[Tarantool-patches] [PATCH luajit 2/5] Remove pow() splitting and cleanup backends.

[Sergey Kaplun via Tarantool-patches <tarantool-patches@dev.tarantool.org>]

From: Mike Pall <mike>

(cherry-picked from commit b2307c8ad817e350d65cc909a579ca2f77439682)

The JIT engine tries to split b^c to exp2(c * log2(b)) with attempt to
rejoin them later for some backends. It adds a dependency on C99
exp2() and log2(), which aren't part of some libm implementations.
Also, for some cases for IEEE754 we can see, that exp2(log2(x)) != x,
due to mathematical functions accuracy and double precision
restrictions. So, the values on the JIT slots and Lua stack are
inconsistent.

This patch removes splitting of pow operator, so IR_POW is emitting for
all cases (except power of 0.5 replaced with sqrt operation).

Also this patch does some refactoring:

* Functions `asm_pow()`, `asm_mod()`, `asm_ldexp()`, `asm_div()`
  (replaced with `asm_fpdiv()` for CPU architectures) are moved to the
  <src/lj_asm.c> as far as their implementation is generic for all
  architectures.
* Fusing of IR_HREF + IR_EQ/IR_NE moved to a `asm_fuseequal()`.
* Since `lj_vm_exp2()` subroutine and `IRFPM_EXP2` are removed as no
  longer used.

Sergey Kaplun:
* added the description and the test for the problem

Part of tarantool/tarantool#8825
---
 src/lj_arch.h                                 |   3 -
 src/lj_asm.c                                  | 106 +++++++++++-------
 src/lj_asm_arm.h                              |  10 +-
 src/lj_asm_arm64.h                            |  39 +------
 src/lj_asm_mips.h                             |  38 +------
 src/lj_asm_ppc.h                              |   9 +-
 src/lj_asm_x86.h                              |  37 +-----
 src/lj_ir.h                                   |   2 +-
 src/lj_ircall.h                               |   1 -
 src/lj_opt_fold.c                             |  18 ++-
 src/lj_opt_narrow.c                           |  20 +---
 src/lj_opt_split.c                            |  21 ----
 src/lj_vm.h                                   |   5 -
 src/lj_vmmath.c                               |   8 --
 .../lj-9-pow-inconsistencies.test.lua         |  63 +++++++++++
 15 files changed, 158 insertions(+), 222 deletions(-)
 create mode 100644 test/tarantool-tests/lj-9-pow-inconsistencies.test.lua

diff --git a/src/lj_arch.h b/src/lj_arch.h
index cf31a291..3bdbe84e 100644
--- a/src/lj_arch.h
+++ b/src/lj_arch.h
@@ -607,9 +607,6 @@
 #if defined(__ANDROID__) || defined(__symbian__) || LJ_TARGET_XBOX360 || LJ_TARGET_WINDOWS
 #define LUAJIT_NO_LOG2
 #endif
-#if defined(__symbian__) || LJ_TARGET_WINDOWS
-#define LUAJIT_NO_EXP2
-#endif
 #if LJ_TARGET_CONSOLE || (LJ_TARGET_IOS && __IPHONE_OS_VERSION_MIN_REQUIRED >= __IPHONE_8_0)
 #define LJ_NO_SYSTEM		1
 #endif
diff --git a/src/lj_asm.c b/src/lj_asm.c
index b352fd35..a6906b19 100644
--- a/src/lj_asm.c
+++ b/src/lj_asm.c
@@ -1356,32 +1356,6 @@ static void asm_call(ASMState *as, IRIns *ir)
   asm_gencall(as, ci, args);
 }
 
-#if !LJ_SOFTFP32
-static void asm_fppow(ASMState *as, IRIns *ir, IRRef lref, IRRef rref)
-{
-  const CCallInfo *ci = &lj_ir_callinfo[IRCALL_pow];
-  IRRef args[2];
-  args[0] = lref;
-  args[1] = rref;
-  asm_setupresult(as, ir, ci);
-  asm_gencall(as, ci, args);
-}
-
-static int asm_fpjoin_pow(ASMState *as, IRIns *ir)
-{
-  IRIns *irp = IR(ir->op1);
-  if (irp == ir-1 && irp->o == IR_MUL && !ra_used(irp)) {
-    IRIns *irpp = IR(irp->op1);
-    if (irpp == ir-2 && irpp->o == IR_FPMATH &&
-	irpp->op2 == IRFPM_LOG2 && !ra_used(irpp)) {
-      asm_fppow(as, ir, irpp->op1, irp->op2);
-      return 1;
-    }
-  }
-  return 0;
-}
-#endif
-
 /* -- PHI and loop handling ----------------------------------------------- */
 
 /* Break a PHI cycle by renaming to a free register (evict if needed). */
@@ -1652,6 +1626,62 @@ static void asm_loop(ASMState *as)
 #error "Missing assembler for target CPU"
 #endif
 
+/* -- Common instruction helpers ------------------------------------------ */
+
+#if !LJ_SOFTFP32
+#if !LJ_TARGET_X86ORX64
+#define asm_ldexp(as, ir)	asm_callid(as, ir, IRCALL_ldexp)
+#define asm_fppowi(as, ir)	asm_callid(as, ir, IRCALL_lj_vm_powi)
+#endif
+
+static void asm_pow(ASMState *as, IRIns *ir)
+{
+#if LJ_64 && LJ_HASFFI
+  if (!irt_isnum(ir->t))
+    asm_callid(as, ir, irt_isi64(ir->t) ? IRCALL_lj_carith_powi64 :
+					  IRCALL_lj_carith_powu64);
+  else
+#endif
+  if (irt_isnum(IR(ir->op2)->t))
+    asm_callid(as, ir, IRCALL_pow);
+  else
+    asm_fppowi(as, ir);
+}
+
+static void asm_div(ASMState *as, IRIns *ir)
+{
+#if LJ_64 && LJ_HASFFI
+  if (!irt_isnum(ir->t))
+    asm_callid(as, ir, irt_isi64(ir->t) ? IRCALL_lj_carith_divi64 :
+					  IRCALL_lj_carith_divu64);
+  else
+#endif
+    asm_fpdiv(as, ir);
+}
+#endif
+
+static void asm_mod(ASMState *as, IRIns *ir)
+{
+#if LJ_64 && LJ_HASFFI
+  if (!irt_isint(ir->t))
+    asm_callid(as, ir, irt_isi64(ir->t) ? IRCALL_lj_carith_modi64 :
+					  IRCALL_lj_carith_modu64);
+  else
+#endif
+    asm_callid(as, ir, IRCALL_lj_vm_modi);
+}
+
+static void asm_fuseequal(ASMState *as, IRIns *ir)
+{
+  /* Fuse HREF + EQ/NE. */
+  if ((ir-1)->o == IR_HREF && ir->op1 == as->curins-1) {
+    as->curins--;
+    asm_href(as, ir-1, (IROp)ir->o);
+  } else {
+    asm_equal(as, ir);
+  }
+}
+
 /* -- Instruction dispatch ------------------------------------------------ */
 
 /* Assemble a single instruction. */
@@ -1674,14 +1704,7 @@ static void asm_ir(ASMState *as, IRIns *ir)
   case IR_ABC:
     asm_comp(as, ir);
     break;
-  case IR_EQ: case IR_NE:
-    if ((ir-1)->o == IR_HREF && ir->op1 == as->curins-1) {
-      as->curins--;
-      asm_href(as, ir-1, (IROp)ir->o);
-    } else {
-      asm_equal(as, ir);
-    }
-    break;
+  case IR_EQ: case IR_NE: asm_fuseequal(as, ir); break;
 
   case IR_RETF: asm_retf(as, ir); break;
 
@@ -1750,7 +1773,13 @@ static void asm_ir(ASMState *as, IRIns *ir)
   case IR_SNEW: case IR_XSNEW: asm_snew(as, ir); break;
   case IR_TNEW: asm_tnew(as, ir); break;
   case IR_TDUP: asm_tdup(as, ir); break;
-  case IR_CNEW: case IR_CNEWI: asm_cnew(as, ir); break;
+  case IR_CNEW: case IR_CNEWI:
+#if LJ_HASFFI
+    asm_cnew(as, ir);
+#else
+    lua_assert(0);
+#endif
+    break;
 
