[Tarantool-patches] [PATCH luajit 2/5] Remove pow() splitting and cleanup backends.
Maxim Kokryashkin
m.kokryashkin at tarantool.org
Sun Aug 20 12:48:18 MSK 2023
Hi, Sergey!
Thanks for the fixes!
LGTM now, see my responses below.
On Thu, Aug 17, 2023 at 06:33:31PM +0300, Sergey Kaplun wrote:
> Hi Maxim!
> Thanks for the review!
> Updated considering your comments.
>
> On 17.08.23, Maxim Kokryashkin wrote:
> > Hi, Sergey!
> > Thanks for the patch!
> > Please consider my comments below.
> >
> > On Tue, Aug 15, 2023 at 12:36:28PM +0300, Sergey Kaplun wrote:
> > > 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
> > Typo: s/with attempt/with an attempt/
>
> Fixed.
>
> > > 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.
> >
> > There is a lot to it. There are chnages in emission, fold optimizations,
> > narrowing, etc. Maybe it is worth mentioning some key changes that
> > happened as a result of that? That way, this changeset is easier to absorb.
>
> It's mentioned below, or I don't understand the idea.
Well, I think my brain just short-circuited or somehting. Yep, everything is ok.
>
> >
> > >
> > > This patch removes splitting of pow operator, so IR_POW is emitting for
> > Typo: s/removes/removes the/
>
> Fixed.
>
> > > all cases (except power of 0.5 replaced with sqrt operation).
> > Typo: s/except/except for the/
> > Typo: s/0.5/0.5, which is/
> > Typo: s/with sqrt/with the sqrt/
>
> Fixed all.
>
> > >
> > > 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
> > Typo: s/to the/to/
>
> Fixed.
>
> > > <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()`.
> > Typo: s/moved/was moved/
> > Typo: s/to a/to/
>
> Fixed all.
>
> > > * Since `lj_vm_exp2()` subroutine and `IRFPM_EXP2` are removed as no
> > > longer used.
> > I can't understand what this sentence means, please rephrase it.
>
> Removed "Since" as measleading.
>
> > >
> >
> > What about changes with `asm_cnew`? I think you should mention them too.
>
> Added.
>
> > > 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)
> >
> > I suggest renaming it to `CORNER_CASES`, since `INTERESTING_VALUES`
> > is not very formal.
>
> Renamed.
>
> > Also, please mention that not all of the possible pairs are faulty
> > and most of them are left here for two reasons:
> > 1. Improved readability.
> > 2. More extensive and change-proof testing.
>
> Added the comment.
>
> >
> > > +
> > > +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
> >
> > Is there a specific reason you decided to use while over for?
>
> Since I can't remember, I think no, so I replaced with `for`.
>
> > > +
> > > +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
> > Same question about for and while.
>
> Fixed.
>
> > > + -- Prevent JIT side effects.
> > > + jit.off()
> > > + jit.flush()
> > Also, I think we should move the part from jit.on() to jit.flush() into
> > a separate function.
>
> I don't agree here -- we still use tons of variables from the cycles,
> and I don't want to see any side-effects of the function call in
> traces.
Ok, that is not a big deal.
>
> See other changes in the iterative patch below:
>
> ===================================================================
> diff --git a/test/tarantool-tests/lj-9-pow-inconsistencies.test.lua b/test/tarantool-tests/lj-9-pow-inconsistencies.test.lua
> index 21b3a0d9..6abba07f 100644
> --- a/test/tarantool-tests/lj-9-pow-inconsistencies.test.lua
> +++ b/test/tarantool-tests/lj-9-pow-inconsistencies.test.lua
> @@ -10,14 +10,18 @@ local nan = 0 / 0
> local inf = math.huge
>
> -- Table with some corner cases to check:
> -local INTERESTING_VALUES = {
> +-- Not all of them fail on each CPU architecture, but bruteforce
> +-- is better, than custom enumerated usage for two reasons:
> +-- * Improved readability.
> +-- * More extensive and change-proof testing.
> +local CORNER_CASES = {
> -- 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)
> +test:plan(1 + (#CORNER_CASES) ^ 2)
>
> jit.opt.start('hotloop=1')
>
> @@ -27,28 +31,25 @@ jit.opt.start('hotloop=1')
> -- functions accuracy and double precision restrictions.
> -- Just use some numbers to observe this misbehaviour.
> local res = {}
> -local cnt = 1
> -while cnt < 4 do
> +for i = 1, 4 do
> -- XXX: use local variable to prevent folding via parser.
> local b = -0.90000000001
> - res[cnt] = 1000 ^ b
> - cnt = cnt + 1
> + res[i] = 1000 ^ b
> 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]
> +for i = 1, #CORNER_CASES do
> + for j = 1, #CORNER_CASES do
> + local b = CORNER_CASES[i]
> + local c = CORNER_CASES[j]
> local results = {}
> local counter = 1
> jit.on()
> - while counter < 4 do
> - results[counter] = b ^ c
> - counter = counter + 1
> + for k = 1, 4 do
> + results[k] = b ^ c
> end
> -- Prevent JIT side effects.
> jit.off()
> ===================================================================
>
> <snipped>
> > > --
> > > 2.41.0
> > >
>
> --
> Best regards,
> Sergey Kaplun
More information about the Tarantool-patches
mailing list