Add intrinsics.masked_expand_load and intrinsics.masked_compress_store

This commit is contained in:
gingerBill
2024-08-05 14:54:09 +01:00
parent 78919f8524
commit f56abf3780
5 changed files with 43 additions and 7 deletions

View File

@@ -286,6 +286,10 @@ simd_scatter :: proc(ptr: #simd[N]rawptr, val: #simd[N]T, mask: #simd[N]U)
simd_masked_load :: proc(ptr: rawptr, val: #simd[N]T, mask: #simd[N]U) -> #simd[N]T where type_is_integer(U) || type_is_boolean(U) ---
simd_masked_store :: proc(ptr: rawptr, val: #simd[N]T, mask: #simd[N]U) where type_is_integer(U) || type_is_boolean(U) ---
simd_masked_expand_load :: proc(ptr: rawptr, val: #simd[N]T, mask: #simd[N]U) -> #simd[N]T where type_is_integer(U) || type_is_boolean(U) ---
simd_masked_compress_store :: proc(ptr: 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 ---

View File

@@ -108,7 +108,8 @@ gather :: intrinsics.simd_gather
scatter :: intrinsics.simd_scatter
masked_load :: intrinsics.simd_masked_load
masked_store :: intrinsics.simd_masked_store
masked_expand_load :: intrinsics.simd_masked_expand_load
masked_compress_store :: intrinsics.simd_masked_compress_store
// extract :: proc(a: #simd[N]T, idx: uint) -> T
extract :: intrinsics.simd_extract

View File

@@ -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 {

View File

@@ -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},

View File

@@ -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;
}