Uh oh!
There was an error while loading. Please reload this page.
Late cast expansion: more improvements - #97387
Conversation
ghost
commented
Jan 23, 2024
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch Issue DetailsThis PR:
Almost final step prior getting rid of the importer's cast expansion.
|
EgorBo
commented
Jan 23, 2024
/azp run runtime-coreclr pgo, runtime-coreclr libraries-pgo, runtime-coreclr pgostress |
|
Azure Pipelines successfully started running 3 pipeline(s). |
EgorBo
commented
Jan 24, 2024
/azp run runtime-coreclr pgo, runtime-coreclr libraries-pgo, runtime-coreclr pgostress |
|
Azure Pipelines successfully started running 3 pipeline(s). |
EgorBo
commented
Jan 24, 2024
/azp run runtime-coreclr pgo, runtime-coreclr libraries-pgo, runtime-coreclr pgostress |
|
Azure Pipelines successfully started running 3 pipeline(s). |
EgorBo
commented
Jan 24, 2024
PTAL @jakobbotsch since you already reviewed this logc, cc @dotnet/jit-contrib this phase is already capable of expanding all kinds of casts and can replace importer, but I need to resolve a couple of CQ nits first before I remove importer. Quite a few changes but the main idea is to put all legality and profitability checks to CI failures aren't related. it also passes the tests if I disable importer |
| // Note: | ||
| // We are conservative on arrays of primitive types here. | ||
| // | ||
| bool Compiler::eeIsClassExact(CORINFO_CLASS_HANDLE clsHnd) |
There was a problem hiding this comment.
I'll remove this method because Jan already moved it to VM side in #97424
| // Likely class handle or NO_CLASS_HANDLE | ||
| // | ||
| static CORINFO_CLASS_HANDLE PickLikelyClass(Compiler* comp, IL_OFFSET offset, unsigned* likelihood) | ||
| enum TypeCheckFailedAction |
There was a problem hiding this comment.
Make it an enum class and remove the prefixes below?
| enum TypeCheckPassedAction | ||
| { | ||
| P_ReturnObj, | ||
| P_ReturnNull, | ||
| }; |
| // comp - Compiler instance | ||
| // castHelper - Cast helper call to expand | ||
| // commonCls - [out] Common denominator class for the fast and the fallback paths. | ||
| // likelihood - [out] Likelihood of successful type check (0..100) |
There was a problem hiding this comment.
Shouldn't it be inclusive in both ends?
| // These are never expanded: | ||
| // CORINFO_HELP_ISINSTANCEOF_EXCEPTION | ||
| // CORINFO_HELP_CHKCASTCLASS_SPECIAL | ||
| // CORINFO_HELP_READYTORUN_ISINSTANCEOF, | ||
| // CORINFO_HELP_READYTORUN_CHKCAST, |
There was a problem hiding this comment.
Just for my curiosity, why not expand the R2R version if we have static PGO?
There was a problem hiding this comment.
Good point, but importer also always gives up on them currently, but I'll add it to my todo
| if ((comp->info.compCompHnd->getClassAttribs(result) & | ||
| (CORINFO_FLG_INTERFACE | CORINFO_FLG_ABSTRACT | CORINFO_FLG_STATIC)) != 0) |
There was a problem hiding this comment.
Is CORINFO_FLG_STATIC possible on a class? I think it only applies to methods. Static classes in C# are abstract and sealed in IL, I think.
There was a problem hiding this comment.
Absolutely not needed, no idea why I added it here. this check was initially added to handle stale static pgo
| if (typeCheckFailedAction == F_CallHelper_AlwaysThrows) | ||
| { | ||
| // fallback call is used only to throw InvalidCastException | ||
| fallbackBb = fgNewBBFromTreeAfter(BBJ_THROW, typeCheckBb, gtUnusedValNode(call), debugInfo, nullptr, true); |
There was a problem hiding this comment.
