about summary refs log tree commit diff
path: root/sysdeps/ia64/fpu/e_fmodl.S
diff options
context:
space:
mode:
Diffstat (limited to 'sysdeps/ia64/fpu/e_fmodl.S')
-rw-r--r--sysdeps/ia64/fpu/e_fmodl.S865
1 files changed, 471 insertions, 394 deletions
diff --git a/sysdeps/ia64/fpu/e_fmodl.S b/sysdeps/ia64/fpu/e_fmodl.S
index da08ae3f5c..3e87eb090f 100644
--- a/sysdeps/ia64/fpu/e_fmodl.S
+++ b/sysdeps/ia64/fpu/e_fmodl.S
@@ -1,7 +1,7 @@
 .file "fmodl.s"
 
 
-// Copyright (c) 2000 - 2003, Intel Corporation
+// Copyright (c) 2000 - 2004, Intel Corporation
 // All rights reserved.
 //
 // Contributed 2000 by the Intel Numerics Group, Intel Corporation
@@ -43,56 +43,88 @@
 // 03/02/00 New Algorithm
 // 04/04/00 Unwind support added
 // 08/15/00 Bundle added after call to __libm_error_support to properly
-//          set [the previously overwritten] GR_Parameter_RESULT.
+// set [ the previously overwritten ] GR_Parameter_RESULT.
 // 11/28/00 Set FR_Y to f9
-// 03/11/02 Fixed flags for fmodl(qnan,zero)
+// 03/11/02 Fixed flags for fmodl(qnan, zero)
 // 05/20/02 Cleaned up namespace and sf0 syntax
-// 02/10/03 Reordered header: .section, .global, .proc, .align
-// 04/28/03 Fix: fmod(sNaN,0) no longer sets errno
+// 02/10/03 Reordered header:.section,.global,.proc,.align
+// 04/28/03 Fix: fmod(sNaN, 0) no longer sets errno
+// 11/23/04 Reformatted routine and improved speed
 //
 // API
 //====================================================================
-// long double fmodl(long double,long double);
+// long double fmodl(long double, long double);
 //
 // Overview of operation
 //====================================================================
-//  fmod(a,b)=a-i*b,
-//  where i is an integer such that, if b!=0,
-//  |i|<|a/b| and |a/b-i|<1
+// fmod(a, b)= a-i*b,
+// where i is an integer such that, if b!= 0,
+// |i|<|a/b| and |a/b-i|<1
 //
 // Algorithm
 //====================================================================
 // a). if |a|<|b|, return a
 // b). get quotient and reciprocal overestimates accurate to
-//     33 bits (q2,y2)
+// 33 bits (q2, y2)
 // c). if the exponent difference (exponent(a)-exponent(b))
-//     is less than 32, truncate quotient to integer and
-//     finish in one iteration
-// d). if exponent(a)-exponent(b)>=32 (q2>=2^32)
-//     round quotient estimate to single precision (k=RN(q2)),
-//     calculate partial remainder (a'=a-k*b),
-//     get quotient estimate (a'*y2), and repeat from c).
+// is less than 32, truncate quotient to integer and
+// finish in one iteration
+// d). if exponent(a)-exponent(b)>= 32 (q2>= 2^32)
+// round quotient estimate to single precision (k= RN(q2)),
+// calculate partial remainder (a'= a-k*b),
+// get quotient estimate (a'*y2), and repeat from c).
 //
 // Registers used
 //====================================================================
-// Predicate registers: p6-p11
-// General registers:   r2,r29,r32 (ar.pfs), r33-r39
-// Floating point registers: f6-f15
-
-GR_SAVE_B0                    = r33
-GR_SAVE_PFS                   = r34
-GR_SAVE_GP                    = r35
-GR_SAVE_SP                    = r36
-
-GR_Parameter_X                = r37
-GR_Parameter_Y                = r38
-GR_Parameter_RESULT           = r39
-GR_Parameter_TAG              = r40
-
-FR_X             = f10
-FR_Y             = f9
-FR_RESULT        = f8
 
