Attachment 216 Details for Bug 357 – GCC Assembler output for test case.

This bugzilla service is closed. All entries have been migrated to https://gitlab.com/libeigen/eigen

GCC Assembler output for test case.

TestEigen.GCC.ASM (text/plain), 9.84 KB, created by Colm on 2011-10-04 20:48:33 UTC

(hide)

Description:

Filename:

MIME Type:

Creator: Colm

Created: 2011-10-04 20:48:33 UTC

Size: 9.84 KB

patch

obsolete

>	.file	"TestEigen.cpp"
>	.text
>	.p2align 4,,15
>	.def	__tcf_0;	.scl	3;	.type	32;	.endef
>	.seh_proc	__tcf_0
>__tcf_0:
>	.seh_endprologue
>	leaq	_ZStL8__ioinit(%rip), %rcx
>	jmp	_ZNSt8ios_base4InitD1Ev
>	.seh_endproc
>	.section	.text$_ZNK5Eigen17CoeffBasedProductIRKNS_6MatrixIdLi4ELi4ELi0ELi4ELi4EEES4_Li6EEcvS4_Ev,"x"
>	.linkonce discard
>	.align 2
>	.p2align 4,,15
>	.globl	_ZNK5Eigen17CoeffBasedProductIRKNS_6MatrixIdLi4ELi4ELi0ELi4ELi4EEES4_Li6EEcvS4_Ev
>	.def	_ZNK5Eigen17CoeffBasedProductIRKNS_6MatrixIdLi4ELi4ELi0ELi4ELi4EEES4_Li6EEcvS4_Ev;	.scl	2;	.type	32;	.endef
>	.seh_proc	_ZNK5Eigen17CoeffBasedProductIRKNS_6MatrixIdLi4ELi4ELi0ELi4ELi4EEES4_Li6EEcvS4_Ev
>_ZNK5Eigen17CoeffBasedProductIRKNS_6MatrixIdLi4ELi4ELi0ELi4ELi4EEES4_Li6EEcvS4_Ev:
>	.seh_endprologue
>	movq	8(%rcx), %r8
>	movq	(%rcx), %rdx
>	leaq	16(%rcx), %rax
>	movsd	(%r8), %xmm1
>	movsd	8(%r8), %xmm0
>	unpcklpd	%xmm1, %xmm1
>	unpcklpd	%xmm0, %xmm0
>	mulpd	(%rdx), %xmm1
>	mulpd	32(%rdx), %xmm0
>	addpd	%xmm1, %xmm0
>	movsd	16(%r8), %xmm1
>	unpcklpd	%xmm1, %xmm1
>	mulpd	64(%rdx), %xmm1
>	addpd	%xmm0, %xmm1
>	movsd	24(%r8), %xmm0
>	unpcklpd	%xmm0, %xmm0
>	mulpd	96(%rdx), %xmm0
>	addpd	%xmm1, %xmm0
>	movapd	%xmm0, 16(%rcx)
>	movsd	(%r8), %xmm1
>	movsd	8(%r8), %xmm0
>	unpcklpd	%xmm1, %xmm1
>	unpcklpd	%xmm0, %xmm0
>	mulpd	16(%rdx), %xmm1
>	mulpd	48(%rdx), %xmm0
>	addpd	%xmm1, %xmm0
>	movsd	16(%r8), %xmm1
>	unpcklpd	%xmm1, %xmm1
>	mulpd	80(%rdx), %xmm1
>	addpd	%xmm0, %xmm1
>	movsd	24(%r8), %xmm0
>	unpcklpd	%xmm0, %xmm0
>	mulpd	112(%rdx), %xmm0
>	addpd	%xmm1, %xmm0
>	movapd	%xmm0, 32(%rcx)
>	movsd	32(%r8), %xmm1
>	movsd	40(%r8), %xmm0
>	unpcklpd	%xmm1, %xmm1
>	unpcklpd	%xmm0, %xmm0
>	mulpd	(%rdx), %xmm1
>	mulpd	32(%rdx), %xmm0
>	addpd	%xmm1, %xmm0
>	movsd	48(%r8), %xmm1
>	unpcklpd	%xmm1, %xmm1
>	mulpd	64(%rdx), %xmm1
>	addpd	%xmm0, %xmm1
>	movsd	56(%r8), %xmm0
>	unpcklpd	%xmm0, %xmm0
>	mulpd	96(%rdx), %xmm0
>	addpd	%xmm1, %xmm0
>	movapd	%xmm0, 48(%rcx)
>	movsd	32(%r8), %xmm1
>	movsd	40(%r8), %xmm0
>	unpcklpd	%xmm1, %xmm1
>	unpcklpd	%xmm0, %xmm0
>	mulpd	16(%rdx), %xmm1
>	mulpd	48(%rdx), %xmm0
>	addpd	%xmm1, %xmm0
>	movsd	48(%r8), %xmm1
>	unpcklpd	%xmm1, %xmm1
>	mulpd	80(%rdx), %xmm1
>	addpd	%xmm0, %xmm1
>	movsd	56(%r8), %xmm0
>	unpcklpd	%xmm0, %xmm0
>	mulpd	112(%rdx), %xmm0
>	addpd	%xmm1, %xmm0
>	movapd	%xmm0, 64(%rcx)
>	movsd	64(%r8), %xmm1
>	movsd	72(%r8), %xmm0
>	unpcklpd	%xmm1, %xmm1
>	unpcklpd	%xmm0, %xmm0
>	mulpd	(%rdx), %xmm1
>	mulpd	32(%rdx), %xmm0
>	addpd	%xmm1, %xmm0
>	movsd	80(%r8), %xmm1
>	unpcklpd	%xmm1, %xmm1
>	mulpd	64(%rdx), %xmm1
>	addpd	%xmm0, %xmm1
>	movsd	88(%r8), %xmm0
>	unpcklpd	%xmm0, %xmm0
>	mulpd	96(%rdx), %xmm0
>	addpd	%xmm1, %xmm0
>	movapd	%xmm0, 80(%rcx)
>	movsd	64(%r8), %xmm1
>	movsd	72(%r8), %xmm0
>	unpcklpd	%xmm1, %xmm1
>	unpcklpd	%xmm0, %xmm0
>	mulpd	16(%rdx), %xmm1
>	mulpd	48(%rdx), %xmm0
>	addpd	%xmm1, %xmm0
>	movsd	80(%r8), %xmm1
>	unpcklpd	%xmm1, %xmm1
>	mulpd	80(%rdx), %xmm1
>	addpd	%xmm0, %xmm1
>	movsd	88(%r8), %xmm0
>	unpcklpd	%xmm0, %xmm0
>	mulpd	112(%rdx), %xmm0
>	addpd	%xmm1, %xmm0
>	movapd	%xmm0, 96(%rcx)
>	movsd	96(%r8), %xmm1
>	movsd	104(%r8), %xmm0
>	unpcklpd	%xmm1, %xmm1
>	unpcklpd	%xmm0, %xmm0
>	mulpd	(%rdx), %xmm1
>	mulpd	32(%rdx), %xmm0
>	addpd	%xmm1, %xmm0
>	movsd	112(%r8), %xmm1
>	unpcklpd	%xmm1, %xmm1
>	mulpd	64(%rdx), %xmm1
>	addpd	%xmm0, %xmm1
>	movsd	120(%r8), %xmm0
>	unpcklpd	%xmm0, %xmm0
>	mulpd	96(%rdx), %xmm0
>	addpd	%xmm1, %xmm0
>	movapd	%xmm0, 112(%rcx)
>	movsd	96(%r8), %xmm1
>	movsd	104(%r8), %xmm0
>	unpcklpd	%xmm1, %xmm1
>	unpcklpd	%xmm0, %xmm0
>	mulpd	16(%rdx), %xmm1
>	mulpd	48(%rdx), %xmm0
>	addpd	%xmm1, %xmm0
>	movsd	112(%r8), %xmm1
>	unpcklpd	%xmm1, %xmm1
>	mulpd	80(%rdx), %xmm1
>	addpd	%xmm0, %xmm1
>	movsd	120(%r8), %xmm0
>	unpcklpd	%xmm0, %xmm0
>	mulpd	112(%rdx), %xmm0
>	addpd	%xmm1, %xmm0
>	movapd	%xmm0, 128(%rcx)
>	ret
>	.seh_endproc
>	.section	.text$_ZN5Eigen8internal40assign_LinearTraversal_CompleteUnrollingINS_6MatrixIdLi4ELi4ELi0ELi4ELi4EEENS_14CwiseNullaryOpINS0_16scalar_random_opIdEES3_EELi3ELi16EE3runERS3_RKS7_,"x"
