mirror of
https://github.com/AuxXxilium/linux_dsm_epyc7002.git
synced 2024-12-28 11:18:45 +07:00
e00d82d07f
Acked-by: Alan Cox <alan@redhat.com> Signed-off-by: Matt Waddel <Matt.Waddel@freescale.com> Cc: Roman Zippel <zippel@linux-m68k.org> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
269 lines
6.8 KiB
ArmAsm
269 lines
6.8 KiB
ArmAsm
|
|
|
| x_unfl.sa 3.4 7/1/91
|
|
|
|
|
| fpsp_unfl --- FPSP handler for underflow exception
|
|
|
|
|
| Trap disabled results
|
|
| For 881/2 compatibility, sw must denormalize the intermediate
|
|
| result, then store the result. Denormalization is accomplished
|
|
| by taking the intermediate result (which is always normalized) and
|
|
| shifting the mantissa right while incrementing the exponent until
|
|
| it is equal to the denormalized exponent for the destination
|
|
| format. After denormalization, the result is rounded to the
|
|
| destination format.
|
|
|
|
|
| Trap enabled results
|
|
| All trap disabled code applies. In addition the exceptional
|
|
| operand needs to made available to the user with a bias of $6000
|
|
| added to the exponent.
|
|
|
|
|
|
|
| Copyright (C) Motorola, Inc. 1990
|
|
| All Rights Reserved
|
|
|
|
|
| For details on the license for this file, please see the
|
|
| file, README, in this same directory.
|
|
|
|
X_UNFL: |idnt 2,1 | Motorola 040 Floating Point Software Package
|
|
|
|
|section 8
|
|
|
|
#include "fpsp.h"
|
|
|
|
|xref denorm
|
|
|xref round
|
|
|xref store
|
|
|xref g_rndpr
|
|
|xref g_opcls
|
|
|xref g_dfmtou
|
|
|xref real_unfl
|
|
|xref real_inex
|
|
|xref fpsp_done
|
|
|xref b1238_fix
|
|
|
|
.global fpsp_unfl
|
|
fpsp_unfl:
|
|
link %a6,#-LOCAL_SIZE
|
|
fsave -(%a7)
|
|
moveml %d0-%d1/%a0-%a1,USER_DA(%a6)
|
|
fmovemx %fp0-%fp3,USER_FP0(%a6)
|
|
fmoveml %fpcr/%fpsr/%fpiar,USER_FPCR(%a6)
|
|
|
|
|
|
|
bsrl unf_res |denormalize, round & store interm op
|
|
|
|
|
| If underflow exceptions are not enabled, check for inexact
|
|
| exception
|
|
|
|
|
btstb #unfl_bit,FPCR_ENABLE(%a6)
|
|
beqs ck_inex
|
|
|
|
btstb #E3,E_BYTE(%a6)
|
|
beqs no_e3_1
|
|
|
|
|
| Clear dirty bit on dest resister in the frame before branching
|
|
| to b1238_fix.
|
|
|
|
|
bfextu CMDREG3B(%a6){#6:#3},%d0 |get dest reg no
|
|
bclrb %d0,FPR_DIRTY_BITS(%a6) |clr dest dirty bit
|
|
bsrl b1238_fix |test for bug1238 case
|
|
movel USER_FPSR(%a6),FPSR_SHADOW(%a6)
|
|
orl #sx_mask,E_BYTE(%a6)
|
|
no_e3_1:
|
|
moveml USER_DA(%a6),%d0-%d1/%a0-%a1
|
|
fmovemx USER_FP0(%a6),%fp0-%fp3
|
|
fmoveml USER_FPCR(%a6),%fpcr/%fpsr/%fpiar
|
|
frestore (%a7)+
|
|
unlk %a6
|
|
bral real_unfl
|
|
|
|
|
| It is possible to have either inex2 or inex1 exceptions with the
|
|
| unfl. If the inex enable bit is set in the FPCR, and either
|
|
| inex2 or inex1 occurred, we must clean up and branch to the
|
|
| real inex handler.
