ed/Odin
1
0
Fork 0
mirror of https://github.com/Ed94/Odin.git synced 2026-06-20 12:44:59 -07:00
Code Issues Packages Projects Releases Wiki Activity

improve code generation for intrinsics.unaligned_load/store on #simd types

Browse Source 
the default implementation calls memcpy on an `alloca` constant, which
seems to heavily confuse the optimizer and produces overall suboptimal
code.

Introducing this specialization simplifies the intermediate
representation produced, resulting in more efficient code.
This commit is contained in:
Andrea Piseri
2023-04-16 15:01:30 +02:00
parent b7b5043aea
commit af63eff8d7
1 changed files with 19 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llvm_backend_proc.cpp
+19 -5
View File
@@ -2363,9 +2363,15 @@ 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 {};
}
@@ -2373,9 +2379,17 @@ 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: