ld.lld: ALIGN() directive incorrectly increasing LMA after empty section

[renesas-kyle-finch]

An ALIGN() directive applied to the LMA of an empty section will influence the LMA of the next non-empty section that follows.

However, when using GCC/ld, in this scenario, the next non-empty section's LMA will begin at the last available LMA address prior to the ALIGN occurirng on the empty section's LMA.

For a simplified example, I modified a sample project from the embedded toolchain.

Changes:

1. Add map file generation to the build target of the Makefile - $(BIN_PATH)/clang $(MICROBIT_TARGET) $(CRT) -g -Xlinker -Map=hello.map -T ../../ldscripts/microbit.ld -o hello.elf $^
2. Add the align_bug section to /opt/LLVMEmbeddedToolchainForArm-17.0.0-Linux-x86_64/lib/clang-runtimes/arm-none-eabi/armv6m_soft_nofp/lib/picolibcpp.ld
   

Actual Result: LMA of .data is the aligned LMA of empty section .align_bug

    3320     3320       28     4 .except_unordered
    3320     3320        0     1         . = ALIGN ( 8 )
    3320     3320       28     4         <internal>:(.ARM.exidx)
20000000     3350        0    16 .align_bug
20000000     3350        0     1         . = .
20000000     3350       38     4 .data
20000000     3350       30     4         /tmp/hello-6adfd4.o:(.data)
20000000     3350        4     1                 gpio_dir_set
20000004     3354        4     1                 uart_starttx
20000008     3358        4     1                 uart_txdrdy

Expected Result: LMA of .data is the the address immediately following 0x3320 + 0x28 (LMA of .except_unordered + size of .except_unordered)

    3320     3320       28     4 .except_unordered
    3320     3320        0     1         . = ALIGN ( 8 )
    3320     3320       28     4         <internal>:(.ARM.exidx)
20000000     3348       38     4 .data
20000000     3348       30     4         /tmp/hello-dc4e29.o:(.data)
20000000     3348        4     1                 gpio_dir_set
20000004     334c        4     1                 uart_starttx
20000008     3350        4     1                 uart_txdrdy

[smithp35]

A simpler reproducer

       .section .text.01, "ax", %progbits
        .balign 4
        nop

        .section .data.01, "a", %progbits
        .balign 4
        .word 10

clang --target=arm-none-eabi -c ld.s
With linker script

MEMORY {
        flash (rx!w) : ORIGIN = 0x0, LENGTH = 0x1000
        ram (w!rx) : ORIGIN = 0x1000, LENGTH = 0x1000
}

SECTIONS
{
  .text : { *(.text .text.*) } >flash AT>flash
  .align : ALIGN(16) {
    . = .;
    KEEP(*(.align))
  } >ram AT>flash
  .data : { *(.data .data.*) } >ram AT>flash
}

ld.lld ld.o --script=ld.ld --print-map

     VMA      LMA     Size Align Out     In      Symbol
       0        0        4     4 .text
       0        0        0     4         ld.o:(.text)
       0        0        4     4         ld.o:(.text.01)
       0        0        0     1                 $a.0
    1000     1000        0    16 .align
    1000     1000        0     1         . = .
    1000       10        4     4 .data
    1000       10        4     4         ld.o:(.data.01)

When compared to GNU ld

.text           0x00000000        0x4
 *(.text .text.*)
 .text          0x00000000        0x0 ld.o
 .text.01       0x00000000        0x4 ld.o
...
.align          0x00001000        0x0 load address 0x00000010
                0x00001000                        . = .
 *(.align)

.data           0x00001000       0x4 load address 0x00000004
 *(.data .data.*)
 .data.01       0x00001000        0x4 ld.o

It is likely that GNU ld has some special case code for 0 sized sections and LMA.

I'm not sure LLD upstream will consider this a bug, or just a difference in behaviour from GNU ld. The specification/documentation for what ALIGN and ALIGN_WITH_INPUT does to LMA in https://sourceware.org/binutils/docs/ld/Output-Section-Description.html are absent. LLD does describe its behaviour here https://lld.llvm.org/ELF/linker_script.html#output-section-lma

I'm aware of some other differences particularly when ALIGN is not used. LLD will do as its docs say and will align the LMA using the section alignment. GNU ld will not align in LMA at all unless ALIGN_WITH_INPUT is used, in which case it will replicate the amount of padding used by the VMA, which is close to, but not identical to LLD (which does not currently support ALIGN_WITH_INPUT). This means that lld will align more in LMA than GNU ld without ALIGN_WITH_INPUT.

I think some action can be taken for empty output sections, but we'll need to reverse engineer what GNU ld is doing with 0 sized output sections. It is easy to break some program that is depending on the existing behaviour if we're not careful.