https://bugs.kde.org/show_bug.cgi?id=385409
--- Comment #44 from Julian Seward <jsew...@acm.org> --- (In reply to Vadim Barkov from comment #43) > What do you think about it? Honestly, I do not understand. You need to explain much more clearly: * what these instructions actually do * what your implementation strategy is * what the problem cases are I have the impression from comment 34 that vfenezbs %v0, %v0, %v0 finds the index of the first zero in the operand and writes that somewhere in the result register. (But why do they all have to to be v0 ?). We have solved such problems in the past for amd64 and so it would help if you could relate your work to the amd64 implementation. For amd64, we generate the sequence t1 = CmpEQ8x16(vec, zero-vector) t2 = pmovmskb(t1), which moves one bit from each lane into t2 t3 = count-leading-zeroes(t2) Because the count-leading-zeroes operation can be done in a way in which the result is defined even if there are undefined bits to the right of the leftmost 1 bit, this means that the resulting t3 value will be defined if the input vector consisted of defined non-zero bytes terminated by a defined-zero, even if the bytes after that are undefined. -- You are receiving this mail because: You are watching all bug changes.