>	.linkonce discard
>	.p2align 4,,15
>	.globl	_ZN5Eigen8internal40assign_LinearTraversal_CompleteUnrollingINS_6MatrixIdLi4ELi4ELi0ELi4ELi4EEENS_14CwiseNullaryOpINS0_16scalar_random_opIdEES3_EELi3ELi16EE3runERS3_RKS7_
>	.def	_ZN5Eigen8internal40assign_LinearTraversal_CompleteUnrollingINS_6MatrixIdLi4ELi4ELi0ELi4ELi4EEENS_14CwiseNullaryOpINS0_16scalar_random_opIdEES3_EELi3ELi16EE3runERS3_RKS7_;	.scl	2;	.type	32;	.endef
>	.seh_proc	_ZN5Eigen8internal40assign_LinearTraversal_CompleteUnrollingINS_6MatrixIdLi4ELi4ELi0ELi4ELi4EEENS_14CwiseNullaryOpINS0_16scalar_random_opIdEES3_EELi3ELi16EE3runERS3_RKS7_
>_ZN5Eigen8internal40assign_LinearTraversal_CompleteUnrollingINS_6MatrixIdLi4ELi4ELi0ELi4ELi4EEENS_14CwiseNullaryOpINS0_16scalar_random_opIdEES3_EELi3ELi16EE3runERS3_RKS7_:
>	pushq	%rbx
>	.seh_pushreg	%rbx
>	subq	$64, %rsp
>	.seh_stackalloc	64
>	movaps	%xmm6, 32(%rsp)
>	.seh_savexmm	%xmm6, 32
>	movaps	%xmm7, 48(%rsp)
>	.seh_savexmm	%xmm7, 48
>	.seh_endprologue
>	movq	%rcx, %rbx
>	call	rand
>	cvtsi2sd	%eax, %xmm0
>	movsd	.LC0(%rip), %xmm7
>	movsd	.LC1(%rip), %xmm6
>	addsd	%xmm0, %xmm0
>	divsd	%xmm7, %xmm0
>	subsd	%xmm6, %xmm0
>	movsd	%xmm0, 24(%rbx)
>	call	rand
>	cvtsi2sd	%eax, %xmm0
>	addsd	%xmm0, %xmm0
>	divsd	%xmm7, %xmm0
>	subsd	%xmm6, %xmm0
>	movsd	%xmm0, 32(%rbx)
>	call	rand
>	cvtsi2sd	%eax, %xmm0
>	addsd	%xmm0, %xmm0
>	divsd	%xmm7, %xmm0
>	subsd	%xmm6, %xmm0
>	movsd	%xmm0, 40(%rbx)
>	call	rand
>	cvtsi2sd	%eax, %xmm0
>	addsd	%xmm0, %xmm0
>	divsd	%xmm7, %xmm0
>	subsd	%xmm6, %xmm0
>	movsd	%xmm0, 48(%rbx)
>	call	rand
>	cvtsi2sd	%eax, %xmm0
>	addsd	%xmm0, %xmm0
>	divsd	%xmm7, %xmm0
>	subsd	%xmm6, %xmm0
>	movsd	%xmm0, 56(%rbx)
>	call	rand
>	cvtsi2sd	%eax, %xmm0
>	addsd	%xmm0, %xmm0
>	divsd	%xmm7, %xmm0
>	subsd	%xmm6, %xmm0
>	movsd	%xmm0, 64(%rbx)
>	call	rand
>	cvtsi2sd	%eax, %xmm0
>	addsd	%xmm0, %xmm0
>	divsd	%xmm7, %xmm0
>	subsd	%xmm6, %xmm0
>	movsd	%xmm0, 72(%rbx)
>	call	rand
>	cvtsi2sd	%eax, %xmm0
>	addsd	%xmm0, %xmm0
>	divsd	%xmm7, %xmm0
>	subsd	%xmm6, %xmm0
>	movsd	%xmm0, 80(%rbx)
>	call	rand
>	cvtsi2sd	%eax, %xmm0
>	addsd	%xmm0, %xmm0
>	divsd	%xmm7, %xmm0
>	subsd	%xmm6, %xmm0
>	movsd	%xmm0, 88(%rbx)
>	call	rand
>	cvtsi2sd	%eax, %xmm0
>	addsd	%xmm0, %xmm0
>	divsd	%xmm7, %xmm0
>	subsd	%xmm6, %xmm0
>	movsd	%xmm0, 96(%rbx)
>	call	rand
>	cvtsi2sd	%eax, %xmm0
>	addsd	%xmm0, %xmm0
>	divsd	%xmm7, %xmm0
>	subsd	%xmm6, %xmm0
>	movsd	%xmm0, 104(%rbx)
>	call	rand
>	cvtsi2sd	%eax, %xmm0
>	addsd	%xmm0, %xmm0
>	divsd	%xmm7, %xmm0
>	subsd	%xmm6, %xmm0
>	movsd	%xmm0, 112(%rbx)
>	call	rand
>	cvtsi2sd	%eax, %xmm0
>	addsd	%xmm0, %xmm0
>	divsd	%xmm7, %xmm0
>	movaps	48(%rsp), %xmm7
>	subsd	%xmm6, %xmm0
>	movaps	32(%rsp), %xmm6
>	movsd	%xmm0, 120(%rbx)
>	addq	$64, %rsp
>	popq	%rbx
>	ret
>	.seh_endproc
>	.def	__main;	.scl	2;	.type	32;	.endef
>	.section .rdata,"dr"
>.LC3:
>	.ascii "Time taken: \0"
>	.section	.text.startup,"x"
>	.p2align 4,,15
>	.globl	main
>	.def	main;	.scl	2;	.type	32;	.endef
>	.seh_proc	main
>main:
>	pushq	%rsi
>	.seh_pushreg	%rsi
>	pushq	%rbx
>	.seh_pushreg	%rbx
>	subq	$632, %rsp
>	.seh_stackalloc	632
>	movaps	%xmm6, 592(%rsp)
>	.seh_savexmm	%xmm6, 592
>	movaps	%xmm7, 608(%rsp)
>	.seh_savexmm	%xmm7, 608
>	.seh_endprologue
>	call	__main
>	call	rand
>	cvtsi2sd	%eax, %xmm0
>	leaq	320(%rsp), %rbx
>	movsd	.LC0(%rip), %xmm7
>	movsd	.LC1(%rip), %xmm6
>	addsd	%xmm0, %xmm0
>	divsd	%xmm7, %xmm0
>	subsd	%xmm6, %xmm0
>	movsd	%xmm0, 320(%rsp)
>	call	rand
>	cvtsi2sd	%eax, %xmm0
>	addsd	%xmm0, %xmm0
>	divsd	%xmm7, %xmm0
>	subsd	%xmm6, %xmm0
>	movsd	%xmm0, 328(%rsp)
>	call	rand
>	cvtsi2sd	%eax, %xmm0
>	leaq	576(%rsp), %rdx
>	movq	%rbx, %rcx
>	addsd	%xmm0, %xmm0
>	divsd	%xmm7, %xmm0
>	subsd	%xmm6, %xmm0
>	movsd	%xmm0, 336(%rsp)
>	call	_ZN5Eigen8internal40assign_LinearTraversal_CompleteUnrollingINS_6MatrixIdLi4ELi4ELi0ELi4ELi4EEENS_14CwiseNullaryOpINS0_16scalar_random_opIdEES3_EELi3ELi16EE3runERS3_RKS7_
>	leaq	32(%rsp), %rcx
>	movq	%rbx, 32(%rsp)
>	movq	%rbx, 40(%rsp)
>	call	_ZNK5Eigen17CoeffBasedProductIRKNS_6MatrixIdLi4ELi4ELi0ELi4ELi4EEES4_Li6EEcvS4_Ev
>	movapd	(%rax), %xmm0
>	leaq	176(%rsp), %rcx
>	movapd	%xmm0, 448(%rsp)
>	movapd	16(%rax), %xmm0
>	movapd	%xmm0, 464(%rsp)
>	movapd	32(%rax), %xmm0
>	movapd	%xmm0, 480(%rsp)
>	movapd	48(%rax), %xmm0
>	movapd	%xmm0, 496(%rsp)
>	movapd	64(%rax), %xmm0
>	movapd	%xmm0, 512(%rsp)
>	movapd	80(%rax), %xmm0
>	movapd	%xmm0, 528(%rsp)
