Commit 1c35e73b614398529782f8c027366c6d8d51ac4b | zig

author	Alex Rønne Petersen <alex@alexrp.com> 2024-07-25 22:06:29+02:00
committer	Andrew Kelley <andrew@ziglang.org> 2024-07-25 18:04:50-07:00
log	1c35e73b614398529782f8c027366c6d8d51ac4b
tree	250c777f79bd747416f105111592fc7c79a344d6
parent	81a172a5064559d1f968f240204f8ce545852441

llvm: Don't emit safety memset() for stores of undef in Debug with safety off.

Before, this code:

    @setRuntimeSafety(false);
    var arr: [38]elf.Addr = undefined;

would emit a call to memset() in the output code in Debug mode, while in all the
release modes, LLVM optimized the memset() out as expected. Emitting the call in
Debug mode is problematic in some contexts, e.g. in std.os.linux.start_pie where
we are not yet ready to correctly perform calls because relocations haven't been
applied yet, or in the early stages of a dynamic linker, etc.

1 files changed, 15 insertions(+), 4 deletions(-)

src
- codegen
  - llvm.zig

src/codegen/llvm.zig+15-4

...	...	@@ -8997,6 +8997,21 @@ pub const FuncGen = struct {
8997	8997
8998	8998	const val_is_undef = if (try self.air.value(bin_op.rhs, pt)) \|val\| val.isUndefDeep(mod) else false;
8999	8999	if (val_is_undef) {
	9000	const owner_mod = self.dg.ownerModule();
	9001
	9002	// Even if safety is disabled, we still emit a memset to undefined since it conveys
	9003	// extra information to LLVM, and LLVM will optimize it out. Safety makes the difference
	9004	// between using 0xaa or actual undefined for the fill byte.
	9005	//
	9006	// However, for Debug builds specifically, we avoid emitting the memset because LLVM
	9007	// will neither use the information nor get rid of the memset, thus leaving an
	9008	// unexpected call in the user's code. This is problematic if the code in question is
	9009	// not ready to correctly make calls yet, such as in our early PIE startup code, or in
	9010	// the early stages of a dynamic linker, etc.
	9011	if (!safety and owner_mod.optimize_mode == .Debug) {
	9012	return .none;
	9013	}
	9014
9000	9015	const ptr_info = ptr_ty.ptrInfo(mod);
9001	9016	const needs_bitmask = (ptr_info.packed_offset.host_size != 0);
9002	9017	if (needs_bitmask) {
...	...	@@ -9006,9 +9021,6 @@ pub const FuncGen = struct {
9006	9021	return .none;
9007	9022	}
9008	9023
9009		// Even if safety is disabled, we still emit a memset to undefined since it conveys
9010		// extra information to LLVM. However, safety makes the difference between using
9011		// 0xaa or actual undefined for the fill byte.
9012	9024	const len = try o.builder.intValue(try o.lowerType(Type.usize), operand_ty.abiSize(pt));
9013	9025	_ = try self.wip.callMemSet(
9014	9026	dest_ptr,
...	...	@@ -9017,7 +9029,6 @@ pub const FuncGen = struct {
9017	9029	len,
9018	9030	if (ptr_ty.isVolatilePtr(mod)) .@"volatile" else .normal,
9019	9031	);
9020		const owner_mod = self.dg.ownerModule();
9021	9032	if (safety and owner_mod.valgrind) {
9022	9033	try self.valgrindMarkUndef(dest_ptr, len);
9023	9034	}