|
|
|
|
|
ck_inex:
|
|
moveb FPCR_ENABLE(%a6),%d0
|
|
andb FPSR_EXCEPT(%a6),%d0
|
|
andib #0x3,%d0
|
|
beqs unfl_done
|
|
|
|
|
|
|
| Inexact enabled and reported, and we must take an inexact exception
|
|
|
|
|
take_inex:
|
|
btstb #E3,E_BYTE(%a6)
|
|
beqs no_e3_2
|
|
|
|
|
| Clear dirty bit on dest resister in the frame before branching
|
|
| to b1238_fix.
|
|
|
|
|
bfextu CMDREG3B(%a6){#6:#3},%d0 |get dest reg no
|
|
bclrb %d0,FPR_DIRTY_BITS(%a6) |clr dest dirty bit
|
|
bsrl b1238_fix |test for bug1238 case
|
|
movel USER_FPSR(%a6),FPSR_SHADOW(%a6)
|
|
orl #sx_mask,E_BYTE(%a6)
|
|
no_e3_2:
|
|
moveb #INEX_VEC,EXC_VEC+1(%a6)
|
|
moveml USER_DA(%a6),%d0-%d1/%a0-%a1
|
|
fmovemx USER_FP0(%a6),%fp0-%fp3
|
|
fmoveml USER_FPCR(%a6),%fpcr/%fpsr/%fpiar
|
|
frestore (%a7)+
|
|
unlk %a6
|
|
bral real_inex
|
|
|
|
unfl_done:
|
|
bclrb #E3,E_BYTE(%a6)
|
|
beqs e1_set |if set then branch
|
|
|
|
|
| Clear dirty bit on dest resister in the frame before branching
|
|
| to b1238_fix.
|
|
|
|
|
bfextu CMDREG3B(%a6){#6:#3},%d0 |get dest reg no
|
|
bclrb %d0,FPR_DIRTY_BITS(%a6) |clr dest dirty bit
|
|
bsrl b1238_fix |test for bug1238 case
|
|
movel USER_FPSR(%a6),FPSR_SHADOW(%a6)
|
|
orl #sx_mask,E_BYTE(%a6)
|
|
moveml USER_DA(%a6),%d0-%d1/%a0-%a1
|
|
fmovemx USER_FP0(%a6),%fp0-%fp3
|
|
fmoveml USER_FPCR(%a6),%fpcr/%fpsr/%fpiar
|
|
frestore (%a7)+
|
|
unlk %a6
|
|
bral fpsp_done
|
|
e1_set:
|
|
moveml USER_DA(%a6),%d0-%d1/%a0-%a1
|
|
fmovemx USER_FP0(%a6),%fp0-%fp3
|
|
fmoveml USER_FPCR(%a6),%fpcr/%fpsr/%fpiar
|
|
unlk %a6
|
|
bral fpsp_done
|
|
|
|
|
| unf_res --- underflow result calculation
|
|
|
|
|
unf_res:
|
|
bsrl g_rndpr |returns RND_PREC in d0 0=ext,
|
|
| ;1=sgl, 2=dbl
|
|
| ;we need the RND_PREC in the
|
|
| ;upper word for round
|
|
movew #0,-(%a7)
|
|
movew %d0,-(%a7) |copy RND_PREC to stack
|
|
|
|
|
|
|
|
| If the exception bit set is E3, the exceptional operand from the
|
|
| fpu is in WBTEMP; else it is in FPTEMP.
|
|
|
|
|
btstb #E3,E_BYTE(%a6)
|
|
beqs unf_E1
|
|
unf_E3:
|
|
lea WBTEMP(%a6),%a0 |a0 now points to operand
|
|
|
|
|
| Test for fsgldiv and fsglmul. If the inst was one of these, then
|
|
| force the precision to extended for the denorm routine. Use
|
|
| the user's precision for the round routine.