   /* Buffer operations. */
   case IR_BUFHDR: asm_bufhdr(as, ir); break;
@@ -2215,6 +2244,10 @@ static void asm_setup_regsp(ASMState *as)
 	if (inloop)
 	  as->modset |= RSET_SCRATCH;
 #if LJ_TARGET_X86
+	if (irt_isnum(IR(ir->op2)->t)) {
+	  if (as->evenspill < 4)  /* Leave room to call pow(). */
+	    as->evenspill = 4;
+	}
 	break;
 #else
 	ir->prev = REGSP_HINT(RID_FPRET);
@@ -2240,9 +2273,6 @@ static void asm_setup_regsp(ASMState *as)
 	  continue;
 	}
 	break;
-      } else if (ir->op2 == IRFPM_EXP2 && !LJ_64) {
-	if (as->evenspill < 4)  /* Leave room to call pow(). */
-	  as->evenspill = 4;
       }
 #endif
       if (inloop)
diff --git a/src/lj_asm_arm.h b/src/lj_asm_arm.h
index 2894e5c9..29a07c80 100644
--- a/src/lj_asm_arm.h
+++ b/src/lj_asm_arm.h
@@ -1275,8 +1275,6 @@ static void asm_cnew(ASMState *as, IRIns *ir)
   ra_allockreg(as, (int32_t)(sz+sizeof(GCcdata)),
 	       ra_releasetmp(as, ASMREF_TMP1));
 }
-#else
-#define asm_cnew(as, ir)	((void)0)
 #endif
 
 /* -- Write barriers ------------------------------------------------------ */
@@ -1371,8 +1369,6 @@ static void asm_callround(ASMState *as, IRIns *ir, int id)
 
 static void asm_fpmath(ASMState *as, IRIns *ir)
 {
-  if (ir->op2 == IRFPM_EXP2 && asm_fpjoin_pow(as, ir))
-    return;
   if (ir->op2 <= IRFPM_TRUNC)
     asm_callround(as, ir, ir->op2);
   else if (ir->op2 == IRFPM_SQRT)
@@ -1499,14 +1495,10 @@ static void asm_mul(ASMState *as, IRIns *ir)
 #define asm_mulov(as, ir)	asm_mul(as, ir)
 
 #if !LJ_SOFTFP
-#define asm_div(as, ir)		asm_fparith(as, ir, ARMI_VDIV_D)
-#define asm_pow(as, ir)		asm_callid(as, ir, IRCALL_lj_vm_powi)
+#define asm_fpdiv(as, ir)	asm_fparith(as, ir, ARMI_VDIV_D)
 #define asm_abs(as, ir)		asm_fpunary(as, ir, ARMI_VABS_D)
-#define asm_ldexp(as, ir)	asm_callid(as, ir, IRCALL_ldexp)
 #endif
 
-#define asm_mod(as, ir)		asm_callid(as, ir, IRCALL_lj_vm_modi)
-
 static void asm_neg(ASMState *as, IRIns *ir)
 {
 #if !LJ_SOFTFP
diff --git a/src/lj_asm_arm64.h b/src/lj_asm_arm64.h
index aea251a9..c3d6889e 100644
--- a/src/lj_asm_arm64.h
+++ b/src/lj_asm_arm64.h
@@ -1249,8 +1249,6 @@ static void asm_cnew(ASMState *as, IRIns *ir)
   ra_allockreg(as, (int32_t)(sz+sizeof(GCcdata)),
 	       ra_releasetmp(as, ASMREF_TMP1));
 }
-#else
-#define asm_cnew(as, ir)	((void)0)
 #endif
 
 /* -- Write barriers ------------------------------------------------------ */
@@ -1327,8 +1325,6 @@ static void asm_fpmath(ASMState *as, IRIns *ir)
   } else if (fpm <= IRFPM_TRUNC) {
     asm_fpunary(as, ir, fpm == IRFPM_FLOOR ? A64I_FRINTMd :
 			fpm == IRFPM_CEIL ? A64I_FRINTPd : A64I_FRINTZd);
-  } else if (fpm == IRFPM_EXP2 && asm_fpjoin_pow(as, ir)) {
-    return;
   } else {
     asm_callid(as, ir, IRCALL_lj_vm_floor + fpm);
   }
@@ -1435,45 +1431,12 @@ static void asm_mul(ASMState *as, IRIns *ir)
   asm_intmul(as, ir);
 }
 
-static void asm_div(ASMState *as, IRIns *ir)
-{
-#if LJ_HASFFI
-  if (!irt_isnum(ir->t))
-    asm_callid(as, ir, irt_isi64(ir->t) ? IRCALL_lj_carith_divi64 :
-					  IRCALL_lj_carith_divu64);
-  else
-#endif
-    asm_fparith(as, ir, A64I_FDIVd);
-}
-
-static void asm_pow(ASMState *as, IRIns *ir)
-{
-#if LJ_HASFFI
-  if (!irt_isnum(ir->t))
-    asm_callid(as, ir, irt_isi64(ir->t) ? IRCALL_lj_carith_powi64 :
-					  IRCALL_lj_carith_powu64);
-  else
-#endif
-    asm_callid(as, ir, IRCALL_lj_vm_powi);
-}
-
 #define asm_addov(as, ir)	asm_add(as, ir)
 #define asm_subov(as, ir)	asm_sub(as, ir)
 #define asm_mulov(as, ir)	asm_mul(as, ir)
 
