maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/llvm/test/CodeGen/PowerPC/tail-dup-layout.ll

634 lines
18 KiB
LLVM
Raw Normal View History

CodeGen: BlockPlacement: Increase tail duplication size for O3. At O3 we are more willing to increase size if we believe it will improve performance. The current threshold for tail-duplication of 2 instructions is conservative, and can be relaxed at O3. Benchmark results: llvm test-suite: 6% improvement in aha, due to duplication of loop latch 3% improvement in hexxagon 2% slowdown in lpbench. Seems related, but couldn't completely diagnose. Internal google benchmark: Produces 4% improvement on internal google protocol buffer serialization benchmarks. Differential-Revision: https://reviews.llvm.org/D32324 llvm-svn: 303084
2017-05-16 01:30:47 +08:00
; RUN: llc -O2 -o - %s | FileCheck --check-prefix=CHECK --check-prefix=CHECK-O2 %s
; RUN: llc -O3 -o - %s | FileCheck --check-prefix=CHECK --check-prefix=CHECK-O3 %s
Codegen: Tail-duplicate during placement. The tail duplication pass uses an assumed layout when making duplication decisions. This is fine, but passes up duplication opportunities that may arise when blocks are outlined. Because we want the updated CFG to affect subsequent placement decisions, this change must occur during placement. In order to achieve this goal, TailDuplicationPass is split into a utility class, TailDuplicator, and the pass itself. The pass delegates nearly everything to the TailDuplicator object, except for looping over the blocks in a function. This allows the same code to be used for tail duplication in both places. This change, in concert with outlining optional branches, allows triangle shaped code to perform much better, esepecially when the taken/untaken branches are correlated, as it creates a second spine when the tests are small enough. Issue from previous rollback fixed, and a new test was added for that case as well. Issue was worklist/scheduling/taildup issue in layout. Issue from 2nd rollback fixed, with 2 additional tests. Issue was tail merging/loop info/tail-duplication causing issue with loops that share a header block. Issue with early tail-duplication of blocks that branch to a fallthrough predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll Differential revision: https://reviews.llvm.org/D18226 llvm-svn: 283934
2016-10-12 04:36:43 +08:00
target datalayout = "e-m:e-i64:64-n32:64"
target triple = "powerpc64le-grtev4-linux-gnu"
; Intended layout:
Codegen: Make chains from trellis-shaped CFGs Lay out trellis-shaped CFGs optimally. A trellis of the shape below: A B |\ /| | \ / | | X | | / \ | |/ \| C D would be laid out A; B->C ; D by the current layout algorithm. Now we identify trellises and lay them out either A->C; B->D or A->D; B->C. This scales with an increasing number of predecessors. A trellis is a a group of 2 or more predecessor blocks that all have the same successors. because of this we can tail duplicate to extend existing trellises. As an example consider the following CFG: B D F H / \ / \ / \ / \ A---C---E---G---Ret Where A,C,E,G are all small (Currently 2 instructions). The CFG preserving layout is then A,B,C,D,E,F,G,H,Ret. The current code will copy C into B, E into D and G into F and yield the layout A,C,B(C),E,D(E),F(G),G,H,ret define void @straight_test(i32 %tag) { entry: br label %test1 test1: ; A %tagbit1 = and i32 %tag, 1 %tagbit1eq0 = icmp eq i32 %tagbit1, 0 br i1 %tagbit1eq0, label %test2, label %optional1 optional1: ; B call void @a() br label %test2 test2: ; C %tagbit2 = and i32 %tag, 2 %tagbit2eq0 = icmp eq i32 %tagbit2, 0 br i1 %tagbit2eq0, label %test3, label %optional2 optional2: ; D call void @b() br label %test3 test3: ; E %tagbit3 = and i32 %tag, 4 %tagbit3eq0 = icmp eq i32 %tagbit3, 0 br i1 %tagbit3eq0, label %test4, label %optional3 optional3: ; F call void @c() br label %test4 test4: ; G %tagbit4 = and i32 %tag, 8 %tagbit4eq0 = icmp eq i32 %tagbit4, 0 br i1 %tagbit4eq0, label %exit, label %optional4 optional4: ; H call void @d() br label %exit exit: ret void } here is the layout after D27742: straight_test: # @straight_test ; ... Prologue elided ; BB#0: # %entry ; A (merged with test1) ; ... More prologue elided mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_2 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_3 b .LBB0_4 .LBB0_2: # %optional1 ; B (copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_4 .LBB0_3: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_5 b .LBB0_6 .LBB0_4: # %optional2 ; D (copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_5: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 b .LBB0_7 .LBB0_6: # %optional3 ; F (copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit ; Ret ld 30, 96(1) # 8-byte Folded Reload addi 1, 1, 112 ld 0, 16(1) mtlr 0 blr The tail-duplication has produced some benefit, but it has also produced a trellis which is not laid out optimally. With this patch, we improve the layouts of such trellises, and decrease the cost calculation for tail-duplication accordingly. This patch produces the layout A,C,E,G,B,D,F,H,Ret. This layout does have back edges, which is a negative, but it has a bigger compensating positive, which is that it handles the case where there are long strings of skipped blocks much better than the original layout. Both layouts handle runs of executed blocks equally well. Branch prediction also improves if there is any correlation between subsequent optional blocks. Here is the resulting concrete layout: straight_test: # @straight_test ; BB#0: # %entry ; A (merged with test1) mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_4 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_5 .LBB0_2: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_3: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 bne 0, .LBB0_7 b .LBB0_8 .LBB0_4: # %optional1 ; B (Copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_2 .LBB0_5: # %optional2 ; D (Copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_3 .LBB0_6: # %optional3 ; F (Copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit Differential Revision: https://reviews.llvm.org/D28522 llvm-svn: 295223
2017-02-16 03:49:14 +08:00
; The chain-based outlining produces the layout
