Add `intrinsics.masked_expand_load` and `intrinsics.masked_compress_store`

3 files changed, 38 insertions, 7 deletions

diff --git a/src/check_builtin.cpp b/src/check_builtin.cpp
index b5851bc01..bde102a8d 100644
--- a/src/check_builtin.cpp
+++ b/src/check_builtin.cpp
@@ -667,12 +667,16 @@ gb_internal bool check_builtin_simd_operation(CheckerContext *c, Operand *operan
 	case BuiltinProc_simd_scatter:
 	case BuiltinProc_simd_masked_load:
 	case BuiltinProc_simd_masked_store:
+	case BuiltinProc_simd_masked_expand_load:
+	case BuiltinProc_simd_masked_compress_store:
 		{
 			// 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)
 
 			// masked_load (ptr: rawptr, values: #simd[N]T, mask: #simd[N]int_or_bool) -> #simd[N]T
 			// masked_store(ptr: rawptr, values: #simd[N]T, mask: #simd[N]int_or_bool)
+			// masked_expand_load (ptr: rawptr, values: #simd[N]T, mask: #simd[N]int_or_bool) -> #simd[N]T
+			// masked_compress_store(ptr: rawptr, values: #simd[N]T, mask: #simd[N]int_or_bool)
 
 			Operand ptr    = {};
 			Operand values = {};
@@ -733,7 +737,8 @@ gb_internal bool check_builtin_simd_operation(CheckerContext *c, Operand *operan
 			}
 
 			if (id == BuiltinProc_simd_gather ||
-			    id == BuiltinProc_simd_masked_load) {
+			    id == BuiltinProc_simd_masked_load ||
+			    id == BuiltinProc_simd_masked_expand_load) {
 				operand->mode = Addressing_Value;
 				operand->type = values.type;
 			} else {
diff --git a/src/checker_builtin_procs.hpp b/src/checker_builtin_procs.hpp
index a5f688cd8..6245dadaf 100644
--- a/src/checker_builtin_procs.hpp
+++ b/src/checker_builtin_procs.hpp
@@ -195,6 +195,8 @@ BuiltinProc__simd_begin,
 	BuiltinProc_simd_scatter,
 	BuiltinProc_simd_masked_load,
 	BuiltinProc_simd_masked_store,
+	BuiltinProc_simd_masked_expand_load,
+	BuiltinProc_simd_masked_compress_store,
 
 	// Platform specific SIMD intrinsics
 	BuiltinProc_simd_x86__MM_SHUFFLE,
@@ -530,6 +532,8 @@ gb_global BuiltinProc builtin_procs[BuiltinProc_COUNT] = {
 	{STR_LIT("simd_scatter"),      3, false, Expr_Stmt, BuiltinProcPkg_intrinsics},
 	{STR_LIT("simd_masked_load"),  3, false, Expr_Expr, BuiltinProcPkg_intrinsics},
 	{STR_LIT("simd_masked_store"), 3, false, Expr_Stmt, BuiltinProcPkg_intrinsics},
+	{STR_LIT("simd_masked_expand_load"),    3, false, Expr_Expr, BuiltinProcPkg_intrinsics},
+	{STR_LIT("simd_masked_compress_store"), 3, false, Expr_Stmt, BuiltinProcPkg_intrinsics},
 
 	{STR_LIT("simd_x86__MM_SHUFFLE"), 4, false, Expr_Expr, BuiltinProcPkg_intrinsics},
 
diff --git a/src/llvm_backend_proc.cpp b/src/llvm_backend_proc.cpp
index ce1cc8586..ceaeb1aca 100644
--- a/src/llvm_backend_proc.cpp
+++ b/src/llvm_backend_proc.cpp
@@ -1693,6 +1693,8 @@ gb_internal lbValue lb_build_builtin_simd_proc(lbProcedure *p, Ast *expr, TypeAn
 	case BuiltinProc_simd_scatter:
 	case BuiltinProc_simd_masked_load:
 	case BuiltinProc_simd_masked_store:
+	case BuiltinProc_simd_masked_expand_load:
+	case BuiltinProc_simd_masked_compress_store:
 		{
 			LLVMValueRef ptr = arg0.value;
 			LLVMValueRef val = arg1.value;
@@ -1705,11 +1707,14 @@ gb_internal lbValue lb_build_builtin_simd_proc(lbProcedure *p, Ast *expr, TypeAn
 
 			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;
-			case BuiltinProc_simd_masked_load:  name = "llvm.masked.load";    break;
-			case BuiltinProc_simd_masked_store: name = "llvm.masked.store";   break;
-			}
+			case BuiltinProc_simd_gather:                name = "llvm.masked.gather";        break;
+			case BuiltinProc_simd_scatter:               name = "llvm.masked.scatter";       break;
+			case BuiltinProc_simd_masked_load:           name = "llvm.masked.load";          break;
+			case BuiltinProc_simd_masked_store:          name = "llvm.masked.store";         break;
+			case BuiltinProc_simd_masked_expand_load:    name = "llvm.masked.expandload";    break;
+			case BuiltinProc_simd_masked_compress_store: name = "llvm.masked.compressstore"; break;
+			}
+			unsigned type_count = 2;
 			LLVMTypeRef types[2] = {
 				lb_type(p->module, arg1.type),
 				lb_type(p->module, arg0.type)
@@ -1718,6 +1723,7 @@ gb_internal lbValue lb_build_builtin_simd_proc(lbProcedure *p, Ast *expr, TypeAn
 			auto alignment = cast(unsigned long long)type_align_of(base_array_type(arg1.type));
 			LLVMValueRef align = LLVMConstInt(LLVMInt32TypeInContext(p->module->ctx), alignment, false);
 
+			unsigned arg_count = 4;
 			LLVMValueRef args[4] = {};
 			switch (builtin_id) {
 			case BuiltinProc_simd_masked_load:
@@ -1739,9 +1745,25 @@ gb_internal lbValue lb_build_builtin_simd_proc(lbProcedure *p, Ast *expr, TypeAn
 				args[2] = align;
 				args[3] = mask;
 				break;
+
+			case BuiltinProc_simd_masked_expand_load:
+				arg_count = 3;
+				type_count = 1;
+				args[0] = ptr;
+				args[1] = mask;
+				args[2] = val;
+				break;
+
+			case BuiltinProc_simd_masked_compress_store:
+				arg_count = 3;
+				type_count = 1;
+				args[0] = val;
+				args[1] = ptr;
+				args[2] = mask;
+				break;
 			}
 
-			res.value = lb_call_intrinsic(p, name, args, gb_count_of(args), types, gb_count_of(types));
+			res.value = lb_call_intrinsic(p, name, args, arg_count, types, type_count);
 			return res;
 
 		}