To: J3 J3/14-205r1 From: John Reid Subject: FAILED_IMAGES intrinsic Date: 2014 June 26 References: N2007, N2013 Discussion ---------- Malcolm Cohen has suggested that FAILED_IMAGES should (1) talk about the number of images "known by the invoking image as having failed", otherwise this has to return the same results on all images, even ones in different teams which cannot know this fact without either substantial "image failure handling and propagation" machinery (which I think unlikely to exist) or similar; and (2) say that after an image control statement that returned a STAT_FAILED_IMAGE result, the return value includes the index of at least one image in the set of images being communicated with (and that failed). I think that perhaps it should not be required to return the indices of all the images that have failed since the communications protocol is likely to return STAT_FAILED_IMAGE as soon as one image has failed rather than wait to see if there are more that have failed. Corresponding edits to FAILED_IMAGES are included here. Edits to N2007: --------------- [24:14] Change "failed images in the specified team" to "images in the specified team that are known by the invoking image to have failed". [24:16] Change "failed images" to "known failed images". [24:17] Append "If the executing image has previously executed an image control statement whose STAT= specifier assigned the value STAT_FAILED_IMAGE or invoked a collective subroutine whose STAT argument was set to STAT_FAILED_IMAGE and has not meanwhile entered or left a CHANGE TEAM construct, at least one image in the set of images participating in that image control statement or collective invocation shall be known to have failed." [24:19] Change "failed images in the current team" to "images in the current team that are known by the invoking image to have failed".