Apply LLVM patches for deterministic operand order and correct peephole optimization for minsize (#500)

## Description

Applies LLVM patches that have not been backported to the major version
used in our submodule:

* [LLVM PR #163679](https://github.com/llvm/llvm-project/pull/163679)
  * *"[RISCV] Allow large div peephole optimization for minsize"*
* Fixes peephole optimization of large integers when `minsize` function
attribute is set.
* (Our revive codebase used a workaround that can be removed once this
500 PR is merged and we publish a new LLVM release.)
* [LLVM PR #190219](https://github.com/llvm/llvm-project/pull/190219)
* *"[TargetLowering] Speculative fix for a non-determinism issue between
different compilers"*
* Operand ordering of `mul i256` were different depending on the target
ABI when building LLVM, producing different RISC-V and PVM bytecode for
our Windows vs macOS/linux builds.
* The bytecode divergence (yet functionally equivalent) affected
Solidity contracts containing a value [multiplied by
itself](https://github.com/llvm/llvm-project/issues/190148) (e.g. via
explicit squaring or exponentiation), manifesting primarily when using
optimization level `-O0`.

.github/actions/get-llvm/action.yml

@@ -8,7 +8,7 @@ inputs:
   target:
     required: true
   version:
-    description: "LLVM release version (e.g., 'llvm-18.1.8'). If not specified, uses the latest LLVM release."
+    description: "LLVM release version (e.g., 'llvm-18.1.8'). If not specified, uses the latest published LLVM release. To test a specific draft release, use the full name (e.g. 'llvm-21.1.8-revive.0f80912')."
     required: false
     default: ""
 
@@ -114,7 +114,9 @@ runs:
               core.setFailed(`Artifact for '${target}' not found in release ${llvmRelease.tag_name} (${llvmRelease.html_url})`);
               process.exit();
             }
-            return asset.browser_download_url;
+            // Use the API URL (rather than `browser_download_url`) so that
+            // the authorization header is respected for draft releases.
+            return asset.url;
           }
 
           if (versionPrefix) {
@@ -127,7 +129,8 @@ runs:
     - name: download
       shell: bash
       run: |
-        curl -sSLo llvm.tar.gz ${{ steps.find.outputs.result }}
+        # Pass the authorization token to allow downloading draft release assets.
+        curl -sSLo llvm.tar.gz -H "Authorization: Bearer ${{ github.token }}" -H "Accept: application/octet-stream" ${{ steps.find.outputs.result }}
 
     - name: unpack
       shell: bash

CHANGELOG.md

@@ -11,6 +11,7 @@ This is a development pre-release.
 
 - npm package: The `--bin` CLI flag was not producing bytecode because the `outputSelection` was not updated to explicitly requested it.
 - Yul mode now properly exits when solc reports validation errors (e.g., invalid hex literals), matching `--standard-json` behavior. [#477](https://github.com/paritytech/revive/issues/477)
+- Fixed a bytecode divergence (yet functionally equivalent) between the resolc Linux/macOS builds and the Windows build for Solidity contracts containing a value multiplied by itself (e.g. via explicit squaring or exponentiation), manifesting primarily when using optimization level `-O0`. [#500](https://github.com/paritytech/revive/pull/500)
 
 ### Changed
 

crates/llvm-builder/src/lib.rs

@@ -82,6 +82,8 @@ pub fn build(
         )
     }
 
+    utils::apply_patches()?;
+
     std::fs::create_dir_all(llvm_path::DIRECTORY_LLVM_TARGET.get().unwrap())?;
 
     if cfg!(target_arch = "x86_64") {

crates/llvm-builder/src/llvm_path.rs

@@ -20,6 +20,9 @@ impl LLVMPath {
     /// The Emscripten SDK source directory.
     pub const DIRECTORY_EMSDK_SOURCE: &'static str = "./emsdk/";
 
+    /// The directory containing patches to apply to the LLVM source.
+    pub const DIRECTORY_PATCHES: &'static str = "./patches/llvm/";
+
     /// Returns the path to the `llvm` stage 1 host LLVM source module directory.
     pub fn llvm_host_module_llvm() -> anyhow::Result<PathBuf> {
         let mut path = PathBuf::from(Self::DIRECTORY_LLVM_HOST_SOURCE);

crates/llvm-builder/src/utils.rs

@@ -1,6 +1,6 @@
 //! The LLVM builder utilities.
 
