📄 zscal_sse2_core2.s
字号:
movapd %xmm2, (XX) addq INCX, XX movapd %xmm4, (XX) addq INCX, XX movapd %xmm6, (XX) addq INCX, XX decq I jg .L121 ALIGN_4.L125: testq $7, M je .L999 testq $4, M je .L126 movapd (X), %xmm0 pshufd $0x4e, %xmm0, %xmm1 addq INCX, X movapd (X), %xmm2 pshufd $0x4e, %xmm2, %xmm3 addq INCX, X movapd (X), %xmm4 pshufd $0x4e, %xmm4, %xmm5 addq INCX, X movapd (X), %xmm6 pshufd $0x4e, %xmm6, %xmm7 addq INCX, X mulpd %xmm14, %xmm0 mulpd %xmm15, %xmm1 mulpd %xmm14, %xmm2 mulpd %xmm15, %xmm3 mulpd %xmm14, %xmm4 mulpd %xmm15, %xmm5 mulpd %xmm14, %xmm6 mulpd %xmm15, %xmm7 addpd %xmm1, %xmm0 addpd %xmm3, %xmm2 addpd %xmm5, %xmm4 addpd %xmm7, %xmm6 movapd %xmm0, (XX) addq INCX, XX movapd %xmm2, (XX) addq INCX, XX movapd %xmm4, (XX) addq INCX, XX movapd %xmm6, (XX) addq INCX, XX ALIGN_3.L126: testq $2, M je .L127 movapd (X), %xmm0 pshufd $0x4e, %xmm0, %xmm1 addq INCX, X movapd (X), %xmm2 pshufd $0x4e, %xmm2, %xmm3 addq INCX, X mulpd %xmm14, %xmm0 mulpd %xmm15, %xmm1 mulpd %xmm14, %xmm2 mulpd %xmm15, %xmm3 addpd %xmm1, %xmm0 addpd %xmm3, %xmm2 movapd %xmm0, (XX) addq INCX, XX movapd %xmm2, (XX) addq INCX, XX ALIGN_3.L127: testq $1, M je .L999 movapd (X), %xmm0 pshufd $0x4e, %xmm0, %xmm1 mulpd %xmm14, %xmm0 mulpd %xmm15, %xmm1 addpd %xmm1, %xmm0 movapd %xmm0, (XX) jmp .L999 ALIGN_3/* Unaligned */.L200: cmpq $2 * SIZE, INCX jne .L220.L210: movq M, I # rcx = n sarq $3, I jle .L215 ALIGN_4.L211: movsd 0 * SIZE(X), %xmm0 movhpd 1 * SIZE(X), %xmm0 pshufd $0x4e, %xmm0, %xmm1 movsd 2 * SIZE(X), %xmm2 movhpd 3 * SIZE(X), %xmm2 pshufd $0x4e, %xmm2, %xmm3 movsd 4 * SIZE(X), %xmm4 movhpd 5 * SIZE(X), %xmm4 pshufd $0x4e, %xmm4, %xmm5 movsd 6 * SIZE(X), %xmm6 movhpd 7 * SIZE(X), %xmm6 pshufd $0x4e, %xmm6, %xmm7 mulpd %xmm14, %xmm0 mulpd %xmm15, %xmm1 mulpd %xmm14, %xmm2 mulpd %xmm15, %xmm3 mulpd %xmm14, %xmm4 mulpd %xmm15, %xmm5 mulpd %xmm14, %xmm6 mulpd %xmm15, %xmm7 addpd %xmm1, %xmm0 addpd %xmm3, %xmm2 addpd %xmm5, %xmm4 addpd %xmm7, %xmm6 movlpd %xmm0, 0 * SIZE(X) movhpd %xmm0, 1 * SIZE(X) movlpd %xmm2, 2 * SIZE(X) movhpd %xmm2, 3 * SIZE(X) movlpd %xmm4, 4 * SIZE(X) movhpd %xmm4, 5 * SIZE(X) movlpd %xmm6, 6 * SIZE(X) movhpd %xmm6, 7 * SIZE(X) movsd 8 * SIZE(X), %xmm0 movhpd 9 * SIZE(X), %xmm0 pshufd $0x4e, %xmm0, %xmm1 movsd 10 * SIZE(X), %xmm2 movhpd 11 * SIZE(X), %xmm2 pshufd $0x4e, %xmm2, %xmm3 movsd 12 * SIZE(X), %xmm4 movhpd 13 * SIZE(X), %xmm4 pshufd $0x4e, %xmm4, %xmm5 movsd 14 * SIZE(X), %xmm6 movhpd 15 * SIZE(X), %xmm6 pshufd $0x4e, %xmm6, %xmm7 mulpd %xmm14, %xmm0 mulpd %xmm15, %xmm1 mulpd %xmm14, %xmm2 mulpd %xmm15, %xmm3 mulpd %xmm14, %xmm4 mulpd %xmm15, %xmm5 mulpd %xmm14, %xmm6 mulpd %xmm15, %xmm7 addpd %xmm1, %xmm0 addpd %xmm3, %xmm2 addpd %xmm5, %xmm4 addpd %xmm7, %xmm6 movlpd %xmm0, 8 * SIZE(X) movhpd %xmm0, 9 * SIZE(X) movlpd %xmm2, 10 * SIZE(X) movhpd %xmm2, 11 * SIZE(X) movlpd %xmm4, 12 * SIZE(X) movhpd %xmm4, 13 * SIZE(X) movlpd %xmm6, 14 * SIZE(X) movhpd %xmm6, 15 * SIZE(X) addq $16 * SIZE, X decq I jg .L211 ALIGN_4.L215: testq $7, M je .L999 testq $4, M je .L216 movsd 0 * SIZE(X), %xmm0 movhpd 1 * SIZE(X), %xmm0 pshufd $0x4e, %xmm0, %xmm1 movsd 2 * SIZE(X), %xmm2 movhpd 3 * SIZE(X), %xmm2 pshufd $0x4e, %xmm2, %xmm3 movsd 4 * SIZE(X), %xmm4 movhpd 5 * SIZE(X), %xmm4 pshufd $0x4e, %xmm4, %xmm5 movsd 6 * SIZE(X), %xmm6 movhpd 7 * SIZE(X), %xmm6 pshufd $0x4e, %xmm6, %xmm7 mulpd %xmm14, %xmm0 mulpd %xmm15, %xmm1 mulpd %xmm14, %xmm2 mulpd %xmm15, %xmm3 mulpd %xmm14, %xmm4 mulpd %xmm15, %xmm5 mulpd %xmm14, %xmm6 mulpd %xmm15, %xmm7 addpd %xmm1, %xmm0 addpd %xmm3, %xmm2 addpd %xmm5, %xmm4 addpd %xmm7, %xmm6 movlpd %xmm0, 0 * SIZE(X) movhpd %xmm0, 1 * SIZE(X) movlpd %xmm2, 2 * SIZE(X) movhpd %xmm2, 3 * SIZE(X) movlpd %xmm4, 4 * SIZE(X) movhpd %xmm4, 5 * SIZE(X) movlpd %xmm6, 6 * SIZE(X) movhpd %xmm6, 7 * SIZE(X) addq $8 * SIZE, X ALIGN_3.L216: testq $2, M je .L217 movsd 0 * SIZE(X), %xmm0 movhpd 1 * SIZE(X), %xmm0 pshufd $0x4e, %xmm0, %xmm1 movsd 2 * SIZE(X), %xmm2 movhpd 3 * SIZE(X), %xmm2 pshufd $0x4e, %xmm2, %xmm3 mulpd %xmm14, %xmm0 mulpd %xmm15, %xmm1 mulpd %xmm14, %xmm2 mulpd %xmm15, %xmm3 addpd %xmm1, %xmm0 addpd %xmm3, %xmm2 movlpd %xmm0, 0 * SIZE(X) movhpd %xmm0, 1 * SIZE(X) movlpd %xmm2, 2 * SIZE(X) movhpd %xmm2, 3 * SIZE(X) addq $4 * SIZE, X ALIGN_3.L217: testq $1, M je .L999 movsd 0 * SIZE(X), %xmm0 movhpd 1 * SIZE(X), %xmm0 pshufd $0x4e, %xmm0, %xmm1 mulpd %xmm14, %xmm0 mulpd %xmm15, %xmm1 addpd %xmm1, %xmm0 movlpd %xmm0, 0 * SIZE(X) movhpd %xmm0, 1 * SIZE(X) jmp .L999 ALIGN_3.L220: movq X, XX movq M, I # rcx = n sarq $3, I jle .L225 ALIGN_4.L221: movsd 0 * SIZE(X), %xmm0 movhpd 1 * SIZE(X), %xmm0 pshufd $0x4e, %xmm0, %xmm1 addq INCX, X movsd 0 * SIZE(X), %xmm2 movhpd 1 * SIZE(X), %xmm2 pshufd $0x4e, %xmm2, %xmm3 addq INCX, X movsd 0 * SIZE(X), %xmm4 movhpd 1 * SIZE(X), %xmm4 pshufd $0x4e, %xmm4, %xmm5 addq INCX, X movsd 0 * SIZE(X), %xmm6 movhpd 1 * SIZE(X), %xmm6 pshufd $0x4e, %xmm6, %xmm7 addq INCX, X mulpd %xmm14, %xmm0 mulpd %xmm15, %xmm1 mulpd %xmm14, %xmm2 mulpd %xmm15, %xmm3 mulpd %xmm14, %xmm4 mulpd %xmm15, %xmm5 mulpd %xmm14, %xmm6 mulpd %xmm15, %xmm7 addpd %xmm1, %xmm0 addpd %xmm3, %xmm2 addpd %xmm5, %xmm4 addpd %xmm7, %xmm6 movlpd %xmm0, 0 * SIZE(XX) movhpd %xmm0, 1 * SIZE(XX) addq INCX, XX movlpd %xmm2, 0 * SIZE(XX) movhpd %xmm2, 1 * SIZE(XX) addq INCX, XX movlpd %xmm4, 0 * SIZE(XX) movhpd %xmm4, 1 * SIZE(XX) addq INCX, XX movlpd %xmm6, 