[PATCH] D118549: apply two more cases for store combine

[Baoshan Pang via Phabricator via llvm-commits]

BaoshanPang added a comment.

In D118549#3304383 <https://reviews.llvm.org/D118549#3304383>, @spatel wrote:

> Adding more potential reviewers. It might help if the baseline tests are committed, so we can see the asm difference for each test. Is it the alignment or something else that prevents combining these stores currently?

The .ll test case is derived from this C file:

  struct ab {
      char a;
      char b;
      char c;
      char d;
  } __attribute__ ((aligned (2)));
  
  void store0(struct ab *_ab, unsigned v)
  {
      _ab->a = v&0xff;
      _ab->b = (v>>8)&0xff;
      _ab->c = (v>>16)&0xff;
      _ab->d = (v>>24)&0xff;
  }
  void store1(struct ab *_ab, unsigned v)
  {
      _ab->a = v&0xff;
      _ab->b = (v>>8)&0xff;
  }
  void store2(struct ab *_ab, unsigned v)
  {
      _ab->c = (v>>16)&0xff;
      _ab->d = (v>>24)&0xff;
  }
  void store3(struct ab *_ab, unsigned v)
  {
      _ab->c = (v)&0xff;
      _ab->d = (v>>8)&0xff;
  }

Before my change, the generated asm code is like this:

      .text
      .attribute  4, 16
      .attribute  5, "rv32i2p0_m2p0_a2p0_c2p0"
      .file   "t.c"
      .globl  store0
      .p2align    1
      .type   store0, at function
  store0:
      sb  a1, 0(a0)
      srli    a2, a1, 8
      sb  a2, 1(a0)
      srli    a2, a1, 16
      sb  a2, 2(a0)
      srli    a1, a1, 24
      sb  a1, 3(a0)
      ret
  .Lfunc_end0:
      .size   store0, .Lfunc_end0-store0
  
      .globl  store1
      .p2align    1
      .type   store1, at function
  store1:
      sh  a1, 0(a0)
      ret
  .Lfunc_end1:
      .size   store1, .Lfunc_end1-store1
  
      .globl  store2
      .p2align    1
      .type   store2, at function
  store2:
      srli    a2, a1, 16
      sb  a2, 2(a0)
      srli    a1, a1, 24
      sb  a1, 3(a0)
      ret
  .Lfunc_end2:
      .size   store2, .Lfunc_end2-store2
  
      .globl  store3
      .p2align    1
      .type   store3, at function
  store3:
      sh  a1, 2(a0)
      ret
  .Lfunc_end3:
      .size   store3, .Lfunc_end3-store3
  
      .ident  "clang version 15.0.0 (git at 172.30.80.200:/opt/repos/fortirisc_llvm.git efb383266d04c70b8adf4ffb3f4872b36bc4653f)"
      .section    ".note.GNU-stack","", at progbits
      .addrsig

With my change the genereated asm code is like this:

      .text
      .attribute  4, 16
      .attribute  5, "rv32i2p0_m2p0_a2p0_c2p0"
      .file   "t.c"
      .globl  store0
      .p2align    1
      .type   store0, at function
  store0:
      srli    a2, a1, 16
      sh  a1, 0(a0)
      sh  a2, 16(a0)
      ret
  .Lfunc_end0:
      .size   store0, .Lfunc_end0-store0
  
      .globl  store1
      .p2align    1
      .type   store1, at function
  store1:
      sh  a1, 0(a0)
      ret
  .Lfunc_end1:
      .size   store1, .Lfunc_end1-store1
  
      .globl  store2
      .p2align    1
      .type   store2, at function
  store2:
      srli    a1, a1, 16
      sh  a1, 2(a0)
      ret
  .Lfunc_end2:
      .size   store2, .Lfunc_end2-store2
  
      .globl  store3
      .p2align    1
      .type   store3, at function
  store3:
      sh  a1, 2(a0)
      ret
  .Lfunc_end3:
      .size   store3, .Lfunc_end3-store3
  
      .ident  "clang version 15.0.0 (https://github.com/llvm/llvm-project.git 59bb18ddfec5e5d91079b163a535cb3a6879093e)"
      .section    ".note.GNU-stack","", at progbits
      .addrsig

The difference are at function store0 and store2. For store0, it was failed to use instruction 'sw' as the alignment is 2, my change detect such case and use 2 'sh' for it; for store2, it was failed to use 'sh' because the extra shift(16), my change make it working for this case too.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D118549/new/

https://reviews.llvm.org/D118549