[jit] Disable the optimization added by 63eb94e451a59491034516f2ec9f27d586b34d86...

[mono.git] / mono / mini / simd-intrinsics.c

diff --git a/mono/mini/simd-intrinsics.c b/mono/mini/simd-intrinsics.c
index 4669dd7c83a10e3c2275916f87350d286f19e4a1..ff737bbe9347a7b0f5d2fce3e44832386bd9c356 100644 (file)
--- a/mono/mini/simd-intrinsics.c
+++ b/mono/mini/simd-intrinsics.c
@@ -12,6 +12,7 @@
 
 #include "mini.h"
 #include "ir-emit.h"
+#include "mono/utils/bsearch.h"
 
 /*
 General notes on SIMD intrinsics
@@ -138,6 +139,9 @@ static const SimdIntrinsc vector4f_intrinsics[] = {
        { SN_CompareNotLessThan, OP_COMPPS, SIMD_VERSION_SSE1, SIMD_EMIT_BINARY, SIMD_COMP_NLT },
        { SN_CompareOrdered, OP_COMPPS, SIMD_VERSION_SSE1, SIMD_EMIT_BINARY, SIMD_COMP_ORD },
        { SN_CompareUnordered, OP_COMPPS, SIMD_VERSION_SSE1, SIMD_EMIT_BINARY, SIMD_COMP_UNORD },
+       { SN_ConvertToDouble, OP_CVTPS2PD, SIMD_VERSION_SSE2, SIMD_EMIT_UNARY },
+       { SN_ConvertToInt, OP_CVTPS2DQ, SIMD_VERSION_SSE2, SIMD_EMIT_UNARY },
+       { SN_ConvertToIntTruncated, OP_CVTTPS2DQ, SIMD_VERSION_SSE2, SIMD_EMIT_UNARY },
        { SN_DuplicateHigh, OP_DUPPS_HIGH, SIMD_VERSION_SSE3, SIMD_EMIT_UNARY },
        { SN_DuplicateLow, OP_DUPPS_LOW, SIMD_VERSION_SSE3, SIMD_EMIT_UNARY },
        { SN_HorizontalAdd, OP_HADDPS, SIMD_VERSION_SSE3, SIMD_EMIT_BINARY },
@@ -189,6 +193,9 @@ static const SimdIntrinsc vector2d_intrinsics[] = {
        { SN_CompareNotLessThan, OP_COMPPD, SIMD_VERSION_SSE1, SIMD_EMIT_BINARY, SIMD_COMP_NLT },
        { SN_CompareOrdered, OP_COMPPD, SIMD_VERSION_SSE1, SIMD_EMIT_BINARY, SIMD_COMP_ORD },
        { SN_CompareUnordered, OP_COMPPD, SIMD_VERSION_SSE1, SIMD_EMIT_BINARY, SIMD_COMP_UNORD },
+       { SN_ConvertToFloat, OP_CVTPD2PS, SIMD_VERSION_SSE2, SIMD_EMIT_UNARY },
+       { SN_ConvertToInt, OP_CVTPD2DQ, SIMD_VERSION_SSE2, SIMD_EMIT_UNARY },
+       { SN_ConvertToIntTruncated, OP_CVTTPD2DQ, SIMD_VERSION_SSE2, SIMD_EMIT_UNARY },
        { SN_Duplicate, OP_DUPPD, SIMD_VERSION_SSE3, SIMD_EMIT_UNARY },
        { SN_HorizontalAdd, OP_HADDPD, SIMD_VERSION_SSE3, SIMD_EMIT_BINARY },
        { SN_HorizontalSub, OP_HSUBPD, SIMD_VERSION_SSE3, SIMD_EMIT_BINARY },   
@@ -315,6 +322,8 @@ static const SimdIntrinsc vector4i_intrinsics[] = {
        { SN_ctor, OP_EXPAND_I4, SIMD_VERSION_SSE1, SIMD_EMIT_CTOR },
        { SN_CompareEqual, OP_PCMPEQD, SIMD_VERSION_SSE1, SIMD_EMIT_BINARY },
        { SN_CompareGreaterThan, OP_PCMPGTD, SIMD_VERSION_SSE1, SIMD_EMIT_BINARY },
+       { SN_ConvertToDouble, OP_CVTDQ2PD, SIMD_VERSION_SSE2, SIMD_EMIT_UNARY },
+       { SN_ConvertToFloat, OP_CVTDQ2PS, SIMD_VERSION_SSE2, SIMD_EMIT_UNARY },
        { SN_LoadAligned, 0, SIMD_VERSION_SSE1, SIMD_EMIT_LOAD_ALIGNED },
        { SN_LogicalRightShift, OP_PSHRD, SIMD_VERSION_SSE1, SIMD_EMIT_SHIFT },
        { SN_Max, OP_PMAXD, SIMD_VERSION_SSE41, SIMD_EMIT_BINARY },
@@ -889,8 +898,9 @@ mono_type_to_expand_op (MonoType *type)
                return OP_EXPAND_R4;
        case MONO_TYPE_R8:
                return OP_EXPAND_R8;
+       default:
+               g_assert_not_reached ();
        }
-       g_assert_not_reached ();
 }
 
 static int
@@ -976,8 +986,9 @@ mono_type_to_extract_op (MonoType *type)
        case MONO_TYPE_U4:
        case MONO_TYPE_R4:
                return OP_EXTRACT_I4;
+       default:
+               g_assert_not_reached ();
        }
-       g_assert_not_reached ();
 }
 
 /*Returns the amount to shift the element index to get the dword it belongs to*/
@@ -995,8 +1006,34 @@ mono_type_elements_shift_bits (MonoType *type)
        case MONO_TYPE_U4:
        case MONO_TYPE_R4:
                return 0;
+       default:
+               g_assert_not_reached ();
+       }
+}
+
+static G_GNUC_UNUSED int
+mono_type_to_insert_op (MonoType *type)
+{
+       switch (type->type) {
+       case MONO_TYPE_I1:
+       case MONO_TYPE_U1:
+               return OP_INSERT_I1;
+       case MONO_TYPE_I2:
+       case MONO_TYPE_U2:
+               return OP_INSERT_I2;
+       case MONO_TYPE_I4:
+       case MONO_TYPE_U4:
+               return OP_INSERT_I4;
+       case MONO_TYPE_I8:
+       case MONO_TYPE_U8:
+               return OP_INSERT_I8;
+       case MONO_TYPE_R4:
+               return OP_INSERT_R4;
+       case MONO_TYPE_R8:
+               return OP_INSERT_R8;
+       default:
+               g_assert_not_reached ();
        }
-       g_assert_not_reached ();
 }
 
 static int
@@ -1019,8 +1056,9 @@ mono_type_to_slow_insert_op (MonoType *type)
                return OP_INSERTX_R4_SLOW;
        case MONO_TYPE_R8:
                return OP_INSERTX_R8_SLOW;
+       default:
+               g_assert_not_reached ();
        }
-       g_assert_not_reached ();
 }
 
 static MonoInst*
@@ -1034,7 +1072,14 @@ simd_intrinsic_emit_setter (const SimdIntrinsc *intrinsic, MonoCompile *cfg, Mon
 
        size = mono_type_size (sig->params [0], &align); 
 
-       if (size == 2 || size == 4 || size == 8) {
+       if (COMPILE_LLVM (cfg)) {
+               MONO_INST_NEW (cfg, ins, mono_type_to_insert_op (sig->params [0]));
+               ins->klass = cmethod->klass;
+               ins->dreg = ins->sreg1 = dreg = load_simd_vreg (cfg, cmethod, args [0], &indirect);
+               ins->sreg2 = args [1]->dreg;
+               ins->inst_c0 = intrinsic->opcode;
+               MONO_ADD_INS (cfg->cbb, ins);
+       } else if (size == 2 || size == 4 || size == 8) {
                MONO_INST_NEW (cfg, ins, mono_type_to_slow_insert_op (sig->params [0]));
                ins->klass = cmethod->klass;
                /*This is a partial load so we encode the dependency on the previous value by setting dreg and sreg1 to the same value.*/
@@ -1433,7 +1478,7 @@ simd_version_name (guint32 version)
 static MonoInst*
 emit_intrinsics (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsig, MonoInst **args, const SimdIntrinsc *intrinsics, guint32 size)
 {
-       const SimdIntrinsc * result = bsearch (cmethod->name, intrinsics, size, sizeof (SimdIntrinsc), &simd_intrinsic_compare_by_name);
+       const SimdIntrinsc * result = mono_binary_search (cmethod->name, intrinsics, size, sizeof (SimdIntrinsc), &simd_intrinsic_compare_by_name);
        if (!result) {
                DEBUG (printf ("function doesn't have a simd intrinsic %s::%s/%d\n", cmethod->klass->name, cmethod->name, fsig->param_count));
                return NULL;