diff options
Diffstat (limited to 'compiler/rustc_codegen_llvm/src')
| -rw-r--r-- | compiler/rustc_codegen_llvm/src/attributes.rs | 51 | ||||
| -rw-r--r-- | compiler/rustc_codegen_llvm/src/back/lto.rs | 14 | ||||
| -rw-r--r-- | compiler/rustc_codegen_llvm/src/builder.rs | 29 | ||||
| -rw-r--r-- | compiler/rustc_codegen_llvm/src/builder/gpu_offload.rs | 2 | ||||
| -rw-r--r-- | compiler/rustc_codegen_llvm/src/context.rs | 14 | ||||
| -rw-r--r-- | compiler/rustc_codegen_llvm/src/intrinsic.rs | 26 |
6 files changed, 97 insertions, 39 deletions
diff --git a/compiler/rustc_codegen_llvm/src/attributes.rs b/compiler/rustc_codegen_llvm/src/attributes.rs index 5affb26483a..573c51a9539 100644 --- a/compiler/rustc_codegen_llvm/src/attributes.rs +++ b/compiler/rustc_codegen_llvm/src/attributes.rs @@ -29,8 +29,18 @@ pub(crate) fn apply_to_callsite(callsite: &Value, idx: AttributePlace, attrs: &[ } /// Get LLVM attribute for the provided inline heuristic. -#[inline] -fn inline_attr<'ll>(cx: &CodegenCx<'ll, '_>, inline: InlineAttr) -> Option<&'ll Attribute> { +pub(crate) fn inline_attr<'ll, 'tcx>( + cx: &CodegenCx<'ll, 'tcx>, + instance: ty::Instance<'tcx>, +) -> Option<&'ll Attribute> { + // `optnone` requires `noinline` + let codegen_fn_attrs = cx.tcx.codegen_fn_attrs(instance.def_id()); + let inline = match (codegen_fn_attrs.inline, &codegen_fn_attrs.optimize) { + (_, OptimizeAttr::DoNotOptimize) => InlineAttr::Never, + (InlineAttr::None, _) if instance.def.requires_inline(cx.tcx) => InlineAttr::Hint, + (inline, _) => inline, + }; + if !cx.tcx.sess.opts.unstable_opts.inline_llvm { // disable LLVM inlining return Some(AttributeKind::NoInline.create_attr(cx.llcx)); @@ -286,6 +296,19 @@ pub(crate) fn tune_cpu_attr<'ll>(cx: &CodegenCx<'ll, '_>) -> Option<&'ll Attribu .map(|tune_cpu| llvm::CreateAttrStringValue(cx.llcx, "tune-cpu", tune_cpu)) } +/// Get the `target-features` LLVM attribute. +pub(crate) fn target_features_attr<'ll>( + cx: &CodegenCx<'ll, '_>, + function_features: Vec<String>, +) -> Option<&'ll Attribute> { + let global_features = cx.tcx.global_backend_features(()).iter().map(String::as_str); + let function_features = function_features.iter().map(String::as_str); + let target_features = + global_features.chain(function_features).intersperse(",").collect::<String>(); + (!target_features.is_empty()) + .then(|| llvm::CreateAttrStringValue(cx.llcx, "target-features", &target_features)) +} + /// Get the `NonLazyBind` LLVM attribute, /// if the codegen options allow skipping the PLT. pub(crate) fn non_lazy_bind_attr<'ll>(cx: &CodegenCx<'ll, '_>) -> Option<&'ll Attribute> { @@ -346,14 +369,6 @@ pub(crate) fn llfn_attrs_from_instance<'ll, 'tcx>( OptimizeAttr::Speed => {} } - // `optnone` requires `noinline` - let inline = match (codegen_fn_attrs.inline, &codegen_fn_attrs.optimize) { - (_, OptimizeAttr::DoNotOptimize) => InlineAttr::Never, - (InlineAttr::None, _) if instance.def.requires_inline(cx.tcx) => InlineAttr::Hint, - (inline, _) => inline, - }; - to_add.extend(inline_attr(cx, inline)); - if cx.sess().must_emit_unwind_tables() { to_add.push(uwtable_attr(cx.llcx, cx.sess().opts.unstable_opts.use_sync_unwind)); } @@ -488,6 +503,14 @@ pub(crate) fn llfn_attrs_from_instance<'ll, 