Each naked function creates separate PE code section

[MauriceKayser]

Description

(At least) on Windows, naked functions are not created in the .text section, but in a .text.First2LettersOfFunctionNameBecause8IsTheSectionNameLengthLimit section, e.g. .text.Ex for a naked function called Example.

Depending on the amount of naked functions, this can bloat the resulting binary file and memory usage.

Maybe the intention was to use the $ sign as a separator instead of a ., which is a magic separation character, as mentioned in #128177, for example .text$FunctionName instead of .text.FunctionName?

Example

1. Create an example.rs:

   Rust code

   #![no_main]
   #![no_std]
   
   #[cfg(not(test))]
   #[panic_handler]
   fn panic(_: &core::panic::PanicInfo) -> ! {
       loop {}
   }
   
   // This ends up in the `.text.Wi` section
   #[unsafe(naked)]
   #[unsafe(no_mangle)]
   extern "system" fn WinMainCRTStartup() -> i32 {
       core::arch::naked_asm!("jmp {}", sym Hello)
   }
   
   // This ends up in the `.text.He` section
   #[unsafe(naked)]
   #[unsafe(no_mangle)]
   extern "system" fn Hello() -> i32 {
       core::arch::naked_asm!("jmp {}", sym Example)
   }
   
   // This ends up in the `.text.Ex` section
   #[unsafe(naked)]
   #[unsafe(no_mangle)]
   extern "system" fn Example() -> i32 {
       core::arch::naked_asm!("
           xor eax, eax
           ret
       ")
   }

2. Compile via: rustc example.rs --emit asm,link,llvm-ir -C panic=abort -C "link-args=/Debug:None /EmitToolVersionInfo:No /NoCoffGrpInfo /NoVcFeature /SubSystem:Windows":

   dumpbin.exe /headers example.exe

   SECTION HEADER #1
   .text.Wi name
          5 virtual size
       1000 virtual address (0000000140001000 to 0000000140001004)
        200 size of raw data
        200 file pointer to raw data (00000200 to 000003FF)
   60000020 flags
            Code
            Execute Read
   
   SECTION HEADER #2
   .text.He name
          5 virtual size
       2000 virtual address (0000000140002000 to 0000000140002004)
        200 size of raw data
        400 file pointer to raw data (00000400 to 000005FF)
   60000020 flags
            Code
            Execute Read
   
   SECTION HEADER #3
   .text.Ex name
          3 virtual size
       3000 virtual address (0000000140003000 to 0000000140003002)
        200 size of raw data
        600 file pointer to raw data (00000600 to 000007FF)
   60000020 flags
            Code
            Execute Read

   LLVM module

   ; ModuleID = 'example.90d8d4ea2542351-cgu.0'
   source_filename = "example.90d8d4ea2542351-cgu.0"
   target datalayout = "e-m:w-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128"
   target triple = "x86_64-pc-windows-msvc"
   
   module asm ".intel_syntax"
   module asm ".pushsection .text.WinMainCRTStartup,\22xr\22"
   module asm ".balign 4"
   module asm ".globl WinMainCRTStartup"
   module asm ".def WinMainCRTStartup"
   module asm ".scl 2"
   module asm ".type 32"
   module asm ".endef"
   module asm "WinMainCRTStartup:"
   module asm "jmp Hello"
   module asm ".popsection"
   module asm ""
   module asm ".att_syntax"
   module asm ".intel_syntax"
   module asm ".pushsection .text.Hello,\22xr\22"
   module asm ".balign 4"
   module asm ".globl Hello"
   module asm ".def Hello"
   module asm ".scl 2"
   module asm ".type 32"
   module asm ".endef"
   module asm "Hello:"
   module asm "jmp Example"
   module asm ".popsection"
   module asm ""
   module asm ".att_syntax"
   module asm ".intel_syntax"
   module asm ".pushsection .text.Example,\22xr\22"
   module asm ".balign 4"
   module asm ".globl Example"
   module asm ".def Example"
   module asm ".scl 2"
   module asm ".type 32"
   module asm ".endef"
   module asm "Example:"
   module asm ""
   module asm "        xor eax, eax"
   module asm "        ret"
   module asm "    "
   module asm ".popsection"
   module asm ""
   module asm ".att_syntax"
   
   @llvm.compiler.used = appending global [2 x ptr] [ptr @Hello, ptr @Example], section "llvm.metadata"
   
   ; __rustc::rust_begin_unwind
   ; Function Attrs: noreturn nounwind uwtable
   define hidden void @_RNvCs691rhTbG0Ee_7___rustc17rust_begin_unwind(ptr align 8 %_1) unnamed_addr #0 {
   start:
     br label %bb1
   
   bb1:                                              ; preds = %bb1, %start
     br label %bb1
   }
   
   ; Function Attrs: noinline nounwind uwtable
   declare i32 @WinMainCRTStartup() unnamed_addr #1
   
   ; Function Attrs: noinline nounwind uwtable
   declare i32 @Hello() unnamed_addr #1
   
   ; Function Attrs: noinline nounwind uwtable
   declare i32 @Example() unnamed_addr #1
   
   attributes #0 = { noreturn nounwind uwtable "target-cpu"="x86-64" "target-features"="+cx16,+sse3,+sahf" }
   attributes #1 = { noinline nounwind uwtable "target-cpu"="x86-64" "target-features"="+cx16,+sse3,+sahf" }
   
   !llvm.module.flags = !{!0, !1}
   !llvm.ident = !{!2}
   
   !0 = !{i32 8, !"PIC Level", i32 2}
   !1 = !{i32 7, !"PIE Level", i32 2}
   !2 = !{!"rustc version 1.88.0 (6b00bc388 2025-06-23)"}

   Assembly module

       .def     @feat.00;
       .scl     3;
       .type    0;
       .endef
       .globl   @feat.00
   .set @feat.00, 0
       .file    "example.90d8d4ea2542351-cgu.0"
   
       .section .text.WinMainCRTStartup,"xr"
       .p2align 2
       .globl   WinMainCRTStartup
       .def     WinMainCRTStartup;
       .scl     2;
       .type    32;
       .endef
   WinMainCRTStartup:
       jmp      Hello
       .text
   
       .section .text.Hello,"xr"
       .p2align 2
       .globl   Hello
       .def     Hello;
       .scl     2;
       .type    32;
       .endef
   Hello:
       jmp      Example
       .text
   
       .section .text.Example,"xr"
       .p2align 2
       .globl   Example
       .def     Example;
       .scl     2;
       .type    32;
       .endef
   Example:
   
       xorl     %eax, %eax
       retq
   
       .text
   
       .def     _RNvCs691rhTbG0Ee_7___rustc17rust_begin_unwind;
       .scl     2;
       .type    32;
       .endef
       .section .text,"xr",one_only,_RNvCs691rhTbG0Ee_7___rustc17rust_begin_unwind
       .globl   _RNvCs691rhTbG0Ee_7___rustc17rust_begin_unwind
       .p2align 4
   _RNvCs691rhTbG0Ee_7___rustc17rust_begin_unwind:
       jmp      .LBB0_1
   .LBB0_1:
       jmp      .LBB0_1

3. Explicitly set the section name with a $ separator to fix the issue:

   Rust code

   - // This ends up in the `.text.Wi` section
   + #[unsafe(link_section = ".text$WinMainCRTStartup")]
   
   - // This ends up in the `.text.He` section
   + #[unsafe(link_section = ".text$Hello")]
   
   - // This ends up in the `.text.Ex` section
   + #[unsafe(link_section = ".text$Example")]

4. Compile again:

   dumpbin.exe /headers example.exe

   SECTION HEADER #1
      .text name
         11 virtual size
       1000 virtual address (0000000140001000 to 0000000140001010)
        200 size of raw data
        200 file pointer to raw data (00000200 to 000003FF)
   60000020 flags
            Code
            Execute Read

   LLVM module

   ; ModuleID = 'example.90d8d4ea2542351-cgu.0'
   source_filename = "example.90d8d4ea2542351-cgu.0"
   target datalayout = "e-m:w-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128"
   target triple = "x86_64-pc-windows-msvc"
   
   module asm ".intel_syntax"
   module asm ".pushsection .text$WinMainCRTStartup,\22xr\22"
   module asm ".balign 4"
   module asm ".globl WinMainCRTStartup"
   module asm ".def WinMainCRTStartup"
   module asm ".scl 2"
   module asm ".type 32"
   module asm ".endef"
   module asm "WinMainCRTStartup:"
   module asm "jmp Hello"
   module asm ".popsection"
   module asm ""
   module asm ".att_syntax"
   module asm ".intel_syntax"
   module asm ".pushsection .text$Hello,\22xr\22"
   module asm ".balign 4"
   module asm ".globl Hello"
   module asm ".def Hello"
   module asm ".scl 2"
   module asm ".type 32"
   module asm ".endef"
   module asm "Hello:"
   module asm "jmp Example"
   module asm ".popsection"
   module asm ""
   module asm ".att_syntax"
   module asm ".intel_syntax"
   module asm ".pushsection .text$Example,\22xr\22"
   module asm ".balign 4"
   module asm ".globl Example"
   module asm ".def Example"
   module asm ".scl 2"
   module asm ".type 32"
   module asm ".endef"
   module asm "Example:"
   module asm ""
   module asm "        xor eax, eax"
   module asm "        ret"
   module asm "    "
   module asm ".popsection"
   module asm ""
   module asm ".att_syntax"
   
   @llvm.compiler.used = appending global [2 x ptr] [ptr @Hello, ptr @Example], section "llvm.metadata"
   
   ; __rustc::rust_begin_unwind
   ; Function Attrs: noreturn nounwind uwtable
   define hidden void @_RNvCs691rhTbG0Ee_7___rustc17rust_begin_unwind(ptr align 8 %_1) unnamed_addr #0 {
   start:
     br label %bb1
   
   bb1:                                              ; preds = %bb1, %start
     br label %bb1
   }
   
   ; Function Attrs: noinline nounwind uwtable
   declare i32 @WinMainCRTStartup() unnamed_addr #1
   
   ; Function Attrs: noinline nounwind uwtable
   declare i32 @Hello() unnamed_addr #1
   
   ; Function Attrs: noinline nounwind uwtable
   declare i32 @Example() unnamed_addr #1
   
   attributes #0 = { noreturn nounwind uwtable "target-cpu"="x86-64" "target-features"="+cx16,+sse3,+sahf" }
   attributes #1 = { noinline nounwind uwtable "target-cpu"="x86-64" "target-features"="+cx16,+sse3,+sahf" }
   
   !llvm.module.flags = !{!0, !1}
   !llvm.ident = !{!2}
   
   !0 = !{i32 8, !"PIC Level", i32 2}
   !1 = !{i32 7, !"PIE Level", i32 2}
   !2 = !{!"rustc version 1.88.0 (6b00bc388 2025-06-23)"}

   Assembly module

       .def     @feat.00;
       .scl     3;
       .type    0;
       .endef
       .globl   @feat.00
   .set @feat.00, 0
       .file    "example.90d8d4ea2542351-cgu.0"
   
       .section .text$WinMainCRTStartup,"xr"
       .p2align 2
       .globl   WinMainCRTStartup
       .def     WinMainCRTStartup;
       .scl     2;
       .type    32;
       .endef
   WinMainCRTStartup:
       jmp      Hello
       .text
   
       .section .text$Hello,"xr"
       .p2align 2
       .globl   Hello
       .def     Hello;
       .scl     2;
       .type    32;
       .endef
   Hello:
       jmp      Example
       .text
   
       .section .text$Example,"xr"
       .p2align 2
       .globl   Example
       .def     Example;
       .scl     2;
       .type    32;
       .endef
   Example:
   
       xorl     %eax, %eax
       retq
   
       .text
   
       .def     _RNvCs691rhTbG0Ee_7___rustc17rust_begin_unwind;
       .scl     2;
       .type    32;
       .endef
       .section .text,"xr",one_only,_RNvCs691rhTbG0Ee_7___rustc17rust_begin_unwind
       .globl   _RNvCs691rhTbG0Ee_7___rustc17rust_begin_unwind
       .p2align 4
   _RNvCs691rhTbG0Ee_7___rustc17rust_begin_unwind:
       jmp      .LBB0_1
   .LBB0_1:
       jmp      .LBB0_1

5. In comparison to naked functions, this does not happen with global_asm. Replace the 3 naked functions and their attributes with the following:

   Rust code

   // These end up in the `.text` section
   core::arch::global_asm!("
       .global WinMainCRTStartup
       WinMainCRTStartup:
           jmp Hello
   ");
   
   core::arch::global_asm!("
       .global Hello
       Hello:
           jmp Example
   ");
   
   core::arch::global_asm!("
       .global Example
       Example:
           xor eax, eax
           ret
   ");

6. Compile again:

   dumpbin.exe /headers example.exe

   SECTION HEADER #1
   .text name
          7 virtual size
       1000 virtual address (0000000140001000 to 0000000140001006)
        200 size of raw data
        200 file pointer to raw data (00000200 to 000003FF)
   60000020 flags
            Code
            Execute Read

   LLVM module

   ; ModuleID = 'example.90d8d4ea2542351-cgu.0'
   source_filename = "example.90d8d4ea2542351-cgu.0"
   target datalayout = "e-m:w-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128"
   target triple = "x86_64-pc-windows-msvc"
   
   module asm ".intel_syntax"
   module asm ""
   module asm "    .global WinMainCRTStartup"
   module asm "    WinMainCRTStartup:"
   module asm "        jmp Hello"
   module asm ""
   module asm ".att_syntax"
   module asm ".intel_syntax"
   module asm ""
   module asm "    .global Hello"
   module asm "    Hello:"
   module asm "        jmp Example"
   module asm ""
   module asm ".att_syntax"
   module asm ".intel_syntax"
   module asm ""
   module asm "    .global Example"
   module asm "    Example:"
   module asm "        xor eax, eax"
   module asm "        ret"
   module asm ""
   module asm ".att_syntax"
   
   ; __rustc::rust_begin_unwind
   ; Function Attrs: noreturn nounwind uwtable
   define hidden void @_RNvCs691rhTbG0Ee_7___rustc17rust_begin_unwind(ptr align 8 %_1) unnamed_addr #0 {
   start:
   br label %bb1
   
   bb1:                                              ; preds = %bb1, %start
   br label %bb1
   }
   
   attributes #0 = { noreturn nounwind uwtable "target-cpu"="x86-64" "target-features"="+cx16,+sse3,+sahf" }
   
   !llvm.module.flags = !{!0, !1}
   !llvm.ident = !{!2}
   
   !0 = !{i32 8, !"PIC Level", i32 2}
   !1 = !{i32 7, !"PIE Level", i32 2}
   !2 = !{!"rustc version 1.88.0 (6b00bc388 2025-06-23)"}

   Assembly module

       .def     @feat.00;
       .scl     3;
       .type    0;
       .endef
       .globl   @feat.00
   .set @feat.00, 0
       .file    "example.90d8d4ea2542351-cgu.0"
   
       .globl   WinMainCRTStartup
   WinMainCRTStartup:
       jmp      Hello
   
       .globl   Hello
   Hello:
       jmp      Example
   
       .globl   Example
   Example:
       xorl     %eax, %eax
       retq
   
       .def     _RNvCs691rhTbG0Ee_7___rustc17rust_begin_unwind;
       .scl     2;
       .type    32;
       .endef
       .section .text,"xr",one_only,_RNvCs691rhTbG0Ee_7___rustc17rust_begin_unwind
       .globl   _RNvCs691rhTbG0Ee_7___rustc17rust_begin_unwind
       .p2align 4
   _RNvCs691rhTbG0Ee_7___rustc17rust_begin_unwind:
       jmp      .LBB0_1
   .LBB0_1:
       jmp      .LBB0_1

Expected result

The first example code should only create one section, named .text, instead of .text.Wi, .text.He, and .text.Ex.

[bjorn3]

cc @folkertdev

[folkertdev]

This choice is based on the behavior for ELF where rust puts each function into its own section, so that the linker can garbage-collect sections whose contents are not used. In other words, in the large, this choice is meant to improve binary size.

Other binary formats and toolchains appear to maybe support this idea less well. Actually for windows there is even a difference between msvc and gnu targets: gnu puts all functions into one section, msvc uses subsections (for gnu you can force subsections with -Zfunction-sections):

https://godbolt.org/z/1K6hEbq6a

based on some really old release notes (for rust 0.11 from 2014), -Zfunction-sections should be the default. I guess that's changed in the meantime.

• Functionality equivalent to GCC & Clang's -ffunction-sections,
  -fdata-sections and --gc-sections has been enabled by default

I'll look into if/how we can match what LLVM does for standard functions more closely.

[ChrisDenton]

So the problem here isn't the separate sections per se. It's the fact they aren't merged into .text in the final binary? Either we need to explicitly tell the linker to merge the sections (/MERGE) or we take advantage of the fact that a $ (and anything after it) is automatically dropped from the section name in the final binary (so .text$ABC will automatically be merged with .text, see grouped sections).

[folkertdev]

I think that is the problem, yes. We can easily use $ (apparently LLVM can't handle it when using inline assembly and intel syntax llvm/llvm-project#101776, but when using module assembly it works as expected).

Should I try to make the gnu/msvc distinction, and not emit the section when targeting gnu? It seems straightforward, but does complicate the code and tests a bit.

[ChrisDenton]

That would seem like a good idea to me but I would just check in with @mati865 about the above

[bjorn3]

You could use sess.opts.unstable_opts.function_sections.unwrap_or(sess.target.function_sections) to match what cg_llvm does for determining if subsections should be used. Just make sure to check that on macOS we do still remove dead naked functions. Mach-O splits a section on symbols for ld64 -dead_strip rather than using subsections, but that requires a certain section flag is specified. I don't know if that flag would be enabled by .pushsection.

[mati865]

I don't have time to look into this right now. Keep in mind, function sections are disabled for windows-gnu (although enabled for windows-gnullvm) because they were causing crashes:

rust/compiler/rustc_target/src/spec/base/windows_gnu.rs

Line 84 in a41214f

function_sections: false,

[folkertdev]

In short the proposal here is to have naked functions follow the configuration for function_sections. Currently each naked function always gets its own section, no matter the target or -Zfunction-sections.