Merge pull request #2464 from ap29600/simd_unaligned_load

Improve code generation for `intrinsics.unaligned_load/store` on `#simd` types
This commit is contained in:
gingerBill
2023-04-20 10:33:32 +01:00
committed by GitHub
+19 -5
View File
@@ -2379,9 +2379,15 @@ gb_internal lbValue lb_build_builtin_proc(lbProcedure *p, Ast *expr, TypeAndValu
{
lbValue dst = lb_build_expr(p, ce->args[0]);
lbValue src = lb_build_expr(p, ce->args[1]);
src = lb_address_from_load_or_generate_local(p, src);
Type *t = type_deref(dst.type);
lb_mem_copy_non_overlapping(p, dst, src, lb_const_int(p->module, t_int, type_size_of(t)), false);
if (is_type_simd_vector(t)) {
LLVMValueRef store = LLVMBuildStore(p->builder, src.value, dst.value);
LLVMSetAlignment(store, 1);
} else {
src = lb_address_from_load_or_generate_local(p, src);
lb_mem_copy_non_overlapping(p, dst, src, lb_const_int(p->module, t_int, type_size_of(t)), false);
}
return {};
}
@@ -2389,9 +2395,17 @@ gb_internal lbValue lb_build_builtin_proc(lbProcedure *p, Ast *expr, TypeAndValu
{
lbValue src = lb_build_expr(p, ce->args[0]);
Type *t = type_deref(src.type);
lbAddr dst = lb_add_local_generated(p, t, false);
lb_mem_copy_non_overlapping(p, dst.addr, src, lb_const_int(p->module, t_int, type_size_of(t)), false);
return lb_addr_load(p, dst);
if (is_type_simd_vector(t)) {
lbValue res = {};
res.type = t;
res.value = LLVMBuildLoad2(p->builder, lb_type(p->module, t), src.value, "");
LLVMSetAlignment(res.value, 1);
return res;
} else {
lbAddr dst = lb_add_local_generated(p, t, false);
lb_mem_copy_non_overlapping(p, dst.addr, src, lb_const_int(p->module, t_int, type_size_of(t)), false);
return lb_addr_load(p, dst);
}
}
case BuiltinProc_atomic_add: