Re: [PATCH RFC] Add support for non-contiguous memory regions

[Simon Richter <Simon dot Richter at hogyros dot de>]

Hi,

On Mon, Dec 02, 2019 at 10:56:22AM +0100, Christophe Lyon wrote:

> > > The attached patches implement support for non-contiguous memory
> > > regions

> > How does this interact with relaxations?

> I haven't tested that specifically yet. Do you have a target /
> testcase in mind that I could check?

A quick testcase would be on PowerPC:

Assembler file "test.S"

		.machine "ppc"

		.section .text.one
		b       2f

		.section .text.two
	2:
		nop

Linker script "test.ld"

	MEMORY {
		one (RXAI) : ORIGIN = 0x00000000, LENGTH = 0x10000000
		two (RXAI) : ORIGIN = 0x20000000, LENGTH = 0x10000000
	}

	SECTIONS {
		one : {
			*(.text.one)
		} > one
		two : {
			*(.text.two)
		} > two
	}

Assemble:

	$ as -o test.o test.S

Disassemble:

	$ objdump -dr test.o

	test.o:     file format elf64-powerpcle


	Disassembly of section .text.one:

	0000000000000000 <.text.one>:
	   0:   00 00 00 48     b       0x0
				0: R_PPC64_REL24        .text.two

	Disassembly of section .text.two:

	0000000000000000 <.text.two>:
	   0:   00 00 00 60     nop

Link:

	$ ld -o test -T test.ld --relax test.o

Disassemble:

	$ objdump -dr test

	test:     file format elf64-powerpcle

	Disassembly of section one:

	0000000000000000 <0000001b.plt_branch.1c:5>:
	   0:   08 80 82 e9     ld      r12,-32760(r2)
	   4:   a6 03 89 7d     mtctr   r12
	   8:   20 04 80 4e     bctr
		...
	  20:   e0 ff ff 4b     b       0 <0000001b.plt_branch.1c:5>

	Disassembly of section two:

	0000000020000000 <two>:
	    20000000:   00 00 00 60     nop

The linker generates the trampoline as the branch displacement doesn't fit
the instruction, extending the section size by 12 bytes. This happens after
addresses are assigned, because that is when the linker learns that the
displacement is too large.

There is also the opposite case, where branch instructions are replaced by
shorter variants on some architectures during relaxation, but I don't have
an example immediately ready.

The desired behaviour for the combination of non-contiguous regions and
relaxations should probably be defined by someone from the core team --
probably locking out the combination (like "-r --relax" on PowerPC) or just
ensuring that it doesn't crash is the sanest thing to do here.

   Simon