR::RealRegister::v0, TR::RealRegister::v31, TR::RealRegister::v0, "4e002be0"),
+    std::make_tuple(TR::InstOpCode::vtrn1_16b, TR::RealRegister::v0, TR::RealRegister::v0, TR::RealRegister::v15, "4e0f2800"),
+    std::make_tuple(TR::InstOpCode::vtrn1_16b, TR::RealRegister::v0, TR::RealRegister::v0, TR::RealRegister::v31, "4e1f2800")
+));
+
+INSTANTIATE_TEST_CASE_P(VectorTrn2, ARM64Trg1Src2EncodingTest, ::testing::Values(
+    std::make_tuple(TR::InstOpCode::vtrn2_8b, TR::RealRegister::v15, TR::RealRegister::v0, TR::RealRegister::v0, "0e00680f"),
+    std::make_tuple(TR::InstOpCode::vtrn2_8b, TR::RealRegister::v31, TR::RealRegister::v0, TR::RealRegister::v0, "0e00681f"),
+    std::make_tuple(TR::InstOpCode::vtrn2_8b, TR::RealRegister::v0, TR::RealRegister::v15, TR::RealRegister::v0, "0e0069e0"),
+    std::make_tuple(TR::InstOpCode::vtrn2_8b, TR::RealRegister::v0, TR::RealRegister::v31, TR::RealRegister::v0, "0e006be0"),
+    std::make_tuple(TR::InstOpCode::vtrn2_8b, TR::RealRegister::v0, TR::RealRegister::v0, TR::RealRegister::v15, "0e0f6800"),
+    std::make_tuple(TR::InstOpCode::vtrn2_8b, TR::RealRegister::v0, TR::RealRegister::v0, TR::RealRegister::v31, "0e1f6800"),
+    std::make_tuple(TR::InstOpCode::vtrn2_16b, TR::RealRegister::v15, TR::RealRegister::v0, TR::RealRegister::v0, "4e00680f"),
+    std::make_tuple(TR::InstOpCode::vtrn2_16b, TR::RealRegister::v31, TR::RealRegister::v0, TR::RealRegister::v0, "4e00681f"),
+    std::make_tuple(TR::InstOpCode::vtrn2_16b, TR::RealRegister::v0, TR::RealRegister::v15, TR::RealRegister::v0, "4e0069e0"),
+    std::make_tuple(TR::InstOpCode::vtrn2_16b, TR::RealRegister::v0, TR::RealRegister::v31, TR::RealRegister::v0, "4e006be0"),
+    std::make_tuple(TR::InstOpCode::vtrn2_16b, TR::RealRegister::v0, TR::RealRegister::v0, TR::RealRegister::v15, "4e0f6800"),
+    std::make_tuple(TR::InstOpCode::vtrn2_16b, TR::RealRegister::v0, TR::RealRegister::v0, TR::RealRegister::v31, "4e1f6800")
+));
+
 INSTANTIATE_TEST_CASE_P(VectorUMLAL, ARM64Trg1Src2EncodingTest, ::testing::Values(
     std::make_tuple(TR::InstOpCode::vumlal_8h, TR::RealRegister::v15, TR::RealRegister::v0, TR::RealRegister::v0, "2e20800f"),
     std::make_tuple(TR::InstOpCode::vumlal_8h, TR::RealRegister::v31, TR::RealRegister::v0, TR::RealRegister::v0, "2e20801f"),