On Thu, 24 Jul 2025, Radek Barton wrote:
> Hello.
>
> Thank you for noticing. Here's the fixed version.
>
> Radek
>
> ---
> From 353e7120b660a5bf5abdff7afbd652c666c66bf8 Mon Sep 17 00:00:00 2001
> From: =?UTF-8?q?Radek=20Barto=C5=88?= <radek.bar...@microsoft.com>
> Date: Sat, 19 Jul 2025 19:17:12 +0200
> Subject: [PATCH v3] Cygwin: mkimport: implement AArch64 +/-4GB relocations
> MIME-Version: 1.0
> Content-Type: text/plain; charset=UTF-8
> Content-Transfer-Encoding: 8bit
>
> Based on https://sourceware.org/pipermail/cygwin-patches/2025q3/014154.html
> suggestion, this patch implements +/-4GB relocations for AArch64 in the
> mkimport
> script by using adrp and ldr instructions. This change required update
> in winsup/cygwin/mm/malloc_wrapper.cc where those instructions are
> decoded to get target import address.
>
> Signed-off-by: Radek BartoĊ <radek.bar...@microsoft.com>
> ---
> winsup/cygwin/mm/malloc_wrapper.cc | 26 +++++++++++++++++---------
> winsup/cygwin/scripts/mkimport | 7 +++++--
> 2 files changed, 22 insertions(+), 11 deletions(-)
>
> diff --git a/winsup/cygwin/mm/malloc_wrapper.cc
> b/winsup/cygwin/mm/malloc_wrapper.cc
> index 863d3089c..20d933bf5 100644
> --- a/winsup/cygwin/mm/malloc_wrapper.cc
> +++ b/winsup/cygwin/mm/malloc_wrapper.cc
> @@ -51,16 +51,24 @@ import_address (void *imp)
> __try
> {
> #if defined(__aarch64__)
> - // If opcode is an adr instruction.
> - uint32_t opcode = *(uint32_t *) imp;
> - if ((opcode & 0x9f000000) == 0x10000000)
> + /* If opcode1 is an adrp and opcode2 is ldr instruction:
> + - https://www.scs.stanford.edu/~zyedidia/arm64/adrp.html
> + - https://www.scs.stanford.edu/~zyedidia/arm64/ldr_imm_gen.html
> + NOTE: This implementation assumes that the relocation table is made
> of
> + those specific AArch64 instructions as generated by the
> + winsup/cygwin/scripts/mkimport script. Please, keep it in sync. */
> + uint32_t opcode1 = *((uint32_t *) imp);
> + uint32_t opcode2 = *(((uint32_t *) imp) + 1);
> + if (((opcode1 & 0x9f000000) == 0x90000000) && ((opcode2 & 0xbfc00000)
> == 0xb9400000))
These checks could be tightened up a bit given the extremely limited
instructions that are used.
+ uint32_t opcode3 = *(((uint32_t *) imp) + 2);
+ if (((opcode1 & 0x9f00001f) == 0x90000010) &&
((opcode2 & 0xffc003ff) == 0xf9400210) &&
(opcode3 == 0xd61f0200))
This ensures:
* xip0 aka x16 is the register used
* the ldr is 64-bit (you already were assuming that by the "64 bit scale"
comment below, but this checks that that bit is set)
* the next instruction after the ldr is a br x16
If you wanted to cover the binutils case, something like:
if (((opcode1 & 0x9f00001f) == 0x90000010) &&
(((opcode2 & 0xffc003ff) == 0xf9400210) &&
(opcode3 == 0xd61f0200)) ||
(((opcode2 & 0xffc003ff) == 0x91000210) &&
(opcode3 == 0xf9400210) &&
(*(((uint32_t *) imp) + 3) == 0xd61f0200))
The 12 bits in the ldr correspond to the same 12 bits in the add, just not
64 bit scaled so that would have to be checked in the below code.
> {
> - uint32_t immhi = (opcode >> 5) & 0x7ffff;
> - uint32_t immlo = (opcode >> 29) & 0x3;
> - int64_t sign_extend = (0l - (immhi >> 18)) << 21;
> - int64_t imm = sign_extend | (immhi << 2) | immlo;
> - uintptr_t jmpto = *(uintptr_t *) ((uint8_t *) imp + imm);
> - return (void *) jmpto;
> + uint32_t immhi = (opcode1 >> 5) & 0x7ffff;
> + uint32_t immlo = (opcode1 >> 29) & 0x3;
> + uint32_t imm12 = ((opcode2 >> 10) & 0xfff) * 8; // 64 bit scale
> + int64_t sign_extend = (0l - ((int64_t) immhi >> 32)) << 33; // sign
> extend from 33 to 64 bits
> + int64_t imm = sign_extend | (((immhi << 2) | immlo) << 12);
> + int64_t base = (int64_t) imp & ~0xfff;
> + uintptr_t* jmpto = (uintptr_t *) (base + imm + imm12);
> + return (void *) *jmpto;
> }
> #else
> if (*((uint16_t *) imp) == 0x25ff)
> diff --git a/winsup/cygwin/scripts/mkimport b/winsup/cygwin/scripts/mkimport
> index 0c1bcafbf..33d8b08fb 100755
> --- a/winsup/cygwin/scripts/mkimport
> +++ b/winsup/cygwin/scripts/mkimport
> @@ -73,8 +73,11 @@ EOF
> .extern $imp_sym
> .global $glob_sym
> $glob_sym:
> - adr x16, $imp_sym
> - ldr x16, [x16]
> + # NOTE: Using instructions that are used by MSVC and LLVM. Binutils are
> + # using adrp/add/ldr-0-offset though. Please, keep it in sync with
> + # import_address implementation in winsup/cygwin/mm/malloc_wrapper.cc.
> + adrp x16, $imp_sym
> + ldr x16, [x16, #:lo12:$imp_sym]
> br x16
> EOF
> } else {
> --
> 2.50.1.vfs.0.0
>
--
When more and more people are thrown out of work, unemployment
results.
-- Calvin Coolidge