Codegen: Tail-duplicate during placement. The tail duplication pass uses an assumed layout when making duplication decisions. This is fine, but passes up duplication opportunities that may arise when blocks are outlined. Because we want the updated CFG to affect subsequent placement decisions, this change must occur during placement. In order to achieve this goal, TailDuplicationPass is split into a utility class, TailDuplicator, and the pass itself. The pass delegates nearly everything to the TailDuplicator object, except for looping over the blocks in a function. This allows the same code to be used for tail duplication in both places. This change, in concert with outlining optional branches, allows triangle shaped code to perform much better, esepecially when the taken/untaken branches are correlated, as it creates a second spine when the tests are small enough. Issue from previous rollback fixed, and a new test was added for that case as well. Issue was worklist/scheduling/taildup issue in layout. Issue from 2nd rollback fixed, with 2 additional tests. Issue was tail merging/loop info/tail-duplication causing issue with loops that share a header block. Issue with early tail-duplication of blocks that branch to a fallthrough predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll Differential revision: https://reviews.llvm.org/D18226 llvm-svn: 283934
2016-10-12 04:36:43 +08:00
; test1
; test2
; test3
; test4
; optional1
; optional2
; optional3
; optional4
Codegen: Make chains from trellis-shaped CFGs Lay out trellis-shaped CFGs optimally. A trellis of the shape below: A B |\ /| | \ / | | X | | / \ | |/ \| C D would be laid out A; B->C ; D by the current layout algorithm. Now we identify trellises and lay them out either A->C; B->D or A->D; B->C. This scales with an increasing number of predecessors. A trellis is a a group of 2 or more predecessor blocks that all have the same successors. because of this we can tail duplicate to extend existing trellises. As an example consider the following CFG: B D F H / \ / \ / \ / \ A---C---E---G---Ret Where A,C,E,G are all small (Currently 2 instructions). The CFG preserving layout is then A,B,C,D,E,F,G,H,Ret. The current code will copy C into B, E into D and G into F and yield the layout A,C,B(C),E,D(E),F(G),G,H,ret define void @straight_test(i32 %tag) { entry: br label %test1 test1: ; A %tagbit1 = and i32 %tag, 1 %tagbit1eq0 = icmp eq i32 %tagbit1, 0 br i1 %tagbit1eq0, label %test2, label %optional1 optional1: ; B call void @a() br label %test2 test2: ; C %tagbit2 = and i32 %tag, 2 %tagbit2eq0 = icmp eq i32 %tagbit2, 0 br i1 %tagbit2eq0, label %test3, label %optional2 optional2: ; D call void @b() br label %test3 test3: ; E %tagbit3 = and i32 %tag, 4 %tagbit3eq0 = icmp eq i32 %tagbit3, 0 br i1 %tagbit3eq0, label %test4, label %optional3 optional3: ; F call void @c() br label %test4 test4: ; G %tagbit4 = and i32 %tag, 8 %tagbit4eq0 = icmp eq i32 %tagbit4, 0 br i1 %tagbit4eq0, label %exit, label %optional4 optional4: ; H call void @d() br label %exit exit: ret void } here is the layout after D27742: straight_test: # @straight_test ; ... Prologue elided ; BB#0: # %entry ; A (merged with test1) ; ... More prologue elided mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_2 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_3 b .LBB0_4 .LBB0_2: # %optional1 ; B (copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_4 .LBB0_3: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_5 b .LBB0_6 .LBB0_4: # %optional2 ; D (copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_5: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 b .LBB0_7 .LBB0_6: # %optional3 ; F (copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit ; Ret ld 30, 96(1) # 8-byte Folded Reload addi 1, 1, 112 ld 0, 16(1) mtlr 0 blr The tail-duplication has produced some benefit, but it has also produced a trellis which is not laid out optimally. With this patch, we improve the layouts of such trellises, and decrease the cost calculation for tail-duplication accordingly. This patch produces the layout A,C,E,G,B,D,F,H,Ret. This layout does have back edges, which is a negative, but it has a bigger compensating positive, which is that it handles the case where there are long strings of skipped blocks much better than the original layout. Both layouts handle runs of executed blocks equally well. Branch prediction also improves if there is any correlation between subsequent optional blocks. Here is the resulting concrete layout: straight_test: # @straight_test ; BB#0: # %entry ; A (merged with test1) mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_4 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_5 .LBB0_2: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_3: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 bne 0, .LBB0_7 b .LBB0_8 .LBB0_4: # %optional1 ; B (Copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_2 .LBB0_5: # %optional2 ; D (Copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_3 .LBB0_6: # %optional3 ; F (Copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit Differential Revision: https://reviews.llvm.org/D28522 llvm-svn: 295223
2017-02-16 03:49:14 +08:00
; exit
Codegen: Tail-duplicate during placement. The tail duplication pass uses an assumed layout when making duplication decisions. This is fine, but passes up duplication opportunities that may arise when blocks are outlined. Because we want the updated CFG to affect subsequent placement decisions, this change must occur during placement. In order to achieve this goal, TailDuplicationPass is split into a utility class, TailDuplicator, and the pass itself. The pass delegates nearly everything to the TailDuplicator object, except for looping over the blocks in a function. This allows the same code to be used for tail duplication in both places. This change, in concert with outlining optional branches, allows triangle shaped code to perform much better, esepecially when the taken/untaken branches are correlated, as it creates a second spine when the tests are small enough. Issue from previous rollback fixed, and a new test was added for that case as well. Issue was worklist/scheduling/taildup issue in layout. Issue from 2nd rollback fixed, with 2 additional tests. Issue was tail merging/loop info/tail-duplication causing issue with loops that share a header block. Issue with early tail-duplication of blocks that branch to a fallthrough predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll Differential revision: https://reviews.llvm.org/D18226 llvm-svn: 283934
2016-10-12 04:36:43 +08:00
; Tail duplication puts test n+1 at the end of optional n
; so optional1 includes a copy of test2 at the end, and branches
; to test3 (at the top) or falls through to optional 2.
Codegen: Make chains from trellis-shaped CFGs Lay out trellis-shaped CFGs optimally. A trellis of the shape below: A B |\ /| | \ / | | X | | / \ | |/ \| C D would be laid out A; B->C ; D by the current layout algorithm. Now we identify trellises and lay them out either A->C; B->D or A->D; B->C. This scales with an increasing number of predecessors. A trellis is a a group of 2 or more predecessor blocks that all have the same successors. because of this we can tail duplicate to extend existing trellises. As an example consider the following CFG: B D F H / \ / \ / \ / \ A---C---E---G---Ret Where A,C,E,G are all small (Currently 2 instructions). The CFG preserving layout is then A,B,C,D,E,F,G,H,Ret. The current code will copy C into B, E into D and G into F and yield the layout A,C,B(C),E,D(E),F(G),G,H,ret define void @straight_test(i32 %tag) { entry: br label %test1 test1: ; A %tagbit1 = and i32 %tag, 1 %tagbit1eq0 = icmp eq i32 %tagbit1, 0 br i1 %tagbit1eq0, label %test2, label %optional1 optional1: ; B call void @a() br label %test2 test2: ; C %tagbit2 = and i32 %tag, 2 %tagbit2eq0 = icmp eq i32 %tagbit2, 0 br i1 %tagbit2eq0, label %test3, label %optional2 optional2: ; D call void @b() br label %test3 test3: ; E %tagbit3 = and i32 %tag, 4 %tagbit3eq0 = icmp eq i32 %tagbit3, 0 br i1 %tagbit3eq0, label %test4, label %optional3 optional3: ; F call void @c() br label %test4 test4: ; G %tagbit4 = and i32 %tag, 8 %tagbit4eq0 = icmp eq i32 %tagbit4, 0 br i1 %tagbit4eq0, label %exit, label %optional4 optional4: ; H call void @d() br label %exit exit: ret void } here is the layout after D27742: straight_test: # @straight_test ; ... Prologue elided ; BB#0: # %entry ; A (merged with test1) ; ... More prologue elided mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_2 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_3 b .LBB0_4 .LBB0_2: # %optional1 ; B (copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_4 .LBB0_3: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_5 b .LBB0_6 .LBB0_4: # %optional2 ; D (copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_5: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 b .LBB0_7 .LBB0_6: # %optional3 ; F (copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit ; Ret ld 30, 96(1) # 8-byte Folded Reload addi 1, 1, 112 ld 0, 16(1) mtlr 0 blr The tail-duplication has produced some benefit, but it has also produced a trellis which is not laid out optimally. With this patch, we improve the layouts of such trellises, and decrease the cost calculation for tail-duplication accordingly. This patch produces the layout A,C,E,G,B,D,F,H,Ret. This layout does have back edges, which is a negative, but it has a bigger compensating positive, which is that it handles the case where there are long strings of skipped blocks much better than the original layout. Both layouts handle runs of executed blocks equally well. Branch prediction also improves if there is any correlation between subsequent optional blocks. Here is the resulting concrete layout: straight_test: # @straight_test ; BB#0: # %entry ; A (merged with test1) mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_4 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_5 .LBB0_2: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_3: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 bne 0, .LBB0_7 b .LBB0_8 .LBB0_4: # %optional1 ; B (Copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_2 .LBB0_5: # %optional2 ; D (Copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_3 .LBB0_6: # %optional3 ; F (Copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit Differential Revision: https://reviews.llvm.org/D28522 llvm-svn: 295223
2017-02-16 03:49:14 +08:00
; The CHECK statements check for the whole string of tests
Codegen: Tail-duplicate during placement. The tail duplication pass uses an assumed layout when making duplication decisions. This is fine, but passes up duplication opportunities that may arise when blocks are outlined. Because we want the updated CFG to affect subsequent placement decisions, this change must occur during placement. In order to achieve this goal, TailDuplicationPass is split into a utility class, TailDuplicator, and the pass itself. The pass delegates nearly everything to the TailDuplicator object, except for looping over the blocks in a function. This allows the same code to be used for tail duplication in both places. This change, in concert with outlining optional branches, allows triangle shaped code to perform much better, esepecially when the taken/untaken branches are correlated, as it creates a second spine when the tests are small enough. Issue from previous rollback fixed, and a new test was added for that case as well. Issue was worklist/scheduling/taildup issue in layout. Issue from 2nd rollback fixed, with 2 additional tests. Issue was tail merging/loop info/tail-duplication causing issue with loops that share a header block. Issue with early tail-duplication of blocks that branch to a fallthrough predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll Differential revision: https://reviews.llvm.org/D18226 llvm-svn: 283934
2016-10-12 04:36:43 +08:00
; and then check that the correct test has been duplicated into the end of
; the optional blocks and that the optional blocks are in the correct order.
Codegen: Make chains from trellis-shaped CFGs Lay out trellis-shaped CFGs optimally. A trellis of the shape below: A B |\ /| | \ / | | X | | / \ | |/ \| C D would be laid out A; B->C ; D by the current layout algorithm. Now we identify trellises and lay them out either A->C; B->D or A->D; B->C. This scales with an increasing number of predecessors. A trellis is a a group of 2 or more predecessor blocks that all have the same successors. because of this we can tail duplicate to extend existing trellises. As an example consider the following CFG: B D F H / \ / \ / \ / \ A---C---E---G---Ret Where A,C,E,G are all small (Currently 2 instructions). The CFG preserving layout is then A,B,C,D,E,F,G,H,Ret. The current code will copy C into B, E into D and G into F and yield the layout A,C,B(C),E,D(E),F(G),G,H,ret define void @straight_test(i32 %tag) { entry: br label %test1 test1: ; A %tagbit1 = and i32 %tag, 1 %tagbit1eq0 = icmp eq i32 %tagbit1, 0 br i1 %tagbit1eq0, label %test2, label %optional1 optional1: ; B call void @a() br label %test2 test2: ; C %tagbit2 = and i32 %tag, 2 %tagbit2eq0 = icmp eq i32 %tagbit2, 0 br i1 %tagbit2eq0, label %test3, label %optional2 optional2: ; D call void @b() br label %test3 test3: ; E %tagbit3 = and i32 %tag, 4 %tagbit3eq0 = icmp eq i32 %tagbit3, 0 br i1 %tagbit3eq0, label %test4, label %optional3 optional3: ; F call void @c() br label %test4 test4: ; G %tagbit4 = and i32 %tag, 8 %tagbit4eq0 = icmp eq i32 %tagbit4, 0 br i1 %tagbit4eq0, label %exit, label %optional4 optional4: ; H call void @d() br label %exit exit: ret void } here is the layout after D27742: straight_test: # @straight_test ; ... Prologue elided ; BB#0: # %entry ; A (merged with test1) ; ... More prologue elided mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_2 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_3 b .LBB0_4 .LBB0_2: # %optional1 ; B (copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_4 .LBB0_3: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_5 b .LBB0_6 .LBB0_4: # %optional2 ; D (copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_5: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 b .LBB0_7 .LBB0_6: # %optional3 ; F (copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit ; Ret ld 30, 96(1) # 8-byte Folded Reload addi 1, 1, 112 ld 0, 16(1) mtlr 0 blr The tail-duplication has produced some benefit, but it has also produced a trellis which is not laid out optimally. With this patch, we improve the layouts of such trellises, and decrease the cost calculation for tail-duplication accordingly. This patch produces the layout A,C,E,G,B,D,F,H,Ret. This layout does have back edges, which is a negative, but it has a bigger compensating positive, which is that it handles the case where there are long strings of skipped blocks much better than the original layout. Both layouts handle runs of executed blocks equally well. Branch prediction also improves if there is any correlation between subsequent optional blocks. Here is the resulting concrete layout: straight_test: # @straight_test ; BB#0: # %entry ; A (merged with test1) mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_4 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_5 .LBB0_2: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_3: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 bne 0, .LBB0_7 b .LBB0_8 .LBB0_4: # %optional1 ; B (Copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_2 .LBB0_5: # %optional2 ; D (Copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_3 .LBB0_6: # %optional3 ; F (Copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit Differential Revision: https://reviews.llvm.org/D28522 llvm-svn: 295223
2017-02-16 03:49:14 +08:00
;CHECK-LABEL: straight_test:
Codegen: Tail-duplicate during placement. The tail duplication pass uses an assumed layout when making duplication decisions. This is fine, but passes up duplication opportunities that may arise when blocks are outlined. Because we want the updated CFG to affect subsequent placement decisions, this change must occur during placement. In order to achieve this goal, TailDuplicationPass is split into a utility class, TailDuplicator, and the pass itself. The pass delegates nearly everything to the TailDuplicator object, except for looping over the blocks in a function. This allows the same code to be used for tail duplication in both places. This change, in concert with outlining optional branches, allows triangle shaped code to perform much better, esepecially when the taken/untaken branches are correlated, as it creates a second spine when the tests are small enough. Issue from previous rollback fixed, and a new test was added for that case as well. Issue was worklist/scheduling/taildup issue in layout. Issue from 2nd rollback fixed, with 2 additional tests. Issue was tail merging/loop info/tail-duplication causing issue with loops that share a header block. Issue with early tail-duplication of blocks that branch to a fallthrough predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll Differential revision: https://reviews.llvm.org/D18226 llvm-svn: 283934
2016-10-12 04:36:43 +08:00
; test1 may have been merged with entry
;CHECK: mr [[TAGREG:[0-9]+]], 3
;CHECK: andi. {{[0-9]+}}, [[TAGREG]], 1
Codegen: Make chains from trellis-shaped CFGs Lay out trellis-shaped CFGs optimally. A trellis of the shape below: A B |\ /| | \ / | | X | | / \ | |/ \| C D would be laid out A; B->C ; D by the current layout algorithm. Now we identify trellises and lay them out either A->C; B->D or A->D; B->C. This scales with an increasing number of predecessors. A trellis is a a group of 2 or more predecessor blocks that all have the same successors. because of this we can tail duplicate to extend existing trellises. As an example consider the following CFG: B D F H / \ / \ / \ / \ A---C---E---G---Ret Where A,C,E,G are all small (Currently 2 instructions). The CFG preserving layout is then A,B,C,D,E,F,G,H,Ret. The current code will copy C into B, E into D and G into F and yield the layout A,C,B(C),E,D(E),F(G),G,H,ret define void @straight_test(i32 %tag) { entry: br label %test1 test1: ; A %tagbit1 = and i32 %tag, 1 %tagbit1eq0 = icmp eq i32 %tagbit1, 0 br i1 %tagbit1eq0, label %test2, label %optional1 optional1: ; B call void @a() br label %test2 test2: ; C %tagbit2 = and i32 %tag, 2 %tagbit2eq0 = icmp eq i32 %tagbit2, 0 br i1 %tagbit2eq0, label %test3, label %optional2 optional2: ; D call void @b() br label %test3 test3: ; E %tagbit3 = and i32 %tag, 4 %tagbit3eq0 = icmp eq i32 %tagbit3, 0 br i1 %tagbit3eq0, label %test4, label %optional3 optional3: ; F call void @c() br label %test4 test4: ; G %tagbit4 = and i32 %tag, 8 %tagbit4eq0 = icmp eq i32 %tagbit4, 0 br i1 %tagbit4eq0, label %exit, label %optional4 optional4: ; H call void @d() br label %exit exit: ret void } here is the layout after D27742: straight_test: # @straight_test ; ... Prologue elided ; BB#0: # %entry ; A (merged with test1) ; ... More prologue elided mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_2 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_3 b .LBB0_4 .LBB0_2: # %optional1 ; B (copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_4 .LBB0_3: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_5 b .LBB0_6 .LBB0_4: # %optional2 ; D (copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_5: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 b .LBB0_7 .LBB0_6: # %optional3 ; F (copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit ; Ret ld 30, 96(1) # 8-byte Folded Reload addi 1, 1, 112 ld 0, 16(1) mtlr 0 blr The tail-duplication has produced some benefit, but it has also produced a trellis which is not laid out optimally. With this patch, we improve the layouts of such trellises, and decrease the cost calculation for tail-duplication accordingly. This patch produces the layout A,C,E,G,B,D,F,H,Ret. This layout does have back edges, which is a negative, but it has a bigger compensating positive, which is that it handles the case where there are long strings of skipped blocks much better than the original layout. Both layouts handle runs of executed blocks equally well. Branch prediction also improves if there is any correlation between subsequent optional blocks. Here is the resulting concrete layout: straight_test: # @straight_test ; BB#0: # %entry ; A (merged with test1) mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_4 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_5 .LBB0_2: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_3: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 bne 0, .LBB0_7 b .LBB0_8 .LBB0_4: # %optional1 ; B (Copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_2 .LBB0_5: # %optional2 ; D (Copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_3 .LBB0_6: # %optional3 ; F (Copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit Differential Revision: https://reviews.llvm.org/D28522 llvm-svn: 295223
2017-02-16 03:49:14 +08:00
;CHECK-NEXT: bc 12, 1, .[[OPT1LABEL:[_0-9A-Za-z]+]]
;CHECK-NEXT: # %test2
Codegen: Tail-duplicate during placement. The tail duplication pass uses an assumed layout when making duplication decisions. This is fine, but passes up duplication opportunities that may arise when blocks are outlined. Because we want the updated CFG to affect subsequent placement decisions, this change must occur during placement. In order to achieve this goal, TailDuplicationPass is split into a utility class, TailDuplicator, and the pass itself. The pass delegates nearly everything to the TailDuplicator object, except for looping over the blocks in a function. This allows the same code to be used for tail duplication in both places. This change, in concert with outlining optional branches, allows triangle shaped code to perform much better, esepecially when the taken/untaken branches are correlated, as it creates a second spine when the tests are small enough. Issue from previous rollback fixed, and a new test was added for that case as well. Issue was worklist/scheduling/taildup issue in layout. Issue from 2nd rollback fixed, with 2 additional tests. Issue was tail merging/loop info/tail-duplication causing issue with loops that share a header block. Issue with early tail-duplication of blocks that branch to a fallthrough predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll Differential revision: https://reviews.llvm.org/D18226 llvm-svn: 283934
2016-10-12 04:36:43 +08:00
;CHECK-NEXT: rlwinm. {{[0-9]+}}, [[TAGREG]], 0, 30, 30
Codegen: Make chains from trellis-shaped CFGs Lay out trellis-shaped CFGs optimally. A trellis of the shape below: A B |\ /| | \ / | | X | | / \ | |/ \| C D would be laid out A; B->C ; D by the current layout algorithm. Now we identify trellises and lay them out either A->C; B->D or A->D; B->C. This scales with an increasing number of predecessors. A trellis is a a group of 2 or more predecessor blocks that all have the same successors. because of this we can tail duplicate to extend existing trellises. As an example consider the following CFG: B D F H / \ / \ / \ / \ A---C---E---G---Ret Where A,C,E,G are all small (Currently 2 instructions). The CFG preserving layout is then A,B,C,D,E,F,G,H,Ret. The current code will copy C into B, E into D and G into F and yield the layout A,C,B(C),E,D(E),F(G),G,H,ret define void @straight_test(i32 %tag) { entry: br label %test1 test1: ; A %tagbit1 = and i32 %tag, 1 %tagbit1eq0 = icmp eq i32 %tagbit1, 0 br i1 %tagbit1eq0, label %test2, label %optional1 optional1: ; B call void @a() br label %test2 test2: ; C %tagbit2 = and i32 %tag, 2 %tagbit2eq0 = icmp eq i32 %tagbit2, 0 br i1 %tagbit2eq0, label %test3, label %optional2 optional2: ; D call void @b() br label %test3 test3: ; E %tagbit3 = and i32 %tag, 4 %tagbit3eq0 = icmp eq i32 %tagbit3, 0 br i1 %tagbit3eq0, label %test4, label %optional3 optional3: ; F call void @c() br label %test4 test4: ; G %tagbit4 = and i32 %tag, 8 %tagbit4eq0 = icmp eq i32 %tagbit4, 0 br i1 %tagbit4eq0, label %exit, label %optional4 optional4: ; H call void @d() br label %exit exit: ret void } here is the layout after D27742: straight_test: # @straight_test ; ... Prologue elided ; BB#0: # %entry ; A (merged with test1) ; ... More prologue elided mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_2 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_3 b .LBB0_4 .LBB0_2: # %optional1 ; B (copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_4 .LBB0_3: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_5 b .LBB0_6 .LBB0_4: # %optional2 ; D (copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_5: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 b .LBB0_7 .LBB0_6: # %optional3 ; F (copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit ; Ret ld 30, 96(1) # 8-byte Folded Reload addi 1, 1, 112 ld 0, 16(1) mtlr 0 blr The tail-duplication has produced some benefit, but it has also produced a trellis which is not laid out optimally. With this patch, we improve the layouts of such trellises, and decrease the cost calculation for tail-duplication accordingly. This patch produces the layout A,C,E,G,B,D,F,H,Ret. This layout does have back edges, which is a negative, but it has a bigger compensating positive, which is that it handles the case where there are long strings of skipped blocks much better than the original layout. Both layouts handle runs of executed blocks equally well. Branch prediction also improves if there is any correlation between subsequent optional blocks. Here is the resulting concrete layout: straight_test: # @straight_test ; BB#0: # %entry ; A (merged with test1) mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_4 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_5 .LBB0_2: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_3: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 bne 0, .LBB0_7 b .LBB0_8 .LBB0_4: # %optional1 ; B (Copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_2 .LBB0_5: # %optional2 ; D (Copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_3 .LBB0_6: # %optional3 ; F (Copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit Differential Revision: https://reviews.llvm.org/D28522 llvm-svn: 295223
2017-02-16 03:49:14 +08:00
;CHECK-NEXT: bne 0, .[[OPT2LABEL:[_0-9A-Za-z]+]]
;CHECK-NEXT: .[[TEST3LABEL:[_0-9A-Za-z]+]]: # %test3
Codegen: Tail-duplicate during placement. The tail duplication pass uses an assumed layout when making duplication decisions. This is fine, but passes up duplication opportunities that may arise when blocks are outlined. Because we want the updated CFG to affect subsequent placement decisions, this change must occur during placement. In order to achieve this goal, TailDuplicationPass is split into a utility class, TailDuplicator, and the pass itself. The pass delegates nearly everything to the TailDuplicator object, except for looping over the blocks in a function. This allows the same code to be used for tail duplication in both places. This change, in concert with outlining optional branches, allows triangle shaped code to perform much better, esepecially when the taken/untaken branches are correlated, as it creates a second spine when the tests are small enough. Issue from previous rollback fixed, and a new test was added for that case as well. Issue was worklist/scheduling/taildup issue in layout. Issue from 2nd rollback fixed, with 2 additional tests. Issue was tail merging/loop info/tail-duplication causing issue with loops that share a header block. Issue with early tail-duplication of blocks that branch to a fallthrough predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll Differential revision: https://reviews.llvm.org/D18226 llvm-svn: 283934
2016-10-12 04:36:43 +08:00
;CHECK-NEXT: rlwinm. {{[0-9]+}}, [[TAGREG]], 0, 29, 29
Codegen: Make chains from trellis-shaped CFGs Lay out trellis-shaped CFGs optimally. A trellis of the shape below: A B |\ /| | \ / | | X | | / \ | |/ \| C D would be laid out A; B->C ; D by the current layout algorithm. Now we identify trellises and lay them out either A->C; B->D or A->D; B->C. This scales with an increasing number of predecessors. A trellis is a a group of 2 or more predecessor blocks that all have the same successors. because of this we can tail duplicate to extend existing trellises. As an example consider the following CFG: B D F H / \ / \ / \ / \ A---C---E---G---Ret Where A,C,E,G are all small (Currently 2 instructions). The CFG preserving layout is then A,B,C,D,E,F,G,H,Ret. The current code will copy C into B, E into D and G into F and yield the layout A,C,B(C),E,D(E),F(G),G,H,ret define void @straight_test(i32 %tag) { entry: br label %test1 test1: ; A %tagbit1 = and i32 %tag, 1 %tagbit1eq0 = icmp eq i32 %tagbit1, 0 br i1 %tagbit1eq0, label %test2, label %optional1 optional1: ; B call void @a() br label %test2 test2: ; C %tagbit2 = and i32 %tag, 2 %tagbit2eq0 = icmp eq i32 %tagbit2, 0 br i1 %tagbit2eq0, label %test3, label %optional2 optional2: ; D call void @b() br label %test3 test3: ; E %tagbit3 = and i32 %tag, 4 %tagbit3eq0 = icmp eq i32 %tagbit3, 0 br i1 %tagbit3eq0, label %test4, label %optional3 optional3: ; F call void @c() br label %test4 test4: ; G %tagbit4 = and i32 %tag, 8 %tagbit4eq0 = icmp eq i32 %tagbit4, 0 br i1 %tagbit4eq0, label %exit, label %optional4 optional4: ; H call void @d() br label %exit exit: ret void } here is the layout after D27742: straight_test: # @straight_test ; ... Prologue elided ; BB#0: # %entry ; A (merged with test1) ; ... More prologue elided mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_2 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_3 b .LBB0_4 .LBB0_2: # %optional1 ; B (copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_4 .LBB0_3: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_5 b .LBB0_6 .LBB0_4: # %optional2 ; D (copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_5: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 b .LBB0_7 .LBB0_6: # %optional3 ; F (copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit ; Ret ld 30, 96(1) # 8-byte Folded Reload addi 1, 1, 112 ld 0, 16(1) mtlr 0 blr The tail-duplication has produced some benefit, but it has also produced a trellis which is not laid out optimally. With this patch, we improve the layouts of such trellises, and decrease the cost calculation for tail-duplication accordingly. This patch produces the layout A,C,E,G,B,D,F,H,Ret. This layout does have back edges, which is a negative, but it has a bigger compensating positive, which is that it handles the case where there are long strings of skipped blocks much better than the original layout. Both layouts handle runs of executed blocks equally well. Branch prediction also improves if there is any correlation between subsequent optional blocks. Here is the resulting concrete layout: straight_test: # @straight_test ; BB#0: # %entry ; A (merged with test1) mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_4 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_5 .LBB0_2: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_3: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 bne 0, .LBB0_7 b .LBB0_8 .LBB0_4: # %optional1 ; B (Copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_2 .LBB0_5: # %optional2 ; D (Copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_3 .LBB0_6: # %optional3 ; F (Copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit Differential Revision: https://reviews.llvm.org/D28522 llvm-svn: 295223
2017-02-16 03:49:14 +08:00
;CHECK-NEXT: bne 0, .[[OPT3LABEL:[_0-9A-Za-z]+]]
;CHECK-NEXT: .[[TEST4LABEL:[_0-9A-Za-z]+]]: # %test4
Codegen: Tail-duplicate during placement. The tail duplication pass uses an assumed layout when making duplication decisions. This is fine, but passes up duplication opportunities that may arise when blocks are outlined. Because we want the updated CFG to affect subsequent placement decisions, this change must occur during placement. In order to achieve this goal, TailDuplicationPass is split into a utility class, TailDuplicator, and the pass itself. The pass delegates nearly everything to the TailDuplicator object, except for looping over the blocks in a function. This allows the same code to be used for tail duplication in both places. This change, in concert with outlining optional branches, allows triangle shaped code to perform much better, esepecially when the taken/untaken branches are correlated, as it creates a second spine when the tests are small enough. Issue from previous rollback fixed, and a new test was added for that case as well. Issue was worklist/scheduling/taildup issue in layout. Issue from 2nd rollback fixed, with 2 additional tests. Issue was tail merging/loop info/tail-duplication causing issue with loops that share a header block. Issue with early tail-duplication of blocks that branch to a fallthrough predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll Differential revision: https://reviews.llvm.org/D18226 llvm-svn: 283934
2016-10-12 04:36:43 +08:00
;CHECK-NEXT: rlwinm. {{[0-9]+}}, [[TAGREG]], 0, 28, 28
Codegen: Make chains from trellis-shaped CFGs Lay out trellis-shaped CFGs optimally. A trellis of the shape below: A B |\ /| | \ / | | X | | / \ | |/ \| C D would be laid out A; B->C ; D by the current layout algorithm. Now we identify trellises and lay them out either A->C; B->D or A->D; B->C. This scales with an increasing number of predecessors. A trellis is a a group of 2 or more predecessor blocks that all have the same successors. because of this we can tail duplicate to extend existing trellises. As an example consider the following CFG: B D F H / \ / \ / \ / \ A---C---E---G---Ret Where A,C,E,G are all small (Currently 2 instructions). The CFG preserving layout is then A,B,C,D,E,F,G,H,Ret. The current code will copy C into B, E into D and G into F and yield the layout A,C,B(C),E,D(E),F(G),G,H,ret define void @straight_test(i32 %tag) { entry: br label %test1 test1: ; A %tagbit1 = and i32 %tag, 1 %tagbit1eq0 = icmp eq i32 %tagbit1, 0 br i1 %tagbit1eq0, label %test2, label %optional1 optional1: ; B call void @a() br label %test2 test2: ; C %tagbit2 = and i32 %tag, 2 %tagbit2eq0 = icmp eq i32 %tagbit2, 0 br i1 %tagbit2eq0, label %test3, label %optional2 optional2: ; D call void @b() br label %test3 test3: ; E %tagbit3 = and i32 %tag, 4 %tagbit3eq0 = icmp eq i32 %tagbit3, 0 br i1 %tagbit3eq0, label %test4, label %optional3 optional3: ; F call void @c() br label %test4 test4: ; G %tagbit4 = and i32 %tag, 8 %tagbit4eq0 = icmp eq i32 %tagbit4, 0 br i1 %tagbit4eq0, label %exit, label %optional4 optional4: ; H call void @d() br label %exit exit: ret void } here is the layout after D27742: straight_test: # @straight_test ; ... Prologue elided ; BB#0: # %entry ; A (merged with test1) ; ... More prologue elided mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_2 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_3 b .LBB0_4 .LBB0_2: # %optional1 ; B (copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_4 .LBB0_3: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_5 b .LBB0_6 .LBB0_4: # %optional2 ; D (copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_5: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 b .LBB0_7 .LBB0_6: # %optional3 ; F (copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit ; Ret ld 30, 96(1) # 8-byte Folded Reload addi 1, 1, 112 ld 0, 16(1) mtlr 0 blr The tail-duplication has produced some benefit, but it has also produced a trellis which is not laid out optimally. With this patch, we improve the layouts of such trellises, and decrease the cost calculation for tail-duplication accordingly. This patch produces the layout A,C,E,G,B,D,F,H,Ret. This layout does have back edges, which is a negative, but it has a bigger compensating positive, which is that it handles the case where there are long strings of skipped blocks much better than the original layout. Both layouts handle runs of executed blocks equally well. Branch prediction also improves if there is any correlation between subsequent optional blocks. Here is the resulting concrete layout: straight_test: # @straight_test ; BB#0: # %entry ; A (merged with test1) mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_4 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_5 .LBB0_2: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_3: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 bne 0, .LBB0_7 b .LBB0_8 .LBB0_4: # %optional1 ; B (Copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_2 .LBB0_5: # %optional2 ; D (Copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_3 .LBB0_6: # %optional3 ; F (Copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit Differential Revision: https://reviews.llvm.org/D28522 llvm-svn: 295223
2017-02-16 03:49:14 +08:00
;CHECK-NEXT: bne 0, .[[OPT4LABEL:[_0-9A-Za-z]+]]
;CHECK-NEXT: .[[EXITLABEL:[_0-9A-Za-z]+]]: # %exit
Codegen: Tail-duplicate during placement. The tail duplication pass uses an assumed layout when making duplication decisions. This is fine, but passes up duplication opportunities that may arise when blocks are outlined. Because we want the updated CFG to affect subsequent placement decisions, this change must occur during placement. In order to achieve this goal, TailDuplicationPass is split into a utility class, TailDuplicator, and the pass itself. The pass delegates nearly everything to the TailDuplicator object, except for looping over the blocks in a function. This allows the same code to be used for tail duplication in both places. This change, in concert with outlining optional branches, allows triangle shaped code to perform much better, esepecially when the taken/untaken branches are correlated, as it creates a second spine when the tests are small enough. Issue from previous rollback fixed, and a new test was added for that case as well. Issue was worklist/scheduling/taildup issue in layout. Issue from 2nd rollback fixed, with 2 additional tests. Issue was tail merging/loop info/tail-duplication causing issue with loops that share a header block. Issue with early tail-duplication of blocks that branch to a fallthrough predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll Differential revision: https://reviews.llvm.org/D18226 llvm-svn: 283934
2016-10-12 04:36:43 +08:00
;CHECK: blr
