Rollup of 25 pull requests

Cargo.lock

@@ -6320,9 +6320,9 @@ dependencies = [
 
 [[package]]
 name = "wasm-component-ld"
-version = "0.5.27"
+version = "0.5.30"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5fd1ba338b2c06654988a76b13f38cad9a1c738e3275bf71ecf70eeac5d98eb4"
+checksum = "baf08212b996e5acc1338cb354247de0d7cd82ada2e6568cdc5a662a998fddbd"
 dependencies = [
  "anyhow",
  "clap",
@@ -6331,7 +6331,7 @@ dependencies = [
  "libc",
  "tempfile",
  "wasi-preview1-component-adapter-provider",
- "wasmparser 0.254.0",
+ "wasmparser 0.257.1",
  "wat",
  "windows-sys 0.61.2",
  "winsplit",
@@ -6358,24 +6358,24 @@ dependencies = [
 
 [[package]]
 name = "wasm-encoder"
-version = "0.254.0"
+version = "0.257.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "09480d646178e5fdd12bb06e812d0af9a3a191dbc9cd697fdc86687beade7393"
+checksum = "7d8ad9f0a39050867bda22e6486c316e2e52d20a42154d5bc433934bf738d085"
 dependencies = [
  "leb128fmt",
- "wasmparser 0.254.0",
+ "wasmparser 0.257.1",
 ]
 
 [[package]]
 name = "wasm-metadata"
-version = "0.254.0"
+version = "0.257.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b01df5f3b4ca7881e843f3bc0fb8a3905d79c68692250dcb8e33e698705ccdb6"
+checksum = "57460ad9bce753e8a80d47a445cb87c8724ea57f463d9c906a947c41032ce1ac"
 dependencies = [
  "anyhow",
  "indexmap",
- "wasm-encoder 0.254.0",
- "wasmparser 0.254.0",
+ "wasm-encoder 0.257.1",
+ "wasmparser 0.257.1",
 ]
 
 [[package]]
@@ -6391,9 +6391,9 @@ dependencies = [
 
 [[package]]
 name = "wasmparser"
-version = "0.254.0"
+version = "0.257.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d5769a29f799fbab136aaf65b4fe5384cd7d93fe6fc9ba0dcb6c8382a1f16e27"
+checksum = "d92fc335fb6d48f46bda1d8b26b69e28320c15ac3272208333833d6e217e2b4a"
 dependencies = [
  "bitflags",
  "hashbrown",
@@ -6404,22 +6404,22 @@ dependencies = [
 
 [[package]]
 name = "wast"
-version = "254.0.0"
+version = "257.0.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e7ed4dfc8f6b9fc38b231065e2cdfbf7359af5ab945990abf09658dcc63c3e32"
+checksum = "3c416ee8044cf88c5280e3d87c322b009525eb25142cf7428f0a90ea88febc3b"
 dependencies = [
  "bumpalo",
  "leb128fmt",
  "memchr",
  "unicode-width 0.2.2",
- "wasm-encoder 0.254.0",
+ "wasm-encoder 0.257.1",
 ]
 
 [[package]]
 name = "wat"
-version = "1.254.0"
+version = "1.257.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7127f7f9b8f127c879991cecd35f494e4628bae1b0874c681414d8d8831e952c"
+checksum = "672e5495f00c5fc6c6b502205bae4708aa4dd13d969754438743afe0a8af0f1d"
 dependencies = [
  "wast",
 ]
@@ -6791,9 +6791,9 @@ dependencies = [
 
 [[package]]
 name = "wit-component"
-version = "0.254.0"
+version = "0.257.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b0e65bb94c369b3c4741ce3d1d2704b1fec93db7c540df0e521a097e7ceeb5be"
+checksum = "2e8b7d04a4c9491ffea354923ed70c8826d52c699fe20a52e8ceca95017dddb8"
 dependencies = [
  "anyhow",
  "bitflags",
@@ -6802,17 +6802,17 @@ dependencies = [
  "serde",
  "serde_derive",
  "serde_json",
- "wasm-encoder 0.254.0",
+ "wasm-encoder 0.257.1",
  "wasm-metadata",
- "wasmparser 0.254.0",
+ "wasmparser 0.257.1",
  "wit-parser",
 ]
 
