Re: [PATCH 4/4] tcg/optimize: optimize TSTNE using smask and zmask

[Paolo Bonzini <pbonzini@redhat.com>]

On 2/29/24 00:10, Richard Henderson wrote:
> On 2/28/24 01:11, Paolo Bonzini wrote:
>> -    /* TSTNE x,sign -> LT x,0 */
>> -    if (arg_is_const_val(*p2, (ctx->type == TCG_TYPE_I32
>> -                               ? INT32_MIN : INT64_MIN))) {
>> +    /* TSTNE x,i -> LT x,0 if i only includes sign bit copies */
>> +    if (arg_is_const(*p2) && (arg_info(*p2)->val & ~i1->s_mask) == 0) {
> 
> This is a good idea, but s_mask isn't defined like you think -- it is 
> *repetitions* of the sign bit, but not including the sign bit itself.  
> For INT64_MIN, s_mask == 0.
> 
> So for TSTNE min,min, (min & ~0) != 0, so the test won't pass.

Oh! So I have to squash:

diff --git a/tcg/optimize.c b/tcg/optimize.c
index ab976a5bbe7..44d1b1a6d8a 100644
--- a/tcg/optimize.c
+++ b/tcg/optimize.c
@@ -140,6 +140,12 @@ static inline bool arg_is_const_val(TCGArg arg, uint64_t val)
      return ts_is_const_val(arg_temp(arg), val);
  }
  
+/* Calculate all the copies of the sign bit, both redundant and not. */
+static inline uint64_t all_sign_bit_copies(TempOptInfo *info)
+{
+    return (info->s_mask >> 1) | INT64_MIN;
+}
+
  static inline bool ts_is_copy(TCGTemp *ts)
  {
      return ts_info(ts)->next_copy != ts;
@@ -825,7 +831,7 @@ static int do_constant_folding_cond1(OptContext *ctx, TCGOp *op, TCGArg dest,
      }
  
      /* TSTNE x,i -> LT x,0 if i only includes sign bit copies */
-    if (arg_is_const(*p2) && (arg_info(*p2)->val & ~i1->s_mask) == 0) {
+    if (arg_is_const(*p2) && (arg_info(*p2)->val & ~all_sign_bit_copies(i1)) == 0) {
          *p2 = arg_new_constant(ctx, 0);
          *pcond = tcg_tst_ltge_cond(cond);
          return -1;


I tested with

    movq $0xffffffff80000000, %rbx
    test %ebx, %ebx
    js y

and I get

  brcond_i64 cc_dst,$0x80000000,tstne,$L1

which works and matches your explanation:

  i1.s_mask == 0xffffffff00000000
  i2.val == 0x80000000
  all_sign_bit_copies(i1) == 0xffffffff80000000
  u2.val & ~all_sign_bit_copies(i1) == 0

Thanks!

Paolo