#11565: Restore code to handle '-fmax-worker-args' flag -------------------------------------+------------------------------------- Reporter: slyfox | Owner: Type: feature request | Status: new Priority: normal | Milestone: Component: Compiler | Version: 7.10.3 Resolution: | Keywords: Operating System: Unknown/Multiple | Architecture: | Unknown/Multiple Type of failure: None/Unknown | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by slyfox): Current motivating example to fix it is DynFlags example itself. I was profiling perf build of GHC and noticed a function that pushes whole DynFlags from stack to heap. This small piece of code emits 10 pages of mov instructions. https://git.haskell.org/ghc.git/blob/HEAD:/compiler/nativeGen/AsmCodeGen.hs#... {{{#!hs 1086 cmmExprNative :: ReferenceKind -> CmmExpr -> CmmOptM CmmExpr 1087 cmmExprNative referenceKind expr = do 1088 dflags <- getDynFlags 1089 let platform = targetPlatform dflags 1090 arch = platformArch platform 1091 case expr of ... 1106 CmmLit (CmmLabel lbl) 1107 -> do 1108 cmmMakeDynamicReference dflags referenceKind lbl ... }}} {{{ │ cmmExprNative :: ReferenceKind -> CmmExpr -> CmmOptM CmmExpr │ cmmExprNative referenceKind expr = do 0,11 │ cmp $0x3,%rax │ ↑ jb 3ceb930 │ -- we must convert block Ids to CLabels here, because we │ -- might have to do the PIC transformation. Hence we must │ -- not modify BlockIds beyond this point. │ │ CmmLit (CmmLabel lbl) │ -> do 2,02 │ add $0x890,%r12 │ cmp 0x358(%r13),%r12 │ ↑ ja 3cf456f 0,16 │ mov 0x7(%rbx),%rax 0,59 │ lea ghc_DynFlags_DynFlags_con_info,%rbx 0,05 │ mov %rbx,-0x888(%r12) 3,41 │18e9: mov 0x50(%rsp),%rbx 0,05 │ mov %rbx,-0x880(%r12) 0,32 │ mov 0x58(%rsp),%r14 │ mov %r14,-0x878(%r12) │ mov 0x60(%rsp),%rbx │ mov %rbx,-0x870(%r12) 0,05 │ mov 0x68(%rsp),%r14 │ mov %r14,-0x868(%r12) │ mov 0x70(%rsp),%rbx │ mov %rbx,-0x860(%r12) │ mov 0x78(%rsp),%r14 0,11 │ mov %r14,-0x858(%r12) 0,05 │ mov 0x80(%rsp),%rbx │ mov %rbx,-0x850(%r12) 0,05 │ mov 0x88(%rsp),%r14 │ mov %r14,-0x848(%r12) │ mov 0x90(%rsp),%rbx │ mov %rbx,-0x840(%r12) 0,05 │ mov 0x98(%rsp),%r14 0,05 │ mov %r14,-0x838(%r12) 0,11 │ mov 0xa0(%rsp),%rbx │ mov %rbx,-0x830(%r12) │ mov 0xa8(%rsp),%r14 │ mov %r14,-0x828(%r12) 0,05 │ mov 0xb0(%rsp),%rbx │ mov %rbx,-0x820(%r12) │ mov 0xb8(%rsp),%r14 ... <a few more pages of it> }}} On x86_64 register mapping is: '''%r12%''' - heap, '''%rsp''' - machine SP. The suspiction is worker/wrapper optimisation that moves huge 140-field record '''DynFlags''' from heap to stack even its not mutated. Looking at the AsmCodeGen.hs with -ddump-worker-wrapper {{{ "inplace/bin/ghc-stage1" -hisuf hi -osuf o -hcsuf hc -static -O -H64m -g -Wall -this-unit-id ghc-8.1 -hide-all-packages -i -icompiler/basicTypes -icompiler/cmm -icompiler/codeGen -icompiler/coreSyn -icompiler/deSugar -icompiler/ghci -icompiler/hsSyn -icompiler/iface -icompiler/llvmGen -icompiler/main -icompiler/nativeGen -icompiler/parser -icompiler/prelude -icompiler/profiling -icompiler/rename -icompiler/simplCore -icompiler/simplStg -icompiler/specialise -icompiler/stgSyn -icompiler/stranal -icompiler/typecheck -icompiler/types -icompiler/utils -icompiler/vectorise -icompiler/stage2/build -Icompiler/stage2/build -icompiler/stage2/build/./autogen -Icompiler/stage2/build/./autogen -Icompiler/. -Icompiler/parser -Icompiler/utils -Icompiler/../rts/dist/build -Icompiler/stage2 -optP- DGHCI -optP-include -optPcompiler/stage2/build/./autogen/cabal_macros.h -package-id array-0.5.1.1 -package-id base-4.9.0.0 -package-id binary-0.8.3.0 -package-id bytestring-0.10.8.1 -package-id containers-0.5.7.1 -package-id deepseq-1.4.2.0 -package-id directory-1.2.6.2 -package-id filepath-1.4.1.0 -package-id ghc-boot-8.1 -package-id ghci-8.1 -package-id hoopl-3.10.2.1 -package-id hpc-0.6.0.3 -package-id process-1.4.2.0 -package-id template-haskell-2.11.0.0 -package-id time-1.6.0.1 -package-id transformers-0.5.2.0 -package-id unix-2.7.2.0 -Wall -fno-warn-name-shadowing -this-unit-id ghc -XHaskell2010 -optc-DTHREADED_RTS -DGHCI_TABLES_NEXT_TO_CODE -DSTAGE=2 -Rghc-timing -O2 -no-user-package-db -rtsopts -Wnoncanonical-monad- instances -odir compiler/stage2/build -hidir compiler/stage2/build -stubdir compiler/stage2/build -dynamic-too -c compiler/nativeGen/AsmCodeGen.hs -o compiler/stage2/build/AsmCodeGen.o -dyno compiler/stage2/build/AsmCodeGen.dyn_o -ddump-worker-wrapper }}} there is a few places with functions with huge arity-140. One of the first places picked at random: '''dumpIfSet_dyn''' accepts a lot of separate arguments. {{{ case dflags_ab5I of { DynFlags ww1_al11 ww2_al12 ww3_al13 ww4_al14 ww5_al15 ww6_al16 [Dmd=] ww7_al17 ww8_al18 ww9_al19 ww10_al1a ww11_al1b ww12_al1c ww13_al1d ww14_al1e ww15_al1f ww16_al1g ww17_al1h ww18_al1i ww19_al1j ww20_al1k ww21_al1l ww22_al1m ww23_al1n ww24_al1o ww25_al1p ww26_al1q ww27_al1r ww28_al1s ww29_al1t ww30_al1u ww31_al1v ww32_al1w ww33_al1x ww34_al1y ww35_al1z ww36_al1A ww37_al1B ww38_al1C ww39_al1D ww40_al1E ww41_al1F ww42_al1G ww43_al1H ww44_al1I ww45_al1J ww46_al1K ww47_al1L ww48_al1M ww49_al1N ww50_al1O ww51_al1P ww52_al1Q ww53_al1R ww54_al1S ww55_al1T ww56_al1U ww57_al1V ww58_al1W ww59_al1X ww60_al1Y ww61_al1Z ww62_al20 ww63_al21 ww64_al22 ww65_al23 ww66_al24 ww67_al25 ww68_al26 ww69_al27 ww70_al28 ww71_al29 ww72_al2a ww73_al2b ww74_al2c ww75_al2d ww76_al2e ww77_al2f ww78_al2g ww79_al2h ww80_al2i ww81_al2j ww82_al2k ww83_al2l ww84_al2m [Dmd=] ww85_al2n [Dmd=] ww86_al2o ww87_al2p ww88_al2q ww89_al2r ww90_al2s ww91_al2t ww92_al2u ww93_al2v ww94_al2w ww95_al2x ww96_al2y ww97_al2z ww98_al2A ww99_al2B ww100_al2C ww101_al2D ww102_al2E ww103_al2F ww104_al2G ww105_al2H ww106_al2I ww107_al2J ww108_al2K ww109_al2L ww110_al2M ww111_al2N ww112_al2O ww113_al2P ww114_al2Q ww115_al2R ww116_al2S [Dmd=] ww117_al2T ww118_al2U ww119_al2V ww120_al2W ww121_al2X ww122_al2Y ww123_al2Z ww124_al30 ww125_al31 ww126_al32 ww127_al33 ww128_al34 ww129_al35 ww130_al36 ww131_al37 ww132_al38 ww133_al39 ww134_al3a ww135_al3b ww136_al3c -> ErrUtils.