Merge pull request #4418 from laytan/check-packed-on-all-loads

check packed load and set alignment on all loads, not just lb_emit_load
This commit is contained in:
gingerBill
2024-10-25 14:35:32 +01:00
committed by GitHub
5 changed files with 45 additions and 26 deletions

View File

@@ -130,7 +130,7 @@ gb_internal lbValue lb_emit_unary_arith(lbProcedure *p, TokenKind op, lbValue x,
LLVMTypeRef vector_type = nullptr;
if (op != Token_Not && lb_try_vector_cast(p->module, val, &vector_type)) {
LLVMValueRef vp = LLVMBuildPointerCast(p->builder, val.value, LLVMPointerType(vector_type, 0), "");
LLVMValueRef v = LLVMBuildLoad2(p->builder, vector_type, vp, "");
LLVMValueRef v = OdinLLVMBuildLoad(p, vector_type, vp);
LLVMValueRef opv = nullptr;
switch (op) {
@@ -324,8 +324,8 @@ gb_internal bool lb_try_direct_vector_arith(lbProcedure *p, TokenKind op, lbValu
LLVMValueRef lhs_vp = LLVMBuildPointerCast(p->builder, lhs_ptr.value, LLVMPointerType(vector_type, 0), "");
LLVMValueRef rhs_vp = LLVMBuildPointerCast(p->builder, rhs_ptr.value, LLVMPointerType(vector_type, 0), "");
LLVMValueRef x = LLVMBuildLoad2(p->builder, vector_type, lhs_vp, "");
LLVMValueRef y = LLVMBuildLoad2(p->builder, vector_type, rhs_vp, "");
LLVMValueRef x = OdinLLVMBuildLoad(p, vector_type, lhs_vp);
LLVMValueRef y = OdinLLVMBuildLoad(p, vector_type, rhs_vp);
LLVMValueRef z = nullptr;
if (is_type_float(integral_type)) {
@@ -551,15 +551,14 @@ gb_internal LLVMValueRef lb_matrix_to_vector(lbProcedure *p, lbValue matrix) {
Type *mt = base_type(matrix.type);
GB_ASSERT(mt->kind == Type_Matrix);
LLVMTypeRef elem_type = lb_type(p->module, mt->Matrix.elem);
unsigned total_count = cast(unsigned)matrix_type_total_internal_elems(mt);
LLVMTypeRef total_matrix_type = LLVMVectorType(elem_type, total_count);
#if 1
LLVMValueRef ptr = lb_address_from_load_or_generate_local(p, matrix).value;
LLVMValueRef matrix_vector_ptr = LLVMBuildPointerCast(p->builder, ptr, LLVMPointerType(total_matrix_type, 0), "");
LLVMValueRef matrix_vector = LLVMBuildLoad2(p->builder, total_matrix_type, matrix_vector_ptr, "");
LLVMSetAlignment(matrix_vector, cast(unsigned)type_align_of(mt));
LLVMValueRef matrix_vector = OdinLLVMBuildLoadAligned(p, total_matrix_type, matrix_vector_ptr, type_align_of(mt));
return matrix_vector;
#else
LLVMValueRef matrix_vector = LLVMBuildBitCast(p->builder, matrix.value, total_matrix_type, "");

