DrSocalkwe3n/luajit
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

ARM: Add fast assembler implementation of floor/ceil/trunc.

Browse Source
This commit is contained in:
Mike Pall
2011-06-05 23:37:29 +02:00
parent a175754ab7
commit a48058a791
5 changed files with 922 additions and 835 deletions
Download Patch File Download Diff File Expand all files Collapse all files

73
src/buildvm_arm.dasc
View File

@@ -1308,10 +1308,7 @@ static void build_subroutines(BuildCtx *ctx)
| movmi CARG1, #0x80000000
| bmi <1
|4:
| // NYI: Use internal implementation.
| IOS mov RA, BASE
| bl extern func
| IOS mov BASE, RA
| bl ->vm_..func
| b ->fff_restv
|.endmacro
|
@@ -2010,23 +2007,76 @@ static void build_subroutines(BuildCtx *ctx)
|// double lj_vm_floor/ceil/trunc(double x);
|.macro vm_round, func
|->vm_ .. func:
| // NYI: Use internal implementation.
| b extern func
| lsl CARG3, CARG2, #1
| adds RB, CARG3, #0x00200000
| bpl >2 // |x| < 1?
| mvn CARG4, #0x3cc
| subs RB, CARG4, RB, asr #21 // 2^0: RB = 51, 2^51: RB = 0.
| bxlo lr // |x| >= 2^52: done.
| mvn CARG4, #1
| bic CARG3, CARG1, CARG4, lsl RB // ztest = lo & ~lomask
| and CARG1, CARG1, CARG4, lsl RB // lo &= lomask
| subs RB, RB, #32
| bicpl CARG4, CARG2, CARG4, lsl RB // |x| <= 2^20: ztest |= hi & ~himask
| orrpl CARG3, CARG3, CARG4
| mvnpl CARG4, #1
| andpl CARG2, CARG2, CARG4, lsl RB // |x| <= 2^20: hi &= himask
|.if "func" == "floor"
| tst CARG3, CARG2, asr #31 // iszero = ((ztest & signmask) == 0)
|.else
| bics CARG3, CARG3, CARG2, asr #31 // iszero = ((ztest & ~signmask) == 0)
|.endif
| bxeq lr // iszero: done.
| mvn CARG4, #1
| cmp RB, #0
| lslpl CARG3, CARG4, RB
| mvnmi CARG3, #0
| add RB, RB, #32
| subs CARG1, CARG1, CARG4, lsl RB // lo = lo-lomask
| sbc CARG2, CARG2, CARG3 // hi = hi-himask+carry
| bx lr
|
|2: // |x| < 1:
| orr CARG3, CARG3, CARG1 // ztest = (2*hi) | lo
|.if "func" == "floor"
| tst CARG3, CARG2, asr #31 // iszero = ((ztest & signmask) == 0)
|.else
| bics CARG3, CARG3, CARG2, asr #31 // iszero = ((ztest & ~signmask) == 0)
|.endif
| mov CARG1, #0 // lo = 0
| and CARG2, CARG2, #0x80000000
| ldrne CARG4, <9 // hi = sign(x) | (iszero ? 0.0 : 1.0)
| orrne CARG2, CARG2, CARG4
| bx lr
|.endmacro
|
|9:
| .long 0x3ff00000 // hiword(1.0)
| vm_round floor
| vm_round ceil
#if LJ_HASJIT
| vm_round trunc
#else
|
|->vm_trunc:
#if LJ_HASJIT
| lsl CARG3, CARG2, #1
| adds RB, CARG3, #0x00200000
| andpl CARG2, CARG2, #0x80000000 // |x| < 1? hi = sign(x), lo = 0.
| movpl CARG1, #0
| bxpl lr
| mvn CARG4, #0x3cc
| subs RB, CARG4, RB, asr #21 // 2^0: RB = 51, 2^51: RB = 0.
| bxlo lr // |x| >= 2^52: already done.
| mvn CARG4, #1
| and CARG1, CARG1, CARG4, lsl RB // lo &= lomask
| subs RB, RB, #32
| andpl CARG2, CARG2, CARG4, lsl RB // |x| <= 2^20: hi &= himask
| bx lr
#endif
|
| // double lj_vm_mod(double dividend, double divisor);
|->vm_mod:
| push {r0, r1, r2, r3, r4, lr}
| bl extern __aeabi_ddiv
| bl extern floor // NYI: Use internal implementation of floor.
| bl ->vm_floor
| ldrd CARG34, [sp, #8]
| bl extern __aeabi_dmul
| ldrd CARG34, [sp]
@@ -2586,9 +2636,7 @@ static void build_ins(BuildCtx *ctx, BCOp op, int defop)
|5: // FP variant.
| ins_arithfallback ins_arithcheck_num
|.if "intins" == "vm_modi"
| IOS mov RC, BASE
| bl fpcall
| IOS mov BASE, RC // NYI: remove once we use internal impl. of floor.
|.else
| bl fpcall
| ins_next1
@@ -3966,7 +4014,6 @@ static void emit_asm_debug(BuildCtx *ctx)
fprintf(ctx->fp,
"\t.align 2\n"
".LEFDE0:\n\n");
/* NYI: emit ARM.exidx. */
break;
default:
break;