[netcore] Implement missing Bmi1/Bmi2 intrinsics (mono/mono#16919)

* Finalize BMI1

* cleanup

* Update simd-intrinsics-netcore.c

* Update simd-intrinsics-netcore.c

* fix LLVM-AOT

Commit migrated from https://github.com/mono/mono/commit/ed622e91125b5846895d6dcca5366b1775b47d6d

diff --git a/src/mono/mono/mini/mini-llvm.c b/src/mono/mono/mini/mini-llvm.c
index f190ef4..a99eefd 100644 (file)
--- a/src/mono/mono/mini/mini-llvm.c
+++ b/src/mono/mono/mini/mini-llvm.c
@@ -332,6 +332,8 @@ typedef enum {
        INTRINS_PDEP_I64,
        INTRINS_BZHI_I32,
        INTRINS_BZHI_I64,
+       INTRINS_BEXTR_I32,
+       INTRINS_BEXTR_I64,
 #if defined(TARGET_AMD64) || defined(TARGET_X86)
        INTRINS_SSE_PMOVMSKB,
        INTRINS_SSE_PSRLI_W,
@@ -7465,6 +7467,14 @@ process_bb (EmitContext *ctx, MonoBasicBlock *bb)
                        values [ins->dreg] = LLVMBuildCall (builder, get_intrins (ctx, ins->opcode == OP_CTTZ32 ? INTRINS_CTTZ_I32 : INTRINS_CTTZ_I64), args, 2, "");
                        break;
                }
+               case OP_BEXTR32:
+               case OP_BEXTR64: {
+                       LLVMValueRef args [2];
+                       args [0] = lhs;
+                       args [1] = convert (ctx, rhs, ins->opcode == OP_BEXTR32 ? LLVMInt32Type () : LLVMInt64Type ()); // cast ushort to u32/u64
+                       values [ins->dreg] = LLVMBuildCall (builder, get_intrins (ctx, ins->opcode == OP_BEXTR32 ? INTRINS_BEXTR_I32 : INTRINS_BEXTR_I64), args, 2, "");
+                       break;
+               }
                case OP_BZHI32:
                case OP_BZHI64: {
                        LLVMValueRef args [2];
@@ -8742,6 +8752,8 @@ static IntrinsicDesc intrinsics[] = {
        {INTRINS_CTTZ_I64, "llvm.cttz.i64"},
        {INTRINS_BZHI_I32, "llvm.x86.bmi.bzhi.32"},
        {INTRINS_BZHI_I64, "llvm.x86.bmi.bzhi.64"},
+       {INTRINS_BEXTR_I32, "llvm.x86.bmi.bextr.32"},
+       {INTRINS_BEXTR_I64, "llvm.x86.bmi.bextr.64"},
        {INTRINS_PEXT_I32, "llvm.x86.bmi.pext.32"},
        {INTRINS_PEXT_I64, "llvm.x86.bmi.pext.64"},
        {INTRINS_PDEP_I32, "llvm.x86.bmi.pdep.32"},
@@ -8951,11 +8963,13 @@ add_intrinsic (LLVMModuleRef module, int id)
        case INTRINS_CTTZ_I64:
                AddFunc2 (module, name, LLVMInt64Type (), LLVMInt64Type (), LLVMInt1Type ());
                break;
+       case INTRINS_BEXTR_I32:
        case INTRINS_BZHI_I32:
        case INTRINS_PEXT_I32:
        case INTRINS_PDEP_I32:
                AddFunc2 (module, name, LLVMInt32Type (), LLVMInt32Type (), LLVMInt32Type ());
                break;
+       case INTRINS_BEXTR_I64:
        case INTRINS_BZHI_I64:
        case INTRINS_PEXT_I64:
        case INTRINS_PDEP_I64:

diff --git a/src/mono/mono/mini/mini-ops.h b/src/mono/mono/mini/mini-ops.h
index 8b1ef68..da2a320 100644 (file)
--- a/src/mono/mono/mini/mini-ops.h
+++ b/src/mono/mono/mini/mini-ops.h
@@ -1007,6 +1007,21 @@ MINI_OP(OP_DPPS, "dpps", XREG, XREG, XREG)
 /* inst_c0 is the rounding mode: 0 = round, 1 = floor, 2 = ceiling */
 MINI_OP(OP_SSE41_ROUNDPD, "roundpd", XREG, XREG, NONE)
 
+/* Intel BMI1 */
+/* Count trailing zeroes, return 32/64 if the input is 0 */
+MINI_OP(OP_CTTZ32, "cttz32", IREG, IREG, NONE)
+MINI_OP(OP_CTTZ64, "cttz64", LREG, LREG, NONE)
+MINI_OP(OP_BEXTR32, "bextr32", IREG, IREG, IREG)
+MINI_OP(OP_BEXTR64, "bextr64", LREG, LREG, LREG)
+
+/* Intel BMI2 */
+MINI_OP(OP_BZHI32, "bzhi32", IREG, IREG, IREG)
+MINI_OP(OP_BZHI64, "bzhi64", LREG, LREG, LREG)
+MINI_OP(OP_PEXT32, "pext32", IREG, IREG, IREG)
+MINI_OP(OP_PEXT64, "pext64", LREG, LREG, LREG)
+MINI_OP(OP_PDEP32, "pdep32", IREG, IREG, IREG)
+MINI_OP(OP_PDEP64, "pdep64", LREG, LREG, LREG)
+
 #endif
 
 MINI_OP(OP_XMOVE,   "xmove", XREG, XREG, NONE)
@@ -1401,16 +1416,4 @@ MINI_OP(OP_LZCNT64, "lzcnt64", LREG, LREG, NONE)
 MINI_OP(OP_POPCNT32, "popcnt32", IREG, IREG, NONE)
 MINI_OP(OP_POPCNT64, "popcnt64", LREG, LREG, NONE)
 
-/* Intel BMI1 */
-/* Count trailing zeroes, return 32/64 if the input is 0 */
-MINI_OP(OP_CTTZ32, "cttz32", IREG, IREG, NONE)
-MINI_OP(OP_CTTZ64, "cttz64", LREG, LREG, NONE)
-
-/* Intel BMI2 */
-MINI_OP(OP_BZHI32, "bzhi32", IREG, IREG, IREG)
-MINI_OP(OP_BZHI64, "bzhi64", LREG, LREG, LREG)
-MINI_OP(OP_PEXT32, "pext32", IREG, IREG, IREG)
-MINI_OP(OP_PEXT64, "pext64", LREG, LREG, LREG)
-MINI_OP(OP_PDEP32, "pdep32", IREG, IREG, IREG)
-MINI_OP(OP_PDEP64, "pdep64", LREG, LREG, LREG)
 

diff --git a/src/mono/mono/mini/simd-intrinsics-netcore.c b/src/mono/mono/mini/simd-intrinsics-netcore.c
index 9aa1813..dd9620b 100644 (file)
--- a/src/mono/mono/mini/simd-intrinsics-netcore.c
+++ b/src/mono/mono/mini/simd-intrinsics-netcore.c
@@ -552,6 +552,8 @@ static guint16 lzcnt_methods [] = {
 };
 
 static guint16 bmi1_methods [] = {
+       SN_AndNot,
+       SN_BitFieldExtract,
        SN_ExtractLowestSetBit,
        SN_GetMaskUpToLowestSetBit,
        SN_ResetLowestSetBit,
@@ -560,6 +562,7 @@ static guint16 bmi1_methods [] = {
 };
 