-use std::fs::File;
+use std::fs::{read_dir, File};
 use std::path::Path;
 use std::path::PathBuf;
 use std::process::Command;
@@ -10,12 +10,6 @@ use std::time::Duration;
 use anyhow::Context;
 use path_slash::PathBufExt;
 
-/// The LLVM host repository URL.
-pub const LLVM_HOST_SOURCE_URL: &str = "https://github.com/llvm/llvm-project";
-
-/// The LLVM host repository tag.
-pub const LLVM_HOST_SOURCE_TAG: &str = "llvmorg-18.1.8";
-
 /// The minimum required XCode version.
 pub const XCODE_MIN_VERSION: u32 = 11;
 
@@ -220,6 +214,57 @@ pub fn install_emsdk() -> anyhow::Result<()> {
     Ok(())
 }
 
+/// Applies all LLVM `.patch` files to the LLVM source.
+///
+/// Patches that have already been applied are skipped.
+pub fn apply_patches() -> anyhow::Result<()> {
+    let patches_directory = PathBuf::from(crate::LLVMPath::DIRECTORY_PATCHES);
+    if !patches_directory.exists() {
+        log::debug!("no patches directory found, skipping");
+        return Ok(());
+    }
+
+    let llvm_source = PathBuf::from(crate::LLVMPath::DIRECTORY_LLVM_SOURCE);
+    let mut patches: Vec<_> = read_dir(&patches_directory)?
+        .filter_map(|entry| entry.ok())
+        .map(|entry| entry.path())
+        .filter(|path| {
+            path.extension()
+                .is_some_and(|extension| extension == "patch")
+        })
+        .collect();
+    patches.sort();
+
+    for patch in patches {
+        let patch = absolute_path(&patch)?;
+        let already_applied = Command::new("git")
+            .args(["apply", "--reverse", "--check"])
+            .arg(&patch)
+            .current_dir(&llvm_source)
+            .stdout(Stdio::null())
+            .stderr(Stdio::null())
+            .status()
+            .map(|status| status.success())
+            .unwrap_or(false);
+
+        if already_applied {
+            log::info!("patch already applied, skipping: {}", patch.display());
+            continue;
+        }
+
+        log::info!("applying patch: {}", patch.display());
+        command(
+            Command::new("git")
+                .args(["apply"])
+                .arg(&patch)
+                .current_dir(&llvm_source),
+            &format!("applying patch {}", patch.display()),
+        )?;
+    }
+
+    Ok(())
+}
+
 /// The LLVM target directory default path.
 pub fn directory_target_llvm(target_env: crate::target_env::TargetEnv) -> PathBuf {
     crate::llvm_path::DIRECTORY_LLVM_TARGET

crates/llvm-builder/tests/common.rs

@@ -24,7 +24,7 @@ impl TestDir {
                 "submodule",
                 "add",
                 "-b",
-                "release/18.x",
+                "release/21.x",
                 "https://github.com/llvm/llvm-project.git",
                 "llvm",
             ])

patches/.gitattributes

@@ -0,0 +1 @@
+*.patch eol=lf

patches/llvm/llvm-pr-163679.patch

@@ -0,0 +1,15 @@
+diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+index eb875583ff..169465e18f 100644
+--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
++++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+@@ -24830,7 +24830,8 @@ bool RISCVTargetLowering::isIntDivCheap(EVT VT, AttributeList Attr) const {
+   // instruction, as it is usually smaller than the alternative sequence.
+   // TODO: Add vector division?
+   bool OptSize = Attr.hasFnAttr(Attribute::MinSize);
+-  return OptSize && !VT.isVector();
++  return OptSize && !VT.isVector() &&
++         VT.getSizeInBits() <= getMaxDivRemBitWidthSupported();
+ }
+ 
+ bool RISCVTargetLowering::preferScalarizeSplat(SDNode *N) const {
+

patches/llvm/llvm-pr-190219.patch

@@ -0,0 +1,18 @@
+diff --git a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+index e0597988e..56520d911 100644
+--- a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
++++ b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+@@ -11055,10 +11055,10 @@ void TargetLowering::forceExpandMultiply(SelectionDAG &DAG, const SDLoc &dl,
+   // If HiLHS and HiRHS are set, multiply them by the opposite low part and add
+   // the products to Hi.
+   if (HiLHS) {
++    SDValue RHLL = DAG.getNode(ISD::MUL, dl, VT, HiRHS, LHS);
++    SDValue RLLH = DAG.getNode(ISD::MUL, dl, VT, RHS, HiLHS);
+     Hi = DAG.getNode(ISD::ADD, dl, VT, Hi,
+-                     DAG.getNode(ISD::ADD, dl, VT,
+-                                 DAG.getNode(ISD::MUL, dl, VT, HiRHS, LHS),
+-                                 DAG.getNode(ISD::MUL, dl, VT, RHS, HiLHS)));
++                     DAG.getNode(ISD::ADD, dl, VT, RHLL, RLLH));
+   }
+ }
+