llvm: improve IR attributes and cast flags for better optimization

[Jarred-Sumner]

Summary

Emit several LLVM IR attributes and instruction flags that Clang emits for equivalent C/C++ but Zig currently omits. Each was verified to produce measurably better optimized IR/asm (branch-to-select, vectorization, call-slot elimination, function purity).

• dereferenceable(N) / dereferenceable_or_null(N) on *T / ?*T params
• noundef + align + dereferenceable on by-ref aggregate params and sret
• writable + dead_on_unwind on sret (new Builder attrs)
• trunc nuw/nsw / zext nneg for unchecked @intCast instead of llvm.assume (keeps memory(none))
• fast flag on fneg under .optimized float mode
• optsize alongside cold for @branchHint(.cold) functions
• index-counted loop for non-byte @memset (SCEV-friendly)
• 16-byte over-alignment for large anonymous constants

Test plan

• zig build test-llvm-ir (13 existing + 10 new cases)
• Full behavior suite, ReleaseFast: 1982 pass / 0 fail
• Targeted behavior tests (cast, cast_int, pointers, struct, fn, floatop, memset, byval_arg_var) in ReleaseFast + ReleaseSafe
• Bitcode roundtrip via llvm-dis confirms new attrs/flags encode correctly
• Full behavior suite, Debug (running)
• CI cross-target builds

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Pro

Run ID: a1ab361e-58ce-4748-b015-c1665f2202aa

📥 Commits

Reviewing files that changed from the base of the PR and between 3458234 and 01e81b8.

📒 Files selected for processing (1)

• test/llvm_ir.zig

Disabled knowledge base sources:

• Linear integration is disabled

You can enable these sources in your CodeRabbit configuration.

---

Walkthrough

Adds LLVM parameter attributes writable and dead_on_unwind, new cast instruction variants/flags (trunc nsw, trunc nuw, trunc nuw nsw, zext nneg) with bitcode encoding, and updates codegen to emit these attributes/casts, refine pointer-byref attributes, cold-function .optsize, float negation fneg fast, and non-byte memset lowering.

Changes

Cohort / File(s)	Summary
LLVM Builder & IR lib/std/zig/llvm/Builder.zig, lib/std/zig/llvm/ir.zig	Added Attribute.Kind tags writable, dead_on_unwind; introduced new cast instruction tags and mapping to cast opcodes; added FunctionBlock.CastFlags bitcode abbrev with packed flags field and registered it in abbrevs.
Code Generation src/codegen/llvm.zig	Emit/propagate new parameter attributes (including writable/dead_on_unwind for sret), extend by-ref/pointer attrs with dereferenceable/dereferenceable_or_null and .noundef, add .optsize for cold functions, select trunc nuw/nsw and zext nneg when appropriate, emit fneg fast under fast float mode, and change non-byte memset to induction-loop form.
Tests test/llvm_ir.zig	Added IR pattern tests for pointer dereferenceable attributes, sret with writable/dead_on_unwind/noundef, by-ref propagation, cast flags (nuw, nsw, combined; nneg), fneg fast behavior, and cold-function cold + optsize emission.

Possibly related PRs

• llvm: improve IR attributes and cast flags for better optimization #21 — Appears to modify the same LLVM builder/ir and codegen paths: adds writable/dead_on_unwind attributes, new cast tags/flags, and FunctionBlock.CastFlags bitcode form; likely closely related at code level.

🚥 Pre-merge checks | ✅ 4

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title directly summarizes the main changes: improving LLVM IR attributes and cast flags for better optimization, which aligns with the primary content across all modified files.
Description check	✅ Passed	The description is well-related to the changeset, providing a detailed summary of LLVM IR attributes, instruction flags, and codegen patterns being added to match Clang's output and improve optimization.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Actionable comments posted: 2

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@lib/std/zig/llvm/ir.zig`:
- Around line 1537-1541: The packed struct named flags currently uses generic
field names bit0/bit1 which are easy to misuse; rename those fields to semantic
names (e.g., nuw and nsw) and add either a brief comment noting the special-case
mapping for zext (bit0 represents nneg) or add small helper accessors on the
flags type (e.g., isNuw(), isNsw(), isNneg()) so callers use semantic APIs
instead of raw bitN; update all uses of flags.bit0/bit1 to the new names or
helpers to keep opcode→bit mapping clear and maintainable.

In `@test/llvm_ir.zig`:
- Around line 160-175: Add a new test case that exercises the combined trunc
flags by calling cases.addMatches with a unique name like "intCast trunc nuw
nsw", using an export fn entry(x: i64) i32 { return `@intCast`(x); } (or an
appropriate signed/unsigned signature that triggers both nuw and nsw) and
include the expected LLVM match string "trunc nuw nsw i64" in the match array;
place it alongside the existing "intCast trunc nuw" and "intCast trunc nsw"
blocks so the .@"trunc nuw nsw" branch in src/codegen/llvm.zig is covered.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Pro

Run ID: 9494e6cb-b9e2-44c5-8b5f-2905f24a9620

📥 Commits

Reviewing files that changed from the base of the PR and between 24475de and 3458234.

📒 Files selected for processing (4)

• lib/std/zig/llvm/Builder.zig
• lib/std/zig/llvm/ir.zig
• src/codegen/llvm.zig
• test/llvm_ir.zig

[coderabbitai]

🧹 Nitpick | 🔵 Trivial

Consider naming cast flag bits semantically.

bit0 / bit1 are correct but easy to misuse later. Using semantic names (or a small helper type) would make opcode→bit mapping safer to maintain.

Optional readability tweak

-        /// trunc: bit0 = nuw, bit1 = nsw. zext: bit0 = nneg.
-        flags: packed struct(u2) {
-            bit0: bool,
-            bit1: bool,
-        },
+        /// trunc: low = nuw, high = nsw. zext: low = nneg.
+        flags: packed struct(u2) {
+            low: bool,
+            high: bool,
+        },

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	/// trunc: bit0 = nuw, bit1 = nsw. zext: bit0 = nneg.
	flags: packed struct(u2) {
	bit0: bool,
	bit1: bool,
	},
	/// trunc: low = nuw, high = nsw. zext: low = nneg.
	flags: packed struct(u2) {
	low: bool,
	high: bool,
	},

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@lib/std/zig/llvm/ir.zig` around lines 1537 - 1541, The packed struct named
flags currently uses generic field names bit0/bit1 which are easy to misuse;
rename those fields to semantic names (e.g., nuw and nsw) and add either a brief
comment noting the special-case mapping for zext (bit0 represents nneg) or add
small helper accessors on the flags type (e.g., isNuw(), isNsw(), isNneg()) so
callers use semantic APIs instead of raw bitN; update all uses of
flags.bit0/bit1 to the new names or helpers to keep opcode→bit mapping clear and
maintainable.