Fix ReverseElement8, ReverseElement16, ReverseElement32

dotnet · Mar 9, 2021 · f31aec1 · f31aec1
1 parent 5b49456
commit f31aec1
Show file tree

Hide file tree

Showing 3 changed files with 24 additions and 24 deletions.
diff --git a/src/mono/mono/mini/mini-llvm.c b/src/mono/mono/mini/mini-llvm.c
@@ -255,6 +255,7 @@ mini_llvm_ins_info[] = {
 
 enum {
 	MAX_VECTOR_ELEMS = 32, // 2 vectors * 128 bits per vector / 8 bits per element
+	ARM64_MAX_VECTOR_ELEMS = 16,
 };
 
 static LLVMIntPredicate cond_to_llvm_cond [] = {
@@ -10273,26 +10274,27 @@ process_bb (EmitContext *ctx, MonoBasicBlock *bb)
 			values [ins->dreg] = result;
 			break;
 		}
-		case OP_ARM64_REV16:
-		case OP_ARM64_REV32:
-		case OP_ARM64_REV64: {
-			unsigned int tmp_bits = 0;
-			switch (ins->opcode) {
-			case OP_ARM64_REV16: tmp_bits = 8; break;
-			case OP_ARM64_REV32: tmp_bits = 16; break;
-			case OP_ARM64_REV64: tmp_bits = 32; break;
-			}
+		case OP_ARM64_REVN: {
 			LLVMTypeRef t = LLVMTypeOf (lhs);
-			unsigned int t_bits = mono_llvm_get_prim_size_bits (t);
-			unsigned int tmp_elements = t_bits / tmp_bits;
-			LLVMTypeRef tmp_t = LLVMVectorType (LLVMIntType (tmp_bits), tmp_elements);
-			LLVMValueRef tmp = LLVMBuildBitCast (builder, lhs, tmp_t, "arm64_rev");
-			int mask [MAX_VECTOR_ELEMS] = { 0 };
-			for (unsigned int i = 0; i < tmp_elements; i += 2) {
-				mask [i] = i + 1;
-				mask [i + 1] = i;
+			LLVMTypeRef elem_t = LLVMGetElementType (t);
+			unsigned int group_bits = mono_llvm_get_prim_size_bits (elem_t);
+			unsigned int vec_bits = mono_llvm_get_prim_size_bits (t);
+			unsigned int tmp_bits = ins->inst_c0;
+			unsigned int tmp_elements = vec_bits / tmp_bits;
+			const int cycle8 [] = { 7, 6, 5, 4, 3, 2, 1, 0, 15, 14, 13, 12, 11, 10, 9, 8 };
+			const int cycle4 [] = { 3, 2, 1, 0, 7, 6, 5, 4, 11, 10, 9, 8, 15, 14, 13, 12 };
+			const int cycle2 [] = { 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14 };
+			const int *cycle = NULL;
+			switch (group_bits / tmp_bits) {
+			case 2: cycle = cycle2; break;
+			case 4: cycle = cycle4; break;
+			case 8: cycle = cycle8; break;
+			default: g_assert_not_reached ();
 			}
-			LLVMValueRef result = LLVMBuildShuffleVector (builder, tmp, LLVMGetUndef (tmp_t), create_const_vector_i32 (mask, tmp_elements), "");
+			g_assert (tmp_elements <= ARM64_MAX_VECTOR_ELEMS);
+			LLVMTypeRef tmp_t = LLVMVectorType (LLVMIntType (tmp_bits), tmp_elements);
+			LLVMValueRef tmp = LLVMBuildBitCast (builder, lhs, tmp_t, "arm64_revn");
+			LLVMValueRef result = LLVMBuildShuffleVector (builder, tmp, LLVMGetUndef (tmp_t), create_const_vector_i32 (cycle, tmp_elements), "");
 			result = LLVMBuildBitCast (builder, result, t, "");
 			values [ins->dreg] = result;
 			break;

diff --git a/src/mono/mono/mini/mini-ops.h b/src/mono/mono/mini/mini-ops.h
@@ -1661,9 +1661,7 @@ MINI_OP3(OP_ARM64_XNSHIFT2, "arm64_xnshift2", XREG, XREG, XREG, IREG)
 MINI_OP(OP_ARM64_SQSHLU, "arm64_sqshlu", XREG, XREG, IREG)
 MINI_OP(OP_ARM64_SQSHLU_SCALAR, "arm64_sqshlu_scalar", XREG, XREG, IREG)
 
-MINI_OP(OP_ARM64_REV16, "arm64_rev16", XREG, XREG, NONE)
-MINI_OP(OP_ARM64_REV32, "arm64_rev32", XREG, XREG, NONE)
-MINI_OP(OP_ARM64_REV64, "arm64_rev64", XREG, XREG, NONE)
+MINI_OP(OP_ARM64_REVN, "arm64_revn", XREG, XREG, NONE)
 
 MINI_OP(OP_ARM64_PMULL, "arm64_pmull", XREG, XREG, XREG)
 MINI_OP(OP_ARM64_PMULL2, "arm64_pmull2", XREG, XREG, XREG)

diff --git a/src/mono/mono/mini/simd-intrinsics.c b/src/mono/mono/mini/simd-intrinsics.c
@@ -1268,9 +1268,9 @@ static SimdIntrinsic advsimd_methods [] = {
 	{SN_ReciprocalSquareRootStepScalar, OP_XOP_OVR_SCALAR_X_X_X, INTRINS_AARCH64_ADV_SIMD_FRSQRTS},
 	{SN_ReciprocalStep, OP_XOP_OVR_X_X_X, INTRINS_AARCH64_ADV_SIMD_FRECPS},
 	{SN_ReciprocalStepScalar, OP_XOP_OVR_SCALAR_X_X_X, INTRINS_AARCH64_ADV_SIMD_FRECPS},
-	{SN_ReverseElement16, OP_ARM64_REV32},
-	{SN_ReverseElement32, OP_ARM64_REV64},
-	{SN_ReverseElement8, OP_ARM64_REV16},
+	{SN_ReverseElement16, OP_ARM64_REVN, 16},
+	{SN_ReverseElement32, OP_ARM64_REVN, 32},
+	{SN_ReverseElement8, OP_ARM64_REVN, 8},
 	{SN_ReverseElementBits, OP_XOP_OVR_X_X, INTRINS_AARCH64_ADV_SIMD_RBIT},
 	{SN_RoundAwayFromZero, OP_XOP_OVR_X_X, INTRINS_AARCH64_ADV_SIMD_FRINTA},
 	{SN_RoundAwayFromZeroScalar, OP_XOP_OVR_SCALAR_X_X, INTRINS_AARCH64_ADV_SIMD_FRINTA},