Ok, this one definitely caught my eye. :)
On 2/17/25 00:09, Richard Henderson wrote:
+ tcg_gen_add_i32(t0, a, b);
+ tcg_gen_setcond_i32(TCG_COND_LTU, t1, t0, a);
Compare against b instead? If there's an immediate (which could even be
zero) it is there.
+ tcg_gen_add_i32(r, t0, ci);
+ tcg_gen_setcond_i32(TCG_COND_LTU, t0, r, t0);
+ tcg_gen_or_i32(co, t0, t1);
setcond + or can become a movcond:
/* co = t1 | (r < t0) */
tcg_gen_movcond_i32(TCG_COND_LTU, co, r, t0, tcg_constant_i32(1), t1);
+ TCGv_i64 t0 = tcg_temp_ebb_new_i64();
+ TCGv_i64 t1 = tcg_temp_ebb_new_i64();
+
+ tcg_gen_add_i64(t0, a, b);
+ tcg_gen_setcond_i64(TCG_COND_LTU, t1, t0, a);
+ tcg_gen_add_i64(r, t0, ci);
+ tcg_gen_setcond_i64(TCG_COND_LTU, t0, r, t0);
+ tcg_gen_or_i64(co, t0, t1);
The same two observations as above apply here, of course.
+ tcg_temp_free_i64(t0);
+ tcg_temp_free_i64(t1);
+ }
+ } else {
+ if (tcg_op_supported(INDEX_op_addci, TCG_TYPE_I32, 0)) {
+ TCGv_i32 discard = tcg_temp_ebb_new_i32();
+ TCGv_i32 zero = tcg_constant_i32(0);
+ TCGv_i32 mone = tcg_constant_i32(-1);
+
+ tcg_gen_op3_i32(INDEX_op_addco, discard, TCGV_LOW(ci), mone);
+ tcg_gen_op3_i32(INDEX_op_addcio, discard, TCGV_HIGH(ci), mone);
This addcio is unnecessary/incorrect. I think you assume that TCGV_HIGH(ci) =
0,
since that's what you set it below, and then this you can remove it.
But if you really wanted to do it, it should go...
+ tcg_gen_op3_i32(INDEX_op_addcio, TCGV_LOW(r),
+ TCGV_LOW(a), TCGV_LOW(b));
... here, after the low word is all set, to "OR" the input high-carry into the
output low-carry. But again, I think it's not necessary.
+ tcg_gen_op3_i32(INDEX_op_addcio, TCGV_HIGH(r),
+ TCGV_HIGH(a), TCGV_HIGH(b));
+ tcg_gen_op3_i32(INDEX_op_addci, TCGV_LOW(co), zero, zero);
+ tcg_temp_free_i32(discard);
+ } else {
+ TCGv_i32 t0 = tcg_temp_ebb_new_i32();
+ TCGv_i32 c0 = tcg_temp_ebb_new_i32();
+ TCGv_i32 c1 = tcg_temp_ebb_new_i32();
+
+ tcg_gen_or_i32(c1, TCGV_LOW(ci), TCGV_HIGH(ci));
+ tcg_gen_setcondi_i32(TCG_COND_NE, c1, c1, 0);
Likewise, this shouldn't be needed and you can just add TCGV_LOW(ci)
below.
+ tcg_gen_add_i32(t0, TCGV_LOW(a), TCGV_LOW(b));
+ tcg_gen_setcond_i32(TCG_COND_LTU, c0, t0, TCGV_LOW(a));
Here you can also change a to b.
+ tcg_gen_add_i32(TCGV_LOW(r), t0, c1);
+ tcg_gen_setcond_i32(TCG_COND_LTU, c1, TCGV_LOW(r), c1);
+ tcg_gen_or_i32(c1, c1, c0);
/* c1 = c0 | (r < ci) */
tcg_gen_movcond_i32(TCG_COND_LTU, c1, TCGV_LOW(r), TCGV_LOW(ci),
tcg_constant_i32(1), c0);
+ tcg_gen_add_i32(t0, TCGV_HIGH(a), TCGV_HIGH(b));
+ tcg_gen_setcond_i32(TCG_COND_LTU, c0, t0, TCGV_HIGH(a));
Change a to b.
+ tcg_gen_add_i32(TCGV_HIGH(r), t0, c1);
+ tcg_gen_setcond_i32(TCG_COND_LTU, c1, TCGV_HIGH(r), c1);
+ tcg_gen_or_i32(TCGV_LOW(co), c0, c1);
+ /* c1 = c0 | (r < c1) */
+ tcg_gen_movcond_i32(TCG_COND_LTU, c1, TCGV_HIGH(r), c1,
tcg_constant_i32(1), c0);
+ tcg_temp_free_i32(t0);
+ tcg_temp_free_i32(c0);
+ tcg_temp_free_i32(c1);
+ }
+ tcg_gen_movi_i32(TCGV_HIGH(co), 0);
+ }
+}
+
void tcg_gen_sub2_i64(TCGv_i64 rl, TCGv_i64 rh, TCGv_i64 al,
TCGv_i64 ah, TCGv_i64 bl, TCGv_i64 bh)
{