+GR_SMALLBIASEXP     = r2
+GR_2P32             = r3
+GR_SMALLBIASEXP     = r20
+GR_ROUNDCONST       = r21
+GR_SIG_B            = r22
+GR_ARPFS            = r23
+GR_TMP1             = r24
+GR_TMP2             = r25
+GR_TMP3             = r26
+
+GR_SAVE_B0          = r33
+GR_SAVE_PFS         = r34
+GR_SAVE_GP          = r35
+GR_SAVE_SP          = r36
+
+GR_Parameter_X      = r37
+GR_Parameter_Y      = r38
+GR_Parameter_RESULT = r39
+GR_Parameter_TAG    = r40
+
+FR_X                = f10
+FR_Y                = f9
+FR_RESULT           = f8
+
+FR_ABS_A            = f6
+FR_ABS_B            = f7
+FR_Y_INV            = f10
+FR_SMALLBIAS        = f11
+FR_E0               = f12
+FR_Q                = f13
+FR_E1               = f14
+FR_2P32             = f15
+FR_TMPX             = f32
+FR_TMPY             = f33
+FR_ROUNDCONST       = f34
+FR_QINT             = f35
+FR_QRND24           = f36
+FR_NORM_B           = f37
+FR_TMP              = f38
+FR_TMP2             = f39
+FR_DFLAG            = f40
+FR_Y_INV0           = f41
+FR_Y_INV1           = f42
+FR_Q0               = f43
+FR_Q1               = f44
+FR_QINT_Z           = f45
+FR_QREM             = f46
+FR_B_SGN_A          = f47
 
 .section .text
 GLOBAL_IEEE754_ENTRY(fmodl)
@@ -101,495 +133,540 @@ GLOBAL_IEEE754_ENTRY(fmodl)
 // result in f8
 
 { .mfi
-  alloc r32=ar.pfs,1,4,4,0
-  // f6=|a|
-  fmerge.s f6=f0,f8
-  mov r2 = 0x0ffdd
+       getf.sig GR_SIG_B = f9
+       // FR_ABS_A = |a|
+       fmerge.s FR_ABS_A = f0, f8
+       mov GR_SMALLBIASEXP = 0x0ffdd
 }
-  {.mfi
-  getf.sig r29=f9
-  // f7=|b|
-  fmerge.s f7=f0,f9
-  nop.i 0;;
+{ .mfi
+       nop.m 0
+       // FR_ABS_B = |b|
+       fmerge.s FR_ABS_B = f0, f9
+       nop.i 0
 }
+;;
 
 { .mfi
-  setf.exp f11 = r2
-  // (1) y0
-  frcpa.s1 f10,p6=f6,f7
-  nop.i 0;;
+       setf.exp FR_SMALLBIAS = GR_SMALLBIASEXP
+       // (1) y0
+       frcpa.s1 FR_Y_INV0, p6 = FR_ABS_A, FR_ABS_B
+       nop.i 0
+}
+;;
+
+{ .mlx
+       nop.m 0
+       movl GR_ROUNDCONST = 0x33a00000
 }
+;;
 
 // eliminate special cases
-{.mmi
-nop.m 0
-nop.m 0
-// y pseudo-zero ?
-cmp.eq p7,p10=r29,r0;;
+{ .mmi
+       nop.m 0
+       nop.m 0
+       // y pseudo-zero ?
+       cmp.eq p7, p10 = GR_SIG_B, r0
 }
+;;
 
-// Y +-NAN, +-inf, +-0?     p7
+// set p7 if b +/-NAN, +/-inf, +/-0
 { .mfi
-      nop.m 999
-(p10)  fclass.m  p7,p10 = f9, 0xe7
-      nop.i 999;;
+       nop.m 0
+ (p10) fclass.m p7, p10 = f9, 0xe7
+       nop.i 0
 }
+;;
 
-// qnan snan inf norm     unorm 0 -+
-// 1    1    1   0        0     0 11
-// e                      3
-// X +-NAN, +-inf, ?        p9
-
 { .mfi
-      nop.m 999
-      fclass.m.unc  p9,p11 = f8, 0xe3
-      nop.i 999
+       mov GR_2P32 = 0x1001f
+       // (2) q0 = a*y0
+ (p6)  fma.s1 FR_Q0 = FR_ABS_A, FR_Y_INV0, f0
+       nop.i 0
+}
+{ .mfi
+       nop.m 0
+       // (3) e0 = 1 - b * y0
+ (p6)  fnma.s1 FR_E0 = FR_ABS_B, FR_Y_INV0, f1
+       nop.i 0
 }
+;;
 
-// |x| < |y|? Return x p8
+// set p9 if a +/-NAN, +/-inf
+{ .mfi
+       nop.m 0
+       fclass.m.unc p9, p11 = f8, 0xe3
+       nop.i 0
+}
+       // |a| < |b|? Return a, p8=1
 { .mfi
-      nop.m 999
-(p10)  fcmp.lt.unc.s1 p8,p0 = f6,f7
-      nop.i 999 ;;
+       nop.m 0
+ (p10) fcmp.lt.unc.s1 p8, p0 = FR_ABS_A, FR_ABS_B
+       nop.i 0
 }
+;;
 
-  { .mfi
-  mov r2=0x1001f
-  // (2) q0=a*y0
-  (p6) fma.s1 f13=f6,f10,f0
-  nop.i 0
-} { .mfi
-  nop.m 0
-  // (3) e0 = 1 - b * y0
-  (p6) fnma.s1 f12=f7,f10,f1
-  nop.i 0;;
+// set p7 if b +/-NAN, +/-inf, +/-0
+{ .mfi
+       nop.m 0
+       // pseudo-NaN ?
+ (p10) fclass.nm p7, p0 = f9, 0xff
+       nop.i 0
 }
+;;
 
-// Y +-NAN, +-inf, +-0?     p7
+// set p9 if a is +/-NaN, +/-Inf
+{ .mfi
+       nop.m 0
+ (p11) fclass.nm p9, p0 = f8, 0xff
+       nop.i 0
+}
 { .mfi
-      nop.m 999
-      // pseudo-NaN ?
-(p10)  fclass.nm  p7,p0 = f9, 0xff
-      nop.i 999
+       nop.m 0
+       // b denormal ? set D flag (if |a|<|b|)
+ (p8)  fnma.s0 FR_DFLAG = f9, f1, f9
+       nop.i 0
 }
+;;
 
-// qnan snan inf norm     unorm 0 -+
-// 1    1    1   0        0     0 11
-// e                      3
-// X +-NAN, +-inf, ?        p9
+{ .mfi
+       // FR_2P32 = 2^32
+       setf.exp FR_2P32 = GR_2P32
+       // (4) q1 = q0+e0*q0
+ (p6)  fma.s1 FR_Q1 = FR_E0, FR_Q0, FR_Q0
+       nop.i 0
+}
+{ .mfi
+       nop.m 0
+       // (5) e1 = e0 * e0 + 2^-34
+ (p6)  fma.s1 FR_E1 = FR_E0, FR_E0, FR_SMALLBIAS
+       nop.i 0
+}
+;;
 
 { .mfi
-      nop.m 999
-(p11)  fclass.nm  p9,p0 = f8, 0xff
-      nop.i 999;;
+       nop.m 0
+       // normalize a (if |a|<|b|)
+ (p8)  fma.s0 f8 = f8, f1, f0
+       nop.i 0
+}
+{ .bbb
+ (p9) br.cond.spnt FMOD_A_NAN_INF
+ (p7) br.cond.spnt FMOD_B_NAN_INF_ZERO
+       // if |a|<|b|, return
+ (p8) br.ret.spnt b0
 }
+;;
+
 
 { .mfi
-  nop.m 0
-  //  y denormal ? set D flag (if |x|<|y|)
-  (p8) fnma.s0 f10=f9,f1,f9
-  nop.i 0;;
+       nop.m 0
+       // (6) y1 = y0 + e0 * y0
+ (p6)  fma.s1 FR_Y_INV1 = FR_E0, FR_Y_INV0, FR_Y_INV0
+       nop.i 0
 }
+;;
 
+{ .mfi
+       nop.m 0
+       // a denormal ? set D flag
+       // b denormal ? set D flag
+       fcmp.eq.s0 p12,p0 = FR_ABS_A, FR_ABS_B
+       nop.i 0
+}
+{ .mfi
+       // set FR_ROUNDCONST = 1.25*2^{-24}
+       setf.s FR_ROUNDCONST = GR_ROUNDCONST
+       // (7) q2 = q1+e1*q1
+ (p6)  fma.s1 FR_Q = FR_Q1, FR_E1, FR_Q1
+       nop.i 0
+}
+;;
 
-{.mfi
-  nop.m 0
-  // normalize x (if |x|<|y|)
-  (p8) fma.s0 f8=f8,f1,f0
-  nop.i 0
+{ .mfi
+       nop.m 0
+       fmerge.s FR_B_SGN_A = f8, f9
+       nop.i 0
 }
-{.bbb
-  (p9) br.cond.spnt FMOD_X_NAN_INF
-  (p7) br.cond.spnt FMOD_Y_NAN_INF_ZERO
-  // if |x|<|y|, return
-  (p8) br.ret.spnt    b0;;
+{ .mfi
+       nop.m 0
+       // (8) y2 = y1 + e1 * y1
+ (p6)  fma.s1 FR_Y_INV = FR_E1, FR_Y_INV1, FR_Y_INV1
+       // set p6 = 0, p10 = 0
+       cmp.ne.and p6, p10 = r0, r0
 }
+;;
 