 [[package]]
 name = "wit-parser"
-version = "0.254.0"
+version = "0.257.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1655131e4f7d3f0cb141f6eca71315ca40eff0f3d4de7cff0a82bacedd8c89b4"
+checksum = "6f815340f0bb65d9775ae21e56b503b496683557b9b627b195d2766c25fb24ae"
 dependencies = [
  "anyhow",
  "hashbrown",
@@ -6824,7 +6824,7 @@ dependencies = [
  "serde_derive",
  "serde_json",
  "unicode-ident",
- "wasmparser 0.254.0",
+ "wasmparser 0.257.1",
 ]
 
 [[package]]

compiler/rustc_codegen_llvm/src/builder/gpu_offload.rs

@@ -197,6 +197,20 @@ fn generate_launcher<'ll>(cx: &CodegenCx<'ll, '_>) -> (&'ll llvm::Value, &'ll ll
     (tgt_decl, tgt_fn_ty)
 }
 
+/// Declares the `omp_get_num_devices` runtime function and returns the
+/// declaration together with its type.
+pub(crate) fn declare_omp_get_num_devices<'ll>(
+    cx: &CodegenCx<'ll, '_>,
+) -> (&'ll llvm::Value, &'ll llvm::Type) {
+    let ti32 = cx.type_i32();
+    let tgt_fn_ty = cx.type_func(&[], ti32);
+    let name = "omp_get_num_devices";
+    let tgt_decl = declare_offload_fn(&cx, name, tgt_fn_ty);
+    let nounwind = llvm::AttributeKind::NoUnwind.create_attr(cx.llcx);
+    attributes::apply_to_llfn(tgt_decl, Function, &[nounwind]);
+    (tgt_decl, tgt_fn_ty)
+}
+
 // What is our @1 here? A magic global, used in our data_{begin/update/end}_mapper:
 // @0 = private unnamed_addr constant [23 x i8] c";unknown;unknown;0;0;;\00", align 1
 // @1 = private unnamed_addr constant %struct.ident_t { i32 0, i32 2, i32 0, i32 22, ptr @0 }, align 8
@@ -591,6 +605,7 @@ pub(crate) fn gen_call_handling<'ll, 'tcx>(
     offload_globals: &OffloadGlobals<'ll>,
     offload_dims: &OffloadKernelDims<'ll>,
     dyn_cache: &'ll Value,
+    device_id: &'ll Value,
 ) {
     let cx = builder.cx;
     let OffloadKernelGlobals {
@@ -775,15 +790,8 @@ pub(crate) fn gen_call_handling<'ll, 'tcx>(
         builder.store(value.2, ptr, value.0);
     }
 
-    let args = vec![
-        s_ident_t,
-        // FIXME(offload) give users a way to select which GPU to use.
-        cx.get_const_i64(u64::MAX), // MAX == -1.
-        num_workgroups,
-        threads_per_block,
-        region_id,
-        a5,
-    ];
+    let device_id = builder.sext(device_id, cx.type_i64());
+    let args = vec![s_ident_t, device_id, num_workgroups, threads_per_block, region_id, a5];
     builder.call(tgt_target_kernel_ty, None, None, tgt_decl, &args, None, None);
     // %41 = call i32 @__tgt_target_kernel(ptr @1, i64 -1, i32 2097152, i32 256, ptr @.kernel_1.region_id, ptr %kernel_args)
 