View File

@@ -774,6 +774,36 @@ gb_internal bool lb_try_vector_cast(lbModule *m, lbValue ptr, LLVMTypeRef *vecto
return false;
}
gb_internal LLVMValueRef OdinLLVMBuildLoad(lbProcedure *p, LLVMTypeRef type, LLVMValueRef value) {
LLVMValueRef result = LLVMBuildLoad2(p->builder, type, value, "");
// If it is not an instruction it isn't a GEP, so we don't need to track alignment in the metadata,
// which is not possible anyway (only LLVM instructions can have metadata).
if (LLVMIsAInstruction(value)) {
u64 is_packed = lb_get_metadata_custom_u64(p->module, value, ODIN_METADATA_IS_PACKED);
if (is_packed != 0) {
LLVMSetAlignment(result, 1);
}
}
return result;
}
gb_internal LLVMValueRef OdinLLVMBuildLoadAligned(lbProcedure *p, LLVMTypeRef type, LLVMValueRef value, i64 alignment) {
LLVMValueRef result = LLVMBuildLoad2(p->builder, type, value, "");
LLVMSetAlignment(result, cast(unsigned)alignment);
if (LLVMIsAInstruction(value)) {
u64 is_packed = lb_get_metadata_custom_u64(p->module, value, ODIN_METADATA_IS_PACKED);
if (is_packed != 0) {
LLVMSetAlignment(result, 1);
}
}
return result;
}
gb_internal void lb_addr_store(lbProcedure *p, lbAddr addr, lbValue value) {
if (addr.addr.value == nullptr) {
return;
@@ -1119,7 +1149,7 @@ gb_internal lbValue lb_emit_load(lbProcedure *p, lbValue value) {
Type *vt = base_type(value.type);
GB_ASSERT(vt->kind == Type_MultiPointer);
Type *t = vt->MultiPointer.elem;
LLVMValueRef v = LLVMBuildLoad2(p->builder, lb_type(p->module, t), value.value, "");
LLVMValueRef v = OdinLLVMBuildLoad(p, lb_type(p->module, t), value.value);
return lbValue{v, t};
} else if (is_type_soa_pointer(value.type)) {
lbValue ptr = lb_emit_struct_ev(p, value, 0);
@@ -1130,16 +1160,7 @@ gb_internal lbValue lb_emit_load(lbProcedure *p, lbValue value) {
GB_ASSERT_MSG(is_type_pointer(value.type), "%s", type_to_string(value.type));
Type *t = type_deref(value.type);
LLVMValueRef v = LLVMBuildLoad2(p->builder, lb_type(p->module, t), value.value, "");
// If it is not an instruction it isn't a GEP, so we don't need to track alignment in the metadata,
// which is not possible anyway (only LLVM instructions can have metadata).
if (LLVMIsAInstruction(value.value)) {
u64 is_packed = lb_get_metadata_custom_u64(p->module, value.value, ODIN_METADATA_IS_PACKED);
if (is_packed != 0) {
LLVMSetAlignment(v, 1);
}
}
LLVMValueRef v = OdinLLVMBuildLoad(p, lb_type(p->module, t), value.value);
return lbValue{v, t};
}
@@ -1413,7 +1434,7 @@ gb_internal lbValue lb_addr_load(lbProcedure *p, lbAddr const &addr) {
LLVMTypeRef vector_type = nullptr;
if (lb_try_vector_cast(p->module, addr.addr, &vector_type)) {
LLVMValueRef vp = LLVMBuildPointerCast(p->builder, addr.addr.value, LLVMPointerType(vector_type, 0), "");
LLVMValueRef v = LLVMBuildLoad2(p->builder, vector_type, vp, "");
LLVMValueRef v = OdinLLVMBuildLoad(p, vector_type, vp);
LLVMValueRef scalars[4] = {};
for (u8 i = 0; i < addr.swizzle.count; i++) {
scalars[i] = LLVMConstInt(lb_type(p->module, t_u32), addr.swizzle.indices[i], false);
@@ -2710,7 +2731,7 @@ general_end:;
if (LLVMIsALoadInst(val) && (src_size >= dst_size && src_align >= dst_align)) {
LLVMValueRef val_ptr = LLVMGetOperand(val, 0);
val_ptr = LLVMBuildPointerCast(p->builder, val_ptr, LLVMPointerType(dst_type, 0), "");
LLVMValueRef loaded_val = LLVMBuildLoad2(p->builder, dst_type, val_ptr, "");
LLVMValueRef loaded_val = OdinLLVMBuildLoad(p, dst_type, val_ptr);
// LLVMSetAlignment(loaded_val, gb_min(src_align, dst_align));
@@ -2726,7 +2747,7 @@ general_end:;
LLVMValueRef nptr = LLVMBuildPointerCast(p->builder, ptr, LLVMPointerType(src_type, 0), "");
LLVMBuildStore(p->builder, val, nptr);
return LLVMBuildLoad2(p->builder, dst_type, ptr, "");
return OdinLLVMBuildLoad(p, dst_type, ptr);
}
}

View File

@@ -2568,7 +2568,7 @@ gb_internal lbValue lb_build_builtin_proc(lbProcedure *p, Ast *expr, TypeAndValu
case BuiltinProc_atomic_load_explicit: {
lbValue dst = lb_build_expr(p, ce->args[0]);
LLVMValueRef instr = LLVMBuildLoad2(p->builder, lb_type(p->module, type_deref(dst.type)), dst.value, "");
LLVMValueRef instr = OdinLLVMBuildLoad(p, lb_type(p->module, type_deref(dst.type)), dst.value);
switch (id) {
case BuiltinProc_non_temporal_load:
{
@@ -2621,8 +2621,7 @@ gb_internal lbValue lb_build_builtin_proc(lbProcedure *p, Ast *expr, TypeAndValu
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);
res.value = OdinLLVMBuildLoadAligned(p, lb_type(p->module, t), src.value, 1);
return res;
} else {
lbAddr dst = lb_add_local_generated(p, t, false);

View File

@@ -2001,7 +2001,7 @@ gb_internal void lb_build_return_stmt_internal(lbProcedure *p, lbValue res) {
LLVMValueRef ptr = p->temp_callee_return_struct_memory;
LLVMValueRef nptr = LLVMBuildPointerCast(p->builder, ptr, LLVMPointerType(src_type, 0), "");
LLVMBuildStore(p->builder, ret_val, nptr);
ret_val = LLVMBuildLoad2(p->builder, ret_type, ptr, "");
ret_val = OdinLLVMBuildLoad(p, ret_type, ptr);
} else {
ret_val = OdinLLVMBuildTransmute(p, ret_val, ret_type);
}

View File

@@ -269,7 +269,7 @@ gb_internal lbValue lb_emit_transmute(lbProcedure *p, lbValue value, Type *t) {
if (lb_try_update_alignment(ptr, align)) {
LLVMTypeRef result_type = lb_type(p->module, t);
res.value = LLVMBuildPointerCast(p->builder, ptr.value, LLVMPointerType(result_type, 0), "");
res.value = LLVMBuildLoad2(p->builder, result_type, res.value, "");
res.value = OdinLLVMBuildLoad(p, result_type, res.value);
return res;
}
lbAddr addr = lb_add_local_generated(p, t, false);