-  {.mfi
-  nop.m 0
-  // x denormal ? set D flag
-  fnma.s0 f32=f6,f1,f6
-  nop.i 0
+//   will compute integer quotient bits (24 bits per iteration)
+.align 32
+loop64:
+{ .mfi
+       nop.m 0
+       // compare q2, 2^32
+       fcmp.lt.unc.s1 p8, p7 = FR_Q, FR_2P32
+       nop.i 0
 }
-{.mfi
-  nop.m 0
-  // y denormal ? set D flag
-  fnma.s0 f33=f7,f1,f7
-  nop.i 0;;
+{ .mfi
+       nop.m 0
+       // will truncate quotient to integer, if exponent<32 (in advance)
+       fcvt.fx.trunc.s1 FR_QINT = FR_Q
+       nop.i 0
 }
+;;
 
-  {.mfi
-  // f15=2^32
-  setf.exp f15=r2
-  // (4) q1=q0+e0*q0
-  (p6) fma.s1 f13=f12,f13,f13
-  nop.i 0
+{ .mfi
+       nop.m 0
+       // if exponent>32 round quotient to single precision (perform in advance)
+       fma.s.s1 FR_QRND24 = FR_Q, f1, f0
+       nop.i 0
 }
+;;
+
 { .mfi
-  nop.m 0
-  // (5) e1 = e0 * e0 + 2^-34
-  (p6) fma.s1 f14=f12,f12,f11
-  nop.i 0;;
+       nop.m 0
+       // set FR_ROUNDCONST = sgn(a)
+ (p8)  fmerge.s FR_ROUNDCONST = f8, f1
+       nop.i 0
 }
-{.mlx
-  nop.m 0
-  movl r2=0x33a00000;;
+{ .mfi
+       nop.m 0
+       // normalize truncated quotient
+ (p8)  fcvt.xf FR_QRND24 = FR_QINT
+       nop.i 0
 }
+;;
+
 { .mfi
-  nop.m 0
-  // (6) y1 = y0 + e0 * y0
-  (p6) fma.s1 f10=f12,f10,f10
-  nop.i 0;;
+       nop.m 0
+       // calculate remainder (assuming FR_QRND24 = RZ(Q))
+ (p7)  fnma.s1 FR_E1 = FR_QRND24, FR_ABS_B, FR_ABS_A
+       nop.i 0
 }
-{.mfi
-  // set f12=1.25*2^{-24}
-  setf.s f12=r2
-  // (7) q2=q1+e1*q1
-  (p6) fma.s1 f13=f13,f14,f13
-  nop.i 0;;
+{ .mfi
+       nop.m 0
+       // also if exponent>32, round quotient to single precision
+       // and subtract 1 ulp: q = q-q*(1.25*2^{-24})
+ (p7)  fnma.s.s1 FR_QINT_Z = FR_QRND24, FR_ROUNDCONST, FR_QRND24
+       nop.i 0
 }
-{.mfi
-  nop.m 0
-  fmerge.s f9=f8,f9
-  nop.i 0
+;;
+
+{ .mfi
+       nop.m 0
+       // (p8) calculate remainder (82-bit format)
+ (p8)  fnma.s1 FR_QREM = FR_QRND24, FR_ABS_B, FR_ABS_A
+       nop.i 0
 }
 { .mfi
-  nop.m 0
-  // (8) y2 = y1 + e1 * y1
-  (p6) fma.s1 f10=f14,f10,f10
-  // set p6=0, p10=0
-  cmp.ne.and p6,p10=r0,r0;;
+       nop.m 0
+       // (p7) calculate remainder (assuming FR_QINT_Z = RZ(Q))
+ (p7)  fnma.s1 FR_ABS_A = FR_QINT_Z, FR_ABS_B, FR_ABS_A
+       nop.i 0
 }
+;;
 
+{ .mfi
+       nop.m 0
+       // Final iteration (p8): is FR_ABS_A the correct remainder 
+       // (quotient was not overestimated) ?
+ (p8)  fcmp.lt.unc.s1 p6, p10 = FR_QREM, f0
+       nop.i 0
+}
+;;
 
