blob: 843e88b3cab84c80c1f1fffd28d59c14c9e33073 [file] [log] [blame]
//===----------------------Hexagon builtin routine ------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
#define Q6_ALIAS(TAG) .global __qdsp_##TAG ; .set __qdsp_##TAG, __hexagon_##TAG
#define END(TAG) .size TAG,.-TAG
// Double Precision Multiply
#define A r1:0
#define AH r1
#define AL r0
#define B r3:2
#define BH r3
#define BL r2
#define C r5:4
#define CH r5
#define CL r4
#define BTMP r15:14
#define BTMPH r15
#define BTMPL r14
#define ATMP r13:12
#define ATMPH r13
#define ATMPL r12
#define CTMP r11:10
#define CTMPH r11
#define CTMPL r10
#define PP_LL r9:8
#define PP_LL_H r9
#define PP_LL_L r8
#define PP_ODD r7:6
#define PP_ODD_H r7
#define PP_ODD_L r6
#define PP_HH r17:16
#define PP_HH_H r17
#define PP_HH_L r16
#define EXPA r18
#define EXPB r19
#define EXPBA r19:18
#define TMP r28
#define P_TMP p0
#define PROD_NEG p3
#define EXACT p2
#define SWAP p1
#define MANTBITS 52
#define HI_MANTBITS 20
#define EXPBITS 11
#define BIAS 1023
#define STACKSPACE 32
#define ADJUST 4
#define FUDGE 7
#define FUDGE2 3
#ifndef SR_ROUND_OFF
#define SR_ROUND_OFF 22
#endif
// First, classify for normal values, and abort if abnormal
//
// Next, unpack mantissa into 0x1000_0000_0000_0000 + mant<<8
//
// Since we know that the 2 MSBs of the H registers is zero, we should never carry
// the partial products that involve the H registers
//
// Try to buy X slots, at the expense of latency if needed
//
// We will have PP_HH with the upper bits of the product, PP_LL with the lower
// PP_HH can have a maximum of 0x03FF_FFFF_FFFF_FFFF or thereabouts
// PP_HH can have a minimum of 0x0100_0000_0000_0000
//
// 0x0100_0000_0000_0000 has EXP of EXPA+EXPB-BIAS
//
// We need to align CTMP.
// If CTMP >> PP, convert PP to 64 bit with sticky, align CTMP, and follow normal add
// If CTMP << PP align CTMP and add 128 bits. Then compute sticky
// If CTMP ~= PP, align CTMP and add 128 bits. May have massive cancellation.
//
// Convert partial product and CTMP to 2's complement prior to addition
//
// After we add, we need to normalize into upper 64 bits, then compute sticky.
.text
.global __hexagon_fmadf4
.type __hexagon_fmadf4,@function
.global __hexagon_fmadf5
.type __hexagon_fmadf5,@function
Q6_ALIAS(fmadf5)
.p2align 5
__hexagon_fmadf4:
__hexagon_fmadf5:
.Lfma_begin:
{
P_TMP = dfclass(A,#2)
P_TMP = dfclass(B,#2)
ATMP = #0
BTMP = #0
}
{
ATMP = insert(A,#MANTBITS,#EXPBITS-3)
BTMP = insert(B,#MANTBITS,#EXPBITS-3)
PP_ODD_H = ##0x10000000
allocframe(#STACKSPACE)
}
{
PP_LL = mpyu(ATMPL,BTMPL)
if (!P_TMP) jump .Lfma_abnormal_ab
ATMPH = or(ATMPH,PP_ODD_H)
BTMPH = or(BTMPH,PP_ODD_H)
}
{
P_TMP = dfclass(C,#2)
if (!P_TMP.new) jump:nt .Lfma_abnormal_c
CTMP = combine(PP_ODD_H,#0)
PP_ODD = combine(#0,PP_LL_H)
