Tarantool development patches archive
 help / color / mirror / Atom feed
* [Tarantool-patches] [PATCH luajit v2] Cleanup math function compilation and fix inconsistencies.
@ 2022-08-18 23:28 Maxim Kokryashkin via Tarantool-patches
  2022-08-22  9:40 ` Sergey Kaplun via Tarantool-patches
  0 siblings, 1 reply; 2+ messages in thread
From: Maxim Kokryashkin via Tarantool-patches @ 2022-08-18 23:28 UTC (permalink / raw)
  To: tarantool-patches, imun, skaplun

This patch changes 'math_unary', `math_htrig` and `math_atrig`
to `math_call` for math functions compilation. Now all of
math functions in IRs are called with `CALLN` instead of
`FPMATH` (`ATAN2` for `math.atan2`).

The `ATAN2` instruction itself is removed, as well as the fold
optimization for it.

Also, this patch adds new fold optimizations for `CALLN`.

Part of tarantool/tarantool#7230
---
As we decided offline, no tests are provided with this patch, because
of their extreme complexity with no real benefits.

Branch: https://github.com/tarantool/luajit/tree/fckxorg/gh-7230-cleanup-math-inconsistencies
PR: https://github.com/tarantool/tarantool/pull/7586

 src/lib_math.c      | 22 +++++++++++-----------
 src/lj_asm.c        |  6 ------
 src/lj_asm_arm.h    |  1 -
 src/lj_asm_arm64.h  |  1 -
 src/lj_asm_x86.h    |  2 --
 src/lj_ffrecord.c   | 19 ++-----------------
 src/lj_ir.h         |  4 +---
 src/lj_ircall.h     | 14 +++++++++-----
 src/lj_opt_fold.c   | 25 ++++++++++++++++++++++++-
 src/lj_opt_split.c  |  3 ---
 src/lj_target_x86.h |  6 ------
 src/lj_vmmath.c     |  6 ------
 12 files changed, 47 insertions(+), 62 deletions(-)

diff --git a/src/lib_math.c b/src/lib_math.c
index ef9dda2d..e9cea9ca 100644
--- a/src/lib_math.c
+++ b/src/lib_math.c
@@ -33,17 +33,17 @@ LJLIB_ASM(math_sqrt)		LJLIB_REC(math_unary IRFPM_SQRT)
   lj_lib_checknum(L, 1);
   return FFH_RETRY;
 }
-LJLIB_ASM_(math_log10)		LJLIB_REC(math_unary IRFPM_LOG10)
-LJLIB_ASM_(math_exp)		LJLIB_REC(math_unary IRFPM_EXP)
-LJLIB_ASM_(math_sin)		LJLIB_REC(math_unary IRFPM_SIN)
-LJLIB_ASM_(math_cos)		LJLIB_REC(math_unary IRFPM_COS)
-LJLIB_ASM_(math_tan)		LJLIB_REC(math_unary IRFPM_TAN)
-LJLIB_ASM_(math_asin)		LJLIB_REC(math_atrig FF_math_asin)
-LJLIB_ASM_(math_acos)		LJLIB_REC(math_atrig FF_math_acos)
-LJLIB_ASM_(math_atan)		LJLIB_REC(math_atrig FF_math_atan)
-LJLIB_ASM_(math_sinh)		LJLIB_REC(math_htrig IRCALL_sinh)
-LJLIB_ASM_(math_cosh)		LJLIB_REC(math_htrig IRCALL_cosh)
-LJLIB_ASM_(math_tanh)		LJLIB_REC(math_htrig IRCALL_tanh)
+LJLIB_ASM_(math_log10)		LJLIB_REC(math_call IRCALL_log10)
+LJLIB_ASM_(math_exp)		LJLIB_REC(math_call IRCALL_exp)
+LJLIB_ASM_(math_sin)		LJLIB_REC(math_call IRCALL_sin)
+LJLIB_ASM_(math_cos)		LJLIB_REC(math_call IRCALL_cos)
+LJLIB_ASM_(math_tan)		LJLIB_REC(math_call IRCALL_tan)
+LJLIB_ASM_(math_asin)		LJLIB_REC(math_call IRCALL_asin)
+LJLIB_ASM_(math_acos)		LJLIB_REC(math_call IRCALL_acos)
+LJLIB_ASM_(math_atan)		LJLIB_REC(math_call IRCALL_atan)
+LJLIB_ASM_(math_sinh)		LJLIB_REC(math_call IRCALL_sinh)
+LJLIB_ASM_(math_cosh)		LJLIB_REC(math_call IRCALL_cosh)
+LJLIB_ASM_(math_tanh)		LJLIB_REC(math_call IRCALL_tanh)
 LJLIB_ASM_(math_frexp)
 LJLIB_ASM_(math_modf)		LJLIB_REC(.)
 