-.align 32
-loop64:
-  {.mfi
-  nop.m 0
-  // compare q2, 2^32
-  fcmp.lt.unc.s1 p8,p7=f13,f15
-  nop.i 0
-}
-  {.mfi
-  nop.m 0
-  // will truncate quotient to integer, if exponent<32 (in advance)
-  fcvt.fx.trunc.s1 f11=f13
-  nop.i 0;;
-}
-  {.mfi
-  nop.m 0
-  // if exponent>32, round quotient to single precision (perform in advance)
-  fma.s.s1 f13=f13,f1,f0
-  nop.i 0;;
-}
-
-
-  {.mfi
-  nop.m 0
-  // set f12=sgn(a)
-  (p8) fmerge.s f12=f8,f1
-  nop.i 0
-}
-  {.mfi
-  nop.m 0
-  // normalize truncated quotient
-  (p8) fcvt.xf f13=f11
-  nop.i 0;;
-}
-  { .mfi
-  nop.m 0
-  // calculate remainder (assuming f13=RZ(Q))
-  (p7) fnma.s1 f14=f13,f7,f6
-  nop.i 0
-}
-  {.mfi
-  nop.m 0
-  // also if exponent>32, round quotient to single precision
-  // and subtract 1 ulp: q=q-q*(1.25*2^{-24})
-  (p7) fnma.s.s1 f11=f13,f12,f13
-  nop.i 0;;
-}
-
-  {.mfi
-  nop.m 0
-  // (p8) calculate remainder (82-bit format)
-  (p8) fnma.s1 f11=f13,f7,f6
-  nop.i 0
-}
-  {.mfi
-  nop.m 0
-  // (p7) calculate remainder (assuming f11=RZ(Q))
-  (p7) fnma.s1 f6=f11,f7,f6
-  nop.i 0;;
-}
-
-
-  {.mfi
-  nop.m 0
-  // Final iteration (p8): is f6 the correct remainder (quotient was not overestimated) ?
-  (p8) fcmp.lt.unc.s1 p6,p10=f11,f0
-  nop.i 0;;
-}
-  {.mfi
-  nop.m 0
-  // get new quotient estimation: a'*y2
-  (p7) fma.s1 f13=f14,f10,f0
-  nop.i 0
-}
-  {.mfb
-  nop.m 0
-  // was f13=RZ(Q) ? (then new remainder f14>=0)
-  (p7) fcmp.lt.unc.s1 p7,p9=f14,f0
-  nop.b 0;;
-}
-
-
-.pred.rel "mutex",p6,p10
-  {.mfb
-  nop.m 0
-  // add b to estimated remainder (to cover the case when the quotient was overestimated)
-  // also set correct sign by using f9=|b|*sgn(a), f12=sgn(a)
-  (p6) fma.s0 f8=f11,f12,f9
-  nop.b 0
-}
-  {.mfb
-  nop.m 0
-  // set correct sign of result before returning: f12=sgn(a)
-  (p10) fma.s0 f8=f11,f12,f0
-  (p8) br.ret.sptk b0;;
-}
-  {.mfi
-  nop.m 0
-  // if f13!=RZ(Q), get alternative quotient estimation: a''*y2
-  (p7) fma.s1 f13=f6,f10,f0
-  nop.i 0
-}
-  {.mfb
-  nop.m 0
-  // if f14 was RZ(Q), set remainder to f14
-  (p9) mov f6=f14
-  br.cond.sptk loop64;;
+{ .mfi
+       nop.m 0
+       // get new quotient estimation: a'*y2
+ (p7)  fma.s1 FR_Q = FR_E1, FR_Y_INV, f0
+       nop.i 0
 }
+{ .mfb
+       nop.m 0
+       // was FR_Q = RZ(Q) ? (then new remainder FR_E1> = 0)
+ (p7)  fcmp.lt.unc.s1 p7, p9 = FR_E1, f0
+       nop.b 0
+}
+;;
 
+.pred.rel "mutex", p6, p10
+{ .mfb
+       nop.m 0
+       // add b to estimated remainder (to cover the case when the quotient was
+       // overestimated)
+       // also set correct sign by using 
+       // FR_B_SGN_A = |b|*sgn(a), FR_ROUNDCONST = sgn(a)
+ (p6)  fma.s0 f8 = FR_QREM, FR_ROUNDCONST, FR_B_SGN_A
+       nop.b 0
+}
+{ .mfb
+       nop.m 0
+       // set correct sign of result before returning: FR_ROUNDCONST = sgn(a)
+ (p10) fma.s0 f8 = FR_QREM, FR_ROUNDCONST, f0
+ (p8)  br.ret.sptk b0
+}
+;;
 