}
.Lfma_abnormal_c_restart:
{
PP_ODD += mpyu(BTMPL,ATMPH)
CTMP = insert(C,#MANTBITS,#EXPBITS-3)
memd(r29+#0) = PP_HH
memd(r29+#8) = EXPBA
}
{
PP_ODD += mpyu(ATMPL,BTMPH)
EXPBA = neg(CTMP)
P_TMP = cmp.gt(CH,#-1)
TMP = xor(AH,BH)
}
{
EXPA = extractu(AH,#EXPBITS,#HI_MANTBITS)
EXPB = extractu(BH,#EXPBITS,#HI_MANTBITS)
PP_HH = combine(#0,PP_ODD_H)
if (!P_TMP) CTMP = EXPBA
}
{
PP_HH += mpyu(ATMPH,BTMPH)
PP_LL = combine(PP_ODD_L,PP_LL_L)
#undef PP_ODD
#undef PP_ODD_H
#undef PP_ODD_L
#undef ATMP
#undef ATMPL
#undef ATMPH
#undef BTMP
#undef BTMPL
#undef BTMPH
#define RIGHTLEFTSHIFT r13:12
#define RIGHTSHIFT r13
#define LEFTSHIFT r12
EXPA = add(EXPA,EXPB)
#undef EXPB
#undef EXPBA
#define EXPC r19
#define EXPCA r19:18
EXPC = extractu(CH,#EXPBITS,#HI_MANTBITS)
}
// PP_HH:PP_LL now has product
// CTMP is negated
// EXPA,B,C are extracted
// We need to negate PP
// Since we will be adding with carry later, if we need to negate,
// just invert all bits now, which we can do conditionally and in parallel
#define PP_HH_TMP r15:14
#define PP_LL_TMP r7:6
{
EXPA = add(EXPA,#-BIAS+(ADJUST))
PROD_NEG = !cmp.gt(TMP,#-1)
PP_LL_TMP = #0
PP_HH_TMP = #0
}
{
PP_LL_TMP = sub(PP_LL_TMP,PP_LL,PROD_NEG):carry
P_TMP = !cmp.gt(TMP,#-1)
SWAP = cmp.gt(EXPC,EXPA) // If C >> PP
if (SWAP.new) EXPCA = combine(EXPA,EXPC)
}
{
PP_HH_TMP = sub(PP_HH_TMP,PP_HH,PROD_NEG):carry
if (P_TMP) PP_LL = PP_LL_TMP
#undef PP_LL_TMP
#define CTMP2 r7:6
#define CTMP2H r7
#define CTMP2L r6
CTMP2 = #0
EXPC = sub(EXPA,EXPC)
}
{
if (P_TMP) PP_HH = PP_HH_TMP
P_TMP = cmp.gt(EXPC,#63)
if (SWAP) PP_LL = CTMP2
if (SWAP) CTMP2 = PP_LL
}
#undef PP_HH_TMP
//#define ONE r15:14
//#define S_ONE r14
#define ZERO r15:14
#define S_ZERO r15
#undef PROD_NEG
#define P_CARRY p3
{
if (SWAP) PP_HH = CTMP // Swap C and PP
if (SWAP) CTMP = PP_HH
if (P_TMP) EXPC = add(EXPC,#-64)
TMP = #63
}
{
// If diff > 63, pre-shift-right by 64...
if (P_TMP) CTMP2 = CTMP
TMP = asr(CTMPH,#31)
RIGHTSHIFT = min(EXPC,TMP)
LEFTSHIFT = #0
}
#undef C
#undef CH
#undef CL
#define STICKIES r5:4
#define STICKIESH r5
#define STICKIESL r4
{
if (P_TMP) CTMP = combine(TMP,TMP) // sign extension of pre-shift-right-64
STICKIES = extract(CTMP2,RIGHTLEFTSHIFT)
CTMP2 = lsr(CTMP2,RIGHTSHIFT)
LEFTSHIFT = sub(#64,RIGHTSHIFT)
}
{
ZERO = #0
TMP = #-2
CTMP2 |= lsl(CTMP,LEFTSHIFT)
CTMP = asr(CTMP,RIGHTSHIFT)
}
{
P_CARRY = cmp.gtu(STICKIES,ZERO) // If we have sticky bits from C shift
if (P_CARRY.new) CTMP2L = and(CTMP2L,TMP) // make sure adding 1 == OR
#undef ZERO
#define ONE r15:14
#define S_ONE r14
ONE = #1
STICKIES = #0
}
{
PP_LL = add(CTMP2,PP_LL,P_CARRY):carry // use the carry to add the sticky
}
{
PP_HH = add(CTMP,PP_HH,P_CARRY):carry
TMP = #62
}
// PP_HH:PP_LL now holds the sum
// We may need to normalize left, up to ??? bits.
//
// I think that if we have massive cancellation, the range we normalize by
// is still limited
{
LEFTSHIFT = add(clb(PP_HH),#-2)
if (!cmp.eq(LEFTSHIFT.new,TMP)) jump:t 1f // all sign bits?
}
// We had all sign bits, shift left by 62.
{
CTMP = extractu(PP_LL,#62,#2)
PP_LL = asl(PP_LL,#62)
EXPA = add(EXPA,#-62) // And adjust exponent of result
}
{
PP_HH = insert(CTMP,#62,#0) // Then shift 63
}
{
LEFTSHIFT = add(clb(PP_HH),#-2)
}
.falign
1:
{
CTMP = asl(PP_HH,LEFTSHIFT)
STICKIES |= asl(PP_LL,LEFTSHIFT)
RIGHTSHIFT = sub(#64,LEFTSHIFT)
EXPA = sub(EXPA,LEFTSHIFT)
}
{
CTMP |= lsr(PP_LL,RIGHTSHIFT)
EXACT = cmp.gtu(ONE,STICKIES)
TMP = #BIAS+BIAS-2
}
{
if (!EXACT) CTMPL = or(CTMPL,S_ONE)
// If EXPA is overflow/underflow, jump to ovf_unf
P_TMP = !cmp.gt(EXPA,TMP)
P_TMP = cmp.gt(EXPA,#1)
if (!P_TMP.new) jump:nt .Lfma_ovf_unf
}
{
// XXX: FIXME: should PP_HH for check of zero be CTMP?
P_TMP = cmp.gtu(ONE,CTMP) // is result true zero?
A = convert_d2df(CTMP)
EXPA = add(EXPA,#-BIAS-60)
PP_HH = memd(r29+#0)
}
{
AH += asl(EXPA,#HI_MANTBITS)
EXPCA = memd(r29+#8)
if (!P_TMP) dealloc_return // not zero, return
}
.Ladd_yields_zero:
// We had full cancellation. Return +/- zero (-0 when round-down)
{
TMP = USR
A = #0
}
{
TMP = extractu(TMP,#2,#SR_ROUND_OFF)
PP_HH = memd(r29+#0)
EXPCA = memd(r29+#8)
}
{
p0 = cmp.eq(TMP,#2)
if (p0.new) AH = ##0x80000000
dealloc_return
}
#undef RIGHTLEFTSHIFT
#undef RIGHTSHIFT
#undef LEFTSHIFT
#undef CTMP2
#undef CTMP2H
#undef CTMP2L
.Lfma_ovf_unf:
{
p0 = cmp.gtu(ONE,CTMP)
if (p0.new) jump:nt .Ladd_yields_zero
}
{
A = convert_d2df(CTMP)
EXPA = add(EXPA,#-BIAS-60)
TMP = EXPA
}
#define NEW_EXPB r7
#define NEW_EXPA r6
{
AH += asl(EXPA,#HI_MANTBITS)
NEW_EXPB = extractu(AH,#EXPBITS,#HI_MANTBITS)
}
{
NEW_EXPA = add(EXPA,NEW_EXPB)
PP_HH = memd(r29+#0)
EXPCA = memd(r29+#8)
#undef PP_HH
#undef PP_HH_H
#undef PP_HH_L
#undef EXPCA
#undef EXPC
#undef EXPA
#undef PP_LL
#undef PP_LL_H
#undef PP_LL_L
#define EXPA r6
#define EXPB r7
#define EXPBA r7:6
#define ATMP r9:8
#define ATMPH r9
#define ATMPL r8
#undef NEW_EXPB
#undef NEW_EXPA
ATMP = abs(CTMP)
}
{
p0 = cmp.gt(EXPA,##BIAS+BIAS)
if (p0.new) jump:nt .Lfma_ovf
}
{
p0 = cmp.gt(EXPA,#0)
if (p0.new) jump:nt .Lpossible_unf
}
{
// TMP has original EXPA.
