Add intrinsics.simd_gather and `intrinsics.simd_scatter

base/intrinsics/intrinsics.odin

@@ -283,6 +283,10 @@ simd_reduce_any         :: proc(a: #simd[N]T) -> T where type_is_boolean(T) ---
 simd_reduce_all         :: proc(a: #simd[N]T) -> T where type_is_boolean(T) ---
 
 
+simd_gather  :: proc(ptr: #simd[N]rawptr, val: #simd[N]T, mask: #simd[N]U) -> #simd[N]T where type_is_integer(U) || type_is_boolean(U) ---
+simd_scatter :: proc(ptr: #simd[N]rawptr, val: #simd[N]T, mask: #simd[N]U)              where type_is_integer(U) || type_is_boolean(U) ---
+
+
 simd_shuffle :: proc(a, b: #simd[N]T, indices: ..int) -> #simd[len(indices)]T ---
 simd_select  :: proc(cond: #simd[N]boolean_or_integer, true, false: #simd[N]T) -> #simd[N]T ---
 

core/simd/simd.odin

@@ -102,6 +102,11 @@ lanes_le :: intrinsics.simd_lanes_le
 lanes_gt :: intrinsics.simd_lanes_gt
 lanes_ge :: intrinsics.simd_lanes_ge
 
+
+// Gather and Scatter intrinsics
+gather  :: intrinsics.simd_gather
+scatter :: intrinsics.simd_scatter
+
 // extract :: proc(a: #simd[N]T, idx: uint) -> T
 extract :: intrinsics.simd_extract
 // replace :: proc(a: #simd[N]T, idx: uint, elem: T) -> #simd[N]T

src/check_builtin.cpp

@@ -663,6 +663,60 @@ gb_internal bool check_builtin_simd_operation(CheckerContext *c, Operand *operan
 			return true;
 		}
 
+	case BuiltinProc_simd_gather:
+	case BuiltinProc_simd_scatter:
+		{
+			// gather (ptr: #simd[N]rawptr, values: #simd[N]T, mask: #simd[N]int_or_bool) -> #simd[N]T
+			// scatter(ptr: #simd[N]rawptr, values: #simd[N]T, mask: #simd[N]int_or_bool)
+
+			Operand ptr    = {};
+			Operand values = {};
+			Operand mask   = {};
+			check_expr(c, &ptr,    ce->args[0]); if (ptr.mode    == Addressing_Invalid) return false;
+			check_expr(c, &values, ce->args[1]); if (values.mode == Addressing_Invalid) return false;
+			check_expr(c, &mask,   ce->args[2]); if (mask.mode   == Addressing_Invalid) return false;
+			if (!is_type_simd_vector(ptr.type))    { error(ptr.expr,    "'%.*s' expected a simd vector type", LIT(builtin_name)); return false; }
+			if (!is_type_simd_vector(values.type)) { error(values.expr, "'%.*s' expected a simd vector type", LIT(builtin_name)); return false; }
+			if (!is_type_simd_vector(mask.type))   { error(mask.expr,   "'%.*s' expected a simd vector type", LIT(builtin_name)); return false; }
+
+			Type *ptr_elem = base_array_type(ptr.type);
+			if (!is_type_rawptr(ptr_elem)) {
+				gbString s = type_to_string(ptr.type);
+				error(ptr.expr, "Expected a simd vector of 'rawptr' for the addresses, got %s", s);
+				gb_string_free(s);
+				return false;
+			}
+			Type *mask_elem = base_array_type(mask.type);
+
+			if (!is_type_integer(mask_elem) && !is_type_boolean(mask_elem)) {
+				gbString s = type_to_string(mask.type);
+				error(mask.expr, "Expected a simd vector of integers or booleans for the mask, got %s", s);
+				gb_string_free(s);
+				return false;
+			}
+
+			i64 ptr_count    = get_array_type_count(ptr.type);
+			i64 values_count = get_array_type_count(values.type);
+			i64 mask_count   = get_array_type_count(mask.type);
+			if (ptr_count != values_count ||
+			    values_count != mask_count ||
+			    mask_count != ptr_count) {
+				gbString s = type_to_string(mask.type);
+				error(mask.expr, "All simd vectors must be of the same length, got %lld vs %lld vs %lld", cast(long long)ptr_count, cast(long long)values_count, cast(long long)mask_count);
+				gb_string_free(s);
+				return false;
+			}
+
+			if (id == BuiltinProc_simd_gather) {
+				operand->mode = Addressing_Value;
+				operand->type = values.type;
+			} else {
+				operand->mode = Addressing_NoValue;
+				operand->type = nullptr;
+			}
+			return true;
+		}
+
 	case BuiltinProc_simd_extract:
 		{
 			Operand x = {};

