Merge commit '93a5433f17ab5ed48cc88f1e69b0713b16183373' into sync_cg_clif-2023-10-24

Author: bjorn3

Diff for: .vscode/settings.json

@@ -33,7 +33,7 @@
             ]
         },
         {
-            "sysroot_src": "./download/sysroot/sysroot_src/library",
+            "sysroot_src": "./build/stdlib/library",
             "crates": [
                 {
                     "root_module": "./example/std_example.rs",

Diff for: Cargo.lock

@@ -8,12 +8,12 @@ crate-type = ["dylib"]
 
 [dependencies]
 # These have to be in sync with each other
-cranelift-codegen = { version = "0.101", features = ["unwind", "all-arch"] }
-cranelift-frontend = { version = "0.101" }
-cranelift-module = { version = "0.101" }
-cranelift-native = { version = "0.101" }
-cranelift-jit = { version = "0.101", optional = true }
-cranelift-object = { version = "0.101" }
+cranelift-codegen = { version = "0.101.1", features = ["unwind", "all-arch"] }
+cranelift-frontend = { version = "0.101.1" }
+cranelift-module = { version = "0.101.1" }
+cranelift-native = { version = "0.101.1" }
+cranelift-jit = { version = "0.101.1", optional = true }
+cranelift-object = { version = "0.101.1" }
 target-lexicon = "0.12.0"
 gimli = { version = "0.28", default-features = false, features = ["write"]}
 object = { version = "0.32", default-features = false, features = ["std", "read_core", "write", "archive", "coff", "elf", "macho", "pe"] }

