1
Fork 0

NVPTX support for new asm!

This commit is contained in:
Michal Sudwoj 2020-05-21 23:01:22 +02:00
parent 3137f8e2d1
commit 58fdc43e03
4 changed files with 238 additions and 0 deletions

View file

@ -254,6 +254,7 @@ impl AsmBuilderMethods<'tcx> for Builder<'a, 'll, 'tcx> {
]);
}
InlineAsmArch::RiscV32 | InlineAsmArch::RiscV64 => {}
InlineAsmArch::Nvptx64 => {}
}
}
if !options.contains(InlineAsmOptions::NOMEM) {
@ -410,6 +411,11 @@ fn reg_to_llvm(reg: InlineAsmRegOrRegClass) -> String {
| InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low4) => "x",
InlineAsmRegClass::Arm(ArmInlineAsmRegClass::dreg)
| InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg) => "w",
InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg16) => "h",
InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg32) => "r",
InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg64) => "l",
InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::freg32) => "f",
InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::freg64) => "d",
InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::reg) => "r",
InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::freg) => "f",
InlineAsmRegClass::X86(X86InlineAsmRegClass::reg) => "r",
@ -452,6 +458,7 @@ fn modifier_to_llvm(
modifier
}
}
InlineAsmRegClass::Nvptx(_) => None,
InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::reg)
| InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::freg) => None,
InlineAsmRegClass::X86(X86InlineAsmRegClass::reg)
@ -502,6 +509,11 @@ fn dummy_output_type(cx: &CodegenCx<'ll, 'tcx>, reg: InlineAsmRegClass) -> &'ll
| InlineAsmRegClass::Arm(ArmInlineAsmRegClass::qreg_low4) => {
cx.type_vector(cx.type_i64(), 2)
}
InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg16) => cx.type_i16(),
InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg32) => cx.type_i32(),
InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::reg64) => cx.type_i64(),
InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::freg32) => cx.type_f32(),
InlineAsmRegClass::Nvptx(NvptxInlineAsmRegClass::freg64) => cx.type_f64(),
InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::reg) => cx.type_i32(),
InlineAsmRegClass::RiscV(RiscVInlineAsmRegClass::freg) => cx.type_f32(),
InlineAsmRegClass::X86(X86InlineAsmRegClass::reg)

