22531 2008-11-27 03:50:01 -0800 Stack overflow on PowerPC on fast/workers/use-machine-stack.html 2009-03-19 21:38:41 -0700 1 1 1 Unclassified WebKit JavaScriptCore 528+ (Nightly build) Mac OS X 10.5 RESOLVED FIXED InRadar P1 Normal --- 1 ap ggaren darin ggaren mrowe oldest_to_newest 100530 0 ap 2008-11-27 03:50:01 -0800 Apparently, non-CTI interpreter uses a lot of stack for native calls, exhausting the whole 512K stack. 100531 1 ap 2008-11-27 04:04:46 -0800 Also, I'm building with gcc 4.0.1 on Tiger - maybe gcc 4.2 learned to reuse stack space used by scoped variables? There's lots of src and dst variables in privateExecute(). 100535 2 ap 2008-11-27 06:10:52 -0800 "A lot" means ~7K per recursive call on this test. 100594 3 ap 2008-11-28 03:45:06 -0800 <rdar://problem/6406045> 100595 4 ap 2008-11-28 03:47:11 -0800 This is not specific to worker threads - the same crash would happen for a thread created via JavaScriptCore API (even worse, in that case we don't control stack size of the thread). 100846 5 ap 2008-12-01 12:05:35 -0800 Disabled the test in r38862. 106287 6 darin 2009-01-19 14:28:31 -0800 7KB per native call? Are you serious? That seems absurd. 106309 7 ap 2009-01-19 23:05:42 -0800 I measured 7K on a debug build - release may be slightly better, but not much better: JSC::Interpreter::privateExecute(JSC::Interpreter::ExecutionFlag, JSC::RegisterFile*, JSC::ExecState*, JSC::JSValuePtr* ): +00000 000cf090 7C0802A6 mflr r0 +00004 000cf094 DBE1FFF8 stfd f31,0xfff8(SP) +00008 000cf098 2F850001 cmpwi cr7,r5,0x1 +0000c 000cf09c BDA1FFAC stmw r13,0xffac(SP) +00010 000cf0a0 429F0005 bcl 20,31,$+4 +00014 000cf0a4 90010008 stw r0,0x8(SP) +00018 000cf0a8 9421ECC0 stwu SP,0xecc0(SP) 0xecc0 is -4928, so privateExecute alone takes about 5K, and there are other methods involved in each call. 114439 8 28779 ggaren 2009-03-19 21:25:33 -0700 Created attachment 28779 patch 114440 9 28779 zwarich 2009-03-19 21:27:46 -0700 Comment on attachment 28779 patch r=me 114441 10 28779 oliver 2009-03-19 21:27:51 -0700 Comment on attachment 28779 patch r=me, though i'd like you to compare degree of reentry between ffx and us.. 114442 11 ggaren 2009-03-19 21:32:02 -0700 Committed revision 41852. I'm not sure how important it is for JavaScript to work well on threads where the client has artificially shrunk the stack size from the default. If that turns out to be an important use case, I guess we'll want an API for specifying how much stack JS gets to use, or a mechanism for asking the OS what the stack boundary is. 114443 12 ggaren 2009-03-19 21:38:41 -0700 > r=me, though i'd like you to compare degree of reentry between ffx and us.. Firefox allows 400-500 -- far more than both our old 128 and our new 64. If we discover that this is a compatibility problem, we should be able to reduce the stack used by Interpreter::execute and dial up the recursion limit. 28779 2009-03-19 21:25:33 -0700 2009-03-19 21:27:46 -0700 patch patch.txt text/plain 5162 ggaren SW5kZXg6IEphdmFTY3JpcHRDb3JlL0NoYW5nZUxvZwo9PT09PT09PT09PT09PT09PT09PT09PT09 PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09Ci0tLSBKYXZhU2NyaXB0 Q29yZS9DaGFuZ2VMb2cJKHJldmlzaW9uIDQxODUxKQorKysgSmF2YVNjcmlwdENvcmUvQ2hhbmdl TG9nCSh3b3JraW5nIGNvcHkpCkBAIC0xLDMgKzEsMjcgQEAKKzIwMDktMDMtMTkgIEdlb2ZmcmV5 IEdhcmVuICA8Z2dhcmVuQGFwcGxlLmNvbT4KKworICAgICAgICBSZXZpZXdlZCBieSBOT0JPRFkg KE9PUFMhKS4KKworICAgICAgICBGaXhlZCA8cmRhcjovL3Byb2JsZW0vNjQwNjA0NT4gUkVHUkVT U0lPTjogU3RhY2sgb3ZlcmZsb3cgb24gUG93ZXJQQyBvbgorICAgICAgICBmYXN0L3dvcmtlcnMv dXNlLW1hY2hpbmUtc3RhY2suaHRtbCAoMjI1MzEpCisKKyAgICAgICAgRGlhbGVkIGRvd24gdGhl IHJlLWVudHJ5IGFsbG93YW5jZSB0byA2NCAoZnJvbSAxMjgpLgorICAgICAgICAKKyAgICAgICAg T24gYSA1MTJLIHN0YWNrLCB0aGlzIGxlYXZlcyBhYm91dCA2NEsgZm9yIG90aGVyIGNvZGUgb24g dGhlIHN0YWNrIHdoaWxlCisgICAgICAgIEphdmFTY3JpcHQgaXMgcnVubmluZy4gTm90IHBlcmZl Y3QsIGJ1dCBpdCBzb2x2ZXMgb3VyIGNyYXNoIG9uIFBQQy4KKworICAgICAgICBEaWZmZXJlbnQg cGxhdGZvcm1zIG1heSB3YW50IHRvIGRpYWwgdGhpcyBkb3duIGV2ZW4gbW9yZS4KKworICAgICAg ICBBbHNvLCBzdWJzdGFudGlhbGx5IHNocnVuayBCeXRlY29kZUdlbmVyYXRvci4gU2luY2Ugd2Ug YWxsb2NhdGUgb25lIG9uCisgICAgICAgIHRoZSBzdGFjayBpbiBvcmRlciB0byB0aHJvdyBhIHN0 YWNrIG92ZXJmbG93IGV4Y2VwdGlvbiAtLSB3ZWxsLCBsZXQncworICAgICAgICBqdXN0IHNheSB0 aGUgb2xkIGNvZGUgaGFkIGFuIGFwcHJlY2lhdGlvbiBmb3IgaXJvbnkuCisKKyAgICAgICAgU3Vu U3BpZGVyIHJlcG9ydHMgbm8gY2hhbmdlLgorCisgICAgICAgICogYnl0ZWNvbXBpbGVyL0J5dGVj b2RlR2VuZXJhdG9yLmg6CisgICAgICAgICogaW50ZXJwcmV0ZXIvSW50ZXJwcmV0ZXIuaDoKKyAg ICAgICAgKEpTQzo6KToKKwogMjAwOS0wMy0xOSAgQ2FtZXJvbiBad2FyaWNoICA8Y3d6d2FyaWNo QHV3YXRlcmxvby5jYT4KIAogICAgICAgICBSZXZpZXdlZCBieSBPbGl2ZXIgSHVudC4KSW5kZXg6 IEphdmFTY3JpcHRDb3JlL2J5dGVjb21waWxlci9CeXRlY29kZUdlbmVyYXRvci5oCj09PT09PT09 PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09 PT0KLS0tIEphdmFTY3JpcHRDb3JlL2J5dGVjb21waWxlci9CeXRlY29kZUdlbmVyYXRvci5oCShy ZXZpc2lvbiA0MTg0OCkKKysrIEphdmFTY3JpcHRDb3JlL2J5dGVjb21waWxlci9CeXRlY29kZUdl bmVyYXRvci5oCSh3b3JraW5nIGNvcHkpCkBAIC00MzgsMTEgKzQzOCwxMSBAQCBuYW1lc3BhY2Ug SlNDIHsKICAgICAgICAgUmVnaXN0ZXJJRCBtX3RoaXNSZWdpc3RlcjsKICAgICAgICAgUmVnaXN0 ZXJJRCBtX2FyZ3VtZW50c1JlZ2lzdGVyOwogICAgICAgICBpbnQgbV9hY3RpdmF0aW9uUmVnaXN0 ZXJJbmRleDsKLSAgICAgICAgU2VnbWVudGVkVmVjdG9yPFJlZ2lzdGVySUQsIDUxMj4gbV9jYWxs ZWVSZWdpc3RlcnM7Ci0gICAgICAgIFNlZ21lbnRlZFZlY3RvcjxSZWdpc3RlcklELCA1MTI+IG1f cGFyYW1ldGVyczsKLSAgICAgICAgU2VnbWVudGVkVmVjdG9yPFJlZ2lzdGVySUQsIDUxMj4gbV9n bG9iYWxzOwotICAgICAgICBTZWdtZW50ZWRWZWN0b3I8TGFiZWxTY29wZSwgMjU2PiBtX2xhYmVs U2NvcGVzOwotICAgICAgICBTZWdtZW50ZWRWZWN0b3I8TGFiZWwsIDI1Nj4gbV9sYWJlbHM7Cisg ICAgICAgIFNlZ21lbnRlZFZlY3RvcjxSZWdpc3RlcklELCAzMj4gbV9jYWxsZWVSZWdpc3RlcnM7 CisgICAgICAgIFNlZ21lbnRlZFZlY3RvcjxSZWdpc3RlcklELCAzMj4gbV9wYXJhbWV0ZXJzOwor ICAgICAgICBTZWdtZW50ZWRWZWN0b3I8UmVnaXN0ZXJJRCwgMzI+IG1fZ2xvYmFsczsKKyAgICAg ICAgU2VnbWVudGVkVmVjdG9yPExhYmVsU2NvcGUsIDg+IG1fbGFiZWxTY29wZXM7CisgICAgICAg IFNlZ21lbnRlZFZlY3RvcjxMYWJlbCwgMzI+IG1fbGFiZWxzOwogICAgICAgICBSZWZQdHI8UmVn aXN0ZXJJRD4gbV9sYXN0Q29uc3RhbnQ7CiAgICAgICAgIGludCBtX2ZpbmFsbHlEZXB0aDsKICAg ICAgICAgaW50IG1fZHluYW1pY1Njb3BlRGVwdGg7CkluZGV4OiBKYXZhU2NyaXB0Q29yZS9pbnRl cnByZXRlci9JbnRlcnByZXRlci5oCj09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09 PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT0KLS0tIEphdmFTY3JpcHRDb3JlL2ludGVy cHJldGVyL0ludGVycHJldGVyLmgJKHJldmlzaW9uIDQxODQ4KQorKysgSmF2YVNjcmlwdENvcmUv aW50ZXJwcmV0ZXIvSW50ZXJwcmV0ZXIuaAkod29ya2luZyBjb3B5KQpAQCAtNjIsNyArNjIsNyBA QCBuYW1lc3BhY2UgSlNDIHsKICAgICAgICAgV2lsbEV4ZWN1dGVTdGF0ZW1lbnQKICAgICB9Owog Ci0gICAgZW51bSB7IE1heFJlZW50cnlEZXB0aCA9IDEyOCB9OworICAgIGVudW0geyBNYXhSZWVu dHJ5RGVwdGggPSA2NCB9OwogCiAgICAgY2xhc3MgSW50ZXJwcmV0ZXIgewogICAgICAgICBmcmll bmQgY2xhc3MgSklUOwpJbmRleDogTGF5b3V0VGVzdHMvQ2hhbmdlTG9nCj09PT09PT09PT09PT09 PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT0KLS0t IExheW91dFRlc3RzL0NoYW5nZUxvZwkocmV2aXNpb24gNDE4NTEpCisrKyBMYXlvdXRUZXN0cy9D aGFuZ2VMb2cJKHdvcmtpbmcgY29weSkKQEAgLTEsMyArMSwxMyBAQAorMjAwOS0wMy0xOSAgR2Vv ZmZyZXkgR2FyZW4gIDxnZ2FyZW5AYXBwbGUuY29tPgorCisgICAgICAgIFJldmlld2VkIGJ5IE5P Qk9EWSAoT09QUyEpLgorICAgICAgICAKKyAgICAgICAgUmUtZW5hYmxlZCB0aGlzIHRlc3QsIHNp bmNlIGl0IG5vIGxvbmdlciBjcmFzaGVzIHdoZW4gSmF2YVNjcmlwdENvcmUgaXMKKyAgICAgICAg Y29tcGlsZWQgaW4gbm9uLUpJVCBtb2RlLgorCisgICAgICAgICogZmFzdC93b3JrZXJzL3VzZS1t YWNoaW5lLXN0YWNrLmh0bWw6IENvcGllZCBmcm9tIGZhc3Qvd29ya2Vycy91c2UtbWFjaGluZS1z dGFjay5odG1sLWRpc2FibGVkLgorICAgICAgICAqIGZhc3Qvd29ya2Vycy91c2UtbWFjaGluZS1z dGFjay5odG1sLWRpc2FibGVkOiBSZW1vdmVkLgorCiAyMDA5LTAzLTE5ICBDYW1lcm9uIFp3YXJp Y2ggIDxjd3p3YXJpY2hAdXdhdGVybG9vLmNhPgogCiAgICAgICAgIFJldmlld2VkIGJ5IE9saXZl ciBIdW50LgpJbmRleDogTGF5b3V0VGVzdHMvZmFzdC93b3JrZXJzL3VzZS1tYWNoaW5lLXN0YWNr Lmh0bWwKPT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09 PT09PT09PT09PT09PT09PQotLS0gTGF5b3V0VGVzdHMvZmFzdC93b3JrZXJzL3VzZS1tYWNoaW5l LXN0YWNrLmh0bWwJKHJldmlzaW9uIDQxODQ4KQkoZnJvbSBMYXlvdXRUZXN0cy9mYXN0L3dvcmtl cnMvdXNlLW1hY2hpbmUtc3RhY2suaHRtbC1kaXNhYmxlZDo0MTg0OCkKKysrIExheW91dFRlc3Rz L2Zhc3Qvd29ya2Vycy91c2UtbWFjaGluZS1zdGFjay5odG1sCSh3b3JraW5nIGNvcHkpCkBAIC0w LDAgKzEsMjMgQEAKKzxib2R5PgorPHA+VGVzdCB3b3JrZXIgdGhyZWFkIHN0YWNrIHVzYWdlLiBT aG91bGQgbm90IGNyYXNoLjwvcD4KKzxkaXYgaWQ9cmVzdWx0PjwvZGl2PgorPHNjcmlwdD4KK2Z1 bmN0aW9uIGxvZyhtZXNzYWdlKQoreworICAgIGRvY3VtZW50LmdldEVsZW1lbnRCeUlkKCJyZXN1 bHQiKS5pbm5lckhUTUwgKz0gbWVzc2FnZSArICI8YnI+IjsKK30KKworaWYgKHdpbmRvdy5sYXlv dXRUZXN0Q29udHJvbGxlcikgeworICAgIGxheW91dFRlc3RDb250cm9sbGVyLmR1bXBBc1RleHQo KTsKKyAgICBsYXlvdXRUZXN0Q29udHJvbGxlci53YWl0VW50aWxEb25lKCk7Cit9CisKK3ZhciB3 b3JrZXIgPSBuZXcgV29ya2VyKCdyZXNvdXJjZXMvdXNlLW1hY2hpbmUtc3RhY2suanMnKTsKK3dv cmtlci5vbm1lc3NhZ2UgPSBmdW5jdGlvbihldnQpIHsKKyAgICBsb2coIlBBU1MgKCIgKyBldnQu ZGF0YSArICIpIik7CisgICAgaWYgKHdpbmRvdy5sYXlvdXRUZXN0Q29udHJvbGxlcikKKyAgICAg ICAgbGF5b3V0VGVzdENvbnRyb2xsZXIubm90aWZ5RG9uZSgpOworfQorPC9zY3JpcHQ+Cis8L2Jv ZHk+Cis8L2h0bWw+CkluZGV4OiBMYXlvdXRUZXN0cy9mYXN0L3dvcmtlcnMvdXNlLW1hY2hpbmUt c3RhY2suaHRtbC1kaXNhYmxlZAo9PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09 PT09PT09PT09PT09PT09PT09PT09PT09PT09PT09Ci0tLSBMYXlvdXRUZXN0cy9mYXN0L3dvcmtl cnMvdXNlLW1hY2hpbmUtc3RhY2suaHRtbC1kaXNhYmxlZAkocmV2aXNpb24gNDE4NDgpCisrKyBM YXlvdXRUZXN0cy9mYXN0L3dvcmtlcnMvdXNlLW1hY2hpbmUtc3RhY2suaHRtbC1kaXNhYmxlZAko d29ya2luZyBjb3B5KQpAQCAtMSwyMyArMCwwIEBACi08Ym9keT4KLTxwPlRlc3Qgd29ya2VyIHRo cmVhZCBzdGFjayB1c2FnZS4gU2hvdWxkIG5vdCBjcmFzaC48L3A+Ci08ZGl2IGlkPXJlc3VsdD48 L2Rpdj4KLTxzY3JpcHQ+Ci1mdW5jdGlvbiBsb2cobWVzc2FnZSkKLXsKLSAgICBkb2N1bWVudC5n ZXRFbGVtZW50QnlJZCgicmVzdWx0IikuaW5uZXJIVE1MICs9IG1lc3NhZ2UgKyAiPGJyPiI7Ci19 Ci0KLWlmICh3aW5kb3cubGF5b3V0VGVzdENvbnRyb2xsZXIpIHsKLSAgICBsYXlvdXRUZXN0Q29u dHJvbGxlci5kdW1wQXNUZXh0KCk7Ci0gICAgbGF5b3V0VGVzdENvbnRyb2xsZXIud2FpdFVudGls RG9uZSgpOwotfQotCi12YXIgd29ya2VyID0gbmV3IFdvcmtlcigncmVzb3VyY2VzL3VzZS1tYWNo aW5lLXN0YWNrLmpzJyk7Ci13b3JrZXIub25tZXNzYWdlID0gZnVuY3Rpb24oZXZ0KSB7Ci0gICAg bG9nKCJQQVNTICgiICsgZXZ0LmRhdGEgKyAiKSIpOwotICAgIGlmICh3aW5kb3cubGF5b3V0VGVz dENvbnRyb2xsZXIpCi0gICAgICAgIGxheW91dFRlc3RDb250cm9sbGVyLm5vdGlmeURvbmUoKTsK LX0KLTwvc2NyaXB0PgotPC9ib2R5PgotPC9odG1sPgo=