>	movapd	96(%rax), %xmm0
>	movapd	%xmm0, 544(%rsp)
>	movapd	112(%rax), %xmm0
>	leaq	448(%rsp), %rax
>	movapd	%xmm0, 560(%rsp)
>	movq	%rax, 176(%rsp)
>	movq	%rax, 184(%rsp)
>	call	_ZNK5Eigen17CoeffBasedProductIRKNS_6MatrixIdLi4ELi4ELi0ELi4ELi4EEES4_Li6EEcvS4_Ev
>	call	clock
>	movl	%eax, %ebx
>/APP
> # 26 ".\TestEigen.cpp" 1
>	#mybegin
> # 0 "" 2
>/NO_APP
>	movsd	.LC2(%rip), %xmm1
>	xorl	%eax, %eax
>	.p2align 4,,10
>.L5:
>	addl	$1, %eax
>	cvtsi2sd	%eax, %xmm0
>	ucomisd	%xmm0, %xmm1
>	ja	.L5
>/APP
> # 30 ".\TestEigen.cpp" 1
>	#myend
> # 0 "" 2
>/NO_APP
>	call	clock
>	leaq	.LC3(%rip), %rdx
>	movl	%eax, %esi
>	leaq	_ZSt4cout(%rip), %rcx
>	subl	%ebx, %esi
>	call	_ZStlsISt11char_traitsIcEERSt13basic_ostreamIcT_ES5_PKc
>	cvtsi2sd	%esi, %xmm1
>	movq	%rax, %rcx
>	divsd	.LC4(%rip), %xmm1
>	call	_ZNSo9_M_insertIdEERSoT_
>	movq	%rax, %rcx
>	call	_ZSt4endlIcSt11char_traitsIcEERSt13basic_ostreamIT_T0_ES6_
>	nop
>	movaps	592(%rsp), %xmm6
>	xorl	%eax, %eax
>	movaps	608(%rsp), %xmm7
>	addq	$632, %rsp
>	popq	%rbx
>	popq	%rsi
>	ret
>	.seh_endproc
>	.p2align 4,,15
>	.def	_GLOBAL__sub_I_main;	.scl	3;	.type	32;	.endef
>	.seh_proc	_GLOBAL__sub_I_main
>_GLOBAL__sub_I_main:
>	subq	$40, %rsp
>	.seh_stackalloc	40
>	.seh_endprologue
>	leaq	_ZStL8__ioinit(%rip), %rcx
>	call	_ZNSt8ios_base4InitC1Ev
>	leaq	__tcf_0(%rip), %rcx
>	addq	$40, %rsp
>	jmp	atexit
>	.seh_endproc
>	.section	.ctors,"w"
>	.align 8
>	.quad	_GLOBAL__sub_I_main
>.lcomm _ZStL8__ioinit,1,1
>	.section .rdata,"dr"
>	.align 8
>.LC0:
>	.long	0
>	.long	1088421824
>	.align 8
>.LC1:
>	.long	0
>	.long	1072693248
>	.align 8
>.LC2:
>	.long	0
>	.long	1097011920
>	.align 8
>.LC4:
>	.long	0
>	.long	1083129856
>	.def	_ZNSt8ios_base4InitD1Ev;	.scl	2;	.type	32;	.endef
>	.def	rand;	.scl	2;	.type	32;	.endef
>	.def	clock;	.scl	2;	.type	32;	.endef
>	.def	_ZStlsISt11char_traitsIcEERSt13basic_ostreamIcT_ES5_PKc;	.scl	2;	.type	32;	.endef
>	.def	_ZNSo9_M_insertIdEERSoT_;	.scl	2;	.type	32;	.endef
>	.def	_ZSt4endlIcSt11char_traitsIcEERSt13basic_ostreamIT_T0_ES6_;	.scl	2;	.type	32;	.endef
>	.def	_ZNSt8ios_base4InitC1Ev;	.scl	2;	.type	32;	.endef
>	.def	atexit;	.scl	2;	.type	32;	.endef

Actions: View

Attachments on bug 357: 216 | 217