src/checker_builtin_procs.hpp

@@ -191,6 +191,9 @@ BuiltinProc__simd_begin,
 	BuiltinProc_simd_lanes_rotate_left,
 	BuiltinProc_simd_lanes_rotate_right,
 
+	BuiltinProc_simd_gather,
+	BuiltinProc_simd_scatter,
+
 
 	// Platform specific SIMD intrinsics
 	BuiltinProc_simd_x86__MM_SHUFFLE,
@@ -522,6 +525,9 @@ gb_global BuiltinProc builtin_procs[BuiltinProc_COUNT] = {
 	{STR_LIT("simd_lanes_rotate_left"), 2, false, Expr_Expr, BuiltinProcPkg_intrinsics},
 	{STR_LIT("simd_lanes_rotate_right"), 2, false, Expr_Expr, BuiltinProcPkg_intrinsics},
 
+	{STR_LIT("simd_gather"),  3, false, Expr_Expr, BuiltinProcPkg_intrinsics},
+	{STR_LIT("simd_scatter"), 3, false, Expr_Stmt, BuiltinProcPkg_intrinsics},
+
 	{STR_LIT("simd_x86__MM_SHUFFLE"), 4, false, Expr_Expr, BuiltinProcPkg_intrinsics},
 
 	{STR_LIT(""), 0, false, Expr_Stmt, BuiltinProcPkg_intrinsics},

src/llvm_backend_proc.cpp

@@ -1688,6 +1688,52 @@ gb_internal lbValue lb_build_builtin_simd_proc(lbProcedure *p, Ast *expr, TypeAn
 			return res;
 		}
 
+
+	case BuiltinProc_simd_gather:
+	case BuiltinProc_simd_scatter:
+		{
+			LLVMValueRef ptr = arg0.value;
+			LLVMValueRef val = arg1.value;
+			LLVMValueRef mask = arg2.value;
+
+			unsigned count = cast(unsigned)get_array_type_count(arg0.type);
+
+			LLVMTypeRef mask_type = LLVMVectorType(LLVMInt1TypeInContext(p->module->ctx), count);
+			mask = LLVMBuildTrunc(p->builder, mask, mask_type, "");
+
+			char const *name = nullptr;
+			switch (builtin_id) {
+			case BuiltinProc_simd_gather:  name = "llvm.masked.gather";  break;
+			case BuiltinProc_simd_scatter: name = "llvm.masked.scatter"; break;
+			}
+			LLVMTypeRef types[2] = {
+				lb_type(p->module, arg1.type),
+				lb_type(p->module, arg0.type)
+			};
+
+			auto alignment = cast(unsigned long long)type_align_of(base_array_type(arg1.type));
+			LLVMValueRef align = LLVMConstInt(LLVMInt32TypeInContext(p->module->ctx), alignment, false);
+
+			LLVMValueRef args[4] = {};
+			switch (builtin_id) {
+			case BuiltinProc_simd_gather:
+				args[0] = ptr;
+				args[1] = align;
+				args[2] = mask;
+				args[3] = val;
+				break;
+			case BuiltinProc_simd_scatter:
+				args[0] = val;
+				args[1] = ptr;
+				args[2] = align;
+				args[3] = mask;
+				break;
+			}
+
+			res.value = lb_call_intrinsic(p, name, args, gb_count_of(args), types, gb_count_of(types));
+			return res;
+
+		}
 	}
 	GB_PANIC("Unhandled simd intrinsic: '%.*s'", LIT(builtin_procs[builtin_id].name));