View file

@ -146,11 +146,13 @@ macro_rules! types {
mod aarch64;
mod arm;
mod nvptx;
mod riscv;
mod x86;
pub use aarch64::{AArch64InlineAsmReg, AArch64InlineAsmRegClass};
pub use arm::{ArmInlineAsmReg, ArmInlineAsmRegClass};
pub use nvptx::{NvptxInlineAsmReg, NvptxInlineAsmRegClass};
pub use riscv::{RiscVInlineAsmReg, RiscVInlineAsmRegClass};
pub use x86::{X86InlineAsmReg, X86InlineAsmRegClass};
@ -162,6 +164,7 @@ pub enum InlineAsmArch {
AArch64,
RiscV32,
RiscV64,
Nvptx64,
}
impl FromStr for InlineAsmArch {
@ -175,6 +178,7 @@ impl FromStr for InlineAsmArch {
"aarch64" => Ok(Self::AArch64),
"riscv32" => Ok(Self::RiscV32),
"riscv64" => Ok(Self::RiscV64),
"nvptx64" => Ok(Self::Nvptx64),
_ => Err(()),
}
}
@ -196,6 +200,7 @@ pub enum InlineAsmReg {
Arm(ArmInlineAsmReg),
AArch64(AArch64InlineAsmReg),
RiscV(RiscVInlineAsmReg),
Nvptx(NvptxInlineAsmReg),
}
impl InlineAsmReg {
@ -205,6 +210,7 @@ impl InlineAsmReg {
Self::Arm(r) => r.name(),
Self::AArch64(r) => r.name(),
Self::RiscV(r) => r.name(),
Self::Nvptx(r) => r.name(),
}
}
@ -214,6 +220,7 @@ impl InlineAsmReg {
Self::Arm(r) => InlineAsmRegClass::Arm(r.reg_class()),
Self::AArch64(r) => InlineAsmRegClass::AArch64(r.reg_class()),
Self::RiscV(r) => InlineAsmRegClass::RiscV(r.reg_class()),
Self::Nvptx(r) => InlineAsmRegClass::Nvptx(r.reg_class()),
}
}
@ -236,6 +243,9 @@ impl InlineAsmReg {
InlineAsmArch::RiscV32 | InlineAsmArch::RiscV64 => {
Self::RiscV(RiscVInlineAsmReg::parse(arch, has_feature, &name)?)
}
InlineAsmArch::Nvptx64 => {
Self::Nvptx(NvptxInlineAsmReg::parse(arch, has_feature, &name)?)
}
})
}
@ -252,6 +262,7 @@ impl InlineAsmReg {
Self::Arm(r) => r.emit(out, arch, modifier),
Self::AArch64(r) => r.emit(out, arch, modifier),
Self::RiscV(r) => r.emit(out, arch, modifier),
Self::Nvptx(r) => r.emit(out, arch, modifier),
}
}
@ -261,6 +272,7 @@ impl InlineAsmReg {
Self::Arm(r) => r.overlapping_regs(|r| cb(Self::Arm(r))),
Self::AArch64(_) => cb(self),
Self::RiscV(_) => cb(self),
Self::Nvptx(_) => cb(self),
}
}
}
@ -281,6 +293,7 @@ pub enum InlineAsmRegClass {
Arm(ArmInlineAsmRegClass),
AArch64(AArch64InlineAsmRegClass),
RiscV(RiscVInlineAsmRegClass),
Nvptx(NvptxInlineAsmRegClass),
}
impl InlineAsmRegClass {
@ -290,6 +303,7 @@ impl InlineAsmRegClass {
Self::Arm(r) => r.name(),
Self::AArch64(r) => r.name(),
Self::RiscV(r) => r.name(),
Self::Nvptx(r) => r.name(),
}
}
@ -302,6 +316,7 @@ impl InlineAsmRegClass {
Self::Arm(r) => r.suggest_class(arch, ty).map(InlineAsmRegClass::Arm),
Self::AArch64(r) => r.suggest_class(arch, ty).map(InlineAsmRegClass::AArch64),
Self::RiscV(r) => r.suggest_class(arch, ty).map(InlineAsmRegClass::RiscV),
Self::Nvptx(r) => r.suggest_class(arch, ty).map(InlineAsmRegClass::Nvptx),
}
}
@ -321,6 +336,7 @@ impl InlineAsmRegClass {
Self::Arm(r) => r.suggest_modifier(arch, ty),
Self::AArch64(r) => r.suggest_modifier(arch, ty),
Self::RiscV(r) => r.suggest_modifier(arch, ty),
Self::Nvptx(r) => r.suggest_modifier(arch, ty),
}
}
@ -336,6 +352,7 @@ impl InlineAsmRegClass {
Self::Arm(r) => r.default_modifier(arch),
Self::AArch64(r) => r.default_modifier(arch),
Self::RiscV(r) => r.default_modifier(arch),
Self::Nvptx(r) => r.default_modifier(arch),
}
}
@ -350,6 +367,7 @@ impl InlineAsmRegClass {
Self::Arm(r) => r.supported_types(arch),
Self::AArch64(r) => r.supported_types(arch),
Self::RiscV(r) => r.supported_types(arch),
Self::Nvptx(r) => r.supported_types(arch),
}
}
@ -367,6 +385,9 @@ impl InlineAsmRegClass {
InlineAsmArch::RiscV32 | InlineAsmArch::RiscV64 => {
Self::RiscV(RiscVInlineAsmRegClass::parse(arch, name)?)
}
InlineAsmArch::Nvptx64 => {
Self::Nvptx(NvptxInlineAsmRegClass::parse(arch, name)?)
}
})
})
}
@ -379,6 +400,7 @@ impl InlineAsmRegClass {
Self::Arm(r) => r.valid_modifiers(arch),
Self::AArch64(r) => r.valid_modifiers(arch),
Self::RiscV(r) => r.valid_modifiers(arch),
Self::Nvptx(r) => r.valid_modifiers(arch),
}
}
}
@ -518,5 +540,10 @@ pub fn allocatable_registers(
riscv::fill_reg_map(arch, has_feature, &mut map);
map
}
InlineAsmArch::Nvptx64 => {
let mut map = nvptx::regclass_map();
nvptx::fill_reg_map(arch, has_feature, &mut map);
map
}
}
}

