[PATCH] D41982: [mips] Reordering callseq* nodes to be linear

[Simon Dardis via Phabricator via llvm-commits]

sdardis requested changes to this revision.
sdardis added a comment.
This revision now requires changes to proceed.

This patch isn't correct, as you're not passing the pointer to the callee or filling the rest of the argument registers with the object being passed. This results in the callee and caller differing in their interpretation of the arguments. The following test case shows this behaviour:

  struct S1 {
    char t[32000];
  };
  
  extern void f2(char);
  
  void f(int b, struct S1 a) {
    f2(a.t[1] + (char)b) ;
  }
  
  void g(void) {
    struct S1 a;
    f(1, a);
  }

This compiled with r310704 reverted has the correct IR in that the caller and callee agree that the second argument to f is a pointer to a byval structure argument, but g doesn't set $5 to the result of the memcpy:

  addiu $4, $1, 16
  addiu $1, $fp, 32016
  addiu $5, $1, 12
  addiu $6, $zero, 32000
  jal memcpy
  nop
  addiu $4, $zero, 1
  sw  $2, 32012($fp)          # 4-byte Folded Spill
  jal f
  nop

while f believes that the second argument is spread across $5-$7:

  addiu $sp, $sp, -32
  sw  $ra, 28($sp)            # 4-byte Folded Spill
  sw  $fp, 24($sp)            # 4-byte Folded Spill
  move  $fp, $sp
  sw  $7, 44($fp)
  sw  $6, 40($fp)
  sw  $5, 36($fp)
  sw  $4, 20($fp)
  lbu $4, 37($fp)
  lw  $5, 20($fp)
  addu  $4, $4, $5
  seb $4, $4
  jal f2

Can you take a look at this?


Repository:
  rL LLVM

https://reviews.llvm.org/D41982