diff --git a/src/lj_asm.c b/src/lj_asm.c
index 10e5872b..1a7fb0c8 100644
--- a/src/lj_asm.c
+++ b/src/lj_asm.c
@@ -1660,7 +1660,6 @@ static void asm_ir(ASMState *as, IRIns *ir)
   case IR_DIV: asm_div(as, ir); break;
   case IR_POW: asm_pow(as, ir); break;
   case IR_ABS: asm_abs(as, ir); break;
-  case IR_ATAN2: asm_atan2(as, ir); break;
   case IR_LDEXP: asm_ldexp(as, ir); break;
   case IR_FPMATH: asm_fpmath(as, ir); break;
   case IR_TOBIT: asm_tobit(as, ir); break;
@@ -2150,11 +2149,6 @@ static void asm_setup_regsp(ASMState *as)
 	as->modset = RSET_SCRATCH;
       break;
 #if !LJ_SOFTFP
-    case IR_ATAN2:
-#if LJ_TARGET_X86
-      if (as->evenspill < 4)  /* Leave room to call atan2(). */
-	as->evenspill = 4;
-#endif
 #if !LJ_TARGET_X86ORX64
     case IR_LDEXP:
 #endif
diff --git a/src/lj_asm_arm.h b/src/lj_asm_arm.h
index 4fd08b9e..8af19eb9 100644
--- a/src/lj_asm_arm.h
+++ b/src/lj_asm_arm.h
@@ -1502,7 +1502,6 @@ static void asm_mul(ASMState *as, IRIns *ir)
 #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_abs(as, ir)		asm_fpunary(as, ir, ARMI_VABS_D)
-#define asm_atan2(as, ir)	asm_callid(as, ir, IRCALL_atan2)
 #define asm_ldexp(as, ir)	asm_callid(as, ir, IRCALL_ldexp)
 #endif
 
diff --git a/src/lj_asm_arm64.h b/src/lj_asm_arm64.h
index da0ee4bb..4aeb51f3 100644
--- a/src/lj_asm_arm64.h
+++ b/src/lj_asm_arm64.h
@@ -1450,7 +1450,6 @@ static void asm_pow(ASMState *as, IRIns *ir)
 #define asm_mulov(as, ir)	asm_mul(as, ir)
 
 #define asm_abs(as, ir)		asm_fpunary(as, ir, A64I_FABS)
-#define asm_atan2(as, ir)	asm_callid(as, ir, IRCALL_atan2)
 #define asm_ldexp(as, ir)	asm_callid(as, ir, IRCALL_ldexp)
 
 static void asm_mod(ASMState *as, IRIns *ir)
diff --git a/src/lj_asm_x86.h b/src/lj_asm_x86.h
index 2850aea9..8a4d4025 100644
--- a/src/lj_asm_x86.h
+++ b/src/lj_asm_x86.h
@@ -1971,8 +1971,6 @@ static void asm_fpmath(ASMState *as, IRIns *ir)
   }
 }
 
-#define asm_atan2(as, ir)	asm_callid(as, ir, IRCALL_atan2)
-
 static void asm_ldexp(ASMState *as, IRIns *ir)
 {
   int32_t ofs = sps_scale(ir->s);  /* Use spill slot or temp slots. */
diff --git a/src/lj_ffrecord.c b/src/lj_ffrecord.c
index be890a93..0c9c211d 100644
--- a/src/lj_ffrecord.c
+++ b/src/lj_ffrecord.c
@@ -563,7 +563,7 @@ static void LJ_FASTCALL recff_math_atan2(jit_State *J, RecordFFData *rd)
 {
   TRef tr = lj_ir_tonum(J, J->base[0]);
   TRef tr2 = lj_ir_tonum(J, J->base[1]);
-  J->base[0] = emitir(IRTN(IR_ATAN2), tr, tr2);
+  J->base[0] = lj_ir_call(J, IRCALL_atan2, tr, tr2);
   UNUSED(rd);
 }
 
@@ -580,22 +580,7 @@ static void LJ_FASTCALL recff_math_ldexp(jit_State *J, RecordFFData *rd)
   UNUSED(rd);
 }
 
-/* Record math.asin, math.acos, math.atan. */
-static void LJ_FASTCALL recff_math_atrig(jit_State *J, RecordFFData *rd)
-{
-  TRef y = lj_ir_tonum(J, J->base[0]);
-  TRef x = lj_ir_knum_one(J);
-  uint32_t ffid = rd->data;
-  if (ffid != FF_math_atan) {
-    TRef tmp = emitir(IRTN(IR_MUL), y, y);
-    tmp = emitir(IRTN(IR_SUB), x, tmp);
-    tmp = emitir(IRTN(IR_FPMATH), tmp, IRFPM_SQRT);
-    if (ffid == FF_math_asin) { x = tmp; } else { x = y; y = tmp; }
-  }
-  J->base[0] = emitir(IRTN(IR_ATAN2), y, x);
-}
-
-static void LJ_FASTCALL recff_math_htrig(jit_State *J, RecordFFData *rd)
+static void LJ_FASTCALL recff_math_call(jit_State *J, RecordFFData *rd)
 {
   TRef tr = lj_ir_tonum(J, J->base[0]);
   J->base[0] = emitir(IRTN(IR_CALLN), tr, rd->data);
diff --git a/src/lj_ir.h b/src/lj_ir.h
index 3059bf65..4bad47ed 100644
--- a/src/lj_ir.h
+++ b/src/lj_ir.h
@@ -75,7 +75,6 @@
   _(NEG,	N , ref, ref) \
   \
   _(ABS,	N , ref, ref) \
-  _(ATAN2,	N , ref, ref) \
   _(LDEXP,	N , ref, ref) \
   _(MIN,	C , ref, ref) \
   _(MAX,	C , ref, ref) \
@@ -178,8 +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) _(EXP) _(EXP2) _(LOG) _(LOG2) _(LOG10) \
-  _(SIN) _(COS) _(TAN) \
+  _(SQRT) _(EXP2) _(LOG) _(LOG2) \
   _(OTHER)
 
 typedef enum {
diff --git a/src/lj_ircall.h b/src/lj_ircall.h
index 973c36e6..aa06b273 100644
--- a/src/lj_ircall.h
+++ b/src/lj_ircall.h
@@ -21,6 +21,7 @@ typedef struct CCallInfo {
 
 #define CCI_OTSHIFT		16
 #define CCI_OPTYPE(ci)		((ci)->flags >> CCI_OTSHIFT)  /* Get op/type. */
+#define CCI_TYPE(ci)		(((ci)->flags>>CCI_OTSHIFT) & IRT_TYPE)
 #define CCI_OPSHIFT		24
 #define CCI_OP(ci)		((ci)->flags >> CCI_OPSHIFT)  /* Get op. */
 
