[pch/backpressure_2] Merge changes from main and resolve conflicts

.github/actions/install-rust/action.yml

@@ -5,7 +5,7 @@ inputs:
   toolchain:
     description: 'Default toolchan to install'
     required: false
-    default: '1.71.0'
+    default: 'default'
   lockfiles:
     description: 'Path glob for Cargo.lock files to use as cache keys'
     required: false
@@ -14,13 +14,28 @@ inputs:
 runs:
   using: composite
   steps:
+    - name: Install Rust
+      shell: bash
+      id: select
+      run: |
+        # Determine MSRV as N in `1.N.0` by looking at the `rust-version`
+        # located in the root `Cargo.toml`.
+        msrv=$(grep 'rust-version.*1' Cargo.toml | sed 's/.*\.\([0-9]*\)\..*/\1/')
+
+        if [ "${{ inputs.toolchain }}" = "default" ]; then
+          echo "version=1.$((msrv+2)).0" >> "$GITHUB_OUTPUT"
+        elif [ "${{ inputs.toolchain }}" = "msrv" ]; then
+          echo "version=1.$msrv.0" >> "$GITHUB_OUTPUT"
+        else
+          echo "version=${{ inputs.toolchain }}" >> "$GITHUB_OUTPUT"
+        fi
 
     - name: Install Rust
       shell: bash
       run: |
         rustup set profile minimal
-        rustup update "${{ inputs.toolchain }}" --no-self-update
-        rustup default "${{ inputs.toolchain }}"
+        rustup update "${{ steps.select.outputs.version }}" --no-self-update
+        rustup default "${{ steps.select.outputs.version }}"
 
         # Save disk space by avoiding incremental compilation. Also turn down
         # debuginfo from 2 to 0 to help save disk space.
@@ -31,11 +46,7 @@ runs:
         EOF
 
         # Deny warnings on CI to keep our code warning-free as it lands in-tree.
-        # Don't do this on nightly though, since there's a fair amount of
-        # warning churn there.
-        if [[ "${{ inputs.toolchain }}" != nightly* ]]; then
-          echo RUSTFLAGS="-D warnings" >> "$GITHUB_ENV"
-        fi
+        echo RUSTFLAGS="-D warnings" >> "$GITHUB_ENV"
 
         if [[ "${{ runner.os }}" = "macOS" ]]; then
           cat >> "$GITHUB_ENV" <<EOF

.github/workflows/main.yml

@@ -383,6 +383,8 @@ jobs:
       with:
         submodules: true
     - uses: ./.github/actions/install-rust
+      with:
+        toolchain: ${{ matrix.rust }}
 
     # Install targets in order to build various tests throughout the repo
     - run: rustup target add wasm32-wasi wasm32-unknown-unknown ${{ matrix.target }}

Cargo.toml

@@ -123,7 +123,9 @@ exclude = [
 version = "13.0.0"
 authors = ["The Wasmtime Project Developers"]
 edition = "2021"
-rust-version = "1.71.0"
+# Wasmtime's current policy is that this number can be no larger than the
+# current stable release of Rust minus 2.
+rust-version = "1.70.0"
 
 [workspace.dependencies]
 wasmtime-wmemcheck = { path = "crates/wmemcheck", version = "=13.0.0" }

RELEASES.md

@@ -80,13 +80,6 @@ Unreleased.
   These methods do not affect the size of the pre-allocated pool.
   [#6835](https://github.com/bytecodealliance/wasmtime/pull/6835)
 
-* Options to the `wasmtime` CLI for Wasmtime itself must now come before the
-  WebAssembly module. For example `wasmtime run foo.wasm --disable-cache` now
-  must be specified as `wasmtime run --disable-cache foo.wasm`. Any
-  argument/option after the WebAssembly module is now interpreted as an argument
-  to the wasm module itself.
-  [#6737](https://github.com/bytecodealliance/wasmtime/pull/6737)
-
 * Builder methods for WASI contexts onw use `&mut self` instead of `self`.
   [#6770](https://github.com/bytecodealliance/wasmtime/pull/6770)
 

ci/build-test-matrix.js

@@ -32,13 +32,22 @@ const names = fs.readFileSync(process.argv[3]).toString();
 //   on CI.
 // * `isa` - changes to `cranelift/codegen/src/$isa` will automatically run this
 //   test suite.
+// * `rust` - the Rust version to install, and if unset this'll be set to
+//   `default`
 const array = [
   {
     "os": "ubuntu-latest",
     "name": "Test Linux x86_64",
     "filter": "linux-x64",
     "isa": "x64"
   },
+  {
+    "os": "ubuntu-latest",
+    "name": "Test MSRV on Linux x86_64",
+    "filter": "linux-x64",
+    "isa": "x64",
+    "rust": "msrv",
+  },
   {
     "os": "macos-latest",
     "name": "Test macOS x86_64",
@@ -90,6 +99,12 @@ const array = [
   }
 ];
 