// ATMP is corresponding value
// Normalize ATMP and shift right to correct location
EXPB = add(clb(ATMP),#-2) // Amount to left shift to normalize
EXPA = sub(#1+5,TMP) // Amount to right shift to denormalize
p3 = cmp.gt(CTMPH,#-1)
}
// Underflow
// We know that the infinte range exponent should be EXPA
// CTMP is 2's complement, ATMP is abs(CTMP)
{
EXPA = add(EXPA,EXPB) // how much to shift back right
ATMP = asl(ATMP,EXPB) // shift left
AH = USR
TMP = #63
}
{
EXPB = min(EXPA,TMP)
EXPA = #0
AL = #0x0030
}
{
B = extractu(ATMP,EXPBA)
ATMP = asr(ATMP,EXPB)
}
{
p0 = cmp.gtu(ONE,B)
if (!p0.new) ATMPL = or(ATMPL,S_ONE)
ATMPH = setbit(ATMPH,#HI_MANTBITS+FUDGE2)
}
{
CTMP = neg(ATMP)
p1 = bitsclr(ATMPL,#(1<<FUDGE2)-1)
if (!p1.new) AH = or(AH,AL)
B = #0
}
{
if (p3) CTMP = ATMP
USR = AH
TMP = #-BIAS-(MANTBITS+FUDGE2)
}
{
A = convert_d2df(CTMP)
}
{
AH += asl(TMP,#HI_MANTBITS)
dealloc_return
}
.Lpossible_unf:
{
TMP = ##0x7fefffff
ATMP = abs(CTMP)
}
{
p0 = cmp.eq(AL,#0)
p0 = bitsclr(AH,TMP)
if (!p0.new) dealloc_return:t
TMP = #0x7fff
}
{
p0 = bitsset(ATMPH,TMP)
BH = USR
BL = #0x0030
}
{
if (p0) BH = or(BH,BL)
}
{
USR = BH
}
{
p0 = dfcmp.eq(A,A)
dealloc_return
}
.Lfma_ovf:
{
TMP = USR
CTMP = combine(##0x7fefffff,#-1)
A = CTMP
}
{
ATMP = combine(##0x7ff00000,#0)
BH = extractu(TMP,#2,#SR_ROUND_OFF)
TMP = or(TMP,#0x28)
}
{
USR = TMP
BH ^= lsr(AH,#31)
BL = BH
}
{
p0 = !cmp.eq(BL,#1)
p0 = !cmp.eq(BH,#2)
}
{
p0 = dfcmp.eq(ATMP,ATMP)
if (p0.new) CTMP = ATMP
}
{
A = insert(CTMP,#63,#0)
dealloc_return
}
#undef CTMP
#undef CTMPH
#undef CTMPL
#define BTMP r11:10
#define BTMPH r11
#define BTMPL r10
#undef STICKIES
#undef STICKIESH
#undef STICKIESL
#define C r5:4
#define CH r5
#define CL r4
.Lfma_abnormal_ab:
{
ATMP = extractu(A,#63,#0)
BTMP = extractu(B,#63,#0)
deallocframe
}
{
p3 = cmp.gtu(ATMP,BTMP)
if (!p3.new) A = B // sort values
if (!p3.new) B = A
}
{
p0 = dfclass(A,#0x0f) // A NaN?