$wdumpIfSet_dyn ww1_al11 ww2_al12 ww3_al13 ww4_al14 ww5_al15 ww6_al16 ww7_al17 ww8_al18 ww9_al19 ww10_al1a ww11_al1b ww12_al1c ww13_al1d ww14_al1e ww15_al1f ww16_al1g ww17_al1h ww18_al1i ww19_al1j ww20_al1k ... }}} I'll try to craft small example that demonstrates the blowup. -- Ticket URL: GHC The Glasgow Haskell Compiler

#11565: Restore code to handle '-fmax-worker-args' flag -------------------------------------+------------------------------------- Reporter: slyfox | Owner: Type: feature request | Status: new Priority: normal | Milestone: Component: Compiler | Version: 7.10.3 Resolution: | Keywords: Operating System: Unknown/Multiple | Architecture: | Unknown/Multiple Type of failure: None/Unknown | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by slyfox): Here comes minimal example for one direction: heap to stack. The trigger is a function with many USED record fields: show in this case. In case of DynFlags it's a full (or large) subset of fields used in various GHC subsystems. {{{#!hs -- A.hs module A(D) where -- like DynFlgs in GHC data D = D { f_00, f_01, f_02, f_03, f_04 , f_10, f_11, f_12, f_13, f_14 , f_20, f_21, f_22, f_23, f_24 , f_30, f_31, f_32, f_33, f_34 , f_40, f_41, f_42, f_43, f_44 , f_50, f_51, f_52, f_53, f_54 , g_00, g_01, g_02, g_03, g_04 , g_10, g_11, g_12, g_13, g_14 , g_20, g_21, g_22, g_23, g_24 , g_30, g_31, g_32, g_33, g_34 , g_40, g_41, g_42, g_43, g_44 , g_50, g_51, g_52, g_53, g_54 , h_00, h_01, h_02, h_03, h_04 , h_10, h_11, h_12, h_13, h_14 , h_20, h_21, h_22, h_23, h_24 , h_30, h_31, h_32, h_33, h_34 , h_40, h_41, h_42, h_43, h_44 , h_50, h_51, h_52, h_53, h_54 , i_00, i_01, i_02, i_03, i_04 , i_10, i_11, i_12, i_13, i_14 , i_20, i_21, i_22, i_23, i_24 , i_30, i_31, i_32, i_33, i_34 , i_40, i_41, i_42, i_43, i_44 , i_50, i_51, i_52, i_53, i_54 :: Int } deriving Show }}} {{{#!hs -- B.hs module B (tiny_foo) where import qualified A tiny_foo :: A.D -> Bool tiny_foo d = null (show d) }}} Let's look at the size of module B on '''-O0''' and '''-O1''' while A is compiled -O2: '''-O0''', no unboxing happens. {{{#!hs $ ghc -c -O2 A.hs && ghc -c -O0 B.hs -ddump-stg -fforce-recomp compilation IS NOT required ==================== STG syntax: ==================== $trModule1_r3JD :: GHC.Types.TrName [GblId, Caf=NoCafRefs, Str=DmdType, Unf=OtherCon []] = NO_CCS GHC.Types.TrNameS! ["main"#]; $trModule2_r3P0 :: GHC.Types.TrName [GblId, Caf=NoCafRefs, Str=DmdType, Unf=OtherCon []] = NO_CCS GHC.Types.TrNameS! ["B"#]; B.$trModule :: GHC.Types.Module [GblId, Caf=NoCafRefs, Str=DmdType, Unf=OtherCon []] = NO_CCS GHC.Types.Module! [$trModule1_r3JD $trModule2_r3P0]; B.tiny_foo :: A.D -> GHC.Types.Bool [GblId, Arity=1, Str=DmdType, Unf=OtherCon []] = \r srt:SRT:[r30 :-> A.$fShowD, rAj :-> Data.Foldable.$fFoldable[]] [d_s3P4] let { sat_s3P5 [Occ=Once] :: [GHC.Types.Char] [LclId, Str=DmdType] = \u srt:SRT:[r30 :-> A.$fShowD] [] GHC.Show.show A.$fShowD d_s3P4; } in Data.Foldable.null Data.Foldable.$fFoldable[] sat_s3P5; }}} '''-O1''', unboxing hapened: {{{#!hs $ ghc -c -O2 A.hs && ghc -c -O1 B.hs -ddump-stg -fforce-recomp compilation IS NOT required ==================== STG syntax: ==================== B.$trModule2 :: GHC.Types.TrName [GblId, Caf=NoCafRefs, Str=DmdType m1, Unf=OtherCon []] = NO_CCS GHC.Types.TrNameS! ["main"#]; B.$trModule1 :: GHC.Types.TrName [GblId, Caf=NoCafRefs, Str=DmdType m1, Unf=OtherCon []] = NO_CCS GHC.Types.TrNameS! ["B"#]; B.$trModule :: GHC.Types.Module [GblId, Caf=NoCafRefs, Str=DmdType m, Unf=OtherCon []] = NO_CCS GHC.Types.Module! [B.$trModule2 B.$trModule1]; B.$wtiny_foo [InlPrag=[0]] :: GHC.Types.Int -> GHC.Types.Int -> GHC.Types.Int ... -> GHC.Types.Int -> GHC.Types.Bool [GblId, Arity=120, Str=DmdType , Unf=OtherCon []] = \r srt:SRT:[r3a :-> A.$w$cshowsPrec] [ww_s47S ww1_s47T ww2_s47U ww3_s47V ww4_s47W ww5_s47X ww6_s47Y ... ww115_s49J ww116_s49K ww117_s49L ww118_s49M ww119_s49N] case A.$w$cshowsPrec 0# ww_s47S ww1_s47T ww2_s47U ww3_s47V ww4_s47W ww5_s47X ww6_s47Y ww7_s47Z ww8_s480 ww9_s481 ww10_s482 ww11_s483 ww12_s484 ww13_s485 ... ww117_s49L ww118_s49M ww119_s49N GHC.Types.[] of _ [Occ=Dead] { [] -> GHC.Types.True []; : _ [Occ=Dead] _ [Occ=Dead] -> GHC.Types.False []; }; B.tiny_foo [InlPrag=INLINE[0]] :: A.D -> GHC.Types.Bool [GblId, Arity=1, Str=DmdType , Unf=OtherCon []] = \r srt:SRT:[r47O :-> B.$wtiny_foo] [w_s49R] case w_s49R of _ [Occ=Dead] { A.D ww1_s49T [Occ=Once] ww2_s49U [Occ=Once] ww3_s49V [Occ=Once] ww4_s49W [Occ=Once] ww5_s49X [Occ=Once] ... ww118_s4bM [Occ=Once] ww119_s4bN [Occ=Once] ww120_s4bO [Occ=Once] -> B.$wtiny_foo ww1_s49T ww2_s49U ww3_s49V ww4_s49W ww5_s49X ww6_s49Y ww7_s49Z ww8_s4a0 ww9_s4a1 ww10_s4a2 ww11_s4a3 ww12_s4a4 ww13_s4a5 ww14_s4a6 ww15_s4a7 ww16_s4a8 }}} This causes a lot of 'mov' instructions from heap to stack to be generated at each callsite. In this case it's 9 pages: {{{#!hs $ ghc -c -O2 A.hs && ghc -c -O1 B.hs -ddump-asm -fforce-recomp ... movq %rbx,856(%rbp) movq 880(%rbp),%rbx movq %rbx,864(%rbp) movq 888(%rbp),%rbx movq %rbx,872(%rbp) movq 896(%rbp),%rbx movq %rbx,880(%rbp) movq 904(%rbp),%rbx movq %rbx,888(%rbp) movq %rax,896(%rbp) movq $GHC.Types.[]_closure+1,904(%rbp) addq $-24,%rbp jmp A.$w$cshowsPrec_info }}} -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/11565#comment:6 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler

#11565: Restore code to handle '-fmax-worker-args' flag -------------------------------------+------------------------------------- Reporter: slyfox | Owner: Type: feature request | Status: new Priority: normal | Milestone: Component: Compiler | Version: 7.10.3 Resolution: | Keywords: Operating System: Unknown/Multiple | Architecture: | Unknown/Multiple Type of failure: None/Unknown | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by Sergei Trofimovich ): In [changeset:"a48de37dcca98e7d477040b0ed298bcd1b3ab303/ghc" a48de37/ghc]: {{{ #!CommitTicketReference repository="ghc" revision="a48de37dcca98e7d477040b0ed298bcd1b3ab303" restore -fmax-worker-args handling (Trac #11565) maxWorkerArgs handling was accidentally lost 3 years ago in a major update of demand analysis commit 0831a12ea2fc73c33652eeec1adc79fa19700578 Old regression is noticeable as: - code bloat (requires stack reshuffling) - compilation slowdown (more code to optimise/generate) - and increased heap usage (DynFlags unboxing/reboxing?) On a simple compile benchmark this change causes heap allocation drop from 70G don to 67G (ghc perf build). Signed-off-by: Sergei Trofimovich Reviewers: simonpj, ezyang, goldfire, austin, bgamari Reviewed By: simonpj, ezyang Subscribers: thomie Differential Revision: https://phabricator.haskell.org/D2503 GHC Trac Issues: #11565 }}} -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/11565#comment:8 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler

#11565: Restore code to handle '-fmax-worker-args' flag -------------------------------------+------------------------------------- Reporter: slyfox | Owner: Type: feature request | Status: closed Priority: normal | Milestone: 8.2.1 Component: Compiler | Version: 7.10.3 Resolution: fixed | Keywords: Operating System: Unknown/Multiple | Architecture: Type of failure: Runtime | Unknown/Multiple performance bug | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by Sergei Trofimovich ): In [changeset:"f93c363fab1ac8ce6f0b474f5967b0b097995827/ghc" f93c363f/ghc]: {{{ #!CommitTicketReference repository="ghc" revision="f93c363fab1ac8ce6f0b474f5967b0b097995827" extend '-fmax-worker-args' limit to specialiser (Trac #11565) It's a complementary change to a48de37dcca98e7d477040b0ed298bcd1b3ab303 restore -fmax-worker-args handling (Trac #11565) I don't have a small example but I've noticed another discrepancy when was profiling GHC for performance cmmExprNative :: ReferenceKind -> CmmExpr -> CmmOptM CmmExpr was specialised by 'spec_one' down to a function with arity 159. As a result 'perf record' pointed at it as at slowest function in whole ghc library. I've extended -fmax-worker-args effect to 'spec_one' as it does the same worker/wrapper split to push arguments to the heap. The change decreases heap usage on a synth.bash benchmark (Trac #9221) from 67G down to 64G (-4%). Benchmark runtime decreased from 14.5 s down to 14.s (-7%). Signed-off-by: Sergei Trofimovich Reviewers: ezyang, simonpj, austin, goldfire, bgamari Subscribers: thomie Differential Revision: https://phabricator.haskell.org/D2507 GHC Trac Issues: #11565 }}} -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/11565#comment:10 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler