{"id":2335,"date":"2019-01-20T16:12:12","date_gmt":"2019-01-20T16:12:12","guid":{"rendered":"https:\/\/nenadnoveljic.com\/blog\/?p=2335"},"modified":"2019-01-21T14:12:14","modified_gmt":"2019-01-21T14:12:14","slug":"tracking-the-optimizer-with-gdb","status":"publish","type":"post","link":"https:\/\/nenadnoveljic.com\/blog\/tracking-the-optimizer-with-gdb\/","title":{"rendered":"Tracking the Optimizer with gdb"},"content":{"rendered":"<h1>lmtdcrel<\/h1>\n<p>lmtdcrel is an Oracle database C function that compares two double-precision float-point numbers. The input values are passed in the XMM0 and XMM1 CPU registers. The result, which is returned in the $EAX register, is calculated as follows:<\/p>\n<ul>\n<li>1 if XMM0 &gt; XMM1<\/li>\n<li>0 if XMM0 == XMM1<\/li>\n<li>-1 if XMM0 &lt; XMM1<\/li>\n<\/ul>\n<p>The <a href=\"https:\/\/en.wikipedia.org\/wiki\/Significand\">significands<\/a>, a.k.a. mantissas, of both numbers in the modified normalized form are rounded to the 7th decimal place prior to comparison.<\/p>\n<p>Since the Oracle C functions aren&#8217;t documented, I had to verify a lot of executions in order to confirm that this theory holds over a broad range of values. Therefore, I automated this process with some gdb commands and a Python script.<\/p>\n<h1>Automated verification<\/h1>\n<p>The following commands at the entry and exit breakpoints, respectively, print the values of the registers involved:<\/p>\n<pre><code>silent\np $xmm0\np $xmm1\ncontinue\nend<\/code><\/pre>\n<pre><code>silent\np\/d $eax\ncontinue\nend<\/code><\/pre>\n<p>The output is spooled into the log file:<\/p>\n<pre><code>set pagination off\nset logging file lmtdcrel.log\nset logging on<\/code><\/pre>\n<p>In the excerpt from the log file below, <span style=\"color: #ff0000;\">53.161750863128702<\/span>, stored in the XMM0 register, is compared to <span style=\"color: #0000ff;\">53.186755259284759<\/span> which is stored XMM1. In this case, the function returns <span style=\"color: #993300;\">-1<\/span>, because XMM0 contains the lower value.<\/p>\n<pre><code>$5717 = {v4_float = {4.6755867, 3.16532612, 0, 0}, v2_double = {<span style=\"color: #ff0000;\">53.161750863128702<\/span>, 0}, v16_int8 = {104, -98, -107, 64, -76, -108, 74, 64, 0, 0, 0, 0, 0, 0, 0, 0}, v8_int16 = {-24984, 16533, -27468, 16458, 0, 0, 0, 0}, v4_int32 = {1083547240, 1078629556, 0, 0}, v2_int64 = {4632678668602547816, 0}, uint128 = 4632678668602547816}\n$5718 = {v4_float = {-4.38126602e-24, 3.16552138, 0, 0}, v2_double = {<span style=\"color: #0000ff;\">53.186755259284759<\/span>, 0}, v16_int8 = {-11, 125, -87, -104, -25, -105, 74, 64, 0, 0, 0, 0, 0, 0, 0, 0}, v8_int16 = {32245, -26455, -26649, 16458, 0, 0, 0, 0}, v4_int32 = {-1733722635, 1078630375, 0, 0}, v2_int64 = {4632682187658460661, 0}, uint128 = 4632682187658460661}\n$5719 = <span style=\"color: #993300;\">-1<\/span><\/code><\/pre>\n<p>The Python program <a href=\"https:\/\/github.com\/nenadnoveljic\/oradb\/blob\/master\/lmtdcrel_verifier.py\">lmtdcrel_verifier.py<\/a> parses the gdb log file, extracts the inputs and return values, simulates the function lmtdcrel and compares the simulation results with the real ones.<\/p>\n<h1>Practical application<\/h1>\n<p>Unsurprisingly, the function <span style=\"display: inline !important; float: none; background-color: transparent; color: #191e23; cursor: text; font-family: 'Noto Serif'; font-size: 16px; font-style: normal; font-variant: normal; font-weight: 400; letter-spacing: normal; orphans: 2; text-align: left; text-decoration: none; text-indent: 0px; text-transform: none; -webkit-text-stroke-width: 0px; white-space: normal; word-spacing: 0px;\">lmtdcrel<\/span> is extensively used by the cost based optimizer (CBO) for comparing the costs of different access paths. A lot of such decisions are logged into the CBO trace, which is knowingly an invaluable source of information. But the CBO doesn&#8217;t trace everything &#8211; and gdb might fill this gap. By inspecting the lmtdcrel&#8217;s arguments before making a decision we can deduce what&#8217;s driving that decision.<\/p>\n<p>In the example below, the function was called by kkqctUpdBestCostStt:<\/p>\n<pre><code>#0  0x0000000017fac680 in lmtdcrel ()\n#1  0x0000000017111b7b in kkqctUpdBestCostStt ()\n#2  0x00000000171169b4 in kkqctNxtLinearStt ()\n#3  0x000000000a0e88b6 in kkqctdrvJPPD ()\n#4  0x000000000a0e6f6e in kkqjpdctr ()\n#5  0x0000000009e1be76 in qksqbApplyToQbcLoc ()\n#6  0x0000000009e1bbc4 in qksqbApplyToQbc ()\n#7  0x0000000009e1940c in kkqctdrvTD ()\n#8  0x000000000a0e65a4 in kkqjpddrv ()\n#9  0x0000000009e11b63 in kkqdrv ()\n#10 0x0000000009e0f043 in kkqctdrvIT ()\n#11 0x0000000009c306aa in apadrv ()\n#12 0x0000000009c2a606 in opitca ()\n...<\/code><\/pre>\n<p>The function kkqctUpdBestCostStt keeps track of the most efficient query block (QB) transformation state. When doing so, it relies on lmtdcrel for comparing costs.<\/p>\n<p>Anyway, we hit the breakpoint on lmtdcrel in the stack above and are now inspecting the register values: XMM0 = <span style=\"color: #ff0000;\">53.161750863128702<\/span> and XMM1 = <span style=\"color: #0000ff;\">53.186755259284759<\/span>.<\/p>\n<p>Then we&#8217;re looking for the values in the optimizer trace to figure out what&#8217;s being compared. For instance, the value in the XMM1 register 53.1867 is the cost of the non-transformed QB:<\/p>\n<pre><code>JPPD: Starting iteration 1, state space = (2) : (0)\nJPPD: Performing join predicate push-down (<span style=\"color: #0000ff;\">no transformation phase<\/span>) from query block SEL$A1EEA789 (#1) to query block SEL$2 (#2)\n...\nJPPD: Updated best state, Cost = <span style=\"color: #0000ff;\">53.186755<\/span><\/code><\/pre>\n<p>Similarly, 53.161751 is the cost of the transformed QB:<\/p>\n<pre><code>JPPD: Costing <span style=\"color: #ff0000;\">transformed<\/span> query.\n...\nFinal cost for query block SEL$A1EEA789 (#1) - All Rows Plan:\n   2235   Best join order: 1\n   2236   Cost: <span style=\"color: #ff0000;\">53.161751<\/span>  Degree: 1  Card: 1.000000  Bytes: 59.000000<\/code><\/pre>\n<p>Expectedly, the optimizer makes the decision to transform the query block because of the lower cost.<\/p>\n<pre><code>JPPD: Performing join predicate push-down (<span style=\"color: #ff0000;\">final <span style=\"color: #000000;\">phase<\/span><\/span>) from query block SEL$A1EEA789 (#1) to query block SEL$2 (#2)<\/code><\/pre>\n<p>Admittedly, this is a very simple example, but in general, if we&#8217;re not sure which comparison leads to which decision, we can postulate a hypothesis which we can easily verify by manipulating register values (in a lab, of course!) when the program enters lmtdcrel. For instance, in our case we can set XMM0 to a value which is greater than XMM1 and then see what happens:<\/p>\n<pre><code>set $xmm0.uint128=<span style=\"color: #008000;\">0x0000000000000000405b94b440959e70<\/span><\/code><\/pre>\n<p>By the way, <span style=\"color: #008000;\">0x0000000000000000405b94b440959e70<\/span> is just the hexadecimal representation of <span style=\"color: #008000;\">110.32350172625752<\/span>:<\/p>\n<pre><code>p $xmm0\n$84 = {v4_float = {4.67559052, 3.43095112, 0, 0}, v2_double = {<span style=\"color: #008000;\">110.32350172625752<\/span>, 0}, v16_int8 = {112, -98, -107, 64, -76, -108, 91, 64, 0, 0, 0, 0, 0, 0, 0, 0}, v8_int16 = {-24976, 16533, -27468, 16475, 0, 0, 0, 0}, v4_int32 = {1083547248, 1079743668, 0, 0}, v2_int64 = { 4637463743206628976, 0}, uint128 = 4637463743206628976}<\/code><\/pre>\n<p>Finally, we can find the following line in the CBO trace, which confirms that the transformation wasn&#8217;t done:<\/p>\n<pre><code>JPPD: Performing join predicate push-down (<span style=\"color: #008000;\">no transformation<\/span> phase) from query block SEL$A1EEA789 (#1) to query block SEL$2 (#2)<\/code><\/pre>\n<p>In conclusion, observing a low-level Oracle C function, such as lmtdcrel and combining that information with the CBO trace can provide additional insights into the optimizer&#8217;s internal functioning.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Tracing a low-level Oracle C function for getting insights into the optimizer decision making <a href=\"https:\/\/nenadnoveljic.com\/blog\/tracking-the-optimizer-with-gdb\/\" class=\"more-link\">Continue Reading <span class=\"meta-nav\">&rarr;<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"inline_featured_image":false,"footnotes":""},"categories":[11,27,5],"tags":[],"class_list":["post-2335","post","type-post","status-publish","format-standard","hentry","category-cost-based-optimizer","category-gdb","category-oracle"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v27.4 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>Tracking the Optimizer with gdb - All-round Database Topics<\/title>\n<meta name=\"description\" content=\"Tracing a low-level Oracle C function for getting insights into the optimizer decision making\" \/>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/nenadnoveljic.com\/blog\/tracking-the-optimizer-with-gdb\/\" \/>\n<meta property=\"og:locale\" content=\"en_US\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Tracking the Optimizer with gdb - All-round Database Topics\" \/>\n<meta property=\"og:description\" content=\"Tracing a low-level Oracle C function for getting insights into the optimizer decision making\" \/>\n<meta property=\"og:url\" content=\"https:\/\/nenadnoveljic.com\/blog\/tracking-the-optimizer-with-gdb\/\" \/>\n<meta property=\"og:site_name\" content=\"All-round Database Topics\" \/>\n<meta property=\"article:published_time\" content=\"2019-01-20T16:12:12+00:00\" \/>\n<meta property=\"article:modified_time\" content=\"2019-01-21T14:12:14+00:00\" \/>\n<meta name=\"author\" content=\"Nenad Noveljic\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:creator\" content=\"@NenadNoveljic\" \/>\n<meta name=\"twitter:label1\" content=\"Written by\" \/>\n\t<meta name=\"twitter:data1\" content=\"Nenad Noveljic\" \/>\n\t<meta name=\"twitter:label2\" content=\"Est. reading time\" \/>\n\t<meta name=\"twitter:data2\" content=\"4 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/nenadnoveljic.com\\\/blog\\\/tracking-the-optimizer-with-gdb\\\/#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/nenadnoveljic.com\\\/blog\\\/tracking-the-optimizer-with-gdb\\\/\"},\"author\":{\"name\":\"Nenad Noveljic\",\"@id\":\"https:\\\/\\\/nenadnoveljic.com\\\/blog\\\/#\\\/schema\\\/person\\\/51458d9dd86dbbdd19f5add451d44efa\"},\"headline\":\"Tracking the Optimizer with gdb\",\"datePublished\":\"2019-01-20T16:12:12+00:00\",\"dateModified\":\"2019-01-21T14:12:14+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/nenadnoveljic.com\\\/blog\\\/tracking-the-optimizer-with-gdb\\\/\"},\"wordCount\":527,\"commentCount\":0,\"articleSection\":[\"cost based optimizer\",\"gdb\",\"Oracle\"],\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\\\/\\\/nenadnoveljic.com\\\/blog\\\/tracking-the-optimizer-with-gdb\\\/#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/nenadnoveljic.com\\\/blog\\\/tracking-the-optimizer-with-gdb\\\/\",\"url\":\"https:\\\/\\\/nenadnoveljic.com\\\/blog\\\/tracking-the-optimizer-with-gdb\\\/\",\"name\":\"Tracking the Optimizer with gdb - All-round Database Topics\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/nenadnoveljic.com\\\/blog\\\/#website\"},\"datePublished\":\"2019-01-20T16:12:12+00:00\",\"dateModified\":\"2019-01-21T14:12:14+00:00\",\"author\":{\"@id\":\"https:\\\/\\\/nenadnoveljic.com\\\/blog\\\/#\\\/schema\\\/person\\\/51458d9dd86dbbdd19f5add451d44efa\"},\"description\":\"Tracing a low-level Oracle C function for getting insights into the optimizer decision making\",\"breadcrumb\":{\"@id\":\"https:\\\/\\\/nenadnoveljic.com\\\/blog\\\/tracking-the-optimizer-with-gdb\\\/#breadcrumb\"},\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/nenadnoveljic.com\\\/blog\\\/tracking-the-optimizer-with-gdb\\\/\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/nenadnoveljic.com\\\/blog\\\/tracking-the-optimizer-with-gdb\\\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\\\/\\\/nenadnoveljic.com\\\/blog\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Tracking the Optimizer with gdb\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/nenadnoveljic.com\\\/blog\\\/#website\",\"url\":\"https:\\\/\\\/nenadnoveljic.com\\\/blog\\\/\",\"name\":\"All-round Database Topics\",\"description\":\"Nenad Noveljic\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/nenadnoveljic.com\\\/blog\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"en-US\"},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/nenadnoveljic.com\\\/blog\\\/#\\\/schema\\\/person\\\/51458d9dd86dbbdd19f5add451d44efa\",\"name\":\"Nenad Noveljic\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/a97b796613ea48ec8a7b79c8ffe1c685dcffc920c68121f6238d5caab5070670?s=96&d=mm&r=g\",\"url\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/a97b796613ea48ec8a7b79c8ffe1c685dcffc920c68121f6238d5caab5070670?s=96&d=mm&r=g\",\"contentUrl\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/a97b796613ea48ec8a7b79c8ffe1c685dcffc920c68121f6238d5caab5070670?s=96&d=mm&r=g\",\"caption\":\"Nenad Noveljic\"},\"sameAs\":[\"nenad-noveljic-9b746a6\",\"https:\\\/\\\/x.com\\\/NenadNoveljic\"],\"url\":\"https:\\\/\\\/nenadnoveljic.com\\\/blog\\\/author\\\/nenad\\\/\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Tracking the Optimizer with gdb - All-round Database Topics","description":"Tracing a low-level Oracle C function for getting insights into the optimizer decision making","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/nenadnoveljic.com\/blog\/tracking-the-optimizer-with-gdb\/","og_locale":"en_US","og_type":"article","og_title":"Tracking the Optimizer with gdb - All-round Database Topics","og_description":"Tracing a low-level Oracle C function for getting insights into the optimizer decision making","og_url":"https:\/\/nenadnoveljic.com\/blog\/tracking-the-optimizer-with-gdb\/","og_site_name":"All-round Database Topics","article_published_time":"2019-01-20T16:12:12+00:00","article_modified_time":"2019-01-21T14:12:14+00:00","author":"Nenad Noveljic","twitter_card":"summary_large_image","twitter_creator":"@NenadNoveljic","twitter_misc":{"Written by":"Nenad Noveljic","Est. reading time":"4 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/nenadnoveljic.com\/blog\/tracking-the-optimizer-with-gdb\/#article","isPartOf":{"@id":"https:\/\/nenadnoveljic.com\/blog\/tracking-the-optimizer-with-gdb\/"},"author":{"name":"Nenad Noveljic","@id":"https:\/\/nenadnoveljic.com\/blog\/#\/schema\/person\/51458d9dd86dbbdd19f5add451d44efa"},"headline":"Tracking