+for (let config of array) {
+  if (config.rust === undefined) {
+    config.rust = 'default';
+  }
+}
+
 function myFilter(item) {
   if (item.isa && names.includes(`cranelift/codegen/src/isa/${item.isa}`)) {
     return true;

cranelift/codegen/src/isa/aarch64/inst.isle

@@ -3122,15 +3122,37 @@
 (rule 5 (amode ty (iadd (sextend x @ (value_type $I32)) y) offset)
         (AMode.RegExtended (amode_add y offset) x (ExtendOp.SXTW)))
 
+;; `RegScaled*` rules where this matches an addition of an "index register" to a
+;; base register. The index register is shifted by the size of the type loaded
+;; in bytes to enable this mode matching.
+;;
+;; Note that this can additionally bundle an extending operation but the
+;; extension must happen before the shift. This will pattern-match the shift
+;; first and then if that succeeds afterwards try to find an extend.
+(rule 6 (amode ty (iadd x (ishl y (iconst (u64_from_imm64 n)))) offset)
+        (if-let $true (u64_eq (ty_bytes ty) (u64_shl 1 n)))
+        (amode_reg_scaled (amode_add x offset) y ty))
+(rule 7 (amode ty (iadd (ishl y (iconst (u64_from_imm64 n))) x) offset)
+        (if-let $true (u64_eq (ty_bytes ty) (u64_shl 1 n)))
+        (amode_reg_scaled (amode_add x offset) y ty))
+
+(decl amode_reg_scaled (Reg Value Type) AMode)
+(rule 0 (amode_reg_scaled base index ty)
+        (AMode.RegScaled base index ty))
+(rule 1 (amode_reg_scaled base (uextend index @ (value_type $I32)) ty)
+        (AMode.RegScaledExtended base index ty (ExtendOp.UXTW)))
+(rule 2 (amode_reg_scaled base (sextend index @ (value_type $I32)) ty)
+        (AMode.RegScaledExtended base index ty (ExtendOp.SXTW)))
+
 ;; Small optimizations where constants found in `iadd` are folded into the
 ;; `offset` immediate.
 ;;
 ;; NB: this should probably be done by mid-end optimizations rather than here
 ;; in the backend, but currently Cranelift doesn't do that.
-(rule 6 (amode ty (iadd x (iconst (simm32 y))) offset)
+(rule 8 (amode ty (iadd x (iconst (simm32 y))) offset)
         (if-let new_offset (s32_add_fallible y offset))
         (amode ty x new_offset))
-(rule 7 (amode ty (iadd (iconst (simm32 x)) y) offset)
+(rule 9 (amode ty (iadd (iconst (simm32 x)) y) offset)
         (if-let new_offset (s32_add_fallible x offset))
         (amode ty y new_offset))
 

cranelift/codegen/src/isa/riscv64/inst/args.rs

@@ -161,6 +161,18 @@ impl AMode {
         }
     }
 
+    /// Retrieve a MachLabel that corresponds to this addressing mode, if it exists.
+    pub(crate) fn get_label_with_sink(&self, sink: &mut MachBuffer<Inst>) -> Option<MachLabel> {
+        match self {
+            &AMode::Const(addr) => Some(sink.get_label_for_constant(addr)),
+            &AMode::Label(label) => Some(label),
+            &AMode::RegOffset(..)
+            | &AMode::SPOffset(..)
+            | &AMode::FPOffset(..)
+            | &AMode::NominalSPOffset(..) => None,
+        }
+    }
+
     pub(crate) fn to_string_with_alloc(&self, allocs: &mut AllocationConsumer<'_>) -> String {
         format!("{}", self.clone().with_allocs(allocs))
     }

cranelift/codegen/src/isa/riscv64/inst/emit.rs

@@ -620,18 +620,49 @@ impl MachInstEmit for Inst {
                 let base = from.get_base_register();
                 let offset = from.get_offset_with_state(state);
                 let offset_imm12 = Imm12::maybe_from_i64(offset);
+                let label = from.get_label_with_sink(sink);
 
