Multiple improvements to the inline assembly code

Cargo.toml

@@ -42,6 +42,9 @@ inline_asm = []
 abi_x86_interrupt = []
 const_fn = []
 
+[package.metadata.docs.rs]
+rustdoc-args = ["--cfg", "docsrs"]
+
 [package.metadata.release]
 no-dev-version = true
 pre-release-replacements = [

Changelog.md

@@ -1,5 +1,10 @@
 # Unreleased
 
+- Multiple improvements to assembly code ([#251](https://github.com/rust-osdev/x86_64/pull/251))
+  - Added `external_asm` implementations for `bochs_breakpoint` and `XCr0`
+  - Updated `options` for `asm!` blocks (to improve performance)
+  - Updated docs to use [`doc_cfg`](https://doc.rust-lang.org/unstable-book/language-features/doc-cfg.html)
+
 # 0.14.1 – 2021-05-06
 
 - Use new `const_fn_trait_bound` feature to fix build on latest nightly ([#250](https://github.com/rust-osdev/x86_64/pull/250))

src/asm/asm.s

@@ -1,6 +1,10 @@
 .text
 .code64
 
+# REMEMBER: This code uses the AMD64 calling convention:
+#   Arguments: RDI, RSI, RDX, RCX
+#   Return: RAX
+
 .global _x86_64_asm_interrupt_enable
 .p2align 4
 _x86_64_asm_interrupt_enable:
@@ -135,11 +139,8 @@ _x86_64_asm_write_rflags:
 .global _x86_64_asm_read_rflags
 .p2align 4
 _x86_64_asm_read_rflags:
-    pushq   %rbp
-    movq    %rsp, %rbp
     pushfq
     popq    %rax
-    popq    %rbp
     retq
 
 .global _x86_64_asm_load_ss
@@ -223,21 +224,19 @@ _x86_64_asm_write_cr4:
 .global _x86_64_asm_rdmsr
 .p2align 4
 _x86_64_asm_rdmsr:
-    mov   %edi,%ecx
+    mov    %edi, %ecx    # First param is the MSR number
     rdmsr
-    shl    $0x20,%rdx   # shift edx to upper 32bit
-    mov    %eax,%eax    # clear upper 32bit of rax
-    or     %rdx,%rax    # or with rdx
+    shl    $32,  %rdx    # shift edx to upper 32bit
+    mov    %eax, %eax    # clear upper 32bit of rax
+    or     %rdx, %rax    # or with rdx
     retq
 
 .global _x86_64_asm_wrmsr
 .p2align 4
 _x86_64_asm_wrmsr:
-    mov   %edi,%ecx
-    movq  %rsi,%rax
-    movq  %rsi,%rdx
-    shr   $0x20,%rdx
-    wrmsr
+    movl  %edi, %ecx    # First param is the MSR number
+    movl  %esi, %eax    # Second param is the low 32-bits
+    wrmsr               # Third param (high 32-bits) is already in %edx
     retq
 
 .global _x86_64_asm_hlt
@@ -252,6 +251,12 @@ _x86_64_asm_nop:
     nop
     retq
 
+.global _x86_64_asm_bochs
+.p2align 4
+_x86_64_asm_bochs:
+    xchgw %bx, %bx
+    retq
+
 .global _x86_64_asm_rdfsbase
 .p2align 4
 _x86_64_asm_rdfsbase:
@@ -275,3 +280,21 @@ _x86_64_asm_rdgsbase:
 _x86_64_asm_wrgsbase:
     wrgsbase %rdi
     retq
+
+.global _x86_64_asm_xgetbv
+.p2align 4
+_x86_64_asm_xgetbv:
+    mov    %edi, %ecx    # First param is the XCR number
+    xgetbv
+    shl    $32,  %rdx    # shift edx to upper 32bit
+    mov    %eax, %eax    # clear upper 32bit of rax
+    or     %rdx, %rax    # or with rdx
+    retq
+
+.global _x86_64_asm_xsetbv
+.p2align 4
+_x86_64_asm_xsetbv:
+    movl  %edi, %ecx    # First param is the XCR number
+    movl  %esi, %eax    # Second param is the low 32-bits
+    xsetbv              # Third param (high 32-bits) is already in %edx
+    retq

src/asm/mod.rs

@@ -36,6 +36,12 @@ extern "C" {
     )]
     pub(crate) fn x86_64_asm_nop();
 
+    #[cfg_attr(
+        any(target_env = "gnu", target_env = "musl"),
+        link_name = "_x86_64_asm_bochs"
+    )]
+    pub(crate) fn x86_64_asm_bochs();
+
     #[cfg_attr(
         any(target_env = "gnu", target_env = "musl"),
         link_name = "_x86_64_asm_read_from_port_u8"
@@ -208,7 +214,7 @@ extern "C" {
         any(target_env = "gnu", target_env = "musl"),
         link_name = "_x86_64_asm_wrmsr"
     )]
-    pub(crate) fn x86_64_asm_wrmsr(msr: u32, value: u64);
+    pub(crate) fn x86_64_asm_wrmsr(msr: u32, low: u32, high: u32);
 
     #[cfg_attr(
         any(target_env = "gnu", target_env = "musl"),
@@ -245,4 +251,16 @@ extern "C" {
         link_name = "_x86_64_asm_wrgsbase"
     )]
     pub(crate) fn x86_64_asm_wrgsbase(val: u64);
+
+    #[cfg_attr(
+        any(target_env = "gnu", target_env = "musl"),
+        link_name = "_x86_64_asm_xgetbv"
+    )]
+    pub(crate) fn x86_64_asm_xgetbv(xcr: u32) -> u64;
+
+    #[cfg_attr(
+        any(target_env = "gnu", target_env = "musl"),
+        link_name = "_x86_64_asm_xsetbv"
+    )]
+    pub(crate) fn x86_64_asm_xsetbv(xcr: u32, low: u32, high: u32);
 }

src/instructions/interrupts.rs

@@ -13,12 +13,11 @@ pub fn are_enabled() -> bool {
 /// This is a wrapper around the `sti` instruction.
 #[inline]
 pub fn enable() {
-    #[cfg(feature = "inline_asm")]
     unsafe {
+        #[cfg(feature = "inline_asm")]
         asm!("sti", options(nomem, nostack));
-    }
-    #[cfg(not(feature = "inline_asm"))]
-    unsafe {
+
+        #[cfg(not(feature = "inline_asm"))]
         crate::asm::x86_64_asm_interrupt_enable();
     }
 }
@@ -28,13 +27,11 @@ pub fn enable() {
 /// This is a wrapper around the `cli` instruction.
 #[inline]
 pub fn disable() {
-    #[cfg(feature = "inline_asm")]
     unsafe {
+        #[cfg(feature = "inline_asm")]
         asm!("cli", options(nomem, nostack));
-    }
 
-    #[cfg(not(feature = "inline_asm"))]
-    unsafe {
+        #[cfg(not(feature = "inline_asm"))]
         crate::asm::x86_64_asm_interrupt_disable();
     }
 }
@@ -129,26 +126,23 @@ where
 /// information.
 #[inline]
 pub fn enable_and_hlt() {
-    #[cfg(feature = "inline_asm")]
     unsafe {
+        #[cfg(feature = "inline_asm")]
         asm!("sti; hlt", options(nomem, nostack));
-    }
-    #[cfg(not(feature = "inline_asm"))]
-    unsafe {
+
+        #[cfg(not(feature = "inline_asm"))]
         crate::asm::x86_64_asm_interrupt_enable_and_hlt();
     }
 }
 
 /// Cause a breakpoint exception by invoking the `int3` instruction.
 #[inline]
 pub fn int3() {
-    #[cfg(feature = "inline_asm")]
     unsafe {
+        #[cfg(feature = "inline_asm")]
         asm!("int3", options(nomem, nostack));
-    }
 
-    #[cfg(not(feature = "inline_asm"))]
-    unsafe {
+        #[cfg(not(feature = "inline_asm"))]
         crate::asm::x86_64_asm_int3();
     }
 }
@@ -159,18 +153,10 @@ pub fn int3() {
 /// immediate. This macro will be replaced by a generic function when support for
 /// const generics is implemented in Rust.
 #[cfg(feature = "inline_asm")]
+#[cfg_attr(docsrs, doc(cfg(any(feature = "nightly", feature = "inline_asm"))))]
 #[macro_export]
 macro_rules! software_interrupt {
     ($x:expr) => {{
         asm!("int {id}", id = const $x, options(nomem, nostack));
     }};
 }
-
-/// Not implemented
-#[cfg(not(feature = "inline_asm"))]
-#[macro_export]
-macro_rules! software_interrupt {
-    ($x:expr) => {{
-        compile_error!("software_interrupt not implemented for non-nightly");
-    }};
-}

src/instructions/mod.rs

@@ -12,13 +12,11 @@ pub mod tlb;
 /// Halts the CPU until the next interrupt arrives.
 #[inline]
 pub fn hlt() {
-    #[cfg(feature = "inline_asm")]
     unsafe {
-        asm!("hlt", options(nomem, nostack));
-    }
+        #[cfg(feature = "inline_asm")]
+        asm!("hlt", options(nomem, nostack, preserves_flags));
 
-    #[cfg(not(feature = "inline_asm"))]
-    unsafe {
+        #[cfg(not(feature = "inline_asm"))]
         crate::asm::x86_64_asm_hlt();
     }
 }
@@ -31,37 +29,37 @@ pub fn hlt() {
 /// endless loop away.
 #[inline]
 pub fn nop() {
-    #[cfg(feature = "inline_asm")]
     unsafe {
+        #[cfg(feature = "inline_asm")]
         asm!("nop", options(nomem, nostack, preserves_flags));
-    }
 
-    #[cfg(not(feature = "inline_asm"))]
-    unsafe {
+        #[cfg(not(feature = "inline_asm"))]
         crate::asm::x86_64_asm_nop();
     }
 }
 
 /// Emits a '[magic breakpoint](https://wiki.osdev.org/Bochs#Magic_Breakpoint)' instruction for the [Bochs](http://bochs.sourceforge.net/) CPU
 /// emulator. Make sure to set `magic_break: enabled=1` in your `.bochsrc` file.
-#[cfg(feature = "inline_asm")]
 #[inline]
 pub fn bochs_breakpoint() {
     unsafe {
-        asm!("xchg bx, bx", options(nomem, nostack));
+        #[cfg(feature = "inline_asm")]
+        asm!("xchg bx, bx", options(nomem, nostack, preserves_flags));
+
+        #[cfg(not(feature = "inline_asm"))]
+        crate::asm::x86_64_asm_bochs();
     }
 }
 
 /// Gets the current instruction pointer. Note that this is only approximate as it requires a few
 /// instructions to execute.
 #[cfg(feature = "inline_asm")]
+#[cfg_attr(docsrs, doc(cfg(any(feature = "nightly", feature = "inline_asm"))))]
 #[inline(always)]
 pub fn read_rip() -> crate::VirtAddr {
     let rip: u64;
     unsafe {
-        asm!(
-            "lea {}, [rip]", out(reg) rip, options(nostack, nomem)
-        );
+        asm!("lea {}, [rip]", out(reg) rip, options(nostack, nomem, preserves_flags));
     }
     crate::VirtAddr::new(rip)
 }