View file

@ -0,0 +1,90 @@
use super::{InlineAsmArch, InlineAsmType};
use rustc_macros::HashStable_Generic;
use std::fmt;
def_reg_class! {
Nvptx NvptxInlineAsmRegClass {
reg16,
reg32,
reg64,
freg32,
freg64,
}
}
impl NvptxInlineAsmRegClass {
pub fn valid_modifiers(self, _arch: InlineAsmArch) -> &'static [char] {
&[]
}
pub fn suggest_class(self, _arch: InlineAsmArch, _ty: InlineAsmType) -> Option<Self> {
None
}
pub fn suggest_modifier(
self,
_arch: InlineAsmArch,
_ty: InlineAsmType,
) -> Option<(char, &'static str)> {
None
}
pub fn default_modifier(self, _arch: InlineAsmArch) -> Option<(char, &'static str)> {
None
}
pub fn supported_types(
self,
_arch: InlineAsmArch,
) -> &'static [(InlineAsmType, Option<&'static str>)] {
match self {
Self::reg16 => types! { _: I8, I16; },
Self::reg32 => types! { _: I8, I16, I32; },
Self::reg64 => types! { _: I8, I16, I32, I64; },
Self::freg32 => types! { _: F32; },
Self::freg64 => types! { _: F32, F64; },
}
}
}
def_regs! {
Nvptx NvptxInlineAsmReg NvptxInlineAsmRegClass {
// We have to define a register, otherwise we get warnings/errors about unused imports and
// unreachable code. Do what clang does and define r0.
r0: reg32 = ["r0"],
#error = ["tid", "tid.x", "tid.y", "tid.z"] => "tid not supported for inline asm",
#error = ["ntid", "ntid.x", "ntid.y", "ntid.z"] => "ntid not supported for inline asm",
#error = ["laneid"] => "laneid not supported for inline asm",
#error = ["warpid"] => "warpid not supported for inline asm",
#error = ["nwarpid"] => "nwarpid not supported for inline asm",
#error = ["ctaid", "ctaid.x", "ctaid.y", "ctaid.z"] => "ctaid not supported for inline asm",
#error = ["nctaid", "nctaid.x", "nctaid.y", "nctaid.z"] => "nctaid not supported for inline asm",
#error = ["smid"] => "smid not supported for inline asm",
#error = ["nsmid"] => "nsmid not supported for inline asm",
#error = ["gridid"] => "gridid not supported for inline asm",
#error = ["lanemask_eq"] => "lanemask_eq not supported for inline asm",
#error = ["lanemask_le"] => "lanemask_le not supported for inline asm",
#error = ["lanemask_lt"] => "lanemask_lt not supported for inline asm",
#error = ["lanemask_ge"] => "lanemask_ge not supported for inline asm",
#error = ["lanemask_gt"] => "lanemask_gt not supported for inline asm",
#error = ["clock", "clock_hi"] => "clock not supported for inline asm",
#error = ["clock64"] => "clock64 not supported for inline asm",
#error = ["pm0", "pm1", "pm2", "pm3", "pm4", "pm5", "pm6", "pm7"] => "pm not supported for inline asm",
#error = ["pm0_64", "pm1_64", "pm2_64", "pm3_64", "pm4_64", "pm5_64", "pm6_64", "pm7_64"] => "pm_64 not supported for inline asm",
#error = ["envreg0", "envreg1", "envreg2", "envreg3", "envreg4", "envreg5", "envreg6", "envreg7", "envreg8", "envreg9", "envreg10", "envreg11", "envreg12", "envreg13", "envreg14", "envreg15", "envreg16", "envreg17", "envreg18", "envreg19", "envreg20", "envreg21", "envreg22", "envreg23", "envreg24", "envreg25", "envreg26", "envreg27", "envreg28", "envreg29", "envreg30", "envreg31"] => "envreg not supported for inline asm",
#error = ["globaltimer", "globaltimer_lo", "globaltimer_hi"] => "globaltimer not supported for inline asm",
#error = ["total_mem_size"] => "total_mem_size not supported for inline asm",
#error = ["dynamic_mem_size"] => "dynamic_mem_size not supported for inline asm",
}
}
impl NvptxInlineAsmReg {
pub fn emit(
self,
out: &mut dyn fmt::Write,
_arch: InlineAsmArch,
_modifier: Option<char>,
) -> fmt::Result {
out.write_str(self.name())
}
}