-                // TODO: We shouldn't just fall back to `LoadAddr` immediately. For `MachLabel`s
-                // we should try to emit the `auipc` and add a relocation on this load.
-                let (addr, imm12) = match (base, offset_imm12) {
-                    // If the offset fits into an imm12 we can directly encode it.
-                    (Some(base), Some(imm12)) => (base, imm12),
-                    // Otherwise load the address it into a reg and load from it.
-                    _ => {
+                let (addr, imm12) = match (base, offset_imm12, label) {
+                    // When loading from a Reg+Offset, if the offset fits into an imm12 we can directly encode it.
+                    (Some(base), Some(imm12), None) => (base, imm12),
+
+                    // Otherwise, if the offset does not fit into a imm12, we need to materialize it into a
+                    // register and load from that.
+                    (Some(_), None, None) => {
                         let tmp = writable_spilltmp_reg();
                         Inst::LoadAddr { rd: tmp, mem: from }.emit(&[], sink, emit_info, state);
                         (tmp.to_reg(), Imm12::zero())
                     }
+
+                    // If the AMode contains a label we can emit an internal relocation that gets
+                    // resolved with the correct address later.
+                    (None, Some(imm), Some(label)) => {
+                        debug_assert_eq!(imm.as_i16(), 0);
+
+                        // Get the current PC.
+                        sink.use_label_at_offset(sink.cur_offset(), label, LabelUse::PCRelHi20);
+                        Inst::Auipc {
+                            rd,
+                            imm: Imm20::from_bits(0),
+                        }
+                        .emit(&[], sink, emit_info, state);
+
+                        // Emit a relocation for the load. This patches the offset into the instruction.
+                        sink.use_label_at_offset(sink.cur_offset(), label, LabelUse::PCRelLo12I);
+
+                        // Imm12 here is meaningless since it's going to get replaced.
+                        (rd.to_reg(), Imm12::zero())
+                    }
+
+                    // These cases are impossible with the current AModes that we have. We either
+                    // always have a register, or always have a label. Never both, and never neither.
+                    (None, None, None)
+                    | (None, Some(_), None)
+                    | (Some(_), None, Some(_))
+                    | (Some(_), Some(_), Some(_))
+                    | (None, None, Some(_)) => {
+                        unreachable!("Invalid load address")
+                    }
                 };
 
                 let srcloc = state.cur_srcloc();
@@ -650,8 +681,6 @@ impl MachInstEmit for Inst {
                 let offset = to.get_offset_with_state(state);
                 let offset_imm12 = Imm12::maybe_from_i64(offset);
 
-                // TODO: We shouldn't just fall back to `LoadAddr` immediately. For `MachLabel`s
-                // we should try to emit the `auipc` and add a relocation on this store.
                 let (addr, imm12) = match (base, offset_imm12) {
                     // If the offset fits into an imm12 we can directly encode it.
                     (Some(base), Some(imm12)) => (base, imm12),

cranelift/codegen/src/legalizer/mod.rs

@@ -16,7 +16,7 @@
 use crate::cursor::{Cursor, FuncCursor};
 use crate::flowgraph::ControlFlowGraph;
 use crate::ir::immediates::Imm64;
-use crate::ir::types::{I128, I64};
+use crate::ir::types::{self, I128, I64};
 use crate::ir::{self, InstBuilder, InstructionData, MemFlags, Value};
 use crate::isa::TargetIsa;
 use crate::trace;
@@ -38,7 +38,17 @@ fn imm_const(pos: &mut FuncCursor, arg: Value, imm: Imm64, is_signed: bool) -> V
             let imm = pos.ins().iconst(I64, imm);
             pos.ins().uextend(I128, imm)
         }
-        _ => pos.ins().iconst(ty.lane_type(), imm),
+        _ => {
+            let bits = imm.bits();
+            let unsigned = match ty.lane_type() {
+                types::I8 => bits as u8 as i64,
+                types::I16 => bits as u16 as i64,
+                types::I32 => bits as u32 as i64,
+                types::I64 => bits,
+                _ => unreachable!(),
+            };
+            pos.ins().iconst(ty.lane_type(), unsigned)
+        }
     }
 }
 

cranelift/codegen/src/prelude.isle

@@ -87,6 +87,7 @@
 
 (decl pure u16_as_u64 (u16) u64)
 (extern constructor u16_as_u64 u16_as_u64)
+(convert u16 u64 u16_as_u64)
 
 (decl pure u32_as_u64 (u32) u64)
 (extern constructor u32_as_u64 u32_as_u64)