From 7c707e4b9535b70de8c6beac24b7bc05595b1932 Mon Sep 17 00:00:00 2001 From: Commeownist Date: Sun, 5 Sep 2021 18:26:01 +0300 Subject: [PATCH] Implement basic inline asm support (#72) * Implement basic support for inline assembly * Disable LTO We don't support it yet at all * Handle `inout(reg) var` correctly Turns out that `+` readwrite output registers cannot be tied with input variables. * Add limited support for llvm_asm! * Handle CHANNEL correctly * Add support for arbitrary explicit registers * Handle symbols properly * Add rudimentary asm tests * Exclude llvm_asm! tests from tests runs * Insert `__builtin_unreachable()` after diverging asm blocks --- .gitignore | 2 + Cargo.lock | 4 +- build.sh | 7 +- cargo.sh | 2 +- config.sh | 9 +- src/asm.rs | 802 +++++++++++++++++++++++++++++++++-------------- test.sh | 9 +- tests/run/asm.rs | 87 +++++ 8 files changed, 670 insertions(+), 252 deletions(-) diff --git a/.gitignore b/.gitignore index 573ed83e535..1e2f9e3aebb 100644 --- a/.gitignore +++ b/.gitignore @@ -7,9 +7,11 @@ perf.data.old *.events *.string* /build_sysroot/sysroot +/build_sysroot/sysroot_src /build_sysroot/Cargo.lock /build_sysroot/test_target/Cargo.lock /rust +/simple-raytracer /regex gimple* *asm diff --git a/Cargo.lock b/Cargo.lock index 952c03b05c7..f3e07fd08ee 100644 --- a/Cargo.lock +++ b/Cargo.lock @@ -56,7 +56,7 @@ dependencies = [ [[package]] name = "gccjit" version = "1.0.0" -source = "git+https://github.com/antoyo/gccjit.rs#0572117c7ffdfcb0e6c6526d45266c3f34796bea" +source = "git+https://github.com/antoyo/gccjit.rs#54be27e41fff7b6ab532e2e21a82df50a12b9ad3" dependencies = [ "gccjit_sys", ] @@ -64,7 +64,7 @@ dependencies = [ [[package]] name = "gccjit_sys" version = "0.0.1" -source = "git+https://github.com/antoyo/gccjit.rs#0572117c7ffdfcb0e6c6526d45266c3f34796bea" +source = "git+https://github.com/antoyo/gccjit.rs#54be27e41fff7b6ab532e2e21a82df50a12b9ad3" dependencies = [ "libc 0.1.12", ] diff --git a/build.sh b/build.sh index 9f1228687e2..17a0d2ab3f0 100755 --- a/build.sh +++ b/build.sh @@ -3,7 +3,12 @@ #set -x set -e -export GCC_PATH=$(cat gcc_path) +if [ -f ./gcc_path ]; then + export GCC_PATH=$(cat gcc_path) +else + echo 'Please put the path to your custom build of libgccjit in the file `gcc_path`, see Readme.md for details' + exit 1 +fi export LD_LIBRARY_PATH="$GCC_PATH" export LIBRARY_PATH="$GCC_PATH" diff --git a/cargo.sh b/cargo.sh index a85865019c1..1001c522052 100755 --- a/cargo.sh +++ b/cargo.sh @@ -20,4 +20,4 @@ fi cmd=$1 shift -RUSTDOCFLAGS=$RUSTFLAGS cargo +${TOOLCHAIN} $cmd --target $TARGET_TRIPLE $@ +RUSTDOCFLAGS="$RUSTFLAGS" cargo +${TOOLCHAIN} $cmd --target $TARGET_TRIPLE $@ diff --git a/config.sh b/config.sh index bd2d915f589..98caeb7407e 100644 --- a/config.sh +++ b/config.sh @@ -2,7 +2,12 @@ set -e export CARGO_INCREMENTAL=0 -export GCC_PATH=$(cat gcc_path) +if [ -f ./gcc_path ]; then + export GCC_PATH=$(cat gcc_path) +else + echo 'Please put the path to your custom build of libgccjit in the file `gcc_path`, see Readme.md for details' + exit 1 +fi unamestr=`uname` if [[ "$unamestr" == 'Linux' ]]; then @@ -30,7 +35,7 @@ if [[ "$HOST_TRIPLE" != "$TARGET_TRIPLE" ]]; then fi fi -export RUSTFLAGS=$linker' -Cpanic=abort -Cdebuginfo=2 -Zpanic-abort-tests -Zcodegen-backend='$(pwd)'/target/'$CHANNEL'/librustc_codegen_gcc.'$dylib_ext' --sysroot '$(pwd)'/build_sysroot/sysroot' +export RUSTFLAGS="$linker -Cpanic=abort -Cdebuginfo=2 -Clto=off -Zpanic-abort-tests -Zcodegen-backend=$(pwd)/target/${CHANNEL:-debug}/librustc_codegen_gcc.$dylib_ext --sysroot $(pwd)/build_sysroot/sysroot" # FIXME(antoyo): remove once the atomic shim is gone if [[ `uname` == 'Darwin' ]]; then diff --git a/src/asm.rs b/src/asm.rs index e4d57c39de4..512b9f553f5 100644 --- a/src/asm.rs +++ b/src/asm.rs @@ -1,254 +1,558 @@ -use gccjit::{RValue, ToRValue, Type}; +use gccjit::{LValue, RValue, ToRValue, Type}; use rustc_ast::ast::{InlineAsmOptions, InlineAsmTemplatePiece}; use rustc_codegen_ssa::mir::operand::OperandValue; use rustc_codegen_ssa::mir::place::PlaceRef; use rustc_codegen_ssa::traits::{AsmBuilderMethods, AsmMethods, BaseTypeMethods, BuilderMethods, GlobalAsmOperandRef, InlineAsmOperandRef}; -use rustc_data_structures::fx::FxHashMap; + use rustc_hir::LlvmInlineAsmInner; -use rustc_middle::bug; +use rustc_middle::{bug, ty::Instance}; use rustc_span::Span; use rustc_target::asm::*; +use std::borrow::Cow; + use crate::builder::Builder; use crate::context::CodegenCx; use crate::type_of::LayoutGccExt; -impl<'a, 'gcc, 'tcx> AsmBuilderMethods<'tcx> for