View file

@ -0,0 +1,109 @@
// no-system-llvm
// assembly-output: emit-asm
// compile-flags: --target --nvptx64-nvidia-cuda
// only-nvptx64
// ignore-nvptx64
#![feature(no_core, lang_items, rustc_attrs)]
#![crate_type = "rlib"]
#![no_core]
#![allow(asm_sub_register, non_camel_case_types)]
#[rustc_builtin_macro]
macro_rules! asm {
() => {};
}
#[rustc_builtin_macro]
macro_rules! concat {
() => {};
}
#[rustc_builtin_macro]
macro_rules! stringify {
() => {};
}
#[lang = "sized"]
trait Sized {}
#[lang = "copy"]
trait Copy {}
type ptr = *mut u8;
impl Copy for i8 {}
impl Copy for i16 {}
impl Copy for i32 {}
impl Copy for f32 {}
impl Copy for i64 {}
impl Copy for f64 {}
impl Copy for ptr {}
#[no_mangle]
fn extern_func();
// CHECK-LABEL: sym_fn
// CHECK: #APP
// CHECK call extern_func;
// CHECK: #NO_APP
#[no_mangle]
pub unsafe fn sym_fn() {
asm!("call {}", sym extern_func);
}
macro_rules! check {
($func:ident $ty:ident, $class:ident $mov:literal) => {
#[no_mangle]
pub unsafe fn $func(x: $ty) -> $ty {
// Hack to avoid function merging
extern "Rust" {
fn dont_merge(s: &str);
}
dont_merge(stringify!($func));
let y;
asm!(concat!($mov, " {}, {};"), out($class) y, in($class) x);
y
}
};
}
// CHECK-LABEL: reg_i8
// CHECK: #APP
// CHECK: mov.i16 {{[a-z0-9]+}}, {{[a-z0-9]+}};
// CHECK: #NO_APP
check!(reg_i8 i8 reg16 "mov.i16");
// CHECK-LABEL: reg_i16
// CHECK: #APP
// CHECK: mov.i16 {{[a-z0-9]+}}, {{[a-z0-9]+}};
// CHECK: #NO_APP
check!(reg_i16 i16 reg16 "mov.i16");
// CHECK-LABEL: reg_i32
// CHECK: #APP
// CHECK: mov.i32 {{[a-z0-9]+}}, {{[a-z0-9]+}};
// CHECK: #NO_APP
check!(reg_i32 i32 reg32 "mov.i32");
// CHECK-LABEL: reg_f32
// CHECK: #APP
// CHECK: mov.f32 {{[a-z0-9]+}}, {{[a-z0-9]+}};
// CHECK: #NO_APP
check!(reg_f32 f32 freg32 "mov.f32");
// CHECK-LABEL: reg_i54
// CHECK: #APP
// CHECK: mov.i64 {{[a-z0-9]+}}, {{[a-z0-9]+}};
// CHECK: #NO_APP
check!(reg_i64 i64 reg64 "mov.i64");
// CHECK-LABEL: reg_f64
// CHECK: #APP
// CHECK: mov.f64 {{[a-z0-9]+}}, {{[a-z0-9]+}};
// CHECK: #NO_APP
check!(reg_f64 f64 freg64 "mov.f64");
// CHECK-LABEL: reg_ptr
// CHECK: #APP
// CHECK: mov.i64 {{[a-z0-9]+}}, {{[a-z0-9]+}};
// CHECK: #NO_APP
check!(reg_ptr ptr reg64 "mov.i64");