0 * SIZE(XX) movhpd %xmm6, 1 * SIZE(XX) addq INCX, XX movsd 0 * SIZE(X), %xmm0 movhpd 1 * SIZE(X), %xmm0 pshufd $0x4e, %xmm0, %xmm1 addq INCX, X movsd 0 * SIZE(X), %xmm2 movhpd 1 * SIZE(X), %xmm2 pshufd $0x4e, %xmm2, %xmm3 addq INCX, X movsd 0 * SIZE(X), %xmm4 movhpd 1 * SIZE(X), %xmm4 pshufd $0x4e, %xmm4, %xmm5 addq INCX, X movsd 0 * SIZE(X), %xmm6 movhpd 1 * SIZE(X), %xmm6 pshufd $0x4e, %xmm6, %xmm7 addq INCX, X mulpd %xmm14, %xmm0 mulpd %xmm15, %xmm1 mulpd %xmm14, %xmm2 mulpd %xmm15, %xmm3 mulpd %xmm14, %xmm4 mulpd %xmm15, %xmm5 mulpd %xmm14, %xmm6 mulpd %xmm15, %xmm7 addpd %xmm1, %xmm0 addpd %xmm3, %xmm2 addpd %xmm5, %xmm4 addpd %xmm7, %xmm6 movlpd %xmm0, 0 * SIZE(XX) movhpd %xmm0, 1 * SIZE(XX) addq INCX, XX movlpd %xmm2, 0 * SIZE(XX) movhpd %xmm2, 1 * SIZE(XX) addq INCX, XX movlpd %xmm4, 0 * SIZE(XX) movhpd %xmm4, 1 * SIZE(XX) addq INCX, XX movlpd %xmm6, 0 * SIZE(XX) movhpd %xmm6, 1 * SIZE(XX) addq INCX, XX decq I jg .L221 ALIGN_4.L225: testq $7, M je .L999 testq $4, M je .L226 movsd 0 * SIZE(X), %xmm0 movhpd 1 * SIZE(X), %xmm0 pshufd $0x4e, %xmm0, %xmm1 addq INCX, X movsd 0 * SIZE(X), %xmm2 movhpd 1 * SIZE(X), %xmm2 pshufd $0x4e, %xmm2, %xmm3 addq INCX, X movsd 0 * SIZE(X), %xmm4 movhpd 1 * SIZE(X), %xmm4 pshufd $0x4e, %xmm4, %xmm5 addq INCX, X movsd 0 * SIZE(X), %xmm6 movhpd 1 * SIZE(X), %xmm6 pshufd $0x4e, %xmm6, %xmm7 addq INCX, X mulpd %xmm14, %xmm0 mulpd %xmm15, %xmm1 mulpd %xmm14, %xmm2 mulpd %xmm15, %xmm3 mulpd %xmm14, %xmm4 mulpd %xmm15, %xmm5 mulpd %xmm14, %xmm6 mulpd %xmm15, %xmm7 addpd %xmm1, %xmm0 addpd %xmm3, %xmm2 addpd %xmm5, %xmm4 addpd %xmm7, %xmm6 movlpd %xmm0, 0 * SIZE(XX) movhpd %xmm0, 1 * SIZE(XX) addq INCX, XX movlpd %xmm2, 0 * SIZE(XX) movhpd %xmm2, 1 * SIZE(XX) addq INCX, XX movlpd %xmm4, 0 * SIZE(XX) movhpd %xmm4, 1 * SIZE(XX) addq INCX, XX movlpd %xmm6, 0 * SIZE(XX) movhpd %xmm6, 1 * SIZE(XX) addq INCX, XX ALIGN_3.L226: testq $2, M je .L227 movsd 0 * SIZE(X), %xmm0 movhpd 1 * SIZE(X), %xmm0 pshufd $0x4e, %xmm0, %xmm1 addq INCX, X movsd 0 * SIZE(X), %xmm2 movhpd 1 * SIZE(X), %xmm2 pshufd $0x4e, %xmm2, %xmm3 addq INCX, X mulpd %xmm14, %xmm0 mulpd %xmm15, %xmm1 mulpd %xmm14, %xmm2 mulpd %xmm15, %xmm3 addpd %xmm1, %xmm0 addpd %xmm3, %xmm2 movlpd %xmm0, 0 * SIZE(XX) movhpd %xmm0, 1 * SIZE(XX) addq INCX, XX movlpd %xmm2, 0 * SIZE(XX) movhpd %xmm2, 1 * SIZE(XX) addq INCX, XX ALIGN_3.L227: testq $1, M je .L999 movsd 0 * SIZE(X), %xmm0 movhpd 1 * SIZE(X), %xmm0 pshufd $0x4e, %xmm0, %xmm1 mulpd %xmm14, %xmm0 mulpd %xmm15, %xmm1 addpd %xmm1, %xmm0 movlpd %xmm0, 0 * SIZE(XX) movhpd %xmm0, 1 * SIZE(XX) ALIGN_3.L999: xorq %rax, %rax RESTOREREGISTERS ret EPILOGUE
⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -