Introduce log2_min_function_alignment flag

cranelift/codegen/meta/src/shared/settings.rs

@@ -398,6 +398,13 @@ pub(crate) fn define() -> SettingGroup {
         0,
     );
 
+    settings.add_num(
+        "log2_min_function_alignment",
+        "The log2 of the minimum alignment of functions",
+        "The bigger of this value and the default alignment will be used as actual alignment.",
+        0,
+    );
+
     // When adding new settings please check if they can also be added
     // in cranelift/fuzzgen/src/lib.rs for fuzzing.
     settings.build()

cranelift/codegen/src/machinst/buffer.rs

@@ -239,6 +239,8 @@ enum ForceVeneers {
 pub struct MachBuffer<I: VCodeInst> {
     /// The buffer contents, as raw bytes.
     data: SmallVec<[u8; 1024]>,
+    /// The required alignment of this buffer.
+    min_alignment: u32,
     /// Any relocations referring to this code. Note that only *external*
     /// relocations are tracked here; references to labels within the buffer are
     /// resolved before emission.
@@ -433,6 +435,7 @@ impl<I: VCodeInst> MachBuffer<I> {
     pub fn new() -> MachBuffer<I> {
         MachBuffer {
             data: SmallVec::new(),
+            min_alignment: I::function_alignment().minimum,
             relocs: SmallVec::new(),
             traps: SmallVec::new(),
             call_sites: SmallVec::new(),
@@ -607,7 +610,7 @@ impl<I: VCodeInst> MachBuffer<I> {
     /// at the ISA's minimum function alignment and can be increased due to
     /// constant requirements.
     fn finish_constants(&mut self, constants: &VCodeConstants) -> u32 {
-        let mut alignment = I::function_alignment().minimum;
+        let mut alignment = self.min_alignment;
         for (constant, offset) in mem::take(&mut self.used_constants) {
             let constant = constants.get(constant);
             let data = constant.as_slice();
@@ -1665,6 +1668,15 @@ impl<I: VCodeInst> MachBuffer<I> {
         stack_map.finalize(emit_state.frame_layout().sp_to_sized_stack_slots());
         self.user_stack_maps.push((return_addr, span, stack_map));
     }
+
+    /// Increase the alignment of the buffer to the given alignment if bigger
+    /// than the current alignment.
+    pub fn set_log2_min_function_alignment(&mut self, align_to: u8) {
+        self.min_alignment = self.min_alignment.max(
+            1u32.checked_shl(u32::from(align_to))
+                .expect("log2_min_function_alignment too large"),
+        );
+    }
 }
 
 impl<I: VCodeInst> Extend<u8> for MachBuffer<I> {

cranelift/codegen/src/machinst/lower.rs

@@ -382,6 +382,7 @@ impl<'func, I: VCodeInst> Lower<'func, I> {
             block_order,
             constants,
             VCodeBuildDirection::Backward,
+            flags.log2_min_function_alignment(),
         );
 
         // We usually need two VRegs per instruction result, plus extras for

cranelift/codegen/src/machinst/vcode.rs

@@ -195,6 +195,8 @@ pub struct VCode<I: VCodeInst> {
 
     /// Facts on VRegs, for proof-carrying code verification.
     facts: Vec<Option<Fact>>,
+
+    log2_min_function_alignment: u8,
 }
 
 /// The result of `VCode::emit`. Contains all information computed
@@ -281,8 +283,16 @@ impl<I: VCodeInst> VCodeBuilder<I> {
         block_order: BlockLoweringOrder,
         constants: VCodeConstants,
         direction: VCodeBuildDirection,
+        log2_min_function_alignment: u8,
     ) -> Self {
-        let vcode = VCode::new(sigs, abi, emit_info, block_order, constants);
+        let vcode = VCode::new(
+            sigs,
+            abi,
+            emit_info,
+            block_order,
+            constants,
+            log2_min_function_alignment,
+        );
 
         VCodeBuilder {
             vcode,
@@ -602,6 +612,7 @@ impl<I: VCodeInst> VCode<I> {
         emit_info: I::Info,
         block_order: BlockLoweringOrder,
         constants: VCodeConstants,
+        log2_min_function_alignment: u8,
     ) -> Self {
         let n_blocks = block_order.lowered_order().len();
         VCode {
@@ -630,6 +641,7 @@ impl<I: VCodeInst> VCode<I> {
             constants,
             debug_value_labels: vec![],
             facts: vec![],
+            log2_min_function_alignment,
         }
     }
 
@@ -714,6 +726,7 @@ impl<I: VCodeInst> VCode<I> {
 
         let _tt = timing::vcode_emit();
         let mut buffer = MachBuffer::new();
+        buffer.set_log2_min_function_alignment(self.log2_min_function_alignment);
         let mut bb_starts: Vec<Option<CodeOffset>> = vec![];
 
         // The first M MachLabels are reserved for block indices.

cranelift/codegen/src/settings.rs

@@ -517,6 +517,7 @@ libcall_call_conv = "isa_default"
 probestack_size_log2 = 12
 probestack_strategy = "outline"
 bb_padding_log2_minus_one = 0
+log2_min_function_alignment = 0
 regalloc_checker = false
 regalloc_verbose_logs = false
 enable_alias_analysis = true

cranelift/jit/src/backend.rs

@@ -641,9 +641,7 @@ impl Module for JITModule {
         let compiled_code = ctx.compiled_code().unwrap();
 
         let size = compiled_code.code_info().total_size as usize;
-        let align = alignment
-            .max(self.isa.function_alignment().minimum as u64)
-            .max(self.isa.symbol_alignment());
+        let align = alignment.max(self.isa.symbol_alignment());
         let ptr = self
             .memory
             .code
@@ -702,9 +700,7 @@ impl Module for JITModule {
         }
 
         let size = bytes.len();
-        let align = alignment
-            .max(self.isa.function_alignment().minimum as u64)
-            .max(self.isa.symbol_alignment());
+        let align = alignment.max(self.isa.symbol_alignment());
         let ptr = self
             .memory
             .code

cranelift/object/src/backend.rs

@@ -495,9 +495,7 @@ impl ObjectModule {
         }
         *defined = true;
 
-        let align = alignment
-            .max(self.isa.function_alignment().minimum.into())
-            .max(self.isa.symbol_alignment());
+        let align = alignment.max(self.isa.symbol_alignment());
         let section = if self.per_function_section {
             // FIXME pass empty symbol name once add_subsection produces `.text` as section name
             // instead of `.text.` when passed an empty symbol name. (object#748) Until then pass

crates/wasmtime/src/engine.rs

@@ -392,6 +392,7 @@ impl Engine {
             | "regalloc_algorithm"
             | "is_pic"
             | "bb_padding_log2_minus_one"
+            | "log2_min_function_alignment"
             | "machine_code_cfg_info"
             | "tls_model" // wasmtime doesn't use tls right now
             | "stack_switch_model" // wasmtime doesn't use stack switching right now