@@ -158,6 +159,14 @@ typedef struct CCallInfo {
   _(ANY,	lj_mem_newgco,		2,  FS, PGC, CCI_L) \
   _(ANY,	lj_math_random_step, 1, FS, NUM, CCI_CASTU64) \
   _(ANY,	lj_vm_modi,		2,  FN, INT, 0) \
+  _(ANY,	log10,			1,   N, NUM, XA_FP) \
+  _(ANY,	exp,			1,   N, NUM, XA_FP) \
+  _(ANY,	sin,			1,   N, NUM, XA_FP) \
+  _(ANY,	cos,			1,   N, NUM, XA_FP) \
+  _(ANY,	tan,			1,   N, NUM, XA_FP) \
+  _(ANY,	asin,			1,   N, NUM, XA_FP) \
+  _(ANY,	acos,			1,   N, NUM, XA_FP) \
+  _(ANY,	atan,			1,   N, NUM, XA_FP) \
   _(ANY,	sinh,			1,   N, NUM, XA_FP) \
   _(ANY,	cosh,			1,   N, NUM, XA_FP) \
   _(ANY,	tanh,			1,   N, NUM, XA_FP) \
@@ -169,14 +178,9 @@ 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,	exp,			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,	log10,			1,   N, NUM, XA_FP) \
-  _(ANY,	sin,			1,   N, NUM, XA_FP) \
-  _(ANY,	cos,			1,   N, NUM, XA_FP) \
-  _(ANY,	tan,			1,   N, NUM, XA_FP) \
   _(ANY,	lj_vm_powi,		2,   N, NUM, XA_FP) \
   _(ANY,	pow,			2,   N, NUM, XA2_FP) \
   _(ANY,	atan2,			2,   N, NUM, XA2_FP) \
diff --git a/src/lj_opt_fold.c b/src/lj_opt_fold.c
index 276dc040..49f74996 100644
--- a/src/lj_opt_fold.c
+++ b/src/lj_opt_fold.c
@@ -173,7 +173,6 @@ LJFOLD(ADD KNUM KNUM)
 LJFOLD(SUB KNUM KNUM)
 LJFOLD(MUL KNUM KNUM)
 LJFOLD(DIV KNUM KNUM)
-LJFOLD(ATAN2 KNUM KNUM)
 LJFOLD(LDEXP KNUM KNUM)
 LJFOLD(MIN KNUM KNUM)
 LJFOLD(MAX KNUM KNUM)
@@ -213,6 +212,30 @@ LJFOLDF(kfold_fpmath)
   return lj_ir_knum(J, y);
 }
 
+LJFOLD(CALLN KNUM any)
+LJFOLDF(kfold_fpcall1)
+{
+  const CCallInfo *ci = &lj_ir_callinfo[fins->op2];
+  if (CCI_TYPE(ci) == IRT_NUM) {
+    double y = ((double (*)(double))ci->func)(knumleft);
+    return lj_ir_knum(J, y);
+  }
+  return NEXTFOLD;
+}
+
+LJFOLD(CALLN CARG IRCALL_atan2)
+LJFOLDF(kfold_fpcall2)
+{
+  if (irref_isk(fleft->op1) && irref_isk(fleft->op2)) {
+    const CCallInfo *ci = &lj_ir_callinfo[fins->op2];
+    double a = ir_knum(IR(fleft->op1))->n;
+    double b = ir_knum(IR(fleft->op2))->n;
+    double y = ((double (*)(double, double))ci->func)(a, b);
+    return lj_ir_knum(J, y);
+  }
+  return NEXTFOLD;
+}
+
 LJFOLD(POW KNUM KINT)
 LJFOLDF(kfold_numpow)
 {
diff --git a/src/lj_opt_split.c b/src/lj_opt_split.c
index fc935204..c0788106 100644
--- a/src/lj_opt_split.c
+++ b/src/lj_opt_split.c
@@ -426,9 +426,6 @@ static void split_ir(jit_State *J)
 	}
 	hi = split_call_l(J, hisubst, oir, ir, IRCALL_lj_vm_floor + ir->op2);
 	break;
-      case IR_ATAN2:
-	hi = split_call_ll(J, hisubst, oir, ir, IRCALL_atan2);
-	break;
       case IR_LDEXP:
 	hi = split_call_li(J, hisubst, oir, ir, IRCALL_ldexp);
 	break;
diff --git a/src/lj_target_x86.h b/src/lj_target_x86.h
index 356f7924..194f8e70 100644
--- a/src/lj_target_x86.h
+++ b/src/lj_target_x86.h
@@ -228,16 +228,10 @@ typedef enum {
   /* Note: little-endian byte-order! */
   XI_FLDZ =	0xeed9,
   XI_FLD1 =	0xe8d9,
-  XI_FLDLG2 =	0xecd9,
-  XI_FLDLN2 =	0xedd9,
   XI_FDUP =	0xc0d9,  /* Really fld st0. */
   XI_FPOP =	0xd8dd,  /* Really fstp st0. */
   XI_FPOP1 =	0xd9dd,  /* Really fstp st1. */
   XI_FRNDINT =	0xfcd9,
-  XI_FSIN =	0xfed9,
-  XI_FCOS =	0xffd9,
-  XI_FPTAN =	0xf2d9,
-  XI_FPATAN =	0xf3d9,
   XI_FSCALE =	0xfdd9,
   XI_FYL2X =	0xf1d9,
 