Builder<'a, 'gcc, 'tcx> { - fn codegen_llvm_inline_asm(&mut self, _ia: &LlvmInlineAsmInner, _outputs: Vec>>, mut _inputs: Vec>, _span: Span) -> bool { - // TODO(antoyo) - return true; - } - fn codegen_inline_asm(&mut self, template: &[InlineAsmTemplatePiece], operands: &[InlineAsmOperandRef<'tcx, Self>], options: InlineAsmOptions, _span: &[Span]) { - let asm_arch = self.tcx.sess.asm_arch.unwrap(); +// Rust asm! and GCC Extended Asm semantics differ substantially. +// +// 1. Rust asm operands go along as one list of operands. Operands themselves indicate +// if they're "in" or "out". "In" and "out" operands can interleave. One operand can be +// both "in" and "out" (`inout(reg)`). +// +// GCC asm has two different lists for "in" and "out" operands. In terms of gccjit, +// this means that all "out" operands must go before "in" operands. "In" and "out" operands +// cannot interleave. +// +// 2. Operand lists in both Rust and GCC are indexed. Index starts from 0. Indexes are important +// because the asm template refers to operands by index. +// +// Mapping from Rust to GCC index would be 1-1 if it wasn't for... +// +// 3. Clobbers. GCC has a separate list of clobbers, and clobbers don't have indexes. +// Contrary, Rust expresses clobbers through "out" operands that aren't tied to +// a variable (`_`), and such "clobbers" do have index. +// +// 4. Furthermore, GCC Extended Asm does not support explicit register constraints +// (like `out("eax")`) directly, offering so-called "local register variables" +// as a workaround. These variables need to be declared and initialized *before* +// the Extended Asm block but *after* normal local variables +// (see comment in `codegen_inline_asm` for explanation). +// +// With that in mind, let's see how we translate Rust syntax to GCC +// (from now on, `CC` stands for "constraint code"): +// +// * `out(reg_class) var` -> translated to output operand: `"=CC"(var)` +// * `inout(reg_class) var` -> translated to output operand: `"+CC"(var)` +// * `in(reg_class) var` -> translated to input operand: `"CC"(var)` +// +// * `out(reg_class) _` -> translated to one `=r(tmp)`, where "tmp" is a temporary unused variable +// +// * `out("explicit register") _` -> not translated to any operands, register is simply added to clobbers list +// +// * `inout(reg_class) in_var => out_var` -> translated to two operands: +// output: `"=CC"(in_var)` +// input: `"num"(out_var)` where num is the GCC index +// of the corresponding output operand +// +// * `inout(reg_class) in_var => _` -> same as `inout(reg_class) in_var => tmp`, +// where "tmp" is a temporary unused variable +// +// * `out/in/inout("explicit register") var` -> translated to one or two operands as described above +// with `"r"(var)` constraint, +// and one register variable assigned to the desired register. +// - let intel_dialect = - match asm_arch { - InlineAsmArch::X86 | InlineAsmArch::X86_64 if !options.contains(InlineAsmOptions::ATT_SYNTAX) => true, - _ => false, - }; +const ATT_SYNTAX_INS: &str = ".att_syntax noprefix\n\t"; +const INTEL_SYNTAX_INS: &str = "\n\t.intel_syntax noprefix"; - // Collect the types of output operands - // FIXME(antoyo): we do this here instead of later because of a bug in libgccjit where creating the - // variable after the extended asm expression causes a segfault: - // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100380 - let mut output_vars = FxHashMap::default(); - let mut operand_numbers = FxHashMap::default(); - let mut current_number = 0; - for (idx, op) in operands.iter().enumerate() { - match *op { - InlineAsmOperandRef::Out { place, .. } => { - let ty = - match place { - Some(place) => place.layout.gcc_type(self.cx, false), - None => { - // If the output is discarded, we don't really care what - // type is used. We're just using this to tell GCC to - // reserve the register. - //dummy_output_type(self.cx, reg.reg_class()) - // NOTE: if no output value, we should not create one (it will be a - // clobber). - continue; - }, - }; - let var = self.current_func().new_local(None, ty, "output_register"); - operand_numbers.insert(idx, current_number); - current_number += 1; - output_vars.insert(idx, var); - } - InlineAsmOperandRef::InOut { out_place, .. } => { - let ty = - match out_place { - Some(place) => place.layout.gcc_type(self.cx, false), - None => { - // NOTE: if no output value, we should not create one. - continue; - }, - }; - operand_numbers.insert(idx, current_number); - current_number += 1; - let var = self.current_func().new_local(None, ty, "output_register"); - output_vars.insert(idx, var); - } - _ => {} - } +struct AsmOutOperand<'a, 'tcx, 'gcc> { + rust_idx: usize, + constraint: &'a str, + late: bool, + readwrite: bool, + + tmp_var: LValue<'gcc>, + out_place: Option>> +} + +struct AsmInOperand<'a, 'tcx> { + rust_idx: usize, + constraint: Cow<'a, str>, + val: RValue<'tcx> +} + +impl AsmOutOperand<'_, '_, '_> { + fn to_constraint(&self) -> String { + let mut res = String::with_capacity(self.constraint.len() + self.late as usize + 1); + + let sign = if self.readwrite { '+' } else { '=' }; + res.push(sign); + if !self.late { + res.push('&'); } - // All output operands must come before the input operands, hence the 2 loops. - for (idx, op) in operands.iter().enumerate() { - match *op { - InlineAsmOperandRef::In { .. } | InlineAsmOperandRef::InOut { .. } => { - operand_numbers.insert(idx, current_number); - current_number += 1; - }, - _ => (), - } - } - - // Build the template string - let mut template_str = String::new(); - for piece in template { - match *piece { - InlineAsmTemplatePiece::String(ref string) => { - if string.contains('%') { - for c in string.chars() { - if c == '%' { - template_str.push_str("%%"); - } - else { - template_str.push(c); - } - } - } - else { - template_str.push_str(string) - } - } - InlineAsmTemplatePiece::Placeholder { operand_idx, modifier, span: _ } => { - match operands[operand_idx] { - InlineAsmOperandRef::Out { reg, place: Some(_), .. } => { - let modifier = modifier_to_gcc(asm_arch, reg.reg_class(), modifier); - if let Some(modifier) = modifier { - template_str.push_str(&format!("%{}{}", modifier, operand_numbers[&operand_idx])); - } else { - template_str.push_str(&format!("%{}", operand_numbers[&operand_idx])); - } - }, - InlineAsmOperandRef::Out { place: None, .. } => { - unimplemented!("Out None"); - }, - InlineAsmOperandRef::In { reg, .. } - | InlineAsmOperandRef::InOut { reg, .. } => { - let modifier = modifier_to_gcc(asm_arch, reg.reg_class(), modifier); - if let Some(modifier) = modifier { - template_str.push_str(&format!("%{}{}", modifier, operand_numbers[&operand_idx])); - } else { - template_str.push_str(&format!("%{}", operand_numbers[&operand_idx])); - } - } - InlineAsmOperandRef::Const { ref string } => { - // Const operands get injected directly into the template - template_str.push_str(string); - } - InlineAsmOperandRef::SymFn { .. } - | InlineAsmOperandRef::SymStatic { .. } => { - unimplemented!(); - // Only emit the raw symbol name - //template_str.push_str(&format!("${{{}:c}}", op_idx[&operand_idx])); - } - } - } - } - } - - let block = self.llbb(); - let template_str = - if intel_dialect { - template_str - } - else { - // FIXME(antoyo): this might break the "m" memory constraint: - // https://stackoverflow.com/a/9347957/389119 - // TODO(antoyo): only set on x86 platforms. - format!(".att_syntax noprefix\n\t{}\n\t.intel_syntax noprefix", template_str) - }; - let extended_asm = block.add_extended_asm(None, &template_str); - - // Collect the types of output operands - let mut output_types = vec![]; - for (idx, op) in operands.iter().enumerate() { - match *op { - InlineAsmOperandRef::Out { reg, late, place } => { - let ty = - match place { - Some(place) => place.layout.gcc_type(self.cx, false), - None => { - // If the output is discarded, we don't really care what - // type is used. We're just using this to tell GCC to - // reserve the register. - dummy_output_type(self.cx, reg.reg_class()) - }, - }; - output_types.push(ty); - let prefix = if late { "=" } else { "=&" }; - let constraint = format!("{}{}", prefix, reg_to_gcc(reg)); - - if place.is_some() { - let var = output_vars[&idx]; - extended_asm.add_output_operand(None, &constraint, var); - } - else { - // NOTE: reg.to_string() returns the register name with quotes around it so - // remove them. - extended_asm.add_clobber(reg.to_string().trim_matches('"')); - } - } - InlineAsmOperandRef::InOut { reg, late, in_value, out_place } => { - let ty = - match out_place { - Some(out_place) => out_place.layout.gcc_type(self.cx, false), - None => dummy_output_type(self.cx, reg.reg_class()) - }; - output_types.push(ty); - // TODO(antoyo): prefix of "+" for reading and writing? - let prefix = if late { "=" } else { "=&" }; - let constraint = format!("{}{}", prefix, reg_to_gcc(reg)); - - if out_place.is_some() { - let var = output_vars[&idx]; - // TODO(antoyo): also specify an output operand when out_place is none: that would - // be the clobber but clobbers do not support general constraint like reg; - // they only support named registers. - // Not sure how we can do this. And the LLVM backend does not seem to add a - // clobber. - extended_asm.add_output_operand(None, &constraint, var); - } - - let constraint = reg_to_gcc(reg); - extended_asm.add_input_operand(None, &constraint, in_value.immediate()); - } - InlineAsmOperandRef::In { reg, value } => { - let constraint = reg_to_gcc(reg); - extended_asm.add_input_operand(None, &constraint, value.immediate()); - } - _ => {} - } - } - - // Write results to outputs - for (idx, op) in operands.iter().enumerate() { - if let InlineAsmOperandRef::Out { place: Some(place), .. } - | InlineAsmOperandRef::InOut { out_place: Some(place), .. } = *op - { - OperandValue::Immediate(output_vars[&idx].to_rvalue()).store(self, place); - } - } + res.push_str(&self.constraint); + res } } +enum ConstraintOrRegister { + Constraint(&'static str), + Register(&'static str) +} + + +impl<'a, 'gcc, 'tcx> AsmBuilderMethods<'tcx> for Builder<'a, 'gcc, 'tcx> { + fn codegen_llvm_inline_asm(&mut self, ia: &LlvmInlineAsmInner, outputs: Vec>>, inputs: Vec>, span: Span) -> bool { + if ia.asm.as_str().is_empty() && outputs.is_empty() { + // TODO(@Commeownist): there's one use of `llvm_asm` in rustc sysroot we can't get rid of just yet. + // Fortunately, it's used as a simple black box to make sure that inputs are not optimized away. + // Let's just emulate it. + let block = self.llbb(); + let extended_asm = block.add_extended_asm(None, ""); + for input in inputs { + extended_asm.add_input_operand(None, "r", input); + } + extended_asm.add_clobber("memory"); + extended_asm.set_volatile_flag(true); + } + else { + // TODO(@Commeownist): switch to `struct_span_err_with_code` + // once we get this merged into rustc + self.sess().struct_span_err(span, "GCC backend does not support `llvm_asm!`") + .help("consider using the `asm!` macro instead") + .emit(); + } + + // We return `true` even if we've failed to generate the asm + // because we want to suppress the "malformed inline assembly" error + // generated by the frontend. + true + } + + fn codegen_inline_asm(&mut self, template: &[InlineAsmTemplatePiece], rust_operands: &[InlineAsmOperandRef<'tcx, Self>], options: InlineAsmOptions, _span: &[Span]) { + let asm_arch = self.tcx.sess.asm_arch.unwrap(); + let is_x86 = matches!(asm_arch, InlineAsmArch::X86 | InlineAsmArch::X86_64); + let att_dialect = is_x86 && options.contains(InlineAsmOptions::ATT_SYNTAX); + let intel_dialect = is_x86 && !options.contains(InlineAsmOptions::ATT_SYNTAX); + + // GCC index of an output operand equals its position in the array + let mut outputs = vec![]; + + // GCC index of an input operand equals its position in the array + // added to `outputs.len()` + let mut inputs = vec![]; + + // Clobbers collected from `out("explicit register") _` and `inout("expl_reg") var => _` + let mut clobbers = vec![]; + + // We're trying to preallocate space for the template + let mut constants_len = 0; + + // There are rules we must adhere to if we want GCC to do the right thing: + // + // * Every local variable that the asm block uses as an output must be declared *before* + // the asm block. + // * There must be no instructions whatsoever between the register variables and the asm. + // + // Therefore, the backend must generate the instructions strictly in this order: + // + // 1. Output variables. + // 2. Register variables. + // 3. The asm block. + // + // We also must make sure that no input operands are emitted before output operands. + // + // This is why we work in passes, first emitting local vars, then local register vars. + // Also, we don't emit any asm operands immediately; we save them to + // the one of the buffers to be emitted later. + + // 1. Normal variables (and saving operands to buffers). + for (rust_idx, op) in rust_operands.iter().enumerate() { + match *op { + InlineAsmOperandRef::Out { reg, late, place } => { + use ConstraintOrRegister::*; + + let (constraint, ty) = match (reg_to_gcc(reg), place) { + (Constraint(constraint), Some(place)) => (constraint, place.layout.gcc_type(self.cx, false)), + // When `reg` is a class and not an explicit register but the out place is not specified, + // we need to create an unused output variable to assign the output to. This var + // needs to be of a type that's "compatible" with the register class, but specific type + // doesn't matter. + (Constraint(constraint), None) => (constraint, dummy_output_type(self.cx, reg.reg_class())), + (Register(_), Some(_)) => { + // left for the next