SFBCOPYALL1
lmaskd = SFBCOPYALL1 << alignd;
rmaskd = SFBCOPYALL1 >> (-(alignd + w) & SFBCOPYBITMASK);
MEMWRITE32(sp, SFBCOPYALL1); WRITE_MB();
lmaskd = SFBCOPYALL1 << alignd;
rmaskd = SFBCOPYALL1 >> (-width & SFBCOPYBITMASK);
MEMWRITE32(sp, SFBCOPYALL1); WRITE_MB();
MEMWRITE32(sp, SFBCOPYALL1); WRITE_MB();
MEMWRITE32(dp, SFBCOPYALL1); WRITE_MB();
MEMWRITE32(sp, SFBCOPYALL1); WRITE_MB();
lmaskd = SFBCOPYALL1 << alignd;
rmaskd = SFBCOPYALL1 >> (-width & SFBCOPYBITMASK);
MEMWRITE32(sp, SFBCOPYALL1); WRITE_MB();
MEMWRITE32(sp, SFBCOPYALL1); WRITE_MB();
MEMWRITE32(dp, SFBCOPYALL1); WRITE_MB();
MEMWRITE32(sp, SFBCOPYALL1); WRITE_MB();
lmask = SFBCOPYALL1 << align;
rmask = SFBCOPYALL1 >> (-width & SFBCOPYBITMASK);
MEMWRITE32(p, SFBCOPYALL1);
MEMWRITE32(p + offset, SFBCOPYALL1);
#define SFBSTIPPLEALL1 SFBCOPYALL1
lmask = SFBCOPYALL1 << align;
rmask = SFBCOPYALL1 >> (-width & SFBCOPYBITMASK);
*(uint32_t *)p = SFBCOPYALL1;
*(uint32_t *)(p + offset) = SFBCOPYALL1;
lmask = SFBCOPYALL1 << align;
rmask = SFBCOPYALL1 >> (-width & SFBCOPYBITMASK);
*(uint32_t *)p = SFBCOPYALL1;
*(uint32_t *)(p + offset) = SFBCOPYALL1;