+{ .mfi
+       nop.m 0
+       // if f13! = RZ(Q), get alternative quotient estimation: a''*y2
+ (p7)  fma.s1 FR_Q = FR_ABS_A, FR_Y_INV, f0
+       nop.i 0
+}
+{ .mfb
+       nop.m 0
+       // if FR_E1 was RZ(Q), set remainder to FR_E1
+ (p9)  fma.s1 FR_ABS_A = FR_E1, f1, f0
+       br.cond.sptk loop64
+}
+;;
 
-FMOD_X_NAN_INF:
+FMOD_A_NAN_INF:
 
-// Y zero ?
-{.mfi
-  nop.m 0
-  fclass.m p10,p0=f8,0xc3     // Test x=nan
-  nop.i 0
+// b zero ?
+{ .mfi
+       nop.m 0
+       fclass.m p10, p0 = f8, 0xc3 // Test a = nan
+       nop.i 0
 }
-{.mfi
-  nop.m 0
-  fma.s1 f10=f9,f1,f0
-  nop.i 0;;
+{ .mfi
+       nop.m 0
+       fma.s1 FR_NORM_B = f9, f1, f0
+       nop.i 0
 }
+;;
 
-{.mfi
-  nop.m 0
-  fma.s0 f8=f8,f1,f0
-  nop.i 0
+{ .mfi
+       nop.m 0
+       fma.s0 f8 = f8, f1, f0
+       nop.i 0
 }
-{.mfi
-  nop.m 0
-(p10) fclass.m p10,p0=f9,0x07 // Test x=nan, and y=zero
-  nop.i 0;;
+{ .mfi
+       nop.m 0
+ (p10) fclass.m p10, p0 = f9, 0x07 // Test x = nan, and y = zero
+       nop.i 0
 }
-{.mfb
- nop.m 0
- fcmp.eq.unc.s1 p11,p0=f10,f0
-(p10) br.ret.spnt b0;;        // Exit with result=x if x=nan and y=zero
+;;
+
+{ .mfb
+       nop.m 0
+       fcmp.eq.unc.s1 p11, p0 = FR_NORM_B, f0
+ (p10) br.ret.spnt b0 // Exit with result = a if a = nan and b = zero
 }
-{.mib
-  nop.m 0
-  nop.i 0
-  // if Y zero
-  (p11) br.cond.spnt FMOD_Y_ZERO;;
+;;
+
+{ .mib
+       nop.m 0
+       nop.i 0
+       // if Y zero
+ (p11) br.cond.spnt FMOD_B_ZERO
 }
+;;
 
-// X infinity? Return QNAN indefinite
+// a= infinity? Return QNAN indefinite
 { .mfi
-     // set p7 t0 0
-     cmp.ne p7,p0=r0,r0
-     fclass.m.unc  p8,p9 = f8, 0x23
-     nop.i 999;;
+       // set p7 t0 0
+       cmp.ne p7, p0 = r0, r0
+       fclass.m.unc p8, p9 = f8, 0x23
+       nop.i 0
 }
-// Y NaN ?
-{.mfi
-     nop.m 999
-(p8) fclass.m p9,p8=f9,0xc3
-     nop.i 0;;
+;;
+
+// b NaN ?
+{ .mfi
+       nop.m 0
+ (p8)  fclass.m p9, p8 = f9, 0xc3
+       nop.i 0
 }
-// Y not pseudo-zero ? (r29 holds significand)
-{.mii
-     nop.m 999
-(p8) cmp.ne p7,p0=r29,r0
-     nop.i 0;;
+;;
+
+// b not pseudo-zero ? (GR_SIG_B holds significand)
+{ .mii
+       nop.m 0
+ (p8)  cmp.ne p7, p0 = GR_SIG_B, r0
+       nop.i 0
 }
-{.mfi
-    nop.m 999
-(p8)  frcpa.s0 f8,p0 = f8,f8
-    nop.i 0
+;;
+
+{ .mfi
+       nop.m 0
+ (p8)  frcpa.s0 f8, p0 = f8, f8
+       nop.i 0
 }
 { .mfi
-     nop.m 999
-    // also set Denormal flag if necessary
-(p7) fnma.s0 f9=f9,f1,f9
-     nop.i 999 ;;
+       nop.m 0
+       // also set Denormal flag if necessary
+ (p7)  fnma.s0 f9 = f9, f1, f9
+       nop.i 0
 }
+;;
 
