add cfcmov

herumi · Nov 29, 2023 · e2d9685 · e2d9685
1 parent a4ec97c
commit e2d9685
Show file tree

Hide file tree

Showing 4 changed files with 111 additions and 1 deletion.
diff --git a/gen/gen_code.cpp b/gen/gen_code.cpp
@@ -641,6 +641,35 @@ void put()
  printf("void ctest%s(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, %d); }\n", s, p->ext);
  }
  }
+ {
+ // cfcmov
+ const struct Tbl {
+ uint8_t code;
+ const char *suf;
+ } tbl[] = {
+ {0x40, "o"},
+ {0x41, "no"},
+ {0x42, "b"},
+ {0x43, "nb"},
+ {0x44, "z"},
+ {0x45, "nz"},
+ {0x46, "be"},
+ {0x47, "nbe"},
+ {0x48, "s"},
+ {0x49, "ns"},
+ {0x4A, "p"},
+ {0x4B, "np"},
+ {0x4C, "l"},
+ {0x4D, "nl"},
+ {0x4E, "le"},
+ {0x4F, "nle"},
+ };
+ for (size_t i = 0; i < NUM_OF_ARRAY(tbl); i++) {
+ const Tbl *p = &tbl[i];
+ printf("void cfcmov%s(const Operand& op1, const Operand& op2) { opCfcmov(Reg(), op1, op2, 0x%02X); }\n", p->suf, p->code);
+ printf("void cfcmov%s(const Reg& d, const Reg& r, const Operand& op) { opCfcmov(d|T_nf, op, r, 0x%02X); }\n", p->suf, p->code);
+ }
+ }
  {
  const struct Tbl {
  const char *name;

diff --git a/test/apx.cpp b/test/apx.cpp
@@ -1601,3 +1601,38 @@ CYBOZU_TEST_AUTO(ctestb)
  CYBOZU_TEST_EQUAL_ARRAY(c.getCode(), tbl, n);
 }
 
