This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Patch applied to config/sh/lib1funcs.asm.
- From: Joern Rennecke <joern dot rennecke at superh dot com>
- To: gcc-patches at gcc dot gnu dot org
- Date: Tue, 25 Jun 2002 21:44:27 +0100
- Subject: Patch applied to config/sh/lib1funcs.asm.
- Organization: SuperH UK Ltd.
--
--------------------------
SuperH
2430 Aztec West / Almondsbury / BRISTOL / BS32 4AQ
T:+44 1454 462330
Tue Jun 25 20:59:56 2002 J"orn Rennecke <joern.rennecke@superh.com>
* config/sh/lib1funcs.asm (udivdi3): Make first divide step
produce a 32 bit result before normalization, then normalize with a
left shift. Compute approximative error of 2nd reciprocal
approximation in 2's complement. Fix mask generation from upper
longword of second divide stage result.
For large divisor, fix shift count used to truncate first stage
divide result; make decision if to adjust upwards based on comparison
of higher parts of normalized values.
(udivdi): Likewise. Undo normalization of result for large divisor
case.
Index: lib1funcs.asm
===================================================================
RCS file: /cvs/gcc/gcc/gcc/config/sh/lib1funcs.asm,v
retrieving revision 1.24
diff -p -r1.24 lib1funcs.asm
*** lib1funcs.asm 24 Jun 2002 20:22:12 -0000 1.24
--- lib1funcs.asm 25 Jun 2002 19:59:31 -0000
*************** GLOBAL(udivdi3):
*** 1501,1533 ****
shlri r6,32,r7
bgt/u r9,r63,tr0 // large_divisor
mmulfx.w r5,r4,r4
! shlri r2,32,r19
! addi r20,14-1,r0
msub.w r1,r4,r1
mulu.l r1,r7,r4
addi r1,-3,r5
mulu.l r5,r19,r5
shlri r4,2,r4 /* chop off leading %0000000000000000 001.00000000000 - or, as
the case may be, %0000000000000000 000.11111111111, still */
muls.l r1,r4,r4 /* leaving at least one sign bit. */
! shlrd r5,r0,r8
! mulu.l r8,r3,r5
mshalds.l r1,r21,r1
shari r4,26,r4
! shlli r5,32,r5
! sub r1,r4,r1 // 31 bit unsigned reciprocal now in r1 (msb equiv. 0.5)
! sub r2,r5,r2
/* Can do second step of 64 : 32 div now, using r1 and the rest in r2. */
shlri r2,22,r21
mulu.l r21,r1,r21
addi r20,30-22,r0
- shlli r8,32,r8
shlrd r21,r0,r21
mulu.l r21,r3,r5
add r8,r21,r8
! mcmpeq.l r21,r63,r21 // See Note 1
addi r20,30,r0
mshfhi.l r63,r21,r21
sub r2,r5,r2
--- 1501,1533 ----
shlri r6,32,r7
bgt/u r9,r63,tr0 // large_divisor
mmulfx.w r5,r4,r4
! shlri r2,32+14,r19
! addi r22,-31,r0
msub.w r1,r4,r1
mulu.l r1,r7,r4
addi r1,-3,r5
mulu.l r5,r19,r5
+ sub r63,r4,r4 // Negate to make sure r1 ends up <= 1/r2
shlri r4,2,r4 /* chop off leading %0000000000000000 001.00000000000 - or, as
the case may be, %0000000000000000 000.11111111111, still */
muls.l r1,r4,r4 /* leaving at least one sign bit. */
! mulu.l r5,r3,r8
mshalds.l r1,r21,r1
shari r4,26,r4
! shlld r8,r0,r8
! add r1,r4,r1 // 31 bit unsigned reciprocal now in r1 (msb equiv. 0.5)
! sub r2,r8,r2