the Optimizer with gdb","datePublished":"2019-01-20T16:12:12+00:00","dateModified":"2019-01-21T14:12:14+00:00","mainEntityOfPage":{"@id":"https:\/\/nenadnoveljic.com\/blog\/tracking-the-optimizer-with-gdb\/"},"wordCount":527,"commentCount":0,"articleSection":["cost based optimizer","gdb","Oracle"],"inLanguage":"en-US","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/nenadnoveljic.com\/blog\/tracking-the-optimizer-with-gdb\/#respond"]}]},{"@type":"WebPage","@id":"https:\/\/nenadnoveljic.com\/blog\/tracking-the-optimizer-with-gdb\/","url":"https:\/\/nenadnoveljic.com\/blog\/tracking-the-optimizer-with-gdb\/","name":"Tracking the Optimizer with gdb - All-round Database Topics","isPartOf":{"@id":"https:\/\/nenadnoveljic.com\/blog\/#website"},"datePublished":"2019-01-20T16:12:12+00:00","dateModified":"2019-01-21T14:12:14+00:00","author":{"@id":"https:\/\/nenadnoveljic.com\/blog\/#\/schema\/person\/51458d9dd86dbbdd19f5add451d44efa"},"description":"Tracing a low-level Oracle C function for getting insights into the optimizer decision making","breadcrumb":{"@id":"https:\/\/nenadnoveljic.com\/blog\/tracking-the-optimizer-with-gdb\/#breadcrumb"},"inLanguage":"en-US","potentialAction":[{"@type":"ReadAction","target":["https:\/\/nenadnoveljic.com\/blog\/tracking-the-optimizer-with-gdb\/"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/nenadnoveljic.com\/blog\/tracking-the-optimizer-with-gdb\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/nenadnoveljic.com\/blog\/"},{"@type":"ListItem","position":2,"name":"Tracking the Optimizer with gdb"}]},{"@type":"WebSite","@id":"https:\/\/nenadnoveljic.com\/blog\/#website","url":"https:\/\/nenadnoveljic.com\/blog\/","name":"All-round Database Topics","description":"Nenad Noveljic","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/nenadnoveljic.com\/blog\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"en-US"},{"@type":"Person","@id":"https:\/\/nenadnoveljic.com\/blog\/#\/schema\/person\/51458d9dd86dbbdd19f5add451d44efa","name":"Nenad Noveljic","image":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/secure.gravatar.com\/avatar\/a97b796613ea48ec8a7b79c8ffe1c685dcffc920c68121f6238d5caab5070670?s=96&d=mm&r=g","url":"https:\/\/secure.gravatar.com\/avatar\/a97b796613ea48ec8a7b79c8ffe1c685dcffc920c68121f6238d5caab5070670?s=96&d=mm&r=g","contentUrl":"https:\/\/secure.gravatar.com\/avatar\/a97b796613ea48ec8a7b79c8ffe1c685dcffc920c68121f6238d5caab5070670?s=96&d=mm&r=g","caption":"Nenad Noveljic"},"sameAs":["nenad-noveljic-9b746a6","https:\/\/x.com\/NenadNoveljic"],"url":"https:\/\/nenadnoveljic.com\/blog\/author\/nenad\/"}]}},"_links":{"self":[{"href":"https:\/\/nenadnoveljic.com\/blog\/wp-json\/wp\/v2\/posts\/2335","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nenadnoveljic.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nenadnoveljic.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nenadnoveljic.com\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nenadnoveljic.com\/blog\/wp-json\/wp\/v2\/comments?post=2335"}],"version-history":[{"count":1,"href":"https:\/\/nenadnoveljic.com\/blog\/wp-json\/wp\/v2\/posts\/2335\/revisions"}],"predecessor-version":[{"id":2367,"href":"https:\/\/nenadnoveljic.com\/blog\/wp-json\/wp\/v2\/posts\/2335\/revisions\/2367"}],"wp:attachment":[{"href":"https:\/\/nenadnoveljic.com\/blog\/wp-json\/wp\/v2\/media?parent=2335"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nenadnoveljic.com\/blog\/wp-json\/wp\/v2\/categories?post=2335"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nenadnoveljic.com\/blog\/wp-json\/wp\/v2\/tags?post=2335"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}