[Tarantool-patches] [PATCH luajit] ARM64: Fix LDP code generation.

[Sergey Bronnikov]

Hi, Sergey,

thanks for the patch! LGTM

Sergey

On 6/11/25 19:01, Sergey Kaplun wrote:
> From: Mike Pall <mike>
>
> Thanks to Zhongwei Yao.
>
> (cherry picked from commit 9493acc1a28f15b0ac4453e716f33436186c7acd)
>
> When fusing two LDR (STR) instructions to the single LDP (STP)
> instruction, the arm64 emitter shifts the offset value to encode the
> immediate. In the case when the offset is negative, the resulting field
> value exceeds the 7-bit length of the immediate, see [1]. This results
> in the invalid instruction decoding.
>
> This patch fixes this by masking the value with the 7-bit-width mask
> `0x7f`.
>
> Sergey Kaplun:
> * added the description and the test for the problem
>
> [1]:https://developer.arm.com/documentation/ddi0602/2025-03/Base-Instructions/LDP--Load-pair-of-registers-
>
> Part of tarantool/tarantool#11278
> ---
>
> Related issues:
> *https://github.com/LuaJIT/LuaJIT/pull/1028
> *https://github.com/tarantool/tarantool/issues/11278
> Branch:https://github.com/tarantool/luajit/tree/skaplun/lj-1028-ldr-fusion-to-ldp-negative-offset
>
>   src/lj_emit_arm64.h                           |  2 +-
>   ...ldr-fusion-to-ldp-negative-offset.test.lua | 45 +++++++++++++++++++
>   2 files changed, 46 insertions(+), 1 deletion(-)
>   create mode 100644 test/tarantool-tests/lj-1028-ldr-fusion-to-ldp-negative-offset.test.lua
>
> diff --git a/src/lj_emit_arm64.h b/src/lj_emit_arm64.h
> index e1a9d3e4..30cd3505 100644
> --- a/src/lj_emit_arm64.h
> +++ b/src/lj_emit_arm64.h
> @@ -143,7 +143,7 @@ static void emit_lso(ASMState *as, A64Ins ai, Reg rd, Reg rn, int64_t ofs)
>         goto nopair;
>       }
>       if (ofsm >= (int)((unsigned int)-64<<sc) && ofsm <= (63<<sc)) {
> -      *as->mcp = aip | A64F_N(rn) | ((ofsm >> sc) << 15) |
> +      *as->mcp = aip | A64F_N(rn) | (((ofsm >> sc) & 0x7f) << 15) |
>   	(ai ^ ((ai == A64I_LDRx || ai == A64I_STRx) ? 0x50000000 : 0x90000000));
>         return;
>       }
> diff --git a/test/tarantool-tests/lj-1028-ldr-fusion-to-ldp-negative-offset.test.lua b/test/tarantool-tests/lj-1028-ldr-fusion-to-ldp-negative-offset.test.lua
> new file mode 100644
> index 00000000..1ba28449
> --- /dev/null
> +++ b/test/tarantool-tests/lj-1028-ldr-fusion-to-ldp-negative-offset.test.lua
> @@ -0,0 +1,45 @@
> +local tap = require('tap')
> +local ffi = require('ffi')
> +
> +-- This test demonstrates LuaJIT's incorrect emitting of LDP
> +-- instruction with negative offset fused from LDR on arm64.
> +-- See alsohttps://github.com/LuaJIT/LuaJIT/pull/1028.
> +local test = tap.test('lj-1028-ldr-fusion-to-ldp-negative-offset'):skipcond({
> +  ['Test requires JIT enabled'] = not jit.status(),
> +})
> +
> +test:plan(1)
> +
> +-- Amount of iterations to compile and start the trace.
> +local N_ITERATIONS = 4
> +
> +ffi.cdef[[
> +  typedef struct data {
> +    int32_t m1;
> +    int32_t m2;
> +  } data;
> +]]
> +
> +local data_arr = ffi.new('data[' .. N_ITERATIONS .. ']')
> +
> +local const_data_ptr = ffi.typeof('const data *')
> +local data = ffi.cast(const_data_ptr, data_arr)
> +
> +local results = {}
> +
> +jit.opt.start('hotloop=1')
> +
> +for i = 1, N_ITERATIONS do
> +  -- Pair loading from the negative offset generates an invalid
> +  -- instruction on AArch64 before this patch.
> +  local field = data[i - 1]
> +  local m1 = field.m1
> +  local m2 = field.m2
> +
> +  -- Use loaded values to avoid DCE.
> +  results[i] = m1 + m2
> +end
> +
> +test:samevalues(results, 'no invalid instruction')
> +
> +test:done(true)
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.tarantool.org/pipermail/tarantool-patches/attachments/20250625/847a7a47/attachment.htm>