 static guint16 bmi2_methods [] = {
+       //SN_MultiplyNoFlags,
        SN_ParallelBitDeposit,
        SN_ParallelBitExtract,
        SN_ZeroHighBits,
@@ -629,9 +632,6 @@ emit_x86_intrinsics (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature
                }
        }
        if (!strcmp (class_name, "Bmi1") || (!strcmp (class_name, "X64") && cmethod->klass->nested_in && !strcmp (m_class_get_name (cmethod->klass->nested_in), "Bmi1"))) {
-               // We only support the subset used by corelib
-               if (m_class_get_image (cfg->method->klass) != mono_get_corlib ())
-                       return NULL;
                if (!COMPILE_LLVM (cfg))
                        return NULL;
                id = lookup_intrins (bmi1_methods, sizeof (bmi1_methods), cmethod);
@@ -645,6 +645,26 @@ emit_x86_intrinsics (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature
                        EMIT_NEW_ICONST (cfg, ins, supported ? 1 : 0);
                        ins->type = STACK_I4;
                        return ins;
+               case SN_AndNot: {
+                       // (a ^ -1) & b
+                       // LLVM replaces it with `andn`
+                       int tmp_reg = alloc_preg (cfg);
+                       int result_reg = alloc_preg (cfg);
+                       EMIT_NEW_BIALU_IMM (cfg, ins, is_64bit ? OP_LXOR_IMM : OP_IXOR_IMM, tmp_reg, args [0]->dreg, -1);
+                       EMIT_NEW_BIALU (cfg, ins, is_64bit ? OP_LAND : OP_IAND, result_reg, tmp_reg, args [1]->dreg);
+                       return ins;
+               }
+               case SN_BitFieldExtract: {
+                       if (fsig->param_count == 2) {
+                               MONO_INST_NEW (cfg, ins, is_64bit ? OP_BEXTR64 : OP_BEXTR32);
+                               ins->dreg = alloc_ireg (cfg);
+                               ins->sreg1 = args [0]->dreg;
+                               ins->sreg2 = args [1]->dreg;
+                               ins->type = is_64bit ? STACK_I8 : STACK_I4;
+                               MONO_ADD_INS (cfg->cbb, ins);
+                               return ins;
+                       }
+               }
                case SN_GetMaskUpToLowestSetBit: {
                        // x ^ (x - 1)
                        // LLVM replaces it with `blsmsk`
@@ -686,7 +706,7 @@ emit_x86_intrinsics (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature
                }
        }
        if (!strcmp (class_name, "Bmi2") || (!strcmp (class_name, "X64") && cmethod->klass->nested_in && !strcmp (m_class_get_name (cmethod->klass->nested_in), "Bmi2"))) {
-               // We only support the subset used by corelib
+               // We only support the subset used by corelib. Remove this check once MultiplyNoFlags is implemented.
                if (m_class_get_image (cfg->method->klass) != mono_get_corlib ())
                        return NULL;
                if (!COMPILE_LLVM (cfg))
@@ -701,7 +721,14 @@ emit_x86_intrinsics (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature
                        EMIT_NEW_ICONST (cfg, ins, supported ? 1 : 0);
                        ins->type = STACK_I4;
                        return ins;
-               case SN_ZeroHighBits:
+               //case SN_MultiplyNoFlags:
+                       //// TODO: implement using _mulx_u32/u64:
+                       //// ulong MultiplyNoFlags(ulong left, ulong right)
+                       //// ulong MultiplyNoFlags(ulong left, ulong right, ulong* low) => MultiplyNoFlags(left, right, low);
+                       //// uint MultiplyNoFlags(uint left, uint right)
+                       //// uint MultiplyNoFlags(uint left, uint right, uint* low)
+                       //return NULL;
+               //case SN_ZeroHighBits:
                        MONO_INST_NEW (cfg, ins, is_64bit ? OP_BZHI64 : OP_BZHI32);
                        ins->dreg = alloc_ireg (cfg);
                        ins->sreg1 = args [0]->dreg;
@@ -714,7 +741,7 @@ emit_x86_intrinsics (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature
                        ins->dreg = alloc_ireg (cfg);
                        ins->sreg1 = args [0]->dreg;
                        ins->sreg2 = args [1]->dreg;
-                       ins->type = STACK_I4;
+                       ins->type = is_64bit ? STACK_I8 : STACK_I4;
                        MONO_ADD_INS (cfg->cbb, ins);
                        return ins;
                case SN_ParallelBitDeposit:
@@ -722,13 +749,12 @@ emit_x86_intrinsics (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature
                        ins->dreg = alloc_ireg (cfg);
                        ins->sreg1 = args [0]->dreg;
                        ins->sreg2 = args [1]->dreg;
-                       ins->type = STACK_I4;
+                       ins->type = is_64bit ? STACK_I8 : STACK_I4;
                        MONO_ADD_INS (cfg->cbb, ins);
                        return ins;
                default:
                        g_assert_not_reached ();
                }
-               //printf ("%s %s\n", mono_method_get_full_name (cfg->method), mono_method_get_full_name (cmethod));
        }
 
        return NULL;

diff --git a/src/mono/mono/mini/simd-methods-netcore.h b/src/mono/mono/mini/simd-methods-netcore.h
index bde5da8..4fa2af2 100644 (file)
--- a/src/mono/mono/mini/simd-methods-netcore.h
+++ b/src/mono/mono/mini/simd-methods-netcore.h
@@ -27,8 +27,8 @@ METHOD(op_Inequality)
 METHOD(op_Multiply)
 METHOD(op_Subtraction)
 // BMI1
-//METHOD(AndNot)
-//METHOD(BitFieldExtract)
+METHOD(AndNot)
+METHOD(BitFieldExtract)
 METHOD(ExtractLowestSetBit)
 METHOD(GetMaskUpToLowestSetBit)
 METHOD(ResetLowestSetBit)