pass + continue + }, + (Register(reg_name), None) => { + clobbers.push(reg_name); + continue + } + }; + + let tmp_var = self.current_func().new_local(None, ty, "output_register"); + outputs.push(AsmOutOperand { + constraint, + rust_idx, + late, + readwrite: false, + tmp_var, + out_place: place + }); + } + + InlineAsmOperandRef::In { reg, value } => { + if let ConstraintOrRegister::Constraint(constraint) = reg_to_gcc(reg) { + inputs.push(AsmInOperand { + constraint: Cow::Borrowed(constraint), + rust_idx, + val: value.immediate() + }); + } + else { + // left for the next pass + continue + } + } + + InlineAsmOperandRef::InOut { reg, late, in_value, out_place } => { + let constraint = if let ConstraintOrRegister::Constraint(constraint) = reg_to_gcc(reg) { + constraint + } + else { + // left for the next pass + continue + }; + + // Rustc frontend guarantees that input and output types are "compatible", + // so we can just use input var's type for the output variable. + // + // This decision is also backed by the fact that LLVM needs in and out + // values to be of *exactly the same type*, not just "compatible". + // I'm not sure if GCC is so picky too, but better safe than sorry. + let ty = in_value.layout.gcc_type(self.cx, false); + let tmp_var = self.current_func().new_local(None, ty, "output_register"); + + // If the out_place is None (i.e `inout(reg) _` syntax was used), we translate + // it to one "readwrite (+) output variable", otherwise we translate it to two + // "out and tied in" vars as described above. + let readwrite = out_place.is_none(); + outputs.push(AsmOutOperand { + constraint, + rust_idx, + late, + readwrite, + tmp_var, + out_place, + }); + + if !readwrite { + let out_gcc_idx = outputs.len() - 1; + let constraint = Cow::Owned(out_gcc_idx.to_string()); + + inputs.push(AsmInOperand { + constraint, + rust_idx, + val: in_value.immediate() + }); + } + } + + InlineAsmOperandRef::Const { ref string } => { + constants_len += string.len() + att_dialect as usize; + } + + InlineAsmOperandRef::SymFn { instance } => { + constants_len += self.tcx.symbol_name(instance).name.len(); + } + InlineAsmOperandRef::SymStatic { def_id } => { + constants_len += self.tcx.symbol_name(Instance::mono(self.tcx, def_id)).name.len(); + } + } + } + + // 2. Register variables. + for (rust_idx, op) in rust_operands.iter().enumerate() { + match *op { + // `out("explicit register") var` + InlineAsmOperandRef::Out { reg, late, place } => { + if let ConstraintOrRegister::Register(reg_name) = reg_to_gcc(reg) { + let out_place = if let Some(place) = place { + place + } + else { + // processed in the previous pass + continue + }; + + let ty = out_place.layout.gcc_type(self.cx, false); + let tmp_var = self.current_func().new_local(None, ty, "output_register"); + tmp_var.set_register_name(reg_name); + + outputs.push(AsmOutOperand { + constraint: "r".into(), + rust_idx, + late, + readwrite: false, + tmp_var, + out_place: Some(out_place) + }); + } + + // processed in the previous pass + } + + // `in("explicit register") var` + InlineAsmOperandRef::In { reg, value } => { + if let ConstraintOrRegister::Register(reg_name) = reg_to_gcc(reg) { + let ty = value.layout.gcc_type(self.cx, false); + let reg_var = self.current_func().new_local(None, ty, "input_register"); + reg_var.set_register_name(reg_name); + self.llbb().add_assignment(None, reg_var, value.immediate()); + + inputs.push(AsmInOperand { + constraint: "r".into(), + rust_idx, + val: reg_var.to_rvalue() + }); + } + + // processed in the previous pass + } + + // `inout("explicit register") in_var => out_var` + InlineAsmOperandRef::InOut { reg, late, in_value, out_place } => { + if let ConstraintOrRegister::Register(reg_name) = reg_to_gcc(reg) { + let out_place = if let Some(place) = out_place { + place + } + else { + // processed in the previous pass + continue + }; + + // See explanation in the first pass. + let ty = in_value.layout.gcc_type(self.cx, false); + let tmp_var = self.current_func().new_local(None, ty, "output_register"); + tmp_var.set_register_name(reg_name); + + outputs.push(AsmOutOperand { + constraint: "r".into(), + rust_idx, + late, + readwrite: false, + tmp_var, + out_place: Some(out_place) + }); + + let constraint = Cow::Owned((outputs.len() - 1).to_string()); + inputs.push(AsmInOperand { + constraint, + rust_idx, + val: in_value.immediate() + }); + } + + // processed in the previous pass + } + + InlineAsmOperandRef::Const { .. } + | InlineAsmOperandRef::SymFn { .. } + | InlineAsmOperandRef::SymStatic { .. } => { + // processed