aboutsummaryrefslogtreecommitdiffstats
path: root/aarch64/Asmexpand.ml
diff options
context:
space:
mode:
Diffstat (limited to 'aarch64/Asmexpand.ml')
-rw-r--r--aarch64/Asmexpand.ml14
1 files changed, 12 insertions, 2 deletions
diff --git a/aarch64/Asmexpand.ml b/aarch64/Asmexpand.ml
index 1a02b019..8187e077 100644
--- a/aarch64/Asmexpand.ml
+++ b/aarch64/Asmexpand.ml
@@ -337,7 +337,7 @@ let expand_builtin_inline name args res =
| "__builtin_bswap16", [BA(DR(IR(RR1 a1)))], BR(DR(IR(RR1 res))) ->
emit (Prev16(W, res, a1));
emit (Pandimm(W, res, RR0 res, Z.of_uint 0xFFFF))
- (* Count leading zeros and leading sign bits *)
+ (* Count leading zeros, leading sign bits, trailing zeros *)
| "__builtin_clz", [BA(DR(IR(RR1 a1)))], BR(DR(IR(RR1 res))) ->
emit (Pclz(W, res, a1))
| ("__builtin_clzl" | "__builtin_clzll"), [BA(DR(IR(RR1 a1)))], BR(DR(IR(RR1 res))) ->
@@ -346,10 +346,16 @@ let expand_builtin_inline name args res =
emit (Pcls(W, res, a1))
| ("__builtin_clsl" | "__builtin_clsll"), [BA(DR(IR(RR1 a1)))], BR(DR(IR(RR1 res))) ->
emit (Pcls(X, res, a1))
+ | "__builtin_ctz", [BA(DR(IR(RR1 a1)))], BR(DR(IR(RR1 res))) ->
+ emit (Prbit(W, res, a1));
+ emit (Pclz(W, res, res))
+ | ("__builtin_ctzl" | "__builtin_ctzll"), [BA(DR(IR(RR1 a1)))], BR(DR(IR(RR1 res))) ->
+ emit (Prbit(X, res, a1));
+ emit (Pclz(X, res, res))
(* Float arithmetic *)
| "__builtin_fabs", [BA(DR(FR a1))], BR(DR(FR res)) ->
emit (Pfabs(D, res, a1))
- | "__builtin_fsqrt", [BA(DR(FR a1))], BR(DR(FR res)) ->
+ | ("__builtin_fsqrt" | "__builtin_sqrt"), [BA(DR(FR a1))], BR(DR(FR res)) ->
emit (Pfsqrt(D, res, a1))
| "__builtin_fmadd", [BA(DR(FR a1)); BA(DR(FR a2)); BA(DR(FR a3))], BR(DR(FR res)) ->
emit (Pfmadd(D, res, a1, a2, a3))
@@ -359,6 +365,10 @@ let expand_builtin_inline name args res =
emit (Pfnmadd(D, res, a1, a2, a3))
| "__builtin_fnmsub", [BA(DR(FR a1)); BA(DR(FR a2)); BA(DR(FR a3))], BR(DR(FR res)) ->
emit (Pfnmsub(D, res, a1, a2, a3))
+ | "__builtin_fmax", [BA(DR(FR a1)); BA(DR(FR a2))], BR(DR(FR res)) ->
+ emit (Pfmax (D, res, a1, a2))
+ | "__builtin_fmin", [BA(DR(FR a1)); BA(DR(FR a2))], BR(DR(FR res)) ->
+ emit (Pfmin (D, res, a1, a2))
(* Vararg *)
| "__builtin_va_start", [BA(DR(IR(RR1 a)))], _ ->
expand_builtin_va_start a