Codegen: Make chains from trellis-shaped CFGs Lay out trellis-shaped CFGs optimally. A trellis of the shape below: A B |\ /| | \ / | | X | | / \ | |/ \| C D would be laid out A; B->C ; D by the current layout algorithm. Now we identify trellises and lay them out either A->C; B->D or A->D; B->C. This scales with an increasing number of predecessors. A trellis is a a group of 2 or more predecessor blocks that all have the same successors. because of this we can tail duplicate to extend existing trellises. As an example consider the following CFG: B D F H / \ / \ / \ / \ A---C---E---G---Ret Where A,C,E,G are all small (Currently 2 instructions). The CFG preserving layout is then A,B,C,D,E,F,G,H,Ret. The current code will copy C into B, E into D and G into F and yield the layout A,C,B(C),E,D(E),F(G),G,H,ret define void @straight_test(i32 %tag) { entry: br label %test1 test1: ; A %tagbit1 = and i32 %tag, 1 %tagbit1eq0 = icmp eq i32 %tagbit1, 0 br i1 %tagbit1eq0, label %test2, label %optional1 optional1: ; B call void @a() br label %test2 test2: ; C %tagbit2 = and i32 %tag, 2 %tagbit2eq0 = icmp eq i32 %tagbit2, 0 br i1 %tagbit2eq0, label %test3, label %optional2 optional2: ; D call void @b() br label %test3 test3: ; E %tagbit3 = and i32 %tag, 4 %tagbit3eq0 = icmp eq i32 %tagbit3, 0 br i1 %tagbit3eq0, label %test4, label %optional3 optional3: ; F call void @c() br label %test4 test4: ; G %tagbit4 = and i32 %tag, 8 %tagbit4eq0 = icmp eq i32 %tagbit4, 0 br i1 %tagbit4eq0, label %exit, label %optional4 optional4: ; H call void @d() br label %exit exit: ret void } here is the layout after D27742: straight_test: # @straight_test ; ... Prologue elided ; BB#0: # %entry ; A (merged with test1) ; ... More prologue elided mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_2 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_3 b .LBB0_4 .LBB0_2: # %optional1 ; B (copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_4 .LBB0_3: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_5 b .LBB0_6 .LBB0_4: # %optional2 ; D (copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_5: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 b .LBB0_7 .LBB0_6: # %optional3 ; F (copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit ; Ret ld 30, 96(1) # 8-byte Folded Reload addi 1, 1, 112 ld 0, 16(1) mtlr 0 blr The tail-duplication has produced some benefit, but it has also produced a trellis which is not laid out optimally. With this patch, we improve the layouts of such trellises, and decrease the cost calculation for tail-duplication accordingly. This patch produces the layout A,C,E,G,B,D,F,H,Ret. This layout does have back edges, which is a negative, but it has a bigger compensating positive, which is that it handles the case where there are long strings of skipped blocks much better than the original layout. Both layouts handle runs of executed blocks equally well. Branch prediction also improves if there is any correlation between subsequent optional blocks. Here is the resulting concrete layout: straight_test: # @straight_test ; BB#0: # %entry ; A (merged with test1) mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_4 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_5 .LBB0_2: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_3: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 bne 0, .LBB0_7 b .LBB0_8 .LBB0_4: # %optional1 ; B (Copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_2 .LBB0_5: # %optional2 ; D (Copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_3 .LBB0_6: # %optional3 ; F (Copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit Differential Revision: https://reviews.llvm.org/D28522 llvm-svn: 295223
2017-02-16 03:49:14 +08:00
;CHECK-NEXT: .[[OPT1LABEL]]:
Codegen: Tail-duplicate during placement. The tail duplication pass uses an assumed layout when making duplication decisions. This is fine, but passes up duplication opportunities that may arise when blocks are outlined. Because we want the updated CFG to affect subsequent placement decisions, this change must occur during placement. In order to achieve this goal, TailDuplicationPass is split into a utility class, TailDuplicator, and the pass itself. The pass delegates nearly everything to the TailDuplicator object, except for looping over the blocks in a function. This allows the same code to be used for tail duplication in both places. This change, in concert with outlining optional branches, allows triangle shaped code to perform much better, esepecially when the taken/untaken branches are correlated, as it creates a second spine when the tests are small enough. Issue from previous rollback fixed, and a new test was added for that case as well. Issue was worklist/scheduling/taildup issue in layout. Issue from 2nd rollback fixed, with 2 additional tests. Issue was tail merging/loop info/tail-duplication causing issue with loops that share a header block. Issue with early tail-duplication of blocks that branch to a fallthrough predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll Differential revision: https://reviews.llvm.org/D18226 llvm-svn: 283934
2016-10-12 04:36:43 +08:00
;CHECK: rlwinm. {{[0-9]+}}, [[TAGREG]], 0, 30, 30
Codegen: Make chains from trellis-shaped CFGs Lay out trellis-shaped CFGs optimally. A trellis of the shape below: A B |\ /| | \ / | | X | | / \ | |/ \| C D would be laid out A; B->C ; D by the current layout algorithm. Now we identify trellises and lay them out either A->C; B->D or A->D; B->C. This scales with an increasing number of predecessors. A trellis is a a group of 2 or more predecessor blocks that all have the same successors. because of this we can tail duplicate to extend existing trellises. As an example consider the following CFG: B D F H / \ / \ / \ / \ A---C---E---G---Ret Where A,C,E,G are all small (Currently 2 instructions). The CFG preserving layout is then A,B,C,D,E,F,G,H,Ret. The current code will copy C into B, E into D and G into F and yield the layout A,C,B(C),E,D(E),F(G),G,H,ret define void @straight_test(i32 %tag) { entry: br label %test1 test1: ; A %tagbit1 = and i32 %tag, 1 %tagbit1eq0 = icmp eq i32 %tagbit1, 0 br i1 %tagbit1eq0, label %test2, label %optional1 optional1: ; B call void @a() br label %test2 test2: ; C %tagbit2 = and i32 %tag, 2 %tagbit2eq0 = icmp eq i32 %tagbit2, 0 br i1 %tagbit2eq0, label %test3, label %optional2 optional2: ; D call void @b() br label %test3 test3: ; E %tagbit3 = and i32 %tag, 4 %tagbit3eq0 = icmp eq i32 %tagbit3, 0 br i1 %tagbit3eq0, label %test4, label %optional3 optional3: ; F call void @c() br label %test4 test4: ; G %tagbit4 = and i32 %tag, 8 %tagbit4eq0 = icmp eq i32 %tagbit4, 0 br i1 %tagbit4eq0, label %exit, label %optional4 optional4: ; H call void @d() br label %exit exit: ret void } here is the layout after D27742: straight_test: # @straight_test ; ... Prologue elided ; BB#0: # %entry ; A (merged with test1) ; ... More prologue elided mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_2 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_3 b .LBB0_4 .LBB0_2: # %optional1 ; B (copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_4 .LBB0_3: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_5 b .LBB0_6 .LBB0_4: # %optional2 ; D (copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_5: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 b .LBB0_7 .LBB0_6: # %optional3 ; F (copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit ; Ret ld 30, 96(1) # 8-byte Folded Reload addi 1, 1, 112 ld 0, 16(1) mtlr 0 blr The tail-duplication has produced some benefit, but it has also produced a trellis which is not laid out optimally. With this patch, we improve the layouts of such trellises, and decrease the cost calculation for tail-duplication accordingly. This patch produces the layout A,C,E,G,B,D,