I don't think gtUnusedValNode is necessary. Also, shouldn't we mark the call with GTF_CALL_M_DOES_NOT_RETURN?
| else if (typeCheckFailedAction == F_ReturnNull) | ||
| { | ||
| // if fallback call is not needed, we just assign null to tmp | ||
| GenTree* fallbackTree = gtNewTempStore(tmpNum, gtNewNull()); | ||
| fallbackBb = fgNewBBFromTreeAfter(BBJ_ALWAYS, typeCheckBb, fallbackTree, debugInfo, lastBb, true); | ||
| } | ||
| else | ||
| { | ||
| GenTree* fallbackTree = gtNewTempStore(tmpNum, call); | ||
| fallbackBb = fgNewBBFromTreeAfter(BBJ_ALWAYS, typeCheckBb, fallbackTree, debugInfo, lastBb, true); | ||
| } |
There was a problem hiding this comment.
It seems the nice comment above is no longer fully accurate since the fallback BB can also assign null in some cases now.
jakobbotsch
left a comment
There was a problem hiding this comment.
This generally looks great to me, I just had a few nits. You can consider running a SPMI collection on this branch to be able to get proper SPMI diffs (maybe once Jan's PR has been merged)
ryujit-bot
commented
Jan 24, 2024
Diff results for #97387Assembly diffsAssembly diffs for linux/arm64 ran on windows/x64Diffs are based on 2,494,760 contexts (1,003,806 MinOpts, 1,490,954 FullOpts). MISSED contexts: base: 4,060 (0.16%), diff: 10,457 (0.42%) Overall (+80 bytes)
FullOpts (+80 bytes)
Assembly diffs for linux/x64 ran on windows/x64Diffs are based on 2,588,583 contexts (1,052,329 MinOpts, 1,536,254 FullOpts). MISSED contexts: base: 3,628 (0.14%), diff: 10,052 (0.39%) Overall (+230 bytes)
FullOpts (+230 bytes)
Assembly diffs for osx/arm64 ran on windows/x64Diffs are based on 2,256,559 contexts (930,876 MinOpts, 1,325,683 FullOpts). MISSED contexts: base: 3,256 (0.14%), diff: 9,406 (0.42%) Overall (-16 bytes)
FullOpts (-16 bytes)
Assembly diffs for windows/arm64 ran on windows/x64Diffs are based on 2,312,039 contexts (931,543 MinOpts, 1,380,496 FullOpts). MISSED contexts: base: 2,687 (0.12%), diff: 8,855 (0.38%) Overall (-80 bytes)
FullOpts (-80 bytes)
Assembly diffs for windows/x64 ran on windows/x64Diffs are based on 2,486,103 contexts (983,689 MinOpts, 1,502,414 FullOpts). MISSED contexts: base: 3,899 (0.16%), diff: 10,701 (0.43%) Overall (+256 bytes)
FullOpts (+256 bytes)
Details here Assembly diffs for linux/arm ran on windows/x86Diffs are based on 2,231,354 contexts (827,812 MinOpts, 1,403,542 FullOpts). MISSED contexts: base: 74,588 (3.23%), diff: 80,924 (3.50%) Overall (-6 bytes)
FullOpts (-6 bytes)
Assembly diffs for windows/x86 ran on windows/x86Diffs are based on 2,289,781 contexts (841,817 MinOpts, 1,447,964 FullOpts). MISSED contexts: base: 5,093 (0.22%), diff: 11,589 (0.50%) Overall (+144 bytes)
FullOpts (+144 bytes)
Details here |
…st-expansion-2 # Conflicts: # src/coreclr/jit/gentree.cpp # src/coreclr/jit/importer.cpp
ryujit-bot
commented
Jan 24, 2024
Diff results for #97387Assembly diffsAssembly diffs for linux/arm64 ran on windows/x64Diffs are based on 2,494,760 contexts (1,003,806 MinOpts, 1,490,954 FullOpts). MISSED contexts: base: 4,060 (0.16%), diff: 10,457 (0.42%) Overall (+80 bytes)
FullOpts (+80 bytes)
Assembly diffs for linux/x64 ran on windows/x64Diffs are based on 2,588,583 contexts (1,052,329 MinOpts, 1,536,254 FullOpts). MISSED contexts: base: 3,628 (0.14%), diff: 10,052 (0.39%) Overall (+230 bytes)
FullOpts (+230 bytes)
Assembly diffs for osx/arm64 ran on windows/x64Diffs are based on 2,256,559 contexts (930,876 MinOpts, 1,325,683 FullOpts). MISSED contexts: base: 3,256 (0.14%), diff: 9,406 (0.42%) Overall (-16 bytes)
FullOpts (-16 bytes)
Assembly diffs for windows/arm64 ran on windows/x64Diffs are based on 2,312,039 contexts (931,543 MinOpts, 1,380,496 FullOpts). MISSED contexts: base: 2,687 (0.12%), diff: 8,855 (0.38%) Overall (-80 bytes)
FullOpts (-80 bytes)
Assembly diffs for windows/x64 ran on windows/x64Diffs are based on 2,486,103 contexts (983,689 MinOpts, 1,502,414 FullOpts). MISSED contexts: base: 3,899 (0.16%), diff: 10,701 (0.43%) Overall (+256 bytes)
FullOpts (+256 bytes)
Details here Assembly diffs for linux/arm ran on windows/x86Diffs are based on 2,231,354 contexts (827,812 MinOpts, 1,403,542 FullOpts). MISSED contexts: base: 74,588 (3.23%), diff: 80,924 (3.50%) Overall (-6 bytes)
FullOpts (-6 bytes)
Assembly diffs for windows/x86 ran on windows/x86Diffs are based on 2,289,781 contexts (841,817 MinOpts, 1,447,964 FullOpts). MISSED contexts: base: 5,093 (0.22%), diff: 11,589 (0.50%) Overall (+144 bytes)
FullOpts (+144 bytes)
Details here |
ryujit-bot
commented
Jan 24, 2024
Diff results for #97387Assembly diffsAssembly diffs for windows/arm64 ran on linux/x64Diffs are based on 2,308,464 contexts (929,692 MinOpts, 1,378,772 FullOpts). MISSED contexts: base: 0 (0.00%), diff: 6,334 (0.27%) Overall (-68 bytes)
FullOpts (-68 bytes)
Details here |
ryujit-bot
commented
Jan 25, 2024
Diff results for #97387Assembly diffsAssembly diffs for linux/arm64 ran on windows/x64Diffs are based on 2,498,787 contexts (1,011,240 MinOpts, 1,487,547 FullOpts). MISSED contexts: base: 0 (0.00%), diff: 6,564 (0.26%) Overall (-16 bytes)
FullOpts (-16 bytes)
Assembly diffs for linux/x64 ran on windows/x64Diffs are based on 2,505,538 contexts (977,780 MinOpts, 1,527,758 FullOpts). MISSED contexts: base: 0 (0.00%), diff: 6,724 (0.27%) Overall (+159 bytes)
FullOpts (+159 bytes)
Assembly diffs for windows/arm64 ran on windows/x64Diffs are based on 2,308,464 contexts (929,692 MinOpts, 1,378,772 FullOpts). MISSED contexts: base: 0 (0.00%), diff: 6,334 (0.27%) Overall (-68 bytes)
FullOpts (-68 bytes)
Assembly diffs for windows/x64 ran on windows/x64Diffs are based on 2,366,596 contexts (928,756 MinOpts, 1,437,840 FullOpts). MISSED contexts: base: 0 (0.00%), diff: 6,605 (0.28%) Overall (+199 bytes)
FullOpts (+199 bytes)
Details here Assembly diffs for linux/arm ran on windows/x86Diffs are based on 2,230,531 contexts (825,130 MinOpts, 1,405,401 FullOpts). MISSED contexts: base: 70,976 (3.08%), diff: 77,526 (3.36%) Overall (-22 bytes)
FullOpts (-22 bytes)
Assembly diffs for windows/x86 ran on windows/x86Diffs are based on 2,292,458 contexts (840,463 MinOpts, 1,451,995 FullOpts). MISSED contexts: base: 7 (0.00%), diff: 6,670 (0.29%) Overall (+127 bytes)
FullOpts (+127 bytes)
Details here |
This PR:
impIsClassExacttoeeIsClassExactfgLateCastExpansionto move all logic to a separatePickCandidateForTypeCheckthat performs all legality and profitability checks and gives a strategy for thefgLateCastExpansionForCallon how to properly expand everything.Almost final step prior getting rid of the importer's cast expansion.