fix detection of mulx at mulPre2/sqrPre2

include/mcl/fp_generator.hpp

@@ -253,12 +253,12 @@ struct FpGenerator : Xbyak::CodeGenerator {
 			op.fpDbl_mod = getCurr<void2u>();
 			gen_fpDbl_mod();
 		}
-		if (op.N == 2 || op.N == 3 || op.N == 4) {
+		if ((useMulx_ && op.N == 2) || op.N == 3 || op.N == 4) {
 			align(16);
 			op.fpDbl_mulPre = getCurr<void3u>();
 			gen_fpDbl_mulPre();
 		}
-		if (op.N == 2 || op.N == 3 || op.N == 4) {
+		if ((useMulx_ && op.N == 2) || op.N == 3 || op.N == 4) {
 			align(16);
 			op.fpDbl_sqrPre = getCurr<void2u>();
 			gen_fpDbl_sqrPre(op);

include/mcl/fp_proto.hpp

@@ -55,6 +55,7 @@ MCL_FP_DEF_FUNC(576)
 
 #undef MCL_FP_DEF_FUNC
 
+void mcl_fpDbl_mod_NIST_P192(mcl::fp::Unit*, const mcl::fp::Unit*);
 void mcl_fp_mul_NIST_P192(mcl::fp::Unit*, const mcl::fp::Unit*, const mcl::fp::Unit*);
 
 }

src/fp.cpp

@@ -413,6 +413,7 @@ void Op::init(const std::string& mstr, int base, size_t maxBitSize, Mode mode)
 #ifdef MCL_USE_LLVM
 	if ((mode == FP_AUTO || mode == FP_LLVM) && mp == mpz_class("0xfffffffffffffffffffffffffffffffeffffffffffffffff")) {
 		fp_mul = &mcl_fp_mul_NIST_P192;
+//		fpDbl_mod = &mcl_fpDbl_mod_NIST_P192;
 		isMont = false;
 	}
 #endif

src/once.txt

@@ -30,7 +30,7 @@ define internal i64 @extract192to64(i192 %x, i192 %shift) {
 	ret i64 %t1
 }
 
-define internal void @fpDbl_modNIST_P192(i192* %out, i192* %px) {
+define void @mcl_fpDbl_mod_NIST_P192(i192* %out, i192* %px) {
 	%L192 = load i192* %px
 	%L = zext i192 %L192 to i256
 
@@ -69,6 +69,6 @@ define void @mcl_fp_mul_NIST_P192(i192* %pz, i192* %px, i192* %py) {
 	%buf = alloca i192, i32 2
 	%p = bitcast i192* %buf to i$(unit)*
 	call void @mcl_fpDbl_mulPre192(i$(unit)* %p, i192* %px, i192* %py)
-	call void @fpDbl_modNIST_P192(i192* %pz, i192* %buf)
+	call void @mcl_fpDbl_mod_NIST_P192(i192* %pz, i192* %buf)
 	ret void
 }