Diff for: Cargo.toml

@@ -8,7 +8,7 @@ If not please open an issue.
 ## Building and testing
 
 ```bash
-$ git clone https://github.com/bjorn3/rustc_codegen_cranelift
+$ git clone https://github.com/rust-lang/rustc_codegen_cranelift
 $ cd rustc_codegen_cranelift
 $ ./y.sh prepare
 $ ./y.sh build
@@ -29,7 +29,7 @@ Extract the `dist` directory in the archive anywhere you want.
 If you want to use `cargo clif build` instead of having to specify the full path to the `cargo-clif` executable, you can add the `bin` subdirectory of the extracted `dist` directory to your `PATH`.
 (tutorial [for Windows](https://stackoverflow.com/a/44272417), and [for Linux/MacOS](https://unix.stackexchange.com/questions/26047/how-to-correctly-add-a-path-to-path/26059#26059)).
 
-[releases]: https://github.com/bjorn3/rustc_codegen_cranelift/releases/tag/dev
+[releases]: https://github.com/rust-lang/rustc_codegen_cranelift/releases/tag/dev
 
 ## Usage
 
@@ -78,7 +78,7 @@ configuration options.
 
 * Inline assembly ([no cranelift support](https://github.com/bytecodealliance/wasmtime/issues/1041))
     * On UNIX there is support for invoking an external assembler for `global_asm!` and `asm!`.
-* SIMD ([tracked here](https://github.com/bjorn3/rustc_codegen_cranelift/issues/171), `std::simd` fully works, `std::arch` is partially supported)
+* SIMD ([tracked here](https://github.com/rust-lang/rustc_codegen_cranelift/issues/171), `std::simd` fully works, `std::arch` is partially supported)
 * Unwinding on panics ([no cranelift support](https://github.com/bytecodealliance/wasmtime/issues/1677), `-Cpanic=abort` is enabled by default)
 
 ## License

Diff for: Readme.md

@@ -353,6 +353,17 @@ fn main() {
 
     let f = V([0.0, 1.0]);
     let _a = f.0[0];
+
+    stack_val_align();
+}
+
+#[inline(never)]
+fn stack_val_align() {
+    #[repr(align(8192))]
+    struct Foo(u8);
+
+    let a = Foo(0);
+    assert_eq!(&a as *const Foo as usize % 8192, 0);
 }
 
 #[cfg(all(

Diff for: example/mini_core_hello_world.rs

@@ -120,32 +120,25 @@ impl<'tcx> FunctionCx<'_, '_, 'tcx> {
         args: &[Value],
     ) -> Cow<'_, [Value]> {
         if self.tcx.sess.target.is_like_windows {
-            let (mut params, mut args): (Vec<_>, Vec<_>) =
-                params
-                    .into_iter()
-                    .zip(args)
-                    .map(|(param, &arg)| {
-                        if param.value_type == types::I128 {
-                            let arg_ptr = Pointer::stack_slot(self.bcx.create_sized_stack_slot(
-                                StackSlotData { kind: StackSlotKind::ExplicitSlot, size: 16 },
-                            ));
-                            arg_ptr.store(self, arg, MemFlags::trusted());
-                            (AbiParam::new(self.pointer_type), arg_ptr.get_addr(self))
-                        } else {
-                            (param, arg)
-                        }
-                    })
-                    .unzip();
+            let (mut params, mut args): (Vec<_>, Vec<_>) = params
+                .into_iter()
+                .zip(args)
+                .map(|(param, &arg)| {
+                    if param.value_type == types::I128 {
+                        let arg_ptr = self.create_stack_slot(16, 16);
+                        arg_ptr.store(self, arg, MemFlags::trusted());
+                        (AbiParam::new(self.pointer_type), arg_ptr.get_addr(self))
+                    } else {
+                        (param, arg)
+                    }
+                })
+                .unzip();
 
             let indirect_ret_val = returns.len() == 1 && returns[0].value_type == types::I128;
 
             if indirect_ret_val {
                 params.insert(0, AbiParam::new(self.pointer_type));
-                let ret_ptr =
-                    Pointer::stack_slot(self.bcx.create_sized_stack_slot(StackSlotData {
-                        kind: StackSlotKind::ExplicitSlot,
-                        size: 16,
-                    }));
+                let ret_ptr = self.create_stack_slot(16, 16);
                 args.insert(0, ret_ptr.get_addr(self));
                 self.lib_call_unadjusted(name, params, vec![], &args);
                 return Cow::Owned(vec![ret_ptr.load(self, types::I128, MemFlags::trusted())]);

Diff for: patches/0001-regex-Ignore-test-which-gets-miscompiled-with-llvm-sysroot.patch

@@ -189,16 +189,13 @@ pub(super) fn from_casted_value<'tcx>(
     let abi_params = cast_target_to_abi_params(cast);
     let abi_param_size: u32 = abi_params.iter().map(|param| param.value_type.bytes()).sum();
     let layout_size = u32::try_from(layout.size.bytes()).unwrap();
-    let stack_slot = fx.bcx.create_sized_stack_slot(StackSlotData {
-        kind: StackSlotKind::ExplicitSlot,
-        // FIXME Don't force the size to a multiple of 16 bytes once Cranelift gets a way to
-        // specify stack slot alignment.
+    let ptr = fx.create_stack_slot(
         // Stack slot size may be bigger for example `[u8; 3]` which is packed into an `i32`.
         // It may also be smaller for example when the type is a wrapper around an integer with a
         // larger alignment than the integer.
-        size: (std::cmp::max(abi_param_size, layout_size) + 15) / 16 * 16,
-    });
-    let ptr = Pointer::stack_slot(stack_slot);
+        std::cmp::max(abi_param_size, layout_size),
+        u32::try_from(layout.align.pref.bytes()).unwrap(),
+    );
     let mut offset = 0;
     let mut block_params_iter = block_params.iter().copied();
     for param in abi_params {

Diff for: src/abi/mod.rs

@@ -104,11 +104,7 @@ pub(crate) fn clif_int_or_float_cast(
                     &[from],
                 )[0];
                 // FIXME(bytecodealliance/wasmtime#6104) use bitcast instead of store to get from i64x2 to i128
-                let stack_slot = fx.bcx.create_sized_stack_slot(StackSlotData {
-                    kind: StackSlotKind::ExplicitSlot,
-                    size: 16,
-                });
-                let ret_ptr = Pointer::stack_slot(stack_slot);
+                let ret_ptr = fx.create_stack_slot(16, 16);
                 ret_ptr.store(fx, ret, MemFlags::trusted());
                 ret_ptr.load(fx, types::I128, MemFlags::trusted())
             } else {

Diff for: src/abi/pass_mode.rs

@@ -383,6 +383,25 @@ impl<'tcx> FunctionCx<'_, '_, 'tcx> {
         })
     }
 
+    pub(crate) fn create_stack_slot(&mut self, size: u32, align: u32) -> Pointer {
+        if align <= 16 {
+            let stack_slot = self.bcx.create_sized_stack_slot(StackSlotData {
+                kind: StackSlotKind::ExplicitSlot,
+                // FIXME Don't force the size to a multiple of 16 bytes once Cranelift gets a way to
+                // specify stack slot alignment.
+                size: (size + 15) / 16 * 16,
+            });
+            Pointer::stack_slot(stack_slot)
+        } else {
+            // Alignment is too big to handle using the above hack. Dynamically realign a stack slot
+            // instead. This wastes some space for the realignment.
+            let base_ptr = self.create_stack_slot(size + align, 16).get_addr(self);
+            let misalign_offset = self.bcx.ins().urem_imm(base_ptr, i64::from(align));
+            let realign_offset = self.bcx.ins().irsub_imm(misalign_offset, i64::from(align));
+            Pointer::new(self.bcx.ins().iadd(base_ptr, realign_offset))
+        }
+    }
+
     pub(crate) fn set_debug_loc(&mut self, source_info: mir::SourceInfo) {
         if let Some(debug_context) = &mut self.cx.debug_context {
             let (file, line, column) =

Diff for: src/cast.rs

@@ -361,12 +361,26 @@ pub(crate) fn run_aot(
     metadata: EncodedMetadata,
     need_metadata_module: bool,
 ) -> Box<OngoingCodegen> {
+    // FIXME handle `-Ctarget-cpu=native`
+    let target_cpu = match tcx.sess.opts.cg.target_cpu {
+        Some(ref name) => name,
+        None => tcx.sess.target.cpu.as_ref(),
+    }
+    .to_owned();
+
     let cgus = if tcx.sess.opts.output_types.should_codegen() {
         tcx.collect_and_partition_mono_items(()).1
     } else {
         // If only `--emit metadata` is used, we shouldn't perform any codegen.
         // Also `tcx.collect_and_partition_mono_items` may panic in that case.
-        &[]
+        return Box::new(OngoingCodegen {
+            modules: vec![],
+            allocator_module: None,
+            metadata_module: None,
+            metadata,
+            crate_info: CrateInfo::new(tcx, target_cpu),
+            concurrency_limiter: ConcurrencyLimiter::new(tcx.sess, 0),
+        });
     };
 
     if tcx.dep_graph.is_fully_enabled() {
@@ -481,13 +495,6 @@ pub(crate) fn run_aot(
         None
     };
 
-    // FIXME handle `-Ctarget-cpu=native`
-    let target_cpu = match tcx.sess.opts.cg.target_cpu {
-        Some(ref name) => name,
-        None => tcx.sess.target.cpu.as_ref(),
-    }
-    .to_owned();
-
     Box::new(OngoingCodegen {
         modules,
         allocator_module,