+#define asm_fpdiv(as, ir)	asm_fparith(as, ir, A64I_FDIVd)
 #define asm_abs(as, ir)		asm_fpunary(as, ir, A64I_FABS)
-#define asm_ldexp(as, ir)	asm_callid(as, ir, IRCALL_ldexp)
-
-static void asm_mod(ASMState *as, IRIns *ir)
-{
-#if LJ_HASFFI
-  if (!irt_isint(ir->t))
-    asm_callid(as, ir, irt_isi64(ir->t) ? IRCALL_lj_carith_modi64 :
-					  IRCALL_lj_carith_modu64);
-  else
-#endif
-    asm_callid(as, ir, IRCALL_lj_vm_modi);
-}
 
 static void asm_neg(ASMState *as, IRIns *ir)
 {
diff --git a/src/lj_asm_mips.h b/src/lj_asm_mips.h
index 4626507b..0f92959b 100644
--- a/src/lj_asm_mips.h
+++ b/src/lj_asm_mips.h
@@ -1613,8 +1613,6 @@ static void asm_cnew(ASMState *as, IRIns *ir)
   ra_allockreg(as, (int32_t)(sz+sizeof(GCcdata)),
 	       ra_releasetmp(as, ASMREF_TMP1));
 }
-#else
-#define asm_cnew(as, ir)	((void)0)
 #endif
 
 /* -- Write barriers ------------------------------------------------------ */
@@ -1683,8 +1681,6 @@ static void asm_fpunary(ASMState *as, IRIns *ir, MIPSIns mi)
 #if !LJ_SOFTFP32
 static void asm_fpmath(ASMState *as, IRIns *ir)
 {
-  if (ir->op2 == IRFPM_EXP2 && asm_fpjoin_pow(as, ir))
-    return;
 #if !LJ_SOFTFP
   if (ir->op2 <= IRFPM_TRUNC)
     asm_callround(as, ir, IRCALL_lj_vm_floor + ir->op2);
@@ -1772,41 +1768,13 @@ static void asm_mul(ASMState *as, IRIns *ir)
   }
 }
 
-static void asm_mod(ASMState *as, IRIns *ir)
-{
-#if LJ_64 && LJ_HASFFI
-  if (!irt_isint(ir->t))
-    asm_callid(as, ir, irt_isi64(ir->t) ? IRCALL_lj_carith_modi64 :
-					  IRCALL_lj_carith_modu64);
-  else
-#endif
-    asm_callid(as, ir, IRCALL_lj_vm_modi);
-}
-
 #if !LJ_SOFTFP32
-static void asm_pow(ASMState *as, IRIns *ir)
-{
-#if LJ_64 && LJ_HASFFI
-  if (!irt_isnum(ir->t))
-    asm_callid(as, ir, irt_isi64(ir->t) ? IRCALL_lj_carith_powi64 :
-					  IRCALL_lj_carith_powu64);
-  else
-#endif
-    asm_callid(as, ir, IRCALL_lj_vm_powi);
-}
-
-static void asm_div(ASMState *as, IRIns *ir)
+static void asm_fpdiv(ASMState *as, IRIns *ir)
 {
-#if LJ_64 && LJ_HASFFI
-  if (!irt_isnum(ir->t))
-    asm_callid(as, ir, irt_isi64(ir->t) ? IRCALL_lj_carith_divi64 :
-					  IRCALL_lj_carith_divu64);
-  else
-#endif
 #if !LJ_SOFTFP
     asm_fparith(as, ir, MIPSI_DIV_D);
 #else
-  asm_callid(as, ir, IRCALL_softfp_div);
+    asm_callid(as, ir, IRCALL_softfp_div);
 #endif
 }
 #endif
@@ -1844,8 +1812,6 @@ static void asm_abs(ASMState *as, IRIns *ir)
 }
 #endif
 
-#define asm_ldexp(as, ir)	asm_callid(as, ir, IRCALL_ldexp)
-
 static void asm_arithov(ASMState *as, IRIns *ir)
 {
   /* TODO MIPSR6: bovc/bnvc. Caveat: no delay slot to load RID_TMP. */
diff --git a/src/lj_asm_ppc.h b/src/lj_asm_ppc.h
index 6aaed058..62a5c3e2 100644
--- a/src/lj_asm_ppc.h
+++ b/src/lj_asm_ppc.h
@@ -1177,8 +1177,6 @@ static void asm_cnew(ASMState *as, IRIns *ir)
   ra_allockreg(as, (int32_t)(sz+sizeof(GCcdata)),
 	       ra_releasetmp(as, ASMREF_TMP1));
 }
-#else
-#define asm_cnew(as, ir)	((void)0)
 #endif
 
 /* -- Write barriers ------------------------------------------------------ */
@@ -1249,8 +1247,6 @@ static void asm_fpunary(ASMState *as, IRIns *ir, PPCIns pi)
 
 static void asm_fpmath(ASMState *as, IRIns *ir)
 {
-  if (ir->op2 == IRFPM_EXP2 && asm_fpjoin_pow(as, ir))
-    return;
   if (ir->op2 == IRFPM_SQRT && (as->flags & JIT_F_SQRT))
     asm_fpunary(as, ir, PPCI_FSQRT);
   else
@@ -1364,9 +1360,7 @@ static void asm_mul(ASMState *as, IRIns *ir)
   }
 }
 
-#define asm_div(as, ir)		asm_fparith(as, ir, PPCI_FDIV)
-#define asm_mod(as, ir)		asm_callid(as, ir, IRCALL_lj_vm_modi)
-#define asm_pow(as, ir)		asm_callid(as, ir, IRCALL_lj_vm_powi)
+#define asm_fpdiv(as, ir)	asm_fparith(as, ir, PPCI_FDIV)
 
 static void asm_neg(ASMState *as, IRIns *ir)
 {
@@ -1390,7 +1384,6 @@ static void asm_neg(ASMState *as, IRIns *ir)
 }
 
 #define asm_abs(as, ir)		asm_fpunary(as, ir, PPCI_FABS)
-#define asm_ldexp(as, ir)	asm_callid(as, ir, IRCALL_ldexp)
 
 static void asm_arithov(ASMState *as, IRIns *ir, PPCIns pi)
 {
diff --git a/src/lj_asm_x86.h b/src/lj_asm_x86.h
index 63d332ca..5f5fe3cf 100644
--- a/src/lj_asm_x86.h
+++ b/src/lj_asm_x86.h
@@ -1857,8 +1857,6 @@ static void asm_cnew(ASMState *as, IRIns *ir)
   asm_gencall(as, ci, args);
   emit_loadi(as, ra_releasetmp(as, ASMREF_TMP1), (int32_t)(sz+sizeof(GCcdata)));
 }
-#else
-#define asm_cnew(as, ir)	((void)0)
 #endif
 
 /* -- Write barriers ------------------------------------------------------ */
@@ -1964,8 +1962,6 @@ static void asm_fpmath(ASMState *as, IRIns *ir)
 		    fpm == IRFPM_CEIL ? lj_vm_ceil_sse : lj_vm_trunc_sse);
       ra_left(as, RID_XMM0, ir->op1);
     }
-  } else if (fpm == IRFPM_EXP2 && asm_fpjoin_pow(as, ir)) {
-    /* Rejoined to pow(). */
   } else {
     asm_callid(as, ir, IRCALL_lj_vm_floor + fpm);
   }
@@ -2000,17 +1996,6 @@ static void asm_fppowi(ASMState *as, IRIns *ir)
   ra_left(as, RID_EAX, ir->op2);
 }
 
-static void asm_pow(ASMState *as, IRIns *ir)
-{
-#if LJ_64 && LJ_HASFFI
-  if (!irt_isnum(ir->t))
-    asm_callid(as, ir, irt_isi64(ir->t) ? IRCALL_lj_carith_powi64 :
-					  IRCALL_lj_carith_powu64);
-  else
-#endif
-    asm_fppowi(as, ir);
-}
-
 static int asm_swapops(ASMState *as, IRIns *ir)
 {
   IRIns *irl = IR(ir->op1);
@@ -2208,27 +2193,7 @@ static void asm_mul(ASMState *as, IRIns *ir)
     asm_intarith(as, ir, XOg_X_IMUL);
 }
 
-static void asm_div(ASMState *as, IRIns *ir)
-{
-#if LJ_64 && LJ_HASFFI
-  if (!irt_isnum(ir->t))
-    asm_callid(as, ir, irt_isi64(ir->t) ? IRCALL_lj_carith_divi64 :
-					  IRCALL_lj_carith_divu64);
-  else
-#endif
-    asm_fparith(as, ir, XO_DIVSD);
-}
-
-static void asm_mod(ASMState *as, IRIns *ir)
-{
-#if LJ_64 && LJ_HASFFI
-  if (!irt_isint(ir->t))
-    asm_callid(as, ir, irt_isi64(ir->t) ? IRCALL_lj_carith_modi64 :
-					  IRCALL_lj_carith_modu64);
-  else
-#endif
-    asm_callid(as, ir, IRCALL_lj_vm_modi);
-}
+#define asm_fpdiv(as, ir)	asm_fparith(as, ir, XO_DIVSD)
 
 static void asm_neg_not(ASMState *as, IRIns *ir, x86Group3 xg)
 {
diff --git a/src/lj_ir.h b/src/lj_ir.h
index e8bca275..43e55069 100644
--- a/src/lj_ir.h
+++ b/src/lj_ir.h
@@ -177,7 +177,7 @@ LJ_STATIC_ASSERT((int)IR_XLOAD + IRDELTA_L2S == (int)IR_XSTORE);
 /* FPMATH sub-functions. ORDER FPM. */
 #define IRFPMDEF(_) \
   _(FLOOR) _(CEIL) _(TRUNC)  /* Must be first and in this order. */ \
-  _(SQRT) _(EXP2) _(LOG) _(LOG2) \
+  _(SQRT) _(LOG) _(LOG2) \
   _(OTHER)
 
 typedef enum {
diff --git a/src/lj_ircall.h b/src/lj_ircall.h
index bbad35b1..af064a6f 100644
--- a/src/lj_ircall.h
+++ b/src/lj_ircall.h
@@ -192,7 +192,6 @@ typedef struct CCallInfo {
   _(FPMATH,	lj_vm_ceil,		1,   N, NUM, XA_FP) \
   _(FPMATH,	lj_vm_trunc,		1,   N, NUM, XA_FP) \
   _(FPMATH,	sqrt,			1,   N, NUM, XA_FP) \
-  _(ANY,	lj_vm_exp2,		1,   N, NUM, XA_FP) \
   _(ANY,	log,			1,   N, NUM, XA_FP) \
   _(ANY,	lj_vm_log2,		1,   N, NUM, XA_FP) \
   _(ANY,	lj_vm_powi,		2,   N, NUM, XA_FP) \
diff --git a/src/lj_opt_fold.c b/src/lj_opt_fold.c
index 27e489af..cd803d87 100644
--- a/src/lj_opt_fold.c
+++ b/src/lj_opt_fold.c
@@ -237,10 +237,11 @@ LJFOLDF(kfold_fpcall2)
 }
 
 LJFOLD(POW KNUM KINT)
+LJFOLD(POW KNUM KNUM)
 LJFOLDF(kfold_numpow)
 {
   lua_Number a = knumleft;
-  lua_Number b = (lua_Number)fright->i;
+  lua_Number b = fright->o == IR_KINT ? (lua_Number)fright->i : knumright;
   lua_Number y = lj_vm_foldarith(a, b, IR_POW - IR_ADD);
   return lj_ir_knum(J, y);
 }
@@ -1077,7 +1078,7 @@ LJFOLDF(simplify_nummuldiv_negneg)
 }
 
 LJFOLD(POW any KINT)
-LJFOLDF(simplify_numpow_xk)
+LJFOLDF(simplify_numpow_xkint)
 {
   int32_t k = fright->i;
   TRef ref = fins->op1;
@@ -1106,13 +1107,22 @@ LJFOLDF(simplify_numpow_xk)
   return ref;
 }
 
+LJFOLD(POW any KNUM)
+LJFOLDF(simplify_numpow_xknum)
+{
+  if (knumright == 0.5)  /* x ^ 0.5 ==> sqrt(x) */
+    return emitir(IRTN(IR_FPMATH), fins->op1, IRFPM_SQRT);
+  return NEXTFOLD;
+}
+
 LJFOLD(POW KNUM any)
 LJFOLDF(simplify_numpow_kx)
 {
   lua_Number n = knumleft;
-  if (n == 2.0) {  /* 2.0 ^ i ==> ldexp(1.0, tonum(i)) */
-    fins->o = IR_CONV;
+  if (n == 2.0 && irt_isint(fright->t)) {  /* 2.0 ^ i ==> ldexp(1.0, i) */
 #if LJ_TARGET_X86ORX64
+    /* Different IR_LDEXP calling convention on x86/x64 requires conversion. */
+    fins->o = IR_CONV;
     fins->op1 = fins->op2;
     fins->op2 = IRCONV_NUM_INT;
     fins->op2 = (IRRef1)lj_opt_fold(J);
diff --git a/src/lj_opt_narrow.c b/src/lj_opt_narrow.c
index bb61f97b..4f285334 100644
--- a/src/lj_opt_narrow.c
+++ b/src/lj_opt_narrow.c
@@ -593,10 +593,10 @@ TRef lj_opt_narrow_pow(jit_State *J, TRef rb, TRef rc, TValue *vb, TValue *vc)
   /* Narrowing must be unconditional to preserve (-x)^i semantics. */
   if (tvisint(vc) || numisint(numV(vc))) {
     int checkrange = 0;
-    /* Split pow is faster for bigger exponents. But do this only for (+k)^i. */
+    /* pow() is faster for bigger exponents. But do this only for (+k)^i. */
     if (tref_isk(rb) && (int32_t)ir_knum(IR(tref_ref(rb)))->u32.hi >= 0) {
       int32_t k = numberVint(vc);
-      if (!(k >= -65536 && k <= 65536)) goto split_pow;
+      if (!(k >= -65536 && k <= 65536)) goto force_pow_num;
       checkrange = 1;
     }
     if (!tref_isinteger(rc)) {
@@ -607,19 +607,11 @@ TRef lj_opt_narrow_pow(jit_State *J, TRef rb, TRef rc, TValue *vb, TValue *vc)
       TRef tmp = emitir(IRTI(IR_ADD), rc, lj_ir_kint(J, 65536));
       emitir(IRTGI(IR_ULE), tmp, lj_ir_kint(J, 2*65536));
     }
-    return emitir(IRTN(IR_POW), rb, rc);
+  } else {
+force_pow_num:
+    rc = lj_ir_tonum(J, rc);  /* Want POW(num, num), not POW(num, int). */
   }
-split_pow:
-  /* FOLD covers most cases, but some are easier to do here. */
-  if (tref_isk(rb) && tvispone(ir_knum(IR(tref_ref(rb)))))
-    return rb;  /* 1 ^ x ==> 1 */
-  rc = lj_ir_tonum(J, rc);
-  if (tref_isk(rc) && ir_knum(IR(tref_ref(rc)))->n == 0.5)
-    return emitir(IRTN(IR_FPMATH), rb, IRFPM_SQRT);  /* x ^ 0.5 ==> sqrt(x) */
-  /* Split up b^c into exp2(c*log2(b)). Assembler may rejoin later. */
-  rb = emitir(IRTN(IR_FPMATH), rb, IRFPM_LOG2);
-  rc = emitir(IRTN(IR_MUL), rb, rc);
-  return emitir(IRTN(IR_FPMATH), rc, IRFPM_EXP2);
+  return emitir(IRTN(IR_POW), rb, rc);
 }
 
 /* -- Predictive narrowing of induction variables ------------------------- */
diff --git a/src/lj_opt_split.c b/src/lj_opt_split.c
index 2fc36b8d..c10a85cb 100644
--- a/src/lj_opt_split.c
+++ b/src/lj_opt_split.c
@@ -403,27 +403,6 @@ static void split_ir(jit_State *J)
 	hi = split_call_li(J, hisubst, oir, ir, IRCALL_lj_vm_powi);
 	break;
       case IR_FPMATH:
-	/* Try to rejoin pow from EXP2, MUL and LOG2. */
-	if (nir->op2 == IRFPM_EXP2 && nir->op1 > J->loopref) {
-	  IRIns *irp = IR(nir->op1);
-	  if (irp->o == IR_CALLN && irp->op2 == IRCALL_softfp_mul) {
-	    IRIns *irm4 = IR(irp->op1);
-	    IRIns *irm3 = IR(irm4->op1);
-	    IRIns *irm12 = IR(irm3->op1);
-	    IRIns *irl1 = IR(irm12->op1);
-	    if (irm12->op1 > J->loopref && irl1->o == IR_CALLN &&
-		irl1->op2 == IRCALL_lj_vm_log2) {
-	      IRRef tmp = irl1->op1;  /* Recycle first two args from LOG2. */
-	      IRRef arg3 = irm3->op2, arg4 = irm4->op2;
-	      J->cur.nins--;
-	      tmp = split_emit(J, IRT(IR_CARG, IRT_NIL), tmp, arg3);
-	      tmp = split_emit(J, IRT(IR_CARG, IRT_NIL), tmp, arg4);
-	      ir->prev = tmp = split_emit(J, IRTI(IR_CALLN), tmp, IRCALL_pow);
-	      hi = split_emit(J, IRT(IR_HIOP, IRT_SOFTFP), tmp, tmp);
-	      break;
-	    }
-	  }
-	}
 	hi = split_call_l(J, hisubst, oir, ir, IRCALL_lj_vm_floor + ir->op2);
 	break;
       case IR_LDEXP:
diff --git a/src/lj_vm.h b/src/lj_vm.h
index 411caafa..abaa7c52 100644
--- a/src/lj_vm.h
+++ b/src/lj_vm.h
@@ -95,11 +95,6 @@ LJ_ASMF double lj_vm_trunc(double);
 LJ_ASMF double lj_vm_trunc_sf(double);
 #endif
 #endif
-#ifdef LUAJIT_NO_EXP2
-LJ_ASMF double lj_vm_exp2(double);
-#else
-#define lj_vm_exp2	exp2
-#endif
 #if LJ_HASFFI
 LJ_ASMF int lj_vm_errno(void);
 #endif
diff --git a/src/lj_vmmath.c b/src/lj_vmmath.c
index ae4e0f15..9c0d3fde 100644
--- a/src/lj_vmmath.c
+++ b/src/lj_vmmath.c
@@ -79,13 +79,6 @@ double lj_vm_log2(double a)
 }
 #endif
 
