Strip minsize attribute from functions with large div/rem (#390)

That's a workaround to avoid LLVM backend crash while selecting
instruction for 256-bit integer division.
The workaround needs to be removed after we switch to newest inkwell
that has a fix in RISC-V backend

---------

Signed-off-by: kvp <mammal_windier8j@icloud.com>

crates/llvm-context/src/polkavm/context/mod.rs

@@ -8,6 +8,7 @@ use inkwell::debug_info::AsDIScope;
 use inkwell::debug_info::DIScope;
 use inkwell::types::BasicType;
 use inkwell::values::BasicValue;
+use inkwell::values::InstructionOpcode;
 use revive_solc_json_interface::PolkaVMDefaultHeapMemorySize;
 use revive_solc_json_interface::PolkaVMDefaultStackMemorySize;
 use revive_solc_json_interface::SolcStandardJsonInputSettingsPolkaVMMemory;
@@ -288,6 +289,17 @@ impl<'ctx> Context<'ctx> {
                 )
             })?;
 
+        // Remove MinSize on functions that perform large integer div/rem to
+        // avoid compiler crash that happens when large integer div/rem by
+        // power-of-2 are not being expanded by ExpandLargeIntDivRem pass as
+        // it expects peephole from DAGCombine, which doesn't happen due to the
+        // MinSize attribute being set on the function.
+        // NOTE: As soon as it strips attribute from a function where large
+        // integer div/rem is used, it's crucial to call it after inlining.
+        // TODO: Remove this once LLVM fix is backported to LLVM 21 and we
+        // switch to corresponding inkwell version.
+        self.strip_minsize_for_divrem();
+
         self.debug_config
             .dump_llvm_ir_optimized(contract_path, self.module())?;
 
@@ -1394,4 +1406,38 @@ impl<'ctx> Context<'ctx> {
             name.to_string()
         }
     }
+
+    /// Scans all functions in the module and removes the `MinSize` attribute
+    /// if the function contains any large sdiv, udiv, srem, urem instructions with either unknown
+    /// NOTE: The check here could be relaxed by checking denominator: if the denominator is
+    /// unknown or is a power-of-2 constant, then need to strip the `minsize` attribute; otherwise
+    /// instruction can be ignored as backend will expand it correctly.
+    fn strip_minsize_for_divrem(&self) {
+        self.module().get_functions().for_each(|func| {
+            let has_divrem = func.get_basic_block_iter().any(|b| {
+                b.get_instructions().any(|inst| match inst.get_opcode() {
+                    InstructionOpcode::SDiv
+                    | InstructionOpcode::UDiv
+                    | InstructionOpcode::SRem
+                    | InstructionOpcode::URem => {
+                        inst.get_type().into_int_type().get_bit_width() >= 256
+                    }
+                    _ => false,
+                })
+            });
+            if has_divrem
+                && func
+                    .get_enum_attribute(
+                        inkwell::attributes::AttributeLoc::Function,
+                        Attribute::MinSize as u32,
+                    )
+                    .is_some()
+            {
+                func.remove_enum_attribute(
+                    inkwell::attributes::AttributeLoc::Function,
+                    Attribute::MinSize as u32,
+                );
+            }
+        });
+    }
 }

crates/resolc/src/tests/cli/optimization.rs

@@ -2,7 +2,8 @@
 
 use crate::tests::cli::utils::{
     self, assert_command_failure, assert_command_success, assert_equal_exit_codes, execute_resolc,
-    execute_solc, RESOLC_YUL_FLAG, SOLIDITY_CONTRACT_PATH, YUL_MEMSET_CONTRACT_PATH,
+    execute_solc, RESOLC_YUL_FLAG, SOLIDITY_CONTRACT_PATH, SOLIDITY_LARGE_DIV_REM_CONTRACT_PATH,
+    YUL_MEMSET_CONTRACT_PATH,
 };
 
 const LEVELS: &[char] = &['0', '1', '2', '3', 's', 'z'];
@@ -56,3 +57,26 @@ fn disable_solc_optimzer() {
 
     assert_ne!(enabled.stdout, disabled.stdout);
 }
+
+#[test]
+fn test_large_div_rem_expansion() {
+    for level in LEVELS {
+        let optimization_argument = format!("-O{level}");
+        let arguments = &[SOLIDITY_LARGE_DIV_REM_CONTRACT_PATH, &optimization_argument];
+        let resolc_result = utils::execute_resolc(arguments);
+        assert!(
+            resolc_result.success,
+            "Providing the level `{optimization_argument}` should succeed with exit code {}, got {}.\nDetails: {}",
+            revive_common::EXIT_CODE_SUCCESS,
+            resolc_result.code,
+            resolc_result.stderr
+        );
+
+        assert!(
+            resolc_result
+                .stderr
+                .contains("Compiler run successful. No output requested"),
+            "Expected the output to contain a success message when providing the level `{optimization_argument}`."
+        );
+    }
+}

crates/resolc/src/tests/cli/utils.rs

@@ -22,6 +22,10 @@ pub const YUL_MEMSET_CONTRACT_PATH: &str = "src/tests/data/yul/memset.yul";
 pub const STANDARD_JSON_CONTRACTS_PATH: &str =
     "src/tests/data/standard_json/solidity_contracts.json";
 
+/// The simple Solidity contract containing i256 divisions and remains that should be compiled
+/// correctly
+pub const SOLIDITY_LARGE_DIV_REM_CONTRACT_PATH: &str = "src/tests/data/solidity/large_div_rem.sol";
+
 /// The `resolc` YUL mode flag.
 pub const RESOLC_YUL_FLAG: &str = "--yul";
 /// The `--yul` option was deprecated in Solidity 0.8.27 in favor of `--strict-assembly`.

crates/resolc/src/tests/data/solidity/large_div_rem.sol

@@ -0,0 +1,41 @@
+// SPDX-License-Identifier: MIT
+
+pragma solidity ^0.8;
+
+contract LargeDivRem {
+    function rem_2(int n) public pure returns (int q) {
+        assembly {
+            q := smod(n, 2)
+        }
+    }
+
+    function div_2(int n) public pure returns (int q) {
+        assembly {
+            q := sdiv(n, 2)
+        }
+    }
+
+    function rem_7(int n) public pure returns (int q) {
+        assembly {
+            q := smod(n, 7)
+        }
+    }
+
+    function div_7(int n) public pure returns (int q) {
+        assembly {
+            q := sdiv(n, 2)
+        }
+    }
+
+    function rem_k(int n, int k) public pure returns (int q) {
+        assembly {
+            q := smod(n, k)
+        }
+    }
+
+    function div_k(int n, int k) public pure returns (int q) {
+        assembly {
+            q := sdiv(n, k)
+        }
+    }
+}