/* Can do second step of 64 : 32 div now, using r1 and the rest in r2. */
shlri r2,22,r21
mulu.l r21,r1,r21
+ shlld r5,r0,r8
addi r20,30-22,r0
shlrd r21,r0,r21
mulu.l r21,r3,r5
add r8,r21,r8
! mcmpgt.l r21,r63,r21 // See Note 1
addi r20,30,r0
mshfhi.l r63,r21,r21
sub r2,r5,r2
*************** LOCAL(large_divisor):
*** 1555,1568 ****
mulu.l r1,r7,r4
addi r1,-3,r5
mulu.l r5,r8,r5
shlri r4,2,r4 /* chop off leading %0000000000000000 001.00000000000 - or, as
the case may be, %0000000000000000 000.11111111111, still */
muls.l r1,r4,r4 /* leaving at least one sign bit. */
! shlri r5,14-1+32,r8
mulu.l r8,r7,r5
mshalds.l r1,r21,r1
shari r4,26,r4
! sub r1,r4,r1 // 31 bit unsigned reciprocal now in r1 (msb equiv. 0.5)
sub r25,r5,r25
/* Can do second step of 64 : 32 div now, using r1 and the rest in r25. */
--- 1555,1569 ----
mulu.l r1,r7,r4
addi r1,-3,r5
mulu.l r5,r8,r5
+ sub r63,r4,r4 // Negate to make sure r1 ends up <= 1/r2
shlri r4,2,r4 /* chop off leading %0000000000000000 001.00000000000 - or, as
the case may be, %0000000000000000 000.11111111111, still */
muls.l r1,r4,r4 /* leaving at least one sign bit. */
! shlri r5,14-1,r8
mulu.l r8,r7,r5
mshalds.l r1,r21,r1
shari r4,26,r4
! add r1,r4,r1 // 31 bit unsigned reciprocal now in r1 (msb equiv. 0.5)
sub r25,r5,r25
/* Can do second step of 64 : 32 div now, using r1 and the rest in r25. */
*************** LOCAL(large_divisor):
*** 1575,1585 ****
add r8,r21,r8
shlld r2,r0,r2
sub r25,r5,r25
! mextr4 r2,r25,r2
! bgtu/u r6,r2,tr0 // no_lo_adj
addi r8,1,r8
! sub r2,r6,r2
LOCAL(no_lo_adj):
/* large_divisor: only needs a few adjustments. */
mulu.l r8,r6,r5
--- 1576,1586 ----
add r8,r21,r8
shlld r2,r0,r2
sub r25,r5,r25
! bgtu/u r7,r25,tr0 // no_lo_adj
addi r8,1,r8
! sub r25,r7,r25
LOCAL(no_lo_adj):
+ mextr4 r2,r25,r2
/* large_divisor: only needs a few adjustments. */
mulu.l r8,r6,r5
*************** GLOBAL(umoddi3):
*** 1647,1668 ****
shlri r6,32,r7
bgt/u r9,r63,tr0 // large_divisor
mmulfx.w r5,r4,r4
! shlri r2,32,r19
! addi r20,14-1,r0
msub.w r1,r4,r1
mulu.l r1,r7,r4
addi r1,-3,r5
mulu.l r5,r19,r5
shlri r4,2,r4 /* chop off leading %0000000000000000 001.00000000000 - or, as
the case may be, %0000000000000000 000.11111111111, still */
muls.l r1,r4,r4 /* leaving at least one sign bit. */
! shlrd r5,r0,r8
! mulu.l r8,r3,r5
mshalds.l r1,r21,r1
shari r4,26,r4
! shlli r5,32,r5
! sub r1,r4,r1 // 31 bit unsigned reciprocal now in r1 (msb equiv. 0.5)
sub r2,r5,r2
/* Can do second step of 64 : 32 div now, using r1 and the rest in r2. */
--- 1648,1669 ----
shlri r6,32,r7
bgt/u r9,r63,tr0 // large_divisor
mmulfx.w r5,r4,r4
! shlri r2,32+14,r19
! addi r22,-31,r0
msub.w r1,r4,r1
mulu.l r1,r7,r4
addi r1,-3,r5
mulu.l r5,r19,r5
+ sub r63,r4,r4 // Negate to make sure r1 ends up <= 1/r2
shlri r4,2,r4 /* chop off leading %0000000000000000 001.00000000000 - or, as
the case may be, %0000000000000000 000.11111111111, still */
muls.l r1,r4,r4 /* leaving at least one sign bit. */
! mulu.l r5,r3,r5
mshalds.l r1,r21,r1
shari r4,26,r4
! shlld r5,r0,r5
! add r1,r4,r1 // 31 bit unsigned reciprocal now in r1 (msb equiv. 0.5)
sub r2,r5,r2
/* Can do second step of 64 : 32 div now, using r1 and the rest in r2. */
*************** GLOBAL(umoddi3):
*** 1672,1678 ****
/* bubble */ /* could test r3 here to check for divide by zero. */
shlrd r21,r0,r21
mulu.l r21,r3,r5
! mcmpeq.l r21,r63,r21 // See Note 1
addi r20,30,r0
mshfhi.l r63,r21,r21
sub r2,r5,r2
--- 1673,1679 ----
/* bubble */ /* could test r3 here to check for divide by zero. */
shlrd r21,r0,r21
mulu.l r21,r3,r5
! mcmpgt.l r21,r63,r21 // See Note 1
addi r20,30,r0
mshfhi.l r63,r21,r21
sub r2,r5,r2
*************** LOCAL(large_divisor):
*** 1700,1713 ****
mulu.l r1,r7,r4
addi r1,-3,r5
mulu.l r5,r8,r5
shlri r4,2,r4 /* chop off leading %0000000000000000 001.00000000000 - or, as
the case may be, %0000000000000000 000.11111111111, still */
muls.l r1,r4,r4 /* leaving at least one sign bit. */
! shlri r5,14-1+32,r8
mulu.l r8,r7,r5
mshalds.l r1,r21,r1
shari r4,26,r4
! sub r1,r4,r1 // 31 bit unsigned reciprocal now in r1 (msb equiv. 0.5)
sub r25,r5,r25
/* Can do second step of 64 : 32 div now, using r1 and the rest in r25. */
--- 1701,1715 ----
mulu.l r1,r7,r4
addi r1,-3,r5
mulu.l r5,r8,r5
+ sub r63,r4,r4 // Negate to make sure r1 ends up <= 1/r2
shlri r4,2,r4 /* chop off leading %0000000000000000 001.00000000000 - or, as
the case may be, %0000000000000000 000.11111111111, still */
muls.l r1,r4,r4 /* leaving at least one sign bit. */
! shlri r5,14-1,r8
mulu.l r8,r7,r5
mshalds.l r1,r21,r1
shari r4,26,r4
! add r1,r4,r1 // 31 bit unsigned reciprocal now in r1 (msb equiv. 0.5)
sub r25,r5,r25
/* Can do second step of 64 : 32 div now, using r1 and the rest in r25. */
*************** LOCAL(large_divisor):
*** 1720,1738 ****
add r8,r21,r8
shlld r2,r0,r2
sub r25,r5,r25
! mextr4 r2,r25,r2
! bgtu/u r6,r2,tr0 // no_lo_adj
addi r8,1,r8
! sub r2,r6,r2
LOCAL(no_lo_adj):
/* large_divisor: only needs a few adjustments. */
mulu.l r8,r6,r5
ptabs r18,tr0
! add r2,r3,r7
cmpgtu r5,r2,r8
cmvne r8,r7,r2
sub r2,r5,r2
blink tr0,r63
/* Note 1: To shift the result of the second divide stage so that the result
always fits into 32 bits, yet we still reduce the rest sufficiently
--- 1722,1741 ----
add r8,r21,r8
shlld r2,r0,r2
sub r25,r5,r25
! bgtu/u r7,r25,tr0 // no_lo_adj
addi r8,1,r8
! sub r25,r7,r25
LOCAL(no_lo_adj):
+ mextr4 r2,r25,r2
/* large_divisor: only needs a few adjustments. */
mulu.l r8,r6,r5
ptabs r18,tr0
! add r2,r6,r7
cmpgtu r5,r2,r8
cmvne r8,r7,r2
sub r2,r5,r2
+ shlrd r2,r22,r2
blink tr0,r63
/* Note 1: To shift the result of the second divide stage so that the result
always fits into 32 bits, yet we still reduce the rest sufficiently