christoph/wamr
1
Fork 0
Code Releases Activity

aot compiler: Use larger alignment for load/store when possible (#3552)

Browse Source 
Consider the following wasm module:
```wast
(module
  (func (export "foo")
    i32.const 0x104
    i32.const 0x12345678
    i32.store
  )
  (memory 1 1)
)
```

While the address (0x104) is perfectly aligned for i32.store,
as our aot compiler uses 1-byte alignment for load/store LLVM
IR instructions, it often produces inefficient machine code,
especially for alignment-sensitive targets.

For example, the above "foo" function is compiled into the
following xtensa machine code.
```
0000002c <aot_func_internal#0>:
  2c:   004136          entry   a1, 32
  2f:   07a182          movi    a8, 0x107
  32:   828a            add.n   a8, a2, a8
  34:   291c            movi.n  a9, 18
  36:   004892          s8i     a9, a8, 0
  39:   06a182          movi    a8, 0x106
  3c:   828a            add.n   a8, a2, a8
  3e:   ffff91          l32r    a9, 3c <aot_func_internal#0+0x10> (ff91828a <aot_func_internal#0+0xff91825e>)
                        3e: R_XTENSA_SLOT0_OP   .literal+0x8
  41:   004892          s8i     a9, a8, 0
  44:   05a182          movi    a8, 0x105
  47:   828a            add.n   a8, a2, a8
  49:   ffff91          l32r    a9, 48 <aot_func_internal#0+0x1c> (ffff9182 <aot_func_internal#0+0xffff9156>)
                        49: R_XTENSA_SLOT0_OP   .literal+0xc
  4c:   41a890          srli    a10, a9, 8
  4f:   0048a2          s8i     a10, a8, 0
  52:   04a182          movi    a8, 0x104
  55:   828a            add.n   a8, a2, a8
  57:   004892          s8i     a9, a8, 0
  5a:   f01d            retw.n
```

Note that the each four bytes are stored separately using
one-byte-store instruction, s8i.

This commit tries to use larger alignments for load/store LLVM IR
instructions when possible.  with this commit, the above example is
compiled into the following machine code, which seems more reasonable.
```
0000002c <aot_func_internal#0>:
  2c:   004136          entry   a1, 32
  2f:   ffff81          l32r    a8, 2c <aot_func_internal#0> (81004136 <aot_func_internal#0+0x8100410a>)
                        2f: R_XTENSA_SLOT0_OP   .literal+0x8
  32:   416282          s32i    a8, a2, 0x104
  35:   f01d            retw.n
```

Note: this doesn't work well for --xip because aot_load_const_from_table()
hides the constness of the value. Maybe we need our own mechanism to
propagate the constness and the value.
This commit is contained in:
YAMAMOTO Takashi
2024-06-22 11:32:52 +09:00
committed by GitHub
parent 607ae47d7b
commit 4c2af25aff
5 changed files with 65 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
core/iwasm/common/wasm_memory.c
View File

@ -883,6 +883,12 @@ wasm_enlarge_memory_internal(WASMModuleInstance *module, uint32 inc_page_count)
}
#endif /* end of WASM_MEM_ALLOC_WITH_USAGE */
/*
* AOT compiler assumes at least 8 byte alignment.
* see aot_check_memory_overflow.
*/
bh_assert(((uintptr_t)memory->memory_data & 0x7) == 0);
memory->num_bytes_per_page = num_bytes_per_page;
memory->cur_page_count = total_page_count;
memory->max_page_count = max_page_count;
@ -1032,5 +1038,11 @@ wasm_allocate_linear_memory(uint8 **data, bool is_shared_memory,
#endif
}
/*
* AOT compiler assumes at least 8 byte alignment.
* see aot_check_memory_overflow.
*/
bh_assert(((uintptr_t)*data & 0x7) == 0);
return BHT_OK;
}