Build failure for 5.1.0-RC2 on Mac OS 10.8

Wed Dec 12 00:43:18 CET 2012

On Tue, Dec 11, 2012 at 06:20:22PM -0500, Jack Howarth wrote:
> On Tue, Dec 11, 2012 at 11:15:35PM +0100, Torbjorn Granlund wrote:
> > Jack Howarth <howarth at bromo.med.uc.edu> writes:
> > 
> >   +  .set LC1, L1-Ltab
> >   +  .long	LC1
> >   
> >   and his reasoning as follows...
> >   
> >   --------------------------------------------------
> >   
> >   The deal is, you want the assembler to evaluate the deltas and not
> >   pass relocations on to the linker. You can see the compiler doing this
> >   is you just compile a simple switch like:
> >   
> > Thank for the information.
> > 
> > This behaviour is highly unusual, and I know not all Apple assembler
> > releases require that, and I would expect some Apple assemblers to
> > actually reject that syntax.
> 
> Torbjorn,
>    I can confirm that the attached mod_34lsub1.fixed.s can be compiled
> with 'as -arch x86_64' on the following Apple Xcode releases...
> 
> Xcode 4.5.2 on darwin12
> Xcode 4.5.2 on darwin11
> Xcode 4.2 on darwin10
> Xcode 3.2.6 on darwin10
> Xcode 3.1.4 on darwin9
> 
> I don't have darwin8 available to check Xcode 2.5 but David Fang
> can probably check that for us.
>              Jack

Torbjorn,
   I guess this modification isn't quite right anyway. I tried the modification...

--- mpn/x86_64/mod_34lsub1.asm.orig	2012-12-09 16:29:52.000000000 -0500
+++ mpn/x86_64/mod_34lsub1.asm	2012-12-11 18:31:25.000000000 -0500
@@ -107,18 +107,24 @@
 	lea	(%r10, %r8), %r8
 	jmp	*%r8
 
-	RODATA
-L(tab):	.long	L(0)-L(tab)
-	.long	L(1)-L(tab)
-	.long	L(2)-L(tab)
-	.long	L(3)-L(tab)
-	.long	L(4)-L(tab)
-	.long	L(5)-L(tab)
-	.long	L(6)-L(tab)
-	.long	L(7)-L(tab)
-	.long	L(8)-L(tab)
+Ltab:
+	.set LC1, L1-Ltab
+	.long        LC1
+	.set LC2, L2-Ltab
+	.long        LC2
+	.set LC3, L3-Ltab
+	.long        LC3
+	.set LC4, L4-Ltab
+	.long        LC4
+	.set LC5, L5-Ltab
+	.long        LC5
+	.set LC6, L6-Ltab
+	.long        LC6
+	.set LC7, L7-Ltab
+	.long        LC7
+	.set LC8, L8-Ltab
+	.long        LC8
 
-	TEXT
 L(6):	add	(ap), %rax
 	adc	8(ap), %rcx
 	adc	16(ap), %rdx

but this permutation produces a failure in the testsuite at...

mpz_perfect_square_p and mpz_sqrt differ
   x  =0x100000000000000000000
   x2 =0x10000000000000000000000000000000000000000
   x2t=0x10000000000000000000000000000000000000000
   mpz_perfect_square_p 0
   mpz_sqrt             1
/bin/sh: line 1: 12657 Abort trap: 6           ${dir}$tst
FAIL: t-perfsqr

Oh, well.

> 
> > 
> > It should be noted that no other concurrent assemblers fail to compute
> > sym1-sym2 when both symbols are in the same segment.  They don't
> > require any contortions in order to trigger symbol arithmetic that is
> > fully foldable.
> > 
> > We know that the code of RC1 works with some Apple assemblers/linkers,
> > but not with ones older and newer.  This make me take an extremely
> > conservative stance wrt Apple porting.
> > 
> > We will never get stable Apple support, however hard we try.  Only if
> > Apple decided for what to support, and wrote a good test suite
> > enforcing the decision, and educated their staff, than we could reach
> > the same GMP quality with Apple as with other platforms.  But the
> > damage is already done with the mess of inconsistent tools out there.
> > 
> > -- 
> > Torbjörn

> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 	.text
> 	.align	5, 0x90
> 	.globl	___gmpn_mod_34lsub1
> 	
> ___gmpn_mod_34lsub1:
> 
> 	
> 
> 	mov	$0x0000FFFFFFFFFFFF, %r11
> 
> 	mov	(%rdi), %rax
> 
> 	cmp	$2, %rsi
> 	ja	Lgt2
> 
> 	jb	Lone
> 
> 	mov	8(%rdi), %rsi
> 	mov	%rax, %rdx
> 	shr	$48, %rax		
> 
> 	and	%r11, %rdx		
> 	add	%rdx, %rax
> 	mov	%esi, %edx
> 
> 	shr	$32, %rsi		
> 	add	%rsi, %rax
> 
> 	shl	$16, %rdx		
> 	add	%rdx, %rax
> Lone:	
> 	ret
> 
> 
> 
> 
> 
> Lgt2:	mov	8(%rdi), %rcx
> 	mov	16(%rdi), %rdx
> 	xor	%r9, %r9
> 	add	$24, %rdi
> 	sub	$12, %rsi
> 	jc	Lend
> 	.align	4, 0x90
> Ltop:
> 	add	(%rdi), %rax
> 	adc	8(%rdi), %rcx
> 	adc	16(%rdi), %rdx
> 	adc	$0, %r9
> 	add	24(%rdi), %rax
> 	adc	32(%rdi), %rcx
> 	adc	40(%rdi), %rdx
> 	adc	$0, %r9
> 	add	48(%rdi), %rax
> 	adc	56(%rdi), %rcx
> 	adc	64(%rdi), %rdx
> 	adc	$0, %r9
> 	add	$72, %rdi
> 	sub	$9, %rsi
> 	jnc	Ltop
> 
> Lend:	lea	Ltab(%rip), %r8
> 
> 
> 	movslq	36(%r8,%rsi,4), %r10
> 	lea	(%r10, %r8), %r8
> 	jmp	*%r8
> 
> Ltab:	
>   .set LC1, L1-Ltab
>   .long	LC1
>   .set LC2, L2-Ltab
>   .long	LC2
>   .set LC3, L3-Ltab
>   .long	LC3
>   .set LC4, L4-Ltab
>   .long	LC4
>   .set LC5, L5-Ltab
>   .long	LC5
>   .set LC6, L6-Ltab
>   .long	LC6
>   .set LC7, L7-Ltab
>   .long	LC7
>   .set LC8, L8-Ltab
>   .long	LC8
> 
> 
> L6:	add	(%rdi), %rax
> 	adc	8(%rdi), %rcx
> 	adc	16(%rdi), %rdx
> 	adc	$0, %r9
> 	add	$24, %rdi
> L3:	add	(%rdi), %rax
> 	adc	8(%rdi), %rcx
> 	adc	16(%rdi), %rdx
> 	jmp	Lcj1
> 
> L7:	add	(%rdi), %rax
> 	adc	8(%rdi), %rcx
> 	adc	16(%rdi), %rdx
> 	adc	$0, %r9
> 	add	$24, %rdi
> L4:	add	(%rdi), %rax
> 	adc	8(%rdi), %rcx
> 	adc	16(%rdi), %rdx
> 	adc	$0, %r9
> 	add	$24, %rdi
> L1:	add	(%rdi), %rax
> 	adc	$0, %rcx
> 	jmp	Lcj2
> 
> L8:	add	(%rdi), %rax
> 	adc	8(%rdi), %rcx
> 	adc	16(%rdi), %rdx
> 	adc	$0, %r9
> 	add	$24, %rdi
> L5:	add	(%rdi), %rax
> 	adc	8(%rdi), %rcx
> 	adc	16(%rdi), %rdx
> 	adc	$0, %r9
> 	add	$24, %rdi
> L2:	add	(%rdi), %rax
> 	adc	8(%rdi), %rcx
> 
> Lcj2:	adc	$0, %rdx
> Lcj1:	adc	$0, %r9
> L0:	add	%r9, %rax
> 	adc	$0, %rcx
> 	adc	$0, %rdx
> 	adc	$0, %rax
> 
> 	mov	%rax, %rdi		
> 	shr	$48, %rax		
> 
> 	and	%r11, %rdi		
> 	mov	%ecx, %r10d	
> 
> 	shr	$32, %rcx		
> 
> 	add	%rdi, %rax		
> 	movzwl	%dx, %edi		
> 	shl	$16, %r10		
> 
> 	add	%rcx, %rax		
> 	shr	$16, %rdx		
> 
> 	add	%r10, %rax		
> 	shl	$32, %rdi		
> 
> 	add	%rdx, %rax		
> 	add	%rdi, %rax		
> 
> 	
> 	ret
> 	

> _______________________________________________
> gmp-bugs mailing list
> gmp-bugs at gmplib.org
> https://gmplib.org/mailman/listinfo/gmp-bugs