if (!p0.new) jump:nt .Lnan
if (!p3) ATMP = BTMP
if (!p3) BTMP = ATMP
}
{
p1 = dfclass(A,#0x08) // A is infinity
p1 = dfclass(B,#0x0e) // B is nonzero
}
{
p0 = dfclass(A,#0x08) // a is inf
p0 = dfclass(B,#0x01) // b is zero
}
{
if (p1) jump .Lab_inf
p2 = dfclass(B,#0x01)
}
{
if (p0) jump .Linvalid
if (p2) jump .Lab_true_zero
TMP = ##0x7c000000
}
// We are left with a normal or subnormal times a subnormal, A > B
// If A and B are both very small, we will go to a single sticky bit; replace
// A and B lower 63 bits with 0x0010_0000_0000_0000, which yields equivalent results
// if A and B might multiply to something bigger, decrease A exp and increase B exp
// and start over
{
p0 = bitsclr(AH,TMP)
if (p0.new) jump:nt .Lfma_ab_tiny
}
{
TMP = add(clb(BTMP),#-EXPBITS)
}
{
BTMP = asl(BTMP,TMP)
}
{
B = insert(BTMP,#63,#0)
AH -= asl(TMP,#HI_MANTBITS)
}
jump .Lfma_begin
.Lfma_ab_tiny:
ATMP = combine(##0x00100000,#0)
{
A = insert(ATMP,#63,#0)
B = insert(ATMP,#63,#0)
}
jump .Lfma_begin
.Lab_inf:
{
B = lsr(B,#63)
p0 = dfclass(C,#0x10)
}
{
A ^= asl(B,#63)
if (p0) jump .Lnan
}
{
p1 = dfclass(C,#0x08)
if (p1.new) jump:nt .Lfma_inf_plus_inf
}
// A*B is +/- inf, C is finite. Return A
{
jumpr r31
}
.falign
.Lfma_inf_plus_inf:
{ // adding infinities of different signs is invalid
p0 = dfcmp.eq(A,C)
if (!p0.new) jump:nt .Linvalid
}
{
jumpr r31
}
.Lnan:
{
p0 = dfclass(B,#0x10)
p1 = dfclass(C,#0x10)
if (!p0.new) B = A
if (!p1.new) C = A
}
{ // find sNaNs
BH = convert_df2sf(B)
BL = convert_df2sf(C)
}
{
BH = convert_df2sf(A)
A = #-1
jumpr r31
}
.Linvalid:
{
TMP = ##0x7f800001 // sp snan
}
{
A = convert_sf2df(TMP)
jumpr r31
}
.Lab_true_zero:
// B is zero, A is finite number
{
p0 = dfclass(C,#0x10)
if (p0.new) jump:nt .Lnan
if (p0.new) A = C
}
{
p0 = dfcmp.eq(B,C) // is C also zero?
AH = lsr(AH,#31) // get sign
}
{
BH ^= asl(AH,#31) // form correctly signed zero in B
if (!p0) A = C // If C is not zero, return C
if (!p0) jumpr r31
}
// B has correctly signed zero, C is also zero
.Lzero_plus_zero:
{
p0 = cmp.eq(B,C) // yes, scalar equals. +0++0 or -0+-0
if (p0.new) jumpr:t r31
A = B
}
{
TMP = USR
}
{
TMP = extractu(TMP,#2,#SR_ROUND_OFF)
A = #0
}
{
p0 = cmp.eq(TMP,#2)
if (p0.new) AH = ##0x80000000
jumpr r31
}
#undef BTMP
#undef BTMPH
#undef BTMPL
#define CTMP r11:10
.falign
.Lfma_abnormal_c:
// We know that AB is normal * normal
// C is not normal: zero, subnormal, inf, or NaN.
{
p0 = dfclass(C,#0x10) // is C NaN?
if (p0.new) jump:nt .Lnan
if (p0.new) A = C // move NaN to A
deallocframe
}
{
p0 = dfclass(C,#0x08) // is C inf?
if (p0.new) A = C // return C
if (p0.new) jumpr:nt r31
}
// zero or subnormal
// If we have a zero, and we know AB is normal*normal, we can just call normal multiply
{
p0 = dfclass(C,#0x01) // is C zero?
if (p0.new) jump:nt __hexagon_muldf3
TMP = #1
}
// Left with: subnormal
// Adjust C and jump back to restart
{
allocframe(#STACKSPACE) // oops, deallocated above, re-allocate frame
CTMP = #0
CH = insert(TMP,#EXPBITS,#HI_MANTBITS)
jump .Lfma_abnormal_c_restart
}
END(fma)