 { .mfb
-      nop.m 999
-(p8)  fma.s0 f8=f8,f1,f0
-      nop.b 999 ;;
+       nop.m 0
+ (p8)  fma.s0 f8 = f8, f1, f0
+       nop.b 0
 }
+;;
 
 { .mfb
-      nop.m 999
-(p9)  frcpa.s0 f8,p7=f8,f9
-      br.ret.sptk    b0 ;;
+       nop.m 0
+ (p9)  frcpa.s0 f8, p7 = f8, f9
+       br.ret.sptk b0
 }
+;;
 
-
-FMOD_Y_NAN_INF_ZERO:
-// Y INF
+FMOD_B_NAN_INF_ZERO:
+// b INF
 { .mfi
-      nop.m 999
-      fclass.m.unc  p7,p0 = f9, 0x23
-      nop.i 999 ;;
+       nop.m 0
+       fclass.m.unc p7, p0 = f9, 0x23
+       nop.i 0
 }
+;;
 
 { .mfb
-      nop.m 999
-(p7)  fma.s0 f8=f8,f1,f0
-(p7)  br.ret.spnt    b0 ;;
+       nop.m 0
+ (p7)  fma.s0 f8 = f8, f1, f0
+ (p7)  br.ret.spnt b0
 }
+;;
 
-// Y NAN?
+// b NAN?
 { .mfi
-      nop.m 999
-      fclass.m.unc  p9,p10 = f9, 0xc3
-      nop.i 999 ;;
+       nop.m 0
+       fclass.m.unc p9, p10 = f9, 0xc3
+       nop.i 0
 }
+;;
+
 { .mfi
-      nop.m 999
-(p10)  fclass.nm  p9,p0 = f9, 0xff
-      nop.i 999 ;;
+       nop.m 0
+ (p10) fclass.nm p9, p0 = f9, 0xff
+       nop.i 0
 }
+;;
 
 { .mfb
-      nop.m 999
-(p9)  fma.s0 f8=f9,f1,f0
-(p9)  br.ret.spnt    b0 ;;
+       nop.m 0
+ (p9)  fma.s0 f8 = f9, f1, f0
+ (p9)  br.ret.spnt b0
 }
+;;
 
-FMOD_Y_ZERO:
+FMOD_B_ZERO:
 // Y zero? Must be zero at this point
 // because it is the only choice left.
 // Return QNAN indefinite
 
-{.mfi
-  nop.m 0
-  // set Invalid
-  frcpa.s0 f12,p0=f0,f0
-  nop.i 0
-}
-// X NAN?
 { .mfi
-      nop.m 999
-      fclass.m.unc  p9,p10 = f8, 0xc3
-      nop.i 999 ;;
+       nop.m 0
+       // set Invalid
+       frcpa.s0 FR_TMP, p0 = f0, f0
+       nop.i 0
 }
+;;
+
+// a NAN?
 { .mfi
-      nop.m 999
-(p10)  fclass.nm  p9,p10 = f8, 0xff
-      nop.i 999 ;;
+       nop.m 0
+       fclass.m.unc p9, p10 = f8, 0xc3
+       nop.i 0
 }
+;;
 
-{.mfi
- nop.m 999
- (p9) frcpa.s0 f11,p7=f8,f0
- nop.i 0;;
+{ .mfi
+       alloc GR_ARPFS = ar.pfs, 1, 4, 4, 0
+ (p10) fclass.nm p9, p10 = f8, 0xff
+       nop.i 0
 }
-
+;;
 
 { .mfi
-      nop.m 999
-(p10) frcpa.s0  f11,p7 = f9,f9
-      mov    GR_Parameter_TAG = 120 ;;
+       nop.m 0
+ (p9)  frcpa.s0 FR_TMP2, p7 = f8, f0
+       nop.i 0
 }
+;;
 
 { .mfi
-      nop.m 999
-      fmerge.s      f10 = f8, f8
-      nop.i 999
+       nop.m 0
+ (p10) frcpa.s0 FR_TMP2, p7 = f9, f9
+       mov GR_Parameter_TAG = 120
 }
+;;
 