in the previous pass + } + } + } + + // 3. Build the template string + + let mut template_str = String::with_capacity(estimate_template_length(template, constants_len, att_dialect)); + if !intel_dialect { + template_str.push_str(ATT_SYNTAX_INS); + } + + for piece in template { + match *piece { + InlineAsmTemplatePiece::String(ref string) => { + // TODO(@Commeownist): switch to `Iterator::intersperse` once it's stable + let mut iter = string.split('%'); + if let Some(s) = iter.next() { + template_str.push_str(s); + } + + for s in iter { + template_str.push_str("%%"); + template_str.push_str(s); + } + } + InlineAsmTemplatePiece::Placeholder { operand_idx, modifier, span: _ } => { + let mut push_to_template = |modifier, gcc_idx| { + use std::fmt::Write; + + template_str.push('%'); + if let Some(modifier) = modifier { + template_str.push(modifier); + } + write!(template_str, "{}", gcc_idx).expect("pushing to string failed"); + }; + + match rust_operands[operand_idx] { + InlineAsmOperandRef::Out { reg, .. } => { + let modifier = modifier_to_gcc(asm_arch, reg.reg_class(), modifier); + let gcc_index = outputs.iter() + .position(|op| operand_idx == op.rust_idx) + .expect("wrong rust index"); + push_to_template(modifier, gcc_index); + } + + InlineAsmOperandRef::In { reg, .. } => { + let modifier = modifier_to_gcc(asm_arch, reg.reg_class(), modifier); + let in_gcc_index = inputs.iter() + .position(|op| operand_idx == op.rust_idx) + .expect("wrong rust index"); + let gcc_index = in_gcc_index + outputs.len(); + push_to_template(modifier, gcc_index); + } + + InlineAsmOperandRef::InOut { reg, .. } => { + let modifier = modifier_to_gcc(asm_arch, reg.reg_class(), modifier); + + // The input register is tied to the output, so we can just use the index of the output register + let gcc_index = outputs.iter() + .position(|op| operand_idx == op.rust_idx) + .expect("wrong rust index"); + push_to_template(modifier, gcc_index); + } + + InlineAsmOperandRef::SymFn { instance } => { + let name = self.tcx.symbol_name(instance).name; + template_str.push_str(name); + } + + InlineAsmOperandRef::SymStatic { def_id } => { + // TODO(@Commeownist): This may not be sufficient for all kinds of statics. + // Some statics may need the `@plt` suffix, like thread-local vars. + let instance = Instance::mono(self.tcx, def_id); + let name = self.tcx.symbol_name(instance).name; + template_str.push_str(name); + } + + InlineAsmOperandRef::Const { ref string } => { + // Const operands get injected directly into the template + if att_dialect { + template_str.push('$'); + } + template_str.push_str(string); + } + } + } + } + } + + if !intel_dialect { + template_str.push_str(INTEL_SYNTAX_INS); + } + + // 4. Generate Extended Asm block + + let block = self.llbb(); + let extended_asm = block.add_extended_asm(None, &template_str); + + for op in &outputs { + extended_asm.add_output_operand(None, &op.to_constraint(), op.tmp_var); + } + + for op in &inputs { + extended_asm.add_input_operand(None, &op.constraint, op.val); + } + + for clobber in clobbers.iter() { + extended_asm.add_clobber(clobber); + } + + if !options.contains(InlineAsmOptions::PRESERVES_FLAGS) { + // TODO(@Commeownist): I'm not 100% sure this one clobber is sufficient + // on all architectures. For instance, what about FP stack? + extended_asm.add_clobber("cc"); + } + if !options.contains(InlineAsmOptions::NOMEM) { + extended_asm.add_clobber("memory"); + } + if !options.contains(InlineAsmOptions::PURE) { + extended_asm.set_volatile_flag(true); + } + if !options.contains(InlineAsmOptions::NOSTACK) { + // TODO(@Commeownist): figure out how to align stack + } + if options.contains(InlineAsmOptions::NORETURN) { + let builtin_unreachable = self.context.get_builtin_function("__builtin_unreachable"); + let builtin_unreachable: RValue<'gcc> = unsafe { std::mem::transmute(builtin_unreachable) }; + self.call(self.type_void(), builtin_unreachable, &[], None); + } + + // Write results to outputs. + // + // We need to do this because: + // 1. Turning `PlaceRef` into `RValue` is error-prone and has nasty edge cases + // (especially with current `rustc_backend_ssa` API). + // 2. Not every output operand has an `out_place`, and it's required by `add_output_operand`. + // + // Instead, we generate a temporary output variable for each output operand, and then this loop, + // generates `out_place = tmp_var;` assignments if out_place exists. + for op in &outputs { + if let Some(place) = op.out_place { + OperandValue::Immediate(op.tmp_var.to_rvalue()).store(self, place); + } + } + + } +} + +fn