--- rss_glx-0.7.4/reallyslick/c_src/rsMath.c.orig 2003-03-02 01:56:28 +0200 +++ rss_glx-0.7.4/reallyslick/c_src/rsMath.c 2003-04-03 18:57:03 +0300 @@ -89,22 +89,22 @@ " femms - movq (%1), %%mm0 /* 1 | 0 */ - movq 8(%1), %%mm1 /* - | 2 */ + movq (%1), %%mm0 + movq 8(%1), %%mm1 - movq %%mm0, %%mm2 /* 1 | 0 */ - movq %%mm1, %%mm3 /* - | 2 */ + movq %%mm0, %%mm2 + movq %%mm1, %%mm3 - pfmul %%mm0, %%mm0 /* 1 * 1 | 0 * 0 */ - pfmul %%mm1, %%mm1 /* - | 2 * 2 */ + pfmul %%mm0, %%mm0 + pfmul %%mm1, %%mm1 - movq %%mm0, %%mm7 /* 1 * 1 | 0 * 0 */ - punpckhdq %%mm7, %%mm7 /* 1 * 1 | 1 * 1 */ + movq %%mm0, %%mm7 + punpckhdq %%mm7, %%mm7 - pfadd %%mm1, %%mm0 /* - | 0 * 0 + 2 * 2 */ - pfadd %%mm7, %%mm0 /* - | 0 * 0 + 2 * 2 + 1 * 1 */ + pfadd %%mm1, %%mm0 + pfadd %%mm7, %%mm0 - pfrsqrt %%mm0, %%mm1 /* 24-bit sqrt */ + pfrsqrt %%mm0, %%mm1 movq %%mm1, %%mm4 pfmul %%mm1, %%mm1 punpckldq %%mm0, %%mm0 @@ -112,19 +112,19 @@ pfrcpit2 %%mm4, %%mm1 pfmul %%mm1, %%mm0 - movd %%mm0, %0 /* length */ + movd %%mm0, %0 - pfrcp %%mm0, %%mm1 /* 24-bit reciprocal */ + pfrcp %%mm0, %%mm1 movq %%mm0, %%mm4 punpckldq %%mm4, %%mm4 pfrcpit1 %%mm1, %%mm4 pfrcpit2 %%mm1, %%mm4 - pfmul %%mm4, %%mm2 /* 1 / length | 0 / length */ - pfmul %%mm4, %%mm3 /* - | 2 / length */ + pfmul %%mm4, %%mm2 + pfmul %%mm4, %%mm3 - movq %%mm2, (%1) /* 1 | 0 */ - movq %%mm3, 8(%1) /* - | 2 */ + movq %%mm2, (%1) + movq %%mm3, 8(%1) femms " @@ -140,25 +140,25 @@ if (cpuid & FLAG_SSE) { __asm__ ( " - movups (%1), %%xmm0 /* 0 | 1 | 2 | - */ - movups %%xmm0, %%xmm1 /* 0 | 1 | 2 | - */ + movups (%1), %%xmm0 + movups %%xmm0, %%xmm1 - mulps %%xmm0, %%xmm0 /* 0 * 0 | 1 * 1 | 2 * 2 | - */ - movups %%xmm0, %%xmm2 /* 0 * 0 | 1 * 1 | 2 * 2 | - */ + mulps %%xmm0, %%xmm0 + movups %%xmm0, %%xmm2 - shufps $9, %%xmm2, %%xmm2 /* 1 * 1 | 2 * 2 | - | - */ - addss %%xmm2, %%xmm0 /* 0 * 0 + 1 * 1 | - | - | - */ - shufps $1, %%xmm2, %%xmm2 /* 2 * 2 | - | - | - */ - addss %%xmm2, %%xmm0 /* 0 * 0 + 1 * 1 + 2 * 2 | - | - | - */ + shufps $9, %%xmm2, %%xmm2 + addss %%xmm2, %%xmm0 + shufps $1, %%xmm2, %%xmm2 + addss %%xmm2, %%xmm0 - sqrtss %%xmm0, %%xmm0 /* length | - | - | - */ - movss %%xmm0, %0 /* length */ + sqrtss %%xmm0, %%xmm0 + movss %%xmm0, %0 - unpcklps %%xmm0, %%xmm0 /* length | length | - | - */ - unpcklps %%xmm0, %%xmm0 /* length | length | length | length */ - divps %%xmm0, %%xmm1 /* 1 / length | 2 / length | 3 / length | - */ + unpcklps %%xmm0, %%xmm0 + unpcklps %%xmm0, %%xmm0 + divps %%xmm0, %%xmm1 - movups %%xmm1, (%1) /* 1 / length | 2 / length | 3 / length | - */ + movups %%xmm1, (%1) " : "=m" (length) : "d" (v) @@ -190,35 +190,35 @@ " femms - movq (%0), %%mm0 /* 1.1 | 1.0 */ - movq 8(%0), %%mm1 /* - | 1.2 */ + movq (%0), %%mm0 + movq 8(%0), %%mm1 - movq (%1), %%mm2 /* 2.1 | 2.0 */ - movq 8(%1), %%mm3 /* - | 2.2 */ + movq (%1), %%mm2 + movq 8(%1), %%mm3 - movq %%mm0, %%mm4 /* 1.1 | 1.0 */ - punpckhdq %%mm4, %%mm4 /* 1.1 | 1.1 */ - punpckldq %%mm1, %%mm4 /* 1.2 | 1.1 */ + movq %%mm0, %%mm4 + punpckhdq %%mm4, %%mm4 + punpckldq %%mm1, %%mm4 - movq %%mm2, %%mm5 /* 2.1 | 2.0 */ - punpckhdq %%mm2, %%mm5 /* 2.1 | 2.1 */ - punpckldq %%mm3, %%mm5 /* 2.2 | 2.1 */ + movq %%mm2, %%mm5 + punpckhdq %%mm2, %%mm5 + punpckldq %%mm3, %%mm5 - pfmul %%mm2, %%mm4 /* 1.2 * 2.1 | 2.0 * 1.1 */ - pfmul %%mm0, %%mm5 /* 1.1 * 2.2 | 1.0 * 2.1 */ + pfmul %%mm2, %%mm4 + pfmul %%mm0, %%mm5 - pfmul %%mm0, %%mm3 /* - | 1.0 * 2.2 */ - pfmul %%mm2, %%mm1 /* - | 2.0 * 1.2 */ + pfmul %%mm0, %%mm3 + pfmul %%mm2, %%mm1 - pfsub %%mm4, %%mm5 /* 0 | 2 */ - pfsub %%mm3, %%mm1 /* - | 1 */ + pfsub %%mm4, %%mm5 + pfsub %%mm3, %%mm1 - movq %%mm5, %%mm6 /* 0 | 2 */ - punpckldq %%mm1, %%mm1 /* 1 | 1 */ - punpckhdq %%mm1, %%mm6 /* 1 | 0 */ + movq %%mm5, %%mm6 + punpckldq %%mm1, %%mm1 + punpckhdq %%mm1, %%mm6 - movq %%mm6, (%2) /* 1 | 0 */ - movq %%mm5, 8(%2) /* - | 2 */ + movq %%mm6, (%2) + movq %%mm5, 8(%2) femms " @@ -234,22 +234,22 @@ if (cpuid & FLAG_SSE) { __asm__ ( " - movups (%0), %%xmm0 /* 1.0 | 1.1 | 1.2 | - */ - movups (%1), %%xmm1 /* 2.0 | 2.1 | 2.2 | - */ + movups (%0), %%xmm0 + movups (%1), %%xmm1 - movups %%xmm0, %%xmm2 /* 1.0 | 1.1 | 1.2 | - */ - shufps $201, %%xmm2, %%xmm2 /* 1.1 | 1.2 | 1.0 | - */ - movups %%xmm1, %%xmm3 /* 2.0 | 2.1 | 2.2 | - */ - shufps $201, %%xmm3, %%xmm3 /* 2.1 | 2.2 | 2.0 | - */ + movups %%xmm0, %%xmm2 + shufps $201, %%xmm2, %%xmm2 + movups %%xmm1, %%xmm3 + shufps $201, %%xmm3, %%xmm3 - mulps %%xmm1, %%xmm2 /* 1.1 * 2.0 | 1.2 * 2.1 | 1.0 * 2.2 | - */ - mulps %%xmm0, %%xmm3 /* 2.1 * 1.0 | 2.2 * 1.1 | 2.0 * 1.2 | - */ + mulps %%xmm1, %%xmm2 + mulps %%xmm0, %%xmm3 - subps %%xmm2, %%xmm3 /* 2 | 0 | 1 | - */ + subps %%xmm2, %%xmm3 - shufps $201, %%xmm3, %%xmm3 /* 0 | 1 | 2 | - */ + shufps $201, %%xmm3, %%xmm3 - movups %%xmm3, (%2) /* 0 | 1 | 2 | - */ + movups %%xmm3, (%2) " : : "a" (vec1), "c" (vec2), "d" (v)