metag: Optimised library functions
Add optimised library functions for metag.
Signed-off-by: James Hogan <james.hogan@imgtec.com>
diff --git a/arch/metag/lib/divsi3.S b/arch/metag/lib/divsi3.S
new file mode 100644
index 0000000..7c8a8ae
--- /dev/null
+++ b/arch/metag/lib/divsi3.S
@@ -0,0 +1,100 @@
+! Copyright (C) 2001, 2002, 2003, 2004, 2005, 2006, 2007
+! Imagination Technologies Ltd
+!
+! Integer divide routines.
+!
+
+ .text
+ .global ___udivsi3
+ .type ___udivsi3,function
+ .align 2
+___udivsi3:
+!!
+!! Since core is signed divide case, just set control variable
+!!
+ MOV D1Re0,D0Ar2 ! Au already in A1Ar1, Bu -> D1Re0
+ MOV D0Re0,#0 ! Result is 0
+ MOV D0Ar4,#0 ! Return positive result
+ B $LIDMCUStart
+ .size ___udivsi3,.-___udivsi3
+
+!!
+!! 32-bit division signed i/p - passed signed 32-bit numbers
+!!
+ .global ___divsi3
+ .type ___divsi3,function
+ .align 2
+___divsi3:
+!!
+!! A already in D1Ar1, B already in D0Ar2 -> make B abs(B)
+!!
+ MOV D1Re0,D0Ar2 ! A already in A1Ar1, B -> D1Re0
+ MOV D0Re0,#0 ! Result is 0
+ XOR D0Ar4,D1Ar1,D1Re0 ! D0Ar4 -ive if result is -ive
+ ABS D1Ar1,D1Ar1 ! abs(A) -> Au
+ ABS D1Re0,D1Re0 ! abs(B) -> Bu
+$LIDMCUStart:
+ CMP D1Ar1,D1Re0 ! Is ( Au > Bu )?
+ LSR D1Ar3,D1Ar1,#2 ! Calculate (Au & (~3)) >> 2
+ CMPHI D1Re0,D1Ar3 ! OR ( (Au & (~3)) <= (Bu << 2) )?
+ LSLSHI D1Ar3,D1Re0,#1 ! Buq = Bu << 1
+ BLS $LIDMCUSetup ! Yes: Do normal divide
+!!
+!! Quick divide setup can assume that CurBit only needs to start at 2
+!!
+$LIDMCQuick:
+ CMP D1Ar1,D1Ar3 ! ( A >= Buq )?
+ ADDCC D0Re0,D0Re0,#2 ! If yes result += 2
+ SUBCC D1Ar1,D1Ar1,D1Ar3 ! and A -= Buq
+ CMP D1Ar1,D1Re0 ! ( A >= Bu )?
+ ADDCC D0Re0,D0Re0,#1 ! If yes result += 1
+ SUBCC D1Ar1,D1Ar1,D1Re0 ! and A -= Bu
+ ORS D0Ar4,D0Ar4,D0Ar4 ! Return neg result?
+ NEG D0Ar2,D0Re0 ! Calulate neg result
+ MOVMI D0Re0,D0Ar2 ! Yes: Take neg result
+$LIDMCRet:
+ MOV PC,D1RtP
+!!
+!! Setup for general unsigned divide code
+!!
+!! D0Re0 is used to form the result, already set to Zero
+!! D1Re0 is the input Bu value, this gets trashed
+!! D0Ar6 is curbit which is set to 1 at the start and shifted up
+!! D0Ar4 is negative if we should return a negative result
+!! D1Ar1 is the input Au value, eventually this holds the remainder
+!!
+$LIDMCUSetup:
+ CMP D1Ar1,D1Re0 ! Is ( Au < Bu )?
+ MOV D0Ar6,#1 ! Set curbit to 1
+ BCS $LIDMCRet ! Yes: Return 0 remainder Au
+!!
+!! Calculate alignment using FFB instruction
+!!
+ FFB D1Ar5,D1Ar1 ! Find first bit of Au
+ ANDN D1Ar5,D1Ar5,#31 ! Handle exceptional case.
+ ORN D1Ar5,D1Ar5,#31 ! if N bit set, set to 31
+ FFB D1Ar3,D1Re0 ! Find first bit of Bu
+ ANDN D1Ar3,D1Ar3,#31 ! Handle exceptional case.
+ ORN D1Ar3,D1Ar3,#31 ! if N bit set, set to 31
+ SUBS D1Ar3,D1Ar5,D1Ar3 ! calculate diff, ffbA - ffbB
+ MOV D0Ar2,D1Ar3 ! copy into bank 0
+ LSLGT D1Re0,D1Re0,D1Ar3 ! ( > 0) ? left shift B
+ LSLGT D0Ar6,D0Ar6,D0Ar2 ! ( > 0) ? left shift curbit
+!!
+!! Now we start the divide proper, logic is
+!!
+!! if ( A >= B ) add curbit to result and subtract B from A
+!! shift curbit and B down by 1 in either case
+!!
+$LIDMCLoop:
+ CMP D1Ar1, D1Re0 ! ( A >= B )?
+ ADDCC D0Re0, D0Re0, D0Ar6 ! If yes result += curbit
+ SUBCC D1Ar1, D1Ar1, D1Re0 ! and A -= B
+ LSRS D0Ar6, D0Ar6, #1 ! Shift down curbit, is it zero?
+ LSR D1Re0, D1Re0, #1 ! Shift down B
+ BNZ $LIDMCLoop ! Was single bit in curbit lost?
+ ORS D0Ar4,D0Ar4,D0Ar4 ! Return neg result?
+ NEG D0Ar2,D0Re0 ! Calulate neg result
+ MOVMI D0Re0,D0Ar2 ! Yes: Take neg result
+ MOV PC,D1RtP
+ .size ___divsi3,.-___divsi3