estimate_template_length(template: &[InlineAsmTemplatePiece], constants_len: usize, att_dialect: bool) -> usize { + let len: usize = template.iter().map(|piece| { + match *piece { + InlineAsmTemplatePiece::String(ref string) => { + string.len() + } + InlineAsmTemplatePiece::Placeholder { .. } => { + // '%' + 1 char modifier + 1 char index + 3 + } + } + }) + .sum(); + + // increase it by 5% to account for possible '%' signs that'll be duplicated + // I pulled the number out of blue, but should be fair enough + // as the upper bound + let mut res = (len as f32 * 1.05) as usize + constants_len; + + if att_dialect { + res += INTEL_SYNTAX_INS.len() + ATT_SYNTAX_INS.len(); + } + res +} + /// Converts a register class to a GCC constraint code. -// TODO(antoyo): return &'static str instead? -fn reg_to_gcc(reg: InlineAsmRegOrRegClass) -> String { - match reg { +fn reg_to_gcc(reg: InlineAsmRegOrRegClass) -> ConstraintOrRegister { + let constraint = match reg { // For vector registers LLVM wants the register name to match the type size. InlineAsmRegOrRegClass::Reg(reg) => { // TODO(antoyo): add support for vector register. - let constraint = - match reg.name() { - "ax" => "a", - "bx" => "b", - "cx" => "c", - "dx" => "d", - "si" => "S", - "di" => "D", - // TODO(antoyo): for registers like r11, we have to create a register variable: https://stackoverflow.com/a/31774784/389119 - // TODO(antoyo): in this case though, it's a clobber, so it should work as r11. - // Recent nightly supports clobber() syntax, so update to it. It does not seem - // like it's implemented yet. - name => name, // FIXME(antoyo): probably wrong. - }; - constraint.to_string() + match reg.name() { + "ax" => "a", + "bx" => "b", + "cx" => "c", + "dx" => "d", + "si" => "S", + "di" => "D", + // For registers like r11, we have to create a register variable: https://stackoverflow.com/a/31774784/389119 + name => return ConstraintOrRegister::Register(name), + } }, InlineAsmRegOrRegClass::RegClass(reg) => match reg { InlineAsmRegClass::AArch64(AArch64InlineAsmRegClass::preg) => unimplemented!(), @@ -278,23 +582,25 @@ fn reg_to_gcc(reg: InlineAsmRegOrRegClass) -> String { InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::reg) => unimplemented!(), InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::freg) => unimplemented!(), InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::vreg) => unimplemented!(), - InlineAsmRegClass::X86(X86InlineAsmRegClass::mmx_reg) => unimplemented!(), InlineAsmRegClass::X86(X86InlineAsmRegClass::reg) => "r", - InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_abcd) => unimplemented!(), - InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_byte) => unimplemented!(), + InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_abcd) => "Q", + InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_byte) => "q", InlineAsmRegClass::X86(X86InlineAsmRegClass::xmm_reg) - | InlineAsmRegClass::X86(X86InlineAsmRegClass::ymm_reg) => unimplemented!(), - InlineAsmRegClass::X86(X86InlineAsmRegClass::x87_reg) => unimplemented!(), - InlineAsmRegClass::X86(X86InlineAsmRegClass::zmm_reg) => unimplemented!(), + | InlineAsmRegClass::X86(X86InlineAsmRegClass::ymm_reg) => "x", + InlineAsmRegClass::X86(X86InlineAsmRegClass::zmm_reg) => "v", InlineAsmRegClass::X86(X86InlineAsmRegClass::kreg) => unimplemented!(), InlineAsmRegClass::Wasm(WasmInlineAsmRegClass::local) => unimplemented!(), + InlineAsmRegClass::X86( + X86InlineAsmRegClass::x87_reg | X86InlineAsmRegClass::mmx_reg, + ) => unreachable!("clobber-only"), InlineAsmRegClass::SpirV(SpirVInlineAsmRegClass::reg) => { bug!("GCC backend does not support SPIR-V") } InlineAsmRegClass::Err => unreachable!(), } - .to_string(), - } + }; + + ConstraintOrRegister::Constraint(constraint) } /// Type to use for outputs that are discarded. It doesn't really matter what @@ -379,7 +685,7 @@ impl<'gcc, 'tcx> AsmMethods for CodegenCx<'gcc, 'tcx> { match operands[operand_idx] { GlobalAsmOperandRef::Const { ref string } => { // Const operands get injected directly into the - // template. Note that we don't need to escape $ + // template. Note that we don't need to escape % // here unlike normal inline assembly. template_str.push_str(string); } @@ -431,8 +737,7 @@ fn modifier_to_gcc(arch: InlineAsmArch, reg: InlineAsmRegClass, modifier: Option InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::vreg) => unimplemented!(), InlineAsmRegClass::X86(X86InlineAsmRegClass::reg) | InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_abcd) => match modifier { - None if arch == InlineAsmArch::X86_64 => Some('q'), - None => Some('k'), + None => if arch == InlineAsmArch::X86_64 { Some('q') } else { Some('k') }, Some('l') => Some('b'), Some('h') => Some('h'), Some('x') => Some('w'), @@ -440,13 +745,22 @@ fn modifier_to_gcc(arch: InlineAsmArch, reg: InlineAsmRegClass, modifier: Option Some('r') => Some('q'), _ => unreachable!(), }, - InlineAsmRegClass::X86(X86InlineAsmRegClass::mmx_reg) => unimplemented!(), - InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_byte) => unimplemented!(), - InlineAsmRegClass::X86(X86InlineAsmRegClass::xmm_reg) - | InlineAsmRegClass::X86(X86InlineAsmRegClass::ymm_reg) - | InlineAsmRegClass::X86(X86InlineAsmRegClass::zmm_reg) => unimplemented!(), - InlineAsmRegClass::X86(X86InlineAsmRegClass::x87_reg) => unimplemented!(), - InlineAsmRegClass::X86(X86InlineAsmRegClass::kreg) => unimplemented!(), + InlineAsmRegClass::X86(X86InlineAsmRegClass::reg_byte) => None, + InlineAsmRegClass::X86(reg @ X86InlineAsmRegClass::xmm_reg) + | InlineAsmRegClass::X86(reg @ X86InlineAsmRegClass::ymm_reg) + | InlineAsmRegClass::X86(reg @ X86InlineAsmRegClass::zmm_reg) => match (reg, modifier) { + (X86InlineAsmRegClass::xmm_reg, None) => Some('x'), + (X86InlineAsmRegClass::ymm_reg, None) => Some('t'), + (X86InlineAsmRegClass::zmm_reg, None) => Some('g'), + (_, Some('x')) => Some('x'), + (_, Some('y')) => Some('t'), + (_, Some('z')) => Some('g'), + _ => unreachable!(), + }, + InlineAsmRegClass::X86(X86InlineAsmRegClass::kreg) => None, + InlineAsmRegClass::X86(X86InlineAsmRegClass::x87_reg | X86InlineAsmRegClass::mmx_reg) => { + unreachable!("clobber-only") + } InlineAsmRegClass::Wasm(WasmInlineAsmRegClass::local) => unimplemented!(), InlineAsmRegClass::SpirV(SpirVInlineAsmRegClass::reg) => { bug!("LLVM backend does not support SPIR-V") diff --git a/test.sh b/test.sh index 06faa98d9b1..114db7d53f5 100755 --- a/test.sh +++ b/test.sh @@ -4,7 +4,12 @@ set -e -export GCC_PATH=$(cat gcc_path) +if [ -f ./gcc_path ]; then + export GCC_PATH=$(cat gcc_path) +else + echo 'Please put the path to your custom build of libgccjit in the file `gcc_path`, see Readme.md for details' + exit 1 +fi export LD_LIBRARY_PATH="$GCC_PATH" export LIBRARY_PATH="$GCC_PATH" @@ -157,7 +162,7 @@ done git checkout -- src/test/ui/issues/auxiliary/issue-3136-a.rs # contains //~ERROR, but shouldn't be removed -rm -r src/test/ui/{abi*,extern/,panic-runtime/,panics/,unsized-locals/,proc-macro/,threads-sendsync/,thinlto/,simd*,borrowck/,test*,*lto*.rs} || true +rm -r src/test/ui/{abi*,extern/,llvm-asm/,panic-runtime/,panics/,unsized-locals/,proc-macro/,threads-sendsync/,thinlto/,simd*,borrowck/,test*,*lto*.rs} || true for test in $(rg --files-with-matches "catch_unwind|should_panic|thread|lto" src/test/ui); do rm $test done diff --git a/tests/run/asm.rs b/tests/run/asm.rs index bd76153e047..9c0055b0b6b 100644 --- a/tests/run/asm.rs +++ b/tests/run/asm.rs @@ -62,5 +62,92 @@ fn main() { } assert_eq!(x, 43); + // check inout(reg_class) x + let mut x: u64 = 42; + unsafe { + asm!("add {0}, {0}", + inout(reg) x + ); + } + assert_eq!(x, 84); + + // check inout("reg") x + let mut x: u64 = 42; + unsafe { + asm!("add r11, r11", + inout("r11") x + ); + } + assert_eq!(x, 84); + + // check a mix of + // in("reg") + // inout(class) x => y + // inout (class) x + let x: u64 = 702; + let y: u64 = 100; + let res: u64; + let mut rem: u64 = 0; + unsafe { + asm!("div r11", + in("r11") y, + inout("eax") x => res, + inout("edx") rem, + ); + } + assert_eq!(res, 7); + assert_eq!(rem, 2); + + // check const + let mut x: u64 = 42; + unsafe { + asm!("add {}, {}", + inout(reg) x, + const 1 + ); + } + assert_eq!(x, 43); + + // check const (ATT syntax) + let mut x: u64 = 42; + unsafe { + asm!("add {}, {}", + const 1, + inout(reg) x, + options(att_syntax) + ); + } + assert_eq!(x, 43); + + // check sym fn + extern "C" fn foo() -> u64 { 42 } + let x: u64; + unsafe { + asm!("call {}", sym foo, lateout("rax") x); + } + assert_eq!(x, 42); + + // check sym fn (ATT syntax) + let x: u64; + unsafe { + asm!("call {}", sym foo, lateout("rax") x, options(att_syntax)); + } + assert_eq!(x, 42); + + // check sym static + static FOO: u64 = 42; + let x: u64; + unsafe { + asm!("mov {1}, qword ptr [rip + {0}]", sym FOO, lateout(reg) x); + } + assert_eq!(x, 42); + + // check sym static (ATT syntax) + let x: u64; + unsafe { + asm!("movq {0}(%rip), {1}", sym FOO, lateout(reg) x, options(att_syntax)); + } + assert_eq!(x, 42); + assert_eq!(unsafe { add_asm(40, 2) }, 42); }