lib1funcs.asm (udivdi3): Make first divide step produce a 32 bit result before normalization...

	* 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.

From-SVN: r54993

1 files changed, 30 insertions, 27 deletions

diff --git a/gcc/config/sh/lib1funcs.asm b/gcc/config/sh/lib1funcs.asm
index 2364db9db8f..1021c9ba3b5 100644
--- a/gcc/config/sh/lib1funcs.asm
+++ b/gcc/config/sh/lib1funcs.asm
@@ -1501,33 +1501,33 @@ GLOBAL(udivdi3):
 	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
+	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.  */
-	shlrd r5,r0,r8
-	mulu.l r8,r3,r5
+	mulu.l r5,r3,r8
 	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
+	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
-	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
+	mcmpgt.l r21,r63,r21 // See Note 1
 	addi r20,30,r0
 	mshfhi.l r63,r21,r21
 	sub r2,r5,r2
@@ -1555,14 +1555,15 @@ LOCAL(large_divisor):
 	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+32,r8
+	shlri r5,14-1,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)
+	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.  */
 
@@ -1575,11 +1576,11 @@ LOCAL(large_divisor):
 	add r8,r21,r8
 	shlld r2,r0,r2
 	sub r25,r5,r25
-	mextr4 r2,r25,r2
-	bgtu/u r6,r2,tr0 // no_lo_adj
+	bgtu/u r7,r25,tr0 // no_lo_adj
 	addi r8,1,r8
-	sub r2,r6,r2
+	sub r25,r7,r25
 LOCAL(no_lo_adj):
+	mextr4 r2,r25,r2
 
 	/* large_divisor: only needs a few adjustments.  */
 	mulu.l r8,r6,r5
@@ -1647,22 +1648,22 @@ GLOBAL(umoddi3):
 	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
+	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.  */
-	shlrd r5,r0,r8
-	mulu.l r8,r3,r5
+	mulu.l r5,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)
+	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.  */
 
@@ -1672,7 +1673,7 @@ GLOBAL(umoddi3):
 	/* 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
+	mcmpgt.l r21,r63,r21 // See Note 1
 	addi r20,30,r0
 	mshfhi.l r63,r21,r21
 	sub r2,r5,r2
@@ -1700,14 +1701,15 @@ LOCAL(large_divisor):
 	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+32,r8
+	shlri r5,14-1,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)
+	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.  */
 
@@ -1720,19 +1722,20 @@ LOCAL(large_divisor):
 	add r8,r21,r8
 	shlld r2,r0,r2
 	sub r25,r5,r25
-	mextr4 r2,r25,r2
-	bgtu/u r6,r2,tr0 // no_lo_adj
+	bgtu/u r7,r25,tr0 // no_lo_adj
 	addi r8,1,r8
-	sub r2,r6,r2
+	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,r3,r7
+	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