Use three-operand LEA for select of constants on some x86 architectures

[kazutakahirata]

Compile:

// clang -O2 -march=znver3
unsigned select_unsigned_lt_10_8_13(unsigned X) {
  return X < 10 ? 8 : 13;
}

I get:

  31 c0                      xor    %eax,%eax
  83 ff 0a                   cmp    $0xa,%edi
  0f 93 c0                   setae  %al
  8d 04 80                   lea    (%rax,%rax,4),%eax
  83 c8 08                   or     $0x8,%eax

We could generate:

  31 c0                      xor    %eax,%eax
  83 ff 0a                   cmp    $0xa,%edi
  0f 93 c0                   setae  %al
  8d 44 80 08                lea    8(%rax,%rax,4),%eax

saving one instruction and 2 bytes on those x86 architectures where three-operand LEAs are not discouraged.

I thought the x86 backend might be intentionally avoiding the three-operand LEA for -march=znver3, but that doesn't seem to be the case. For return X < 10 ? 9 : 12, I get:

  31 c0                      xor    %eax,%eax
  83 ff 0a                   cmp    $0xa,%edi
  0f 93 c0                   setae  %al
  8d 44 40 09                lea    0x9(%rax,%rax,2),%eax

Even on those x86 architectures where three-operand LEAs are discouraged, this optimization might be useful for size optimization purposes.

[llvmbot]

@llvm/issue-subscribers-backend-x86

[RKSimon]

This looks like its something to do with the different condition codes coming from IR:

define i32 @select_unsigned_lt_10_8_13j(i32 %0) {
  %2 = icmp ult i32 %0, 10
  %3 = select i1 %2, i32 8, i32 13
  ret i32 %3
}

define i32 @select_unsigned_lt_10_9_12j(i32 %0) {
  %2 = icmp ugt i32 %0, 9
  %3 = select i1 %2, i32 12, i32 9
  ret i32 %3
}

[kazutakahirata]

Just posted: https://reviews.llvm.org/D146787