'tcx>( let function_features = codegen_fn_attrs.target_features.iter().map(|f| f.name.as_str()).collect::<Vec<&str>>(); + // Apply function attributes as per usual if there are no user defined + // target features otherwise this will get applied at the callsite. + if function_features.is_empty() { + if let Some(inline_attr) = inline_attr(cx, instance) { + to_add.push(inline_attr); + } + } + let function_features = function_features .iter() // Convert to LLVMFeatures and filter out unavailable ones @@ -513,13 +536,7 @@ pub(crate) fn llfn_attrs_from_instance<'ll, 'tcx>( } } - let global_features = cx.tcx.global_backend_features(()).iter().map(|s| s.as_str()); - let function_features = function_features.iter().map(|s| s.as_str()); - let target_features: String = - global_features.chain(function_features).intersperse(",").collect(); - if !target_features.is_empty() { - to_add.push(llvm::CreateAttrStringValue(cx.llcx, "target-features", &target_features)); - } + to_add.extend(target_features_attr(cx, function_features)); attributes::apply_to_llfn(llfn, Function, &to_add); } diff --git a/compiler/rustc_codegen_llvm/src/back/lto.rs b/compiler/rustc_codegen_llvm/src/back/lto.rs index fc38c4f3e51..ad2e722cfef 100644 --- a/compiler/rustc_codegen_llvm/src/back/lto.rs +++ b/compiler/rustc_codegen_llvm/src/back/lto.rs @@ -43,9 +43,7 @@ fn prepare_lto( .map(|symbol| CString::new(symbol.to_owned()).unwrap()) .collect::<Vec<CString>>(); - if cgcx.regular_module_config.instrument_coverage - || cgcx.regular_module_config.pgo_gen.enabled() - { + if cgcx.module_config.instrument_coverage || cgcx.module_config.pgo_gen.enabled() { // These are weak symbols that point to the profile version and the // profile name, which need to be treated as exported so LTO doesn't nix // them. @@ -55,15 +53,15 @@ fn prepare_lto( symbols_below_threshold.extend(PROFILER_WEAK_SYMBOLS.iter().map(|&sym| sym.to_owned())); } - if cgcx.regular_module_config.sanitizer.contains(SanitizerSet::MEMORY) { + if cgcx.module_config.sanitizer.contains(SanitizerSet::MEMORY) { let mut msan_weak_symbols = Vec::new(); // Similar to profiling, preserve weak msan symbol during LTO. - if cgcx.regular_module_config.sanitizer_recover.contains(SanitizerSet::MEMORY) { + if cgcx.module_config.sanitizer_recover.contains(SanitizerSet::MEMORY) { msan_weak_symbols.push(c"__msan_keep_going"); } - if cgcx.regular_module_config.sanitizer_memory_track_origins != 0 { + if cgcx.module_config.sanitizer_memory_track_origins != 0 { msan_weak_symbols.push(c"__msan_track_origins"); } @@ -583,7 +581,7 @@ pub(crate) fn run_pass_manager( thin: bool, ) { let _timer = cgcx.prof.generic_activity_with_arg("LLVM_lto_optimize", &*module.name); - let config = cgcx.config(module.kind); + let config = &cgcx.module_config; // Now we have one massive module inside of llmod. Time to run the // LTO-specific optimization passes that LLVM provides. @@ -745,7 +743,7 @@ pub(crate) fn optimize_thin_module( let module_llvm = ModuleLlvm::parse(cgcx, module_name, thin_module.data(), dcx); let mut module = ModuleCodegen::new_regular(thin_module.name(), module_llvm); // Given that the newly created module lacks a thinlto buffer for embedding, we need