+CYBOZU_TEST_AUTO(cfcmov)
+{
+ struct Code : Xbyak::CodeGenerator {
+ Code()
+ {
+ cfcmovb(r30w, r31w);
+ cfcmovb(r30d, r31d);
+ cfcmovb(r30, r31);
+ cfcmovb(ptr [r8+r20*4+0x3], r19w);
+ cfcmovb(ptr [r8+r20*4+0x3], r19d);
+ cfcmovb(ptr [r8+r20*4+0x3], r19);
+ cfcmovb(r30w, ptr [r9]);
+ cfcmovb(r30d, ptr [r9]);
+ cfcmovb(r30, ptr [r9]);
+ cfcmovb(r20w, r30w, r31w);
+ cfcmovb(r20d, r30d, r31d);
+ cfcmovb(r20, r30, r31);
+ cfcmovb(r20w, r30w, ptr [r9]);
+ cfcmovb(r20d, r30d, ptr [r9]);
+ cfcmovb(r20, r30, ptr [r9]);
+ }
+ } c;
+ const uint8_t tbl[] = {
+ 0x62, 0x4c, 0x7d, 0x0c, 0x42, 0xfe, 0x62, 0x4c, 0x7c, 0x0c, 0x42, 0xfe, 0x62, 0x4c, 0xfc, 0x0c,
+ 0x42, 0xfe, 0x62, 0xc4, 0x79, 0x0c, 0x42, 0x5c, 0xa0, 0x03, 0x62, 0xc4, 0x78, 0x0c, 0x42, 0x5c,
+ 0xa0, 0x03, 0x62, 0xc4, 0xf8, 0x0c, 0x42, 0x5c, 0xa0, 0x03, 0x62, 0x44, 0x7d, 0x08, 0x42, 0x31,
+ 0x62, 0x44, 0x7c, 0x08, 0x42, 0x31, 0x62, 0x44, 0xfc, 0x08, 0x42, 0x31, 0x62, 0x4c, 0x5d, 0x14,
+ 0x42, 0xf7, 0x62, 0x4c, 0x5c, 0x14, 0x42, 0xf7, 0x62, 0x4c, 0xdc, 0x14, 0x42, 0xf7, 0x62, 0x44,
+ 0x5d, 0x14, 0x42, 0x31, 0x62, 0x44, 0x5c, 0x14, 0x42, 0x31, 0x62, 0x44, 0xdc, 0x14, 0x42, 0x31,
+ };
+ const size_t n = sizeof(tbl);
+ CYBOZU_TEST_EQUAL(c.getSize(), n);
+ CYBOZU_TEST_EQUAL_ARRAY(c.getCode(), tbl, n);
+}
+
diff --git a/xbyak/xbyak.h b/xbyak/xbyak.h
@@ -2192,7 +2192,7 @@ class CodeGenerator : public CodeArray {
  // (r, r, m) or (r, m, r)
  bool opROO(const Reg& d, const Operand& op1, const Operand& op2, uint64_t type, int code, int immSize = 0, int sc = NONE)
  {
- if (!d.isREG() && !(d.hasRex2NFZU() || op1.hasRex2NFZU() || op2.hasRex2NFZU())) return false;
+ if (!(type & T_MUST_EVEX) && !d.isREG() && !(d.hasRex2NFZU() || op1.hasRex2NFZU() || op2.hasRex2NFZU())) return false;
  const Operand *p1 = &op1, *p2 = &op2;
  if (p1->isMEM()) { std::swap(p1, p2); } else { if (p2->isMEM()) code |= 2; }
  if (p1->isMEM()) XBYAK_THROW_RET(ERR_BAD_COMBINATION, false)
@@ -2681,6 +2681,20 @@ class CodeGenerator : public CodeArray {
  opROO(Reg(15 - dfv, Operand::REG, opBit), op, Reg(0, Operand::REG, opBit), T_VEX|T_CODE1_IF1, 0xF6, immBit / 8, sc);
  db(imm, immBit / 8);
  }
+ void opCfcmov(const Reg& d, const Operand& op1, const Operand& op2, int code)
+ {
+ const int dBit = d.getBit();
+ const int op2Bit = op2.getBit();
+ if (dBit > 0 && op2Bit > 0 && dBit != op2Bit) XBYAK_THROW(ERR_BAD_SIZE_OF_REGISTER)
+ if (op1.isBit(8) || op2Bit == 8) XBYAK_THROW(ERR_BAD_SIZE_OF_REGISTER)
+ if (op2.isMEM()) {
+ if (op1.isMEM()) XBYAK_THROW(ERR_BAD_COMBINATION)
+ uint64_t type = dBit > 0 ? (T_MUST_EVEX|T_NF) : T_MUST_EVEX;
+ opROO(d, op2, op1, type, code);
+ } else {
+ opROO(d, op1, static_cast<const Reg&>(op2)|T_nf, T_MUST_EVEX|T_NF, 0x42);
+ }
+ }
 #ifdef XBYAK64
  void opAMX(const Tmm& t1, const Address& addr, uint64_t type, int code)
  {

diff --git a/xbyak/xbyak_mnemonic.h b/xbyak/xbyak_mnemonic.h
@@ -123,6 +123,38 @@ void ccmpt(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, d
 void ccmpz(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 4); }
 void ccmpz(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 4); }
 void cdq() { db(0x99); }
+void cfcmovb(const Operand& op1, const Operand& op2) { opCfcmov(Reg(), op1, op2, 0x42); }
+void cfcmovb(const Reg& d, const Reg& r, const Operand& op) { opCfcmov(d|T_nf, op, r, 0x42); }
+void cfcmovbe(const Operand& op1, const Operand& op2) { opCfcmov(Reg(), op1, op2, 0x46); }
+void cfcmovbe(const Reg& d, const Reg& r, const Operand& op) { opCfcmov(d|T_nf, op, r, 0x46); }
+void cfcmovl(const Operand& op1, const Operand& op2) { opCfcmov(Reg(), op1, op2, 0x4C); }
+void cfcmovl(const Reg& d, const Reg& r, const Operand& op) { opCfcmov(d|T_nf, op, r, 0x4C); }
+void cfcmovle(const Operand& op1, const Operand& op2) { opCfcmov(Reg(), op1, op2, 0x4E); }
+void cfcmovle(const Reg& d, const Reg& r, const Operand& op) { opCfcmov(d|T_nf, op, r, 0x4E); }
+void cfcmovnb(const Operand& op1, const Operand& op2) { opCfcmov(Reg(), op1, op2, 0x43); }
+void cfcmovnb(const Reg& d, const Reg& r, const Operand& op) { opCfcmov(d|T_nf, op, r, 0x43); }
+void cfcmovnbe(const Operand& op1, const Operand& op2) { opCfcmov(Reg(), op1, op2, 0x47); }
+void cfcmovnbe(const Reg& d, const Reg& r, const Operand& op) { opCfcmov(d|T_nf, op, r, 0x47); }
+void cfcmovnl(const Operand& op1, const Operand& op2) { opCfcmov(Reg(), op1, op2, 0x4D); }
+void cfcmovnl(const Reg& d, const Reg& r, const Operand& op) { opCfcmov(d|T_nf, op, r, 0x4D); }
+void cfcmovnle(const Operand& op1, const Operand& op2) { opCfcmov(Reg(), op1, op2, 0x4F); }
+void cfcmovnle(const Reg& d, const Reg& r, const Operand& op) { opCfcmov(d|T_nf, op, r, 0x4F); }
+void cfcmovno(const Operand& op1, const Operand& op2) { opCfcmov(Reg(), op1, op2, 0x41); }
+void cfcmovno(const Reg& d, const Reg& r, const Operand& op) { opCfcmov(d|T_nf, op, r, 0x41); }
+void cfcmovnp(const Operand& op1, const Operand& op2) { opCfcmov(Reg(), op1, op2, 0x4B); }
+void cfcmovnp(const Reg& d, const Reg& r, const Operand& op) { opCfcmov(d|T_nf, op, r, 0x4B); }
+void cfcmovns(const Operand& op1, const Operand& op2) { opCfcmov(Reg(), op1, op2, 0x49); }
+void cfcmovns(const Reg& d, const Reg& r, const Operand& op) { opCfcmov(d|T_nf, op, r, 0x49); }
+void cfcmovnz(const Operand& op1, const Operand& op2) { opCfcmov(Reg(), op1, op2, 0x45); }
+void cfcmovnz(const Reg& d, const Reg& r, const Operand& op) { opCfcmov(d|T_nf, op, r, 0x45); }
+void cfcmovo(const Operand& op1, const Operand& op2) { opCfcmov(Reg(), op1, op2, 0x40); }
+void cfcmovo(const Reg& d, const Reg& r, const Operand& op) { opCfcmov(d|T_nf, op, r, 0x40); }
+void cfcmovp(const Operand& op1, const Operand& op2) { opCfcmov(Reg(), op1, op2, 0x4A); }
+void cfcmovp(const Reg& d, const Reg& r, const Operand& op) { opCfcmov(d|T_nf, op, r, 0x4A); }
+void cfcmovs(const Operand& op1, const Operand& op2) { opCfcmov(Reg(), op1, op2, 0x48); }
+void cfcmovs(const Reg& d, const Reg& r, const Operand& op) { opCfcmov(d|T_nf, op, r, 0x48); }
+void cfcmovz(const Operand& op1, const Operand& op2) { opCfcmov(Reg(), op1, op2, 0x44); }
+void cfcmovz(const Reg& d, const Reg& r, const Operand& op) { opCfcmov(d|T_nf, op, r, 0x44); }
 void clc() { db(0xF8); }
 void cld() { db(0xFC); }
 void cldemote(const Address& addr) { opMR(addr, eax, T_0F, 0x1C); }