diff --git a/src/lj_vmmath.c b/src/lj_vmmath.c
index b231d3e8..c04459bd 100644
--- a/src/lj_vmmath.c
+++ b/src/lj_vmmath.c
@@ -48,7 +48,6 @@ double lj_vm_foldarith(double x, double y, int op)
   case IR_NEG - IR_ADD: return -x; break;
   case IR_ABS - IR_ADD: return fabs(x); break;
 #if LJ_HASJIT
-  case IR_ATAN2 - IR_ADD: return atan2(x, y); break;
   case IR_LDEXP - IR_ADD: return ldexp(x, (int)y); break;
   case IR_MIN - IR_ADD: return x > y ? y : x; break;
   case IR_MAX - IR_ADD: return x < y ? y : x; break;
@@ -129,14 +128,9 @@ 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_EXP: return exp(x);
   case IRFPM_EXP2: return lj_vm_exp2(x);
   case IRFPM_LOG: return log(x);
   case IRFPM_LOG2: return lj_vm_log2(x);
-  case IRFPM_LOG10: return log10(x);
-  case IRFPM_SIN: return sin(x);
-  case IRFPM_COS: return cos(x);
-  case IRFPM_TAN: return tan(x);
   default: lua_assert(0);
   }
   return 0;
-- 
2.36.1


^ permalink raw reply	[flat|nested] 2+ messages in thread

* Re: [Tarantool-patches] [PATCH luajit v2] Cleanup math function compilation and fix inconsistencies.
  2022-08-18 23:28 [Tarantool-patches] [PATCH luajit v2] Cleanup math function compilation and fix inconsistencies Maxim Kokryashkin via Tarantool-patches
@ 2022-08-22  9:40 ` Sergey Kaplun via Tarantool-patches
  0 siblings, 0 replies; 2+ messages in thread
From: Sergey Kaplun via Tarantool-patches @ 2022-08-22  9:40 UTC (permalink / raw)
  To: Maxim Kokryashkin; +Cc: tarantool-patches

Hi, Maxim!

Thanks for the patch!
LGTM.

On 19.08.22, Maxim Kokryashkin wrote:
> This patch changes 'math_unary', `math_htrig` and `math_atrig`
> to `math_call` for math functions compilation. Now all of
> math functions in IRs are called with `CALLN` instead of
> `FPMATH` (`ATAN2` for `math.atan2`).
> 
> The `ATAN2` instruction itself is removed, as well as the fold
> optimization for it.
> 
> Also, this patch adds new fold optimizations for `CALLN`.
> 
> Part of tarantool/tarantool#7230
> ---
> As we decided offline, no tests are provided with this patch, because
> of their extreme complexity with no real benefits.
> 
> Branch: https://github.com/tarantool/luajit/tree/fckxorg/gh-7230-cleanup-math-inconsistencies
> PR: https://github.com/tarantool/tarantool/pull/7586
> 
>  src/lib_math.c      | 22 +++++++++++-----------
>  src/lj_asm.c        |  6 ------
>  src/lj_asm_arm.h    |  1 -
>  src/lj_asm_arm64.h  |  1 -
>  src/lj_asm_x86.h    |  2 --
>  src/lj_ffrecord.c   | 19 ++-----------------
>  src/lj_ir.h         |  4 +---
>  src/lj_ircall.h     | 14 +++++++++-----
>  src/lj_opt_fold.c   | 25 ++++++++++++++++++++++++-
>  src/lj_opt_split.c  |  3 ---
>  src/lj_target_x86.h |  6 ------
>  src/lj_vmmath.c     |  6 ------
>  12 files changed, 47 insertions(+), 62 deletions(-)
> 

<snipped>

> -- 
> 2.36.1
> 

-- 
Best regards,
Sergey Kaplun

^ permalink raw reply	[flat|nested] 2+ messages in thread

end of thread, other threads:[~2022-08-22  9:43 UTC | newest]

Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2022-08-18 23:28 [Tarantool-patches] [PATCH luajit v2] Cleanup math function compilation and fix inconsistencies Maxim Kokryashkin via Tarantool-patches
2022-08-22  9:40 ` Sergey Kaplun via Tarantool-patches

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox