-fsplit-ldp-stp

Description

Each ldp or stp instruction with poor performance is split into two ldr or str instructions.

Usage

Use the -fsplit-ldp-stp option to enable optimization. --param=param-ldp-dependency-search-range=[1,32]: specifies the search scope. The default value is 16.

Note: This optimization requires the -O1 or higher optimization level.

Result

The test case is as follows:

int __RTL (startwith ("split_complex_instructions"))
simple_ldp_after_store ()
{
(function "simple_ldp_after_store"
  (insn-chain
    (block 2
      (edge-from entry (flags "FALLTHRU"))
      (cnote 3 [bb 2] NOTE_INSN_BASIC_BLOCK)
      (cinsn 228 (set (reg/i:DI sp)
                   (reg/i:DI x0)))
      (cinsn 238 (set (reg/i:DI x1)
                   (reg/i:DI x0)))
      (cinsn 101 (set (mem/c:DI
                        (plus:DI (reg/f:DI sp)
                          (const_int 8))[1 S4 A32])(reg:DI x0)))
      (cinsn 10 (parallel [
        (set (reg:DI x29)
          (mem:DI (plus:DI (reg/f:DI sp) (const_int 8)) [1 S4 A32]))
        (set (reg:DI x30)
          (mem:DI (plus:DI (reg/f:DI sp)
            (const_int 16)) [1 S4 A32]))]))
      (cinsn 102 (set (mem/c:DI (plus:DI (reg/f:DI x1)
                                          (const_int -16)) [1 S4 A32])
                      (reg:DI x0)))
      (cinsn 11 (parallel [
        (set (reg:DI x3)
          (mem:DI (plus:DI (reg/f:DI x1) (const_int -16)) [1 S4 A32]))
        (set (reg:DI x4)
          (mem:DI (plus:DI (reg/f:DI x1) (const_int -8)) [1 S4 A32]))
      ]))
      (cinsn 103 (set (mem/c:DI (reg/f:DI x1) [1 S4 A32])
                      (reg:DI x0)))
      (cinsn 12 (parallel [
        (set (reg:DI x5) (mem:DI (reg/f:DI x1) [1 S4 A32]))
        (set (reg:DI x6) (mem:DI (plus:DI (reg/f:DI x1)
                                          (const_int 8)) [1 S4 A32]))
      ]))
      (cinsn 13 (use (reg/i:DI sp)))
      (cinsn 14 (use (reg/i:DI cc)))
      (cinsn 15 (use (reg/i:DI x29)))
      (cinsn 16 (use (reg/i:DI x30)))
      (cinsn 17 (use (reg/i:DI x0)))
      (cinsn 18 (use (reg/i:DI x3)))
      (cinsn 19 (use (reg/i:DI x4)))
      (cinsn 20 (use (reg/i:DI x5)))
      (cinsn 21 (use (reg/i:DI x6)))
      (edge-to exit (flags "FALLTHRU"))
    ) ;; block 2
  ) ;; insn-chain
) ;; function "simple_ldp_after_store"
}

Test command:

gcc -O1 -fsplit-ldp-stp -S test.c -o test.s

Figure 1 Option disabled

Figure 2 Option enabled

After this option is enabled, the generated assembly code does not use the ldp instruction. Instead, it is split into two ldr instructions.

Parent topic: Static Compilation Optimization