|
|
|
|
|
movew CMDREG3B(%a6),%d1 |check for fsgldiv or fsglmul
|
|
andiw #0x7f,%d1
|
|
cmpiw #0x30,%d1 |check for sgldiv
|
|
beqs unf_sgl
|
|
cmpiw #0x33,%d1 |check for sglmul
|
|
bnes unf_cont |if not, use fpcr prec in round
|
|
unf_sgl:
|
|
clrl %d0
|
|
movew #0x1,(%a7) |override g_rndpr precision
|
|
| ;force single
|
|
bras unf_cont
|
|
unf_E1:
|
|
lea FPTEMP(%a6),%a0 |a0 now points to operand
|
|
unf_cont:
|
|
bclrb #sign_bit,LOCAL_EX(%a0) |clear sign bit
|
|
sne LOCAL_SGN(%a0) |store sign
|
|
|
|
bsrl denorm |returns denorm, a0 points to it
|
|
|
|
|
| WARNING:
|
|
| ;d0 has guard,round sticky bit
|
|
| ;make sure that it is not corrupted
|
|
| ;before it reaches the round subroutine
|
|
| ;also ensure that a0 isn't corrupted
|
|
|
|
|
|
|
| Set up d1 for round subroutine d1 contains the PREC/MODE
|
|
| information respectively on upper/lower register halves.
|
|
|
|
|
bfextu FPCR_MODE(%a6){#2:#2},%d1 |get mode from FPCR
|
|
| ;mode in lower d1
|
|
addl (%a7)+,%d1 |merge PREC/MODE
|
|
|
|
|
| WARNING: a0 and d0 are assumed to be intact between the denorm and
|
|
| round subroutines. All code between these two subroutines
|
|
| must not corrupt a0 and d0.
|
|
|
|
|
|
|
|
| Perform Round
|
|
| Input: a0 points to input operand
|
|
| d0{31:29} has guard, round, sticky
|
|
| d1{01:00} has rounding mode
|
|
| d1{17:16} has rounding precision
|
|
| Output: a0 points to rounded operand
|
|
|
|
|
|
|
bsrl round |returns rounded denorm at (a0)
|
|
|
|
|
| Differentiate between store to memory vs. store to register
|
|
|
|
|
unf_store:
|
|
bsrl g_opcls |returns opclass in d0{2:0}
|
|
cmpib #0x3,%d0
|
|
bnes not_opc011
|
|
|
|
|
| At this point, a store to memory is pending
|
|
|
|
|
opc011:
|
|
bsrl g_dfmtou
|
|
tstb %d0
|
|
beqs ext_opc011 |If extended, do not subtract
|
|
| ;If destination format is sgl/dbl,
|
|
tstb LOCAL_HI(%a0) |If rounded result is normal,don't
|
|
| ;subtract
|
|
bmis ext_opc011
|
|
subqw #1,LOCAL_EX(%a0) |account for denorm bias vs.
|
|
| ;normalized bias
|
|
| ; normalized denormalized
|
|
| ;single $7f $7e
|
|
| ;double $3ff $3fe
|
|
|
|
|
ext_opc011:
|
|
bsrl store |stores to memory
|
|
bras unf_done |finish up
|
|
|
|
|
|
|
| At this point, a store to a float register is pending
|
|
|
|
|
not_opc011:
|
|
bsrl store |stores to float register
|
|
| ;a0 is not corrupted on a store to a
|
|
| ;float register.
|
|
|
|
|
| Set the condition codes according to result
|
|
|
|
|
tstl LOCAL_HI(%a0) |check upper mantissa
|
|
bnes ck_sgn
|
|
tstl LOCAL_LO(%a0) |check lower mantissa
|
|
bnes ck_sgn
|
|
bsetb #z_bit,FPSR_CC(%a6) |set condition codes if zero
|
|
ck_sgn:
|
|
btstb #sign_bit,LOCAL_EX(%a0) |check the sign bit
|
|
beqs unf_done
|
|
bsetb #neg_bit,FPSR_CC(%a6)
|
|
|
|
|
|
|
| Finish.
|
|
|
|
|
unf_done:
|
|
btstb #inex2_bit,FPSR_EXCEPT(%a6)
|
|
beqs no_aunfl
|
|
bsetb #aunfl_bit,FPSR_AEXCEPT(%a6)
|
|
no_aunfl:
|
|
rts
|
|
|
|
|end
|