compiler/rustc_codegen_llvm/src/intrinsic.rs

@@ -37,7 +37,7 @@ use crate::abi::FnAbiLlvmExt;
 use crate::builder::Builder;
 use crate::builder::autodiff::{adjust_activity_to_abi, generate_enzyme_call};
 use crate::builder::gpu_offload::{
-    OffloadKernelDims, gen_call_handling, gen_define_handling, register_offload,
+    self, OffloadKernelDims, declare_omp_get_num_devices, register_offload,
 };
 use crate::context::CodegenCx;
 use crate::declare::declare_raw_fn;
@@ -241,6 +241,13 @@ impl<'ll, 'tcx> IntrinsicCallBuilderMethods<'tcx> for Builder<'_, 'll, 'tcx> {
                 // offload *has* a return type, but somehow works without mentioning the place
                 return IntrinsicResult::WroteIntoPlace;
             }
+            sym::offload_get_num_devices => {
+                let (fn_decl, fn_ty) = declare_omp_get_num_devices(self.cx);
+
+                let llval = self.call(fn_ty, None, None, fn_decl, &[], None, None);
+
+                return IntrinsicResult::Operand(OperandValue::Immediate(llval));
+            },
             sym::is_val_statically_known => {
                 if let OperandValue::Immediate(imm) = args[0].val {
                     self.call_intrinsic(
@@ -1851,7 +1858,11 @@ fn codegen_offload<'ll, 'tcx>(
         OperandValue::Immediate(val) => val,
         _ => panic!("unparsable"),
     };
-    let args = get_args_from_tuple(bx, args[4], fn_target);
+    let device_id = match args[4].val {
+        OperandValue::Immediate(val) => val,
+        _ => panic!("unparsable"),
+    };
+    let args = get_args_from_tuple(bx, args[5], fn_target);
     let target_symbol = mangle_offload_export(tcx, fn_target);
 
     let sig = tcx.fn_sig(fn_target.def_id()).instantiate(tcx, fn_target.args).skip_norm_wip();
@@ -1882,8 +1893,9 @@ fn codegen_offload<'ll, 'tcx>(
         }
     };
     register_offload(cx);
-    let offload_data = gen_define_handling(&cx, &metadata, target_symbol, offload_globals);
-    gen_call_handling(
+    let offload_data =
+        gpu_offload::gen_define_handling(&cx, &metadata, target_symbol, offload_globals);
+    gpu_offload::gen_call_handling(
         bx,
         &offload_data,
         &args,
@@ -1892,6 +1904,7 @@ fn codegen_offload<'ll, 'tcx>(
         offload_globals,
         &offload_dims,
         &dyn_cache,
+        &device_id,
     );
 }
 

compiler/rustc_codegen_ssa/src/back/archive.rs

@@ -623,8 +623,9 @@ impl<'a> ArArchiveBuilder<'a> {
                             io::Error::new(
                                 io::ErrorKind::InvalidData,
                                 format!(
-                                    "archive member at offset {start} with size {} \
+                                    "archive member of {} at offset {start} with size {} \
                                          exceeds archive size {} in `{}`",
+                                    src_archive.0.display(),
                                     file_range.1,
                                     archive_data.len(),
                                     src_archive.0.display(),
@@ -642,11 +643,18 @@ impl<'a> ArArchiveBuilder<'a> {
                     }
                 }
                 ArchiveEntrySource::File(file) => unsafe {
-                    let mmap = Mmap::map(
-                        File::open(file)
-                            .map_err(|err| io_error_context("failed to open object file", err))?,
-                    )
-                    .map_err(|err| io_error_context("failed to map object file", err))?;
+                    let mmap = Mmap::map(File::open(&file).map_err(|err| {
+                        io_error_context(
+                            &format!("failed to open object file {}", file.display()),
+                            err,
+                        )
+                    })?)
+                    .map_err(|err| {
+                        io_error_context(
+                            &format!("failed to map object file {}", file.display()),
+                            err,
+                        )
+                    })?;
                     if entry.kind == ArchiveEntryKind::RustObj
                         && let Some(sym) = &symbols
                     {

compiler/rustc_codegen_ssa/src/mir/intrinsic.rs

@@ -135,6 +135,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                 | sym::atomic_fence
                 | sym::atomic_singlethreadfence
                 | sym::caller_location
+                | sym::offload_get_num_devices
                 | sym::return_address => {}
                 _ => {
                     span_bug!(