+{ .mfi
+       nop.m 0
+       fmerge.s FR_X = f8, f8
+       nop.i 0
+}
 { .mfb
-      nop.m 999
-      fma.s0 f8=f11,f1,f0
-      br.sptk __libm_error_region;;
+       nop.m 0
+       fma.s0 f8 = FR_TMP2, f1, f0
+       br.sptk __libm_error_region
 }
+;;
 
 GLOBAL_IEEE754_END(fmodl)
 
-
 LOCAL_LIBM_ENTRY(__libm_error_region)
 .prologue
 { .mfi
-        add   GR_Parameter_Y=-32,sp             // Parameter 2 value
-        nop.f 0
-.save   ar.pfs,GR_SAVE_PFS
-        mov  GR_SAVE_PFS=ar.pfs                 // Save ar.pfs
+       add GR_Parameter_Y = -32, sp // Parameter 2 value
+       nop.f 0
+.save ar.pfs, GR_SAVE_PFS
+       mov GR_SAVE_PFS = ar.pfs     // Save ar.pfs
 }
 { .mfi
 .fframe 64
-        add sp=-64,sp                           // Create new stack
-        nop.f 0
-        mov GR_SAVE_GP=gp                       // Save gp
-};;
+       add sp = -64, sp             // Create new stack
+       nop.f 0
+       mov GR_SAVE_GP = gp          // Save gp
+}
+;;
+
 { .mmi
-        stfe [GR_Parameter_Y] = FR_Y,16         // Save Parameter 2 on stack
-        add GR_Parameter_X = 16,sp              // Parameter 1 address
-.save   b0, GR_SAVE_B0
-        mov GR_SAVE_B0=b0                       // Save b0
-};;
+       stfe [ GR_Parameter_Y ] = FR_Y, 16 // Save Parameter 2 on stack
+       add GR_Parameter_X = 16, sp  // Parameter 1 address
+.save b0, GR_SAVE_B0
+       mov GR_SAVE_B0 = b0          // Save b0
+}
+;;
+
 .body
 { .mib
-        stfe [GR_Parameter_X] = FR_X            // Store Parameter 1 on stack
-        add   GR_Parameter_RESULT = 0,GR_Parameter_Y
-    nop.b 0                                 // Parameter 3 address
+       stfe [ GR_Parameter_X ] = FR_X // Store Parameter 1 on stack
+       add GR_Parameter_RESULT = 0, GR_Parameter_Y
+       nop.b 0                      // Parameter 3 address
 }
 { .mib
-        stfe [GR_Parameter_Y] = FR_RESULT      // Store Parameter 3 on stack
-        add   GR_Parameter_Y = -16,GR_Parameter_Y
-        br.call.sptk b0=__libm_error_support#  // Call error handling function
-};;
+       stfe [ GR_Parameter_Y ] = FR_RESULT // Store Parameter 3 on stack
+       add GR_Parameter_Y = -16, GR_Parameter_Y
+       br.call.sptk b0 = __libm_error_support# // Call error handling function
+}
+;;
+
 { .mmi
-        nop.m 0
-        nop.m 0
-        add   GR_Parameter_RESULT = 48,sp
-};;
+       nop.m 0
+       nop.m 0
+       add GR_Parameter_RESULT = 48, sp
+}
+;;
+
 { .mmi
-        ldfe  f8 = [GR_Parameter_RESULT]       // Get return result off stack
+       ldfe f8 = [ GR_Parameter_RESULT ] // Get return result off stack
 .restore sp
-        add   sp = 64,sp                       // Restore stack pointer
-        mov   b0 = GR_SAVE_B0                  // Restore return address
-};;
+       add sp = 64, sp                   // Restore stack pointer
+       mov b0 = GR_SAVE_B0               // Restore return address
+}
+;;
+
 { .mib
-        mov   gp = GR_SAVE_GP                  // Restore gp
-        mov   ar.pfs = GR_SAVE_PFS             // Restore ar.pfs
-        br.ret.sptk     b0                     // Return
-};;
+       mov gp = GR_SAVE_GP               // Restore gp
+       mov ar.pfs = GR_SAVE_PFS          // Restore ar.pfs
+       br.ret.sptk b0                    // Return
+}
+;;
 
 LOCAL_LIBM_END(__libm_error_region)
 
-
-
-
-.type   __libm_error_support#,@function
+.type __libm_error_support#, @function
 .global __libm_error_support#
-
-