to re-add it here. - if cgcx.config(ModuleKind::Regular).embed_bitcode() { + if cgcx.module_config.embed_bitcode() { module.thin_lto_buffer = Some(thin_module.data().to_vec()); } { diff --git a/compiler/rustc_codegen_llvm/src/builder.rs b/compiler/rustc_codegen_llvm/src/builder.rs index 37379586d58..7d0691366e6 100644 --- a/compiler/rustc_codegen_llvm/src/builder.rs +++ b/compiler/rustc_codegen_llvm/src/builder.rs @@ -1392,7 +1392,7 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> { fn call( &mut self, llty: &'ll Type, - fn_attrs: Option<&CodegenFnAttrs>, + fn_call_attrs: Option<&CodegenFnAttrs>, fn_abi: Option<&FnAbi<'tcx, Ty<'tcx>>>, llfn: &'ll Value, args: &[&'ll Value], @@ -1409,10 +1409,10 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> { } // Emit CFI pointer type membership test - self.cfi_type_test(fn_attrs, fn_abi, instance, llfn); + self.cfi_type_test(fn_call_attrs, fn_abi, instance, llfn); // Emit KCFI operand bundle - let kcfi_bundle = self.kcfi_operand_bundle(fn_attrs, fn_abi, instance, llfn); + let kcfi_bundle = self.kcfi_operand_bundle(fn_call_attrs, fn_abi, instance, llfn); if let Some(kcfi_bundle) = kcfi_bundle.as_ref().map(|b| b.as_ref()) { bundles.push(kcfi_bundle); } @@ -1429,6 +1429,29 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> { c"".as_ptr(), ) }; + + if let Some(instance) = instance { + // Attributes on the function definition being called + let fn_defn_attrs = self.cx.tcx.codegen_fn_attrs(instance.def_id()); + if let Some(fn_call_attrs) = fn_call_attrs + && !fn_call_attrs.target_features.is_empty() + // If there is an inline attribute and a target feature that matches + // we will add the attribute to the callsite otherwise we'll omit + // this and not add the attribute to prevent soundness issues. + && let Some(inlining_rule) = attributes::inline_attr(&self.cx, instance) + && self.cx.tcx.is_target_feature_call_safe( + &fn_call_attrs.target_features, + &fn_defn_attrs.target_features, + ) + { + attributes::apply_to_callsite( + call, + llvm::AttributePlace::Function, + &[inlining_rule], + ); + } + } + if let Some(fn_abi) = fn_abi { fn_abi.apply_attrs_callsite(self, call); } diff --git a/compiler/rustc_codegen_llvm/src/builder/gpu_offload.rs b/compiler/rustc_codegen_llvm/src/builder/gpu_offload.rs index 1280ab1442a..0737a18384b 100644 --- a/compiler/rustc_codegen_llvm/src/builder/gpu_offload.rs +++ b/compiler/rustc_codegen_llvm/src/builder/gpu_offload.rs @@ -193,7 +193,7 @@ fn gen_define_handling<'ll>( // reference) types. let num_ptr_types = types .iter() - .map(|&x| matches!(cx.type_kind(x), rustc_codegen_ssa::common::TypeKind::Pointer)) + .filter(|&x| matches!(cx.type_kind(x), rustc_codegen_ssa::common::TypeKind::Pointer)) .count(); // We do not know their size anymore at this level, so hardcode a placeholder. diff --git a/compiler/rustc_codegen_llvm/src/context.rs b/compiler/rustc_codegen_llvm/src/context.rs index 4fd6110ac4a..a69fa54a54a 100644 --- a/compiler/rustc_codegen_llvm/src/context.rs +++ b/compiler/rustc_codegen_llvm/src/context.rs @@ -217,6 +217,10 @@ pub(crate) unsafe fn create_module<'ll>( // LLVM 22.0 updated the default layout on avr: https://github.com/llvm/llvm-project/pull/153010 target_data_layout = target_data_layout.replace("n8:16", "n8") } + if sess.target.arch == "nvptx64" { + // LLVM 22 updated the NVPTX layout to indicate 256-bit vector load/store: https://github.com/llvm/llvm-project/pull/155198 + target_data_layout = target_data_layout.replace("-i256:256", ""); + } } // Ensure the data-layout values hardcoded remain the defaults. @@ -849,7 +853,7 @@ impl<'ll, 'tcx> MiscCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> { fn declare_c_main(&self, fn_type: Self::Type) -> Option<Self::Function> { let entry_name = self.sess().target.entry_name.as_ref(); if self.get_declared_value(entry_name).is_none() { - Some(self.declare_entry_fn( + let llfn = self.declare_entry_fn( entry_name, llvm::CallConv::from_conv( self.sess().target.entry_abi, @@ -857,7 +861,13 @@ impl<'ll, 'tcx> MiscCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> { ), llvm::UnnamedAddr::Global, fn_type, - )) + ); + attributes::apply_to_llfn( + llfn, + llvm::AttributePlace::Function, + attributes::target_features_attr(self, vec![]).as_slice(), + ); + Some(llfn) } else { // If the symbol already exists, it is an error: for example, the user wrote // #[no_mangle] extern "C" fn main(..) {..} diff --git a/compiler/rustc_codegen_llvm/src/intrinsic.rs b/compiler/rustc_codegen_llvm/src/intrinsic.rs index 49d3dedbeab..85f71f331a4 100644 --- a/compiler/rustc_codegen_llvm/src/intrinsic.rs +++ b/compiler/rustc_codegen_llvm/src/intrinsic.rs @@ -383,7 +383,9 @@ impl<'ll, 'tcx> IntrinsicCallBuilderMethods<'tcx> for Builder<'_, 'll, 'tcx> { | sym::rotate_left | sym::rotate_right | sym::saturating_add - | sym::saturating_sub => { + | sym::saturating_sub + | sym::unchecked_funnel_shl + | sym::unchecked_funnel_shr => { let ty = args[0].layout.ty; if !ty.is_integral() { tcx.dcx().emit_err(InvalidMonomorphization::BasicIntegerType { @@ -424,18 +426,26 @@ impl<'ll, 'tcx> IntrinsicCallBuilderMethods<'tcx> for Builder<'_, 'll, 'tcx> { sym::bitreverse => { self.call_intrinsic("llvm.bitreverse", &[llty], &[args[0].immediate()]) } - sym::rotate_left | sym::rotate_right => { - let is_left = name == sym::rotate_left; - let val = args[0].immediate(); - let raw_shift = args[1].immediate(); - // rotate = funnel shift with first two args the same + sym::rotate_left + | sym::rotate_right + | sym::unchecked_funnel_shl + | sym::unchecked_funnel_shr => { + let is_left = name == sym::rotate_left || name == sym::unchecked_funnel_shl; + let lhs = args[0].immediate(); + let (rhs, raw_shift) = + if name == sym::rotate_left || name == sym::rotate_right { + // rotate = funnel shift with first two args the same + (lhs, args[1].immediate()) + } else { + (args[1].immediate(), args[2].immediate()) + }; let llvm_name = format!("llvm.fsh{}", if is_left { 'l' } else { 'r' }); // llvm expects shift to be the same type as the values, but rust // always uses `u32`. - let raw_shift = self.intcast(raw_shift, self.val_ty(val), false); + let raw_shift = self.intcast(raw_shift, self.val_ty(lhs), false); - self.call_intrinsic(llvm_name, &[llty], &[val, val, raw_shift]) + self.call_intrinsic(llvm_name, &[llty], &[lhs, rhs, raw_shift]) } sym::saturating_add | sym::saturating_sub => { let is_add = name == sym::saturating_add; |