diff options
Diffstat (limited to 'sysdeps/sparc/sparc32/sparcv8')
-rw-r--r-- | sysdeps/sparc/sparc32/sparcv8/addmul_1.S | 85 | ||||
-rw-r--r-- | sysdeps/sparc/sparc32/sparcv8/dotmul.S | 13 | ||||
-rw-r--r-- | sysdeps/sparc/sparc32/sparcv8/mul_1.S | 76 | ||||
-rw-r--r-- | sysdeps/sparc/sparc32/sparcv8/rem.S | 18 | ||||
-rw-r--r-- | sysdeps/sparc/sparc32/sparcv8/sdiv.S | 14 | ||||
-rw-r--r-- | sysdeps/sparc/sparc32/sparcv8/submul_1.S | 14 | ||||
-rw-r--r-- | sysdeps/sparc/sparc32/sparcv8/udiv.S | 13 | ||||
-rw-r--r-- | sysdeps/sparc/sparc32/sparcv8/udiv_qrnnd.S | 157 | ||||
-rw-r--r-- | sysdeps/sparc/sparc32/sparcv8/umul.S | 13 | ||||
-rw-r--r-- | sysdeps/sparc/sparc32/sparcv8/urem.S | 15 |
10 files changed, 266 insertions, 152 deletions
diff --git a/sysdeps/sparc/sparc32/sparcv8/addmul_1.S b/sysdeps/sparc/sparc32/sparcv8/addmul_1.S index fb9ea7cf0e..7b871b8e8b 100644 --- a/sysdeps/sparc/sparc32/sparcv8/addmul_1.S +++ b/sysdeps/sparc/sparc32/sparcv8/addmul_1.S @@ -27,54 +27,52 @@ ! size o2 ! s2_limb o3 -#include "sysdep.h" +#include <sysdep.h> -.text - .align 4 - .global C_SYMBOL_NAME(__mpn_addmul_1) -C_SYMBOL_NAME(__mpn_addmul_1): - orcc %g0,%g0,%g2 +ENTRY(__mpn_addmul_1) ld [%o1+0],%o4 ! 1 - sll %o2,4,%g1 - and %g1,(4-1)<<4,%g1 -#if PIC + orcc %g0,%g0,%g2 mov %o7,%g4 ! Save return address register - call 1f - add %o7,LL-1f,%g3 -1: mov %g4,%o7 ! Restore return address register -#else - sethi %hi(LL),%g3 - or %g3,%lo(LL),%g3 -#endif - jmp %g3+%g1 - nop -LL: -LL00: add %o0,-4,%o0 - b Loop00 /* 4, 8, 12, ... */ - add %o1,-4,%o1 - nop -LL01: b Loop01 /* 1, 5, 9, ... */ + and %g1,(4-1)<<4,%g1 +1: call 2f + add %o7,3f-1b,%g3 +2: jmp %g3+%g1 + mov %g4,%o7 ! Restore return address register + + .align 4 +3: +LOC(00): + add %o0,-4,%o0 + b LOC(loop00) /* 4, 8, 12, ... */ + add %o1,-4,%o1 nop +LOC(01): + b LOC(loop01) /* 1, 5, 9, ... */ + nop nop nop -LL10: add %o0,-12,%o0 /* 2, 6, 10, ... */ - b Loop10 - add %o1,4,%o1 +LOC(10): + add %o0,-12,%o0 /* 2, 6, 10, ... */ + b LOC(loop10) + add %o1,4,%o1 nop -LL11: add %o0,-8,%o0 /* 3, 7, 11, ... */ - b Loop11 - add %o1,-8,%o1 +LOC(11): + add %o0,-8,%o0 /* 3, 7, 11, ... */ + b LOC(loop11) + add %o1,-8,%o1 nop -1: addcc %g3,%g2,%g3 ! 1 +LOC(loop): + addcc %g3,%g2,%g3 ! 1 ld [%o1+4],%o4 ! 2 rd %y,%g2 ! 1 addx %g0,%g2,%g2 ld [%o0+0],%g1 ! 2 addcc %g1,%g3,%g3 st %g3,[%o0+0] ! 1 -Loop00: umul %o4,%o3,%g3 ! 2 +LOC(loop00): + umul %o4,%o3,%g3 ! 2 ld [%o0+4],%g1 ! 2 addxcc %g3,%g2,%g3 ! 2 ld [%o1+8],%o4 ! 3 @@ -83,7 +81,8 @@ Loop00: umul %o4,%o3,%g3 ! 2 nop addcc %g1,%g3,%g3 st %g3,[%o0+4] ! 2 -Loop11: umul %o4,%o3,%g3 ! 3 +LOC(loop11): + umul %o4,%o3,%g3 ! 3 addxcc %g3,%g2,%g3 ! 3 ld [%o1+12],%o4 ! 4 rd %y,%g2 ! 3 @@ -92,7 +91,8 @@ Loop11: umul %o4,%o3,%g3 ! 3 ld [%o0+8],%g1 ! 2 addcc %g1,%g3,%g3 st %g3,[%o0+8] ! 3 -Loop10: umul %o4,%o3,%g3 ! 4 +LOC(loop10): + umul %o4,%o3,%g3 ! 4 addxcc %g3,%g2,%g3 ! 4 ld [%o1+0],%o4 ! 1 rd %y,%g2 ! 4 @@ -102,9 +102,10 @@ Loop10: umul %o4,%o3,%g3 ! 4 st %g3,[%o0+12] ! 4 add %o0,16,%o0 addx %g0,%g2,%g2 -Loop01: addcc %o2,-4,%o2 - bg 1b - umul %o4,%o3,%g3 ! 1 +LOC(loop01): + addcc %o2,-4,%o2 + bg LOC(loop) + umul %o4,%o3,%g3 ! 1 addcc %g3,%g2,%g3 ! 4 rd %y,%g2 ! 4 @@ -112,13 +113,7 @@ Loop01: addcc %o2,-4,%o2 ld [%o0+0],%g1 ! 2 addcc %g1,%g3,%g3 st %g3,[%o0+0] ! 4 - addx %g0,%g2,%o0 - retl - nop - - -! umul, ld, addxcc, rd, st - -! umul, ld, addxcc, rd, ld, addcc, st, addx + addx %g0,%g2,%o0 +END(__mpn_addmul_1) diff --git a/sysdeps/sparc/sparc32/sparcv8/dotmul.S b/sysdeps/sparc/sparc32/sparcv8/dotmul.S new file mode 100644 index 0000000000..9b20cc3684 --- /dev/null +++ b/sysdeps/sparc/sparc32/sparcv8/dotmul.S @@ -0,0 +1,13 @@ +/* + * Sparc v8 has multiply. + */ + +#include <sysdep.h> + +ENTRY(.mul) + + smul %o0, %o1, %o0 + retl + rd %y, %o1 + +END(.mul) diff --git a/sysdeps/sparc/sparc32/sparcv8/mul_1.S b/sysdeps/sparc/sparc32/sparcv8/mul_1.S index b641feb453..801e224c15 100644 --- a/sysdeps/sparc/sparc32/sparcv8/mul_1.S +++ b/sysdeps/sparc/sparc32/sparcv8/mul_1.S @@ -27,73 +27,77 @@ ! size o2 ! s2_limb o3 -#include "sysdep.h" +#include <sysdep.h> -.text - .align 8 - .global C_SYMBOL_NAME(__mpn_mul_1) -C_SYMBOL_NAME(__mpn_mul_1): +ENTRY(__mpn_mul_1) sll %o2,4,%g1 - and %g1,(4-1)<<4,%g1 -#if PIC mov %o7,%g4 ! Save return address register - call 1f - add %o7,LL-1f,%g3 -1: mov %g4,%o7 ! Restore return address register -#else - sethi %hi(LL),%g3 - or %g3,%lo(LL),%g3 -#endif + and %g1,(4-1)<<4,%g1 +1: call 2f + add %o7,3f-1b,%g3 +2: mov %g4,%o7 ! Restore return address register jmp %g3+%g1 - ld [%o1+0],%o4 ! 1 -LL: -LL00: add %o0,-4,%o0 + ld [%o1+0],%o4 ! 1 + + .align 4 +3: +LOC(00): + add %o0,-4,%o0 add %o1,-4,%o1 - b Loop00 /* 4, 8, 12, ... */ - orcc %g0,%g0,%g2 -LL01: b Loop01 /* 1, 5, 9, ... */ - orcc %g0,%g0,%g2 + b LOC(loop00) /* 4, 8, 12, ... */ + orcc %g0,%g0,%g2 +LOC(01): + b LOC(loop01) /* 1, 5, 9, ... */ + orcc %g0,%g0,%g2 nop nop -LL10: add %o0,-12,%o0 /* 2, 6, 10, ... */ +LOC(10): + add %o0,-12,%o0 /* 2, 6, 10, ... */ add %o1,4,%o1 - b Loop10 - orcc %g0,%g0,%g2 + b LOC(loop10) + orcc %g0,%g0,%g2 nop -LL11: add %o0,-8,%o0 /* 3, 7, 11, ... */ +LOC(11): + add %o0,-8,%o0 /* 3, 7, 11, ... */ add %o1,-8,%o1 - b Loop11 - orcc %g0,%g0,%g2 + b LOC(loop11) + orcc %g0,%g0,%g2 -Loop: addcc %g3,%g2,%g3 ! 1 +LOC(loop): + addcc %g3,%g2,%g3 ! 1 ld [%o1+4],%o4 ! 2 st %g3,[%o0+0] ! 1 rd %y,%g2 ! 1 -Loop00: umul %o4,%o3,%g3 ! 2 +LOC(loop00): + umul %o4,%o3,%g3 ! 2 addxcc %g3,%g2,%g3 ! 2 ld [%o1+8],%o4 ! 3 st %g3,[%o0+4] ! 2 rd %y,%g2 ! 2 -Loop11: umul %o4,%o3,%g3 ! 3 +LOC(loop11): + umul %o4,%o3,%g3 ! 3 addxcc %g3,%g2,%g3 ! 3 ld [%o1+12],%o4 ! 4 add %o1,16,%o1 st %g3,[%o0+8] ! 3 rd %y,%g2 ! 3 -Loop10: umul %o4,%o3,%g3 ! 4 +LOC(loop10): + umul %o4,%o3,%g3 ! 4 addxcc %g3,%g2,%g3 ! 4 ld [%o1+0],%o4 ! 1 st %g3,[%o0+12] ! 4 add %o0,16,%o0 rd %y,%g2 ! 4 addx %g0,%g2,%g2 -Loop01: addcc %o2,-4,%o2 - bg Loop - umul %o4,%o3,%g3 ! 1 +LOC(loop01): + addcc %o2,-4,%o2 + bg LOC(loop) + umul %o4,%o3,%g3 ! 1 addcc %g3,%g2,%g3 ! 4 st %g3,[%o0+0] ! 4 rd %y,%g2 ! 4 - retl - addx %g0,%g2,%o0 + addx %g0,%g2,%o0 + +END(__mpn_mul_1) diff --git a/sysdeps/sparc/sparc32/sparcv8/rem.S b/sysdeps/sparc/sparc32/sparcv8/rem.S new file mode 100644 index 0000000000..061e80093c --- /dev/null +++ b/sysdeps/sparc/sparc32/sparcv8/rem.S @@ -0,0 +1,18 @@ +/* + * Sparc v8 has divide. + */ + +#include <sysdep.h> + +ENTRY(.rem) + + sra %o0, 31, %o2 + wr %o2, 0, %y + sdivcc %o0, %o1, %o2 + bvs,a 1f + xnor %o2, %g0, %o2 +1: smul %o2, %o1, %o2 + retl + sub %o0, %o2, %o0 + +END(.rem) diff --git a/sysdeps/sparc/sparc32/sparcv8/sdiv.S b/sysdeps/sparc/sparc32/sparcv8/sdiv.S new file mode 100644 index 0000000000..81c3ac1c97 --- /dev/null +++ b/sysdeps/sparc/sparc32/sparcv8/sdiv.S @@ -0,0 +1,14 @@ +/* + * Sparc v8 has divide. + */ + +#include <sysdep.h> + +ENTRY(.div) + + sra %o0, 31, %o2 + wr %o2, 0, %y + ret + sdiv %o0, %o1, %o0 + +END(.div) diff --git a/sysdeps/sparc/sparc32/sparcv8/submul_1.S b/sysdeps/sparc/sparc32/sparcv8/submul_1.S index e40119d011..c71f228511 100644 --- a/sysdeps/sparc/sparc32/sparcv8/submul_1.S +++ b/sysdeps/sparc/sparc32/sparcv8/submul_1.S @@ -27,12 +27,9 @@ ! size o2 ! s2_limb o3 -#include "sysdep.h" +#include <sysdep.h> -.text - .align 4 - .global C_SYMBOL_NAME(__mpn_submul_1) -C_SYMBOL_NAME(__mpn_submul_1): +ENTRY(__mpn_submul_1) sub %g0,%o2,%o2 ! negate ... sll %o2,2,%o2 ! ... and scale size sub %o1,%o2,%o1 ! o1 is offset s1_ptr @@ -40,7 +37,8 @@ C_SYMBOL_NAME(__mpn_submul_1): mov 0,%o0 ! clear cy_limb -Loop: ld [%o1+%o2],%o4 +LOC(loop): + ld [%o1+%o2],%o4 ld [%g1+%o2],%g2 umul %o4,%o3,%o5 rd %y,%g3 @@ -51,8 +49,10 @@ Loop: ld [%o1+%o2],%o4 st %g2,[%g1+%o2] addcc %o2,4,%o2 - bne Loop + bne LOC(loop) nop retl nop + +END(__mpn_submul_1) diff --git a/sysdeps/sparc/sparc32/sparcv8/udiv.S b/sysdeps/sparc/sparc32/sparcv8/udiv.S new file mode 100644 index 0000000000..4e3cddc63f --- /dev/null +++ b/sysdeps/sparc/sparc32/sparcv8/udiv.S @@ -0,0 +1,13 @@ +/* + * Sparc v8 has divide. + */ + +#include <sysdep.h> + +ENTRY(.udiv) + + wr %g0, 0, %y + retl + udiv %o0, %o1, %o0 + +END(.udiv) diff --git a/sysdeps/sparc/sparc32/sparcv8/udiv_qrnnd.S b/sysdeps/sparc/sparc32/sparcv8/udiv_qrnnd.S index 49c2398806..5b0d42923a 100644 --- a/sysdeps/sparc/sparc32/sparcv8/udiv_qrnnd.S +++ b/sysdeps/sparc/sparc32/sparcv8/udiv_qrnnd.S @@ -27,66 +27,75 @@ #include "sysdep.h" - .text - .align 4 - .global C_SYMBOL_NAME(__udiv_qrnnd) -C_SYMBOL_NAME(__udiv_qrnnd): +ENTRY(__udiv_qrnnd) tst %o3 - bneg Largedivisor + bneg LOC(largedivisor) mov 8,%g1 - b Lp1 + b LOC(p1) addxcc %o2,%o2,%o2 -Lplop: bcc Ln1 +LOC(plop): + bcc LOC(n1) addxcc %o2,%o2,%o2 -Lp1: addx %o1,%o1,%o1 +LOC(p1): + addx %o1,%o1,%o1 subcc %o1,%o3,%o4 - bcc Ln2 + bcc LOC(n2) addxcc %o2,%o2,%o2 -Lp2: addx %o1,%o1,%o1 +LOC(p2): + addx %o1,%o1,%o1 subcc %o1,%o3,%o4 - bcc Ln3 + bcc LOC(n3) addxcc %o2,%o2,%o2 -Lp3: addx %o1,%o1,%o1 +LOC(p3): + addx %o1,%o1,%o1 subcc %o1,%o3,%o4 - bcc Ln4 + bcc LOC(n4) addxcc %o2,%o2,%o2 -Lp4: addx %o1,%o1,%o1 +LOC(p4): + addx %o1,%o1,%o1 addcc %g1,-1,%g1 - bne Lplop + bne LOC(plop) subcc %o1,%o3,%o4 - bcc Ln5 + bcc LOC(n5) addxcc %o2,%o2,%o2 -Lp5: st %o1,[%o0] +LOC(p5): + st %o1,[%o0] retl xnor %g0,%o2,%o0 -Lnlop: bcc Lp1 +LOC(nlop): + bcc LOC(p1) addxcc %o2,%o2,%o2 -Ln1: addx %o4,%o4,%o4 +LOC(n1): + addx %o4,%o4,%o4 subcc %o4,%o3,%o1 - bcc Lp2 + bcc LOC(p2) addxcc %o2,%o2,%o2 -Ln2: addx %o4,%o4,%o4 +LOC(n2): + addx %o4,%o4,%o4 subcc %o4,%o3,%o1 - bcc Lp3 + bcc LOC(p3) addxcc %o2,%o2,%o2 -Ln3: addx %o4,%o4,%o4 +LOC(n3): + addx %o4,%o4,%o4 subcc %o4,%o3,%o1 - bcc Lp4 + bcc LOC(p4) addxcc %o2,%o2,%o2 -Ln4: addx %o4,%o4,%o4 +LOC(n4): + addx %o4,%o4,%o4 addcc %g1,-1,%g1 - bne Lnlop + bne LOC(nlop) subcc %o4,%o3,%o1 - bcc Lp5 + bcc LOC(p5) addxcc %o2,%o2,%o2 -Ln5: st %o4,[%o0] +LOC(n5): + st %o4,[%o0] retl xnor %g0,%o2,%o0 -Largedivisor: +LOC(largedivisor): and %o2,1,%o5 ! %o5 = n0 & 1 srl %o2,1,%o2 @@ -98,89 +107,109 @@ Largedivisor: srl %o3,1,%g3 ! %g3 = floor(d / 2) add %g3,%g2,%g3 ! %g3 = ceil(d / 2) - b LLp1 + b LOC(Lp1) addxcc %o2,%o2,%o2 -LLplop: bcc LLn1 +LOC(Lplop): + bcc LOC(Ln1) addxcc %o2,%o2,%o2 -LLp1: addx %o1,%o1,%o1 +LOC(Lp1): + addx %o1,%o1,%o1 subcc %o1,%g3,%o4 - bcc LLn2 + bcc LOC(Ln2) addxcc %o2,%o2,%o2 -LLp2: addx %o1,%o1,%o1 +LOC(Lp2): + addx %o1,%o1,%o1 subcc %o1,%g3,%o4 - bcc LLn3 + bcc LOC(Ln3) addxcc %o2,%o2,%o2 -LLp3: addx %o1,%o1,%o1 +LOC(Lp3): + addx %o1,%o1,%o1 subcc %o1,%g3,%o4 - bcc LLn4 + bcc LOC(Ln4) addxcc %o2,%o2,%o2 -LLp4: addx %o1,%o1,%o1 +LOC(Lp4): + addx %o1,%o1,%o1 addcc %g1,-1,%g1 - bne LLplop + bne LOC(Lplop) subcc %o1,%g3,%o4 - bcc LLn5 + bcc LOC(Ln5) addxcc %o2,%o2,%o2 -LLp5: add %o1,%o1,%o1 ! << 1 +LOC(Lp5): + add %o1,%o1,%o1 ! << 1 tst %g2 - bne Oddp + bne LOC(Oddp) add %o5,%o1,%o1 st %o1,[%o0] retl xnor %g0,%o2,%o0 -LLnlop: bcc LLp1 +LOC(Lnlop): + bcc LOC(Lp1) addxcc %o2,%o2,%o2 -LLn1: addx %o4,%o4,%o4 +LOC(Ln1): + addx %o4,%o4,%o4 subcc %o4,%g3,%o1 - bcc LLp2 + bcc LOC(Lp2) addxcc %o2,%o2,%o2 -LLn2: addx %o4,%o4,%o4 +LOC(Ln2): + addx %o4,%o4,%o4 subcc %o4,%g3,%o1 - bcc LLp3 + bcc LOC(Lp3) addxcc %o2,%o2,%o2 -LLn3: addx %o4,%o4,%o4 +LOC(Ln3): + addx %o4,%o4,%o4 subcc %o4,%g3,%o1 - bcc LLp4 + bcc LOC(Lp4) addxcc %o2,%o2,%o2 -LLn4: addx %o4,%o4,%o4 +LOC(Ln4): + addx %o4,%o4,%o4 addcc %g1,-1,%g1 - bne LLnlop + bne LOC(Lnlop) subcc %o4,%g3,%o1 - bcc LLp5 + bcc LOC(Lp5) addxcc %o2,%o2,%o2 -LLn5: add %o4,%o4,%o4 ! << 1 +LOC(Ln5): + add %o4,%o4,%o4 ! << 1 tst %g2 - bne Oddn + bne LOC(Oddn) add %o5,%o4,%o4 st %o4,[%o0] retl xnor %g0,%o2,%o0 -Oddp: xnor %g0,%o2,%o2 +LOC(Oddp): + xnor %g0,%o2,%o2 ! q' in %o2. r' in %o1 addcc %o1,%o2,%o1 - bcc LLp6 + bcc LOC(Lp6) addx %o2,0,%o2 sub %o1,%o3,%o1 -LLp6: subcc %o1,%o3,%g0 - bcs LLp7 +LOC(Lp6): + subcc %o1,%o3,%g0 + bcs LOC(Lp7) subx %o2,-1,%o2 sub %o1,%o3,%o1 -LLp7: st %o1,[%o0] +LOC(Lp7): + st %o1,[%o0] retl mov %o2,%o0 -Oddn: xnor %g0,%o2,%o2 +LOC(Oddn): + xnor %g0,%o2,%o2 ! q' in %o2. r' in %o4 addcc %o4,%o2,%o4 - bcc LLn6 + bcc LOC(Ln6) addx %o2,0,%o2 sub %o4,%o3,%o4 -LLn6: subcc %o4,%o3,%g0 - bcs LLn7 +LOC(Ln6): + subcc %o4,%o3,%g0 + bcs LOC(Ln7) subx %o2,-1,%o2 sub %o4,%o3,%o4 -LLn7: st %o4,[%o0] +LOC(Ln7): + st %o4,[%o0] retl mov %o2,%o0 + +END(__udiv_qrnnd) diff --git a/sysdeps/sparc/sparc32/sparcv8/umul.S b/sysdeps/sparc/sparc32/sparcv8/umul.S new file mode 100644 index 0000000000..cec454a7dd --- /dev/null +++ b/sysdeps/sparc/sparc32/sparcv8/umul.S @@ -0,0 +1,13 @@ +/* + * Sparc v8 has multiply. + */ + +#include <sysdep.h> + +ENTRY(.umul) + + umul %o0, %o1, %o0 + retl + rd %y, %o1 + +END(.umul) diff --git a/sysdeps/sparc/sparc32/sparcv8/urem.S b/sysdeps/sparc/sparc32/sparcv8/urem.S new file mode 100644 index 0000000000..81e123f3da --- /dev/null +++ b/sysdeps/sparc/sparc32/sparcv8/urem.S @@ -0,0 +1,15 @@ +/* + * Sparc v8 has divide. + */ + +#include <sysdep.h> + +ENTRY(.urem) + + wr %g0, 0, %y + udiv %o0, %o1, %o2 + umul %o2, %o1, %o2 + retl + sub %o0, %o2, %o0 + +END(.urem) |