-#ifdef LUAJIT_NO_EXP2
-double lj_vm_exp2(double a)
-{
-  return exp(a * 0.6931471805599453);
-}
-#endif
-
 #if !LJ_TARGET_X86ORX64
 /* Unsigned x^k. */
 static double lj_vm_powui(double x, uint32_t k)
@@ -128,7 +121,6 @@ double lj_vm_foldfpm(double x, int fpm)
   case IRFPM_CEIL: return lj_vm_ceil(x);
   case IRFPM_TRUNC: return lj_vm_trunc(x);
   case IRFPM_SQRT: return sqrt(x);
-  case IRFPM_EXP2: return lj_vm_exp2(x);
   case IRFPM_LOG: return log(x);
   case IRFPM_LOG2: return lj_vm_log2(x);
   default: lua_assert(0);
diff --git a/test/tarantool-tests/lj-9-pow-inconsistencies.test.lua b/test/tarantool-tests/lj-9-pow-inconsistencies.test.lua
new file mode 100644
index 00000000..21b3a0d9
--- /dev/null
+++ b/test/tarantool-tests/lj-9-pow-inconsistencies.test.lua
@@ -0,0 +1,63 @@
+local tap = require('tap')
+-- Test to demonstrate the incorrect JIT behaviour when splitting
+-- IR_POW.
+-- See also https://github.com/LuaJIT/LuaJIT/issues/9.
+local test = tap.test('lj-9-pow-inconsistencies'):skipcond({
+  ['Test requires JIT enabled'] = not jit.status(),
+})
+
+local nan = 0 / 0
+local inf = math.huge
+
+-- Table with some corner cases to check:
+local INTERESTING_VALUES = {
+  -- 0, -0, 1, -1 special cases with nan, inf, etc..
+  0, -0, 1, -1, nan, inf, -inf,
+  -- x ^  inf = 0 (inf), if |x| < 1 (|x| > 1).
+  -- x ^ -inf = inf (0), if |x| < 1 (|x| > 1).
+  0.999999, 1.000001, -0.999999, -1.000001,
+}
+test:plan(1 + (#INTERESTING_VALUES) ^ 2)
+
+jit.opt.start('hotloop=1')
+
+-- The JIT engine tries to split b^c to exp2(c * log2(b)).
+-- For some cases for IEEE754 we can see, that
+-- (double)exp2((double)log2(x)) != x, due to mathematical
+-- functions accuracy and double precision restrictions.
+-- Just use some numbers to observe this misbehaviour.
+local res = {}
+local cnt = 1
+while cnt < 4 do
+  -- XXX: use local variable to prevent folding via parser.
+  local b = -0.90000000001
+  res[cnt] = 1000 ^ b
+  cnt = cnt + 1
+end
+
+test:samevalues(res, 'consistent pow operator behaviour for corner case')
+
+-- Prevent JIT side effects for parent loops.
+jit.off()
+for i = 1, #INTERESTING_VALUES do
+  for j = 1, #INTERESTING_VALUES do
+    local b = INTERESTING_VALUES[i]
+    local c = INTERESTING_VALUES[j]
+    local results = {}
+    local counter = 1
+    jit.on()
+    while counter < 4 do
+      results[counter] = b ^ c
+      counter = counter + 1
+    end
+    -- Prevent JIT side effects.
+    jit.off()
+    jit.flush()
+    test:samevalues(
+      results,
+      ('consistent pow operator behaviour for (%s)^(%s)'):format(b, c)
+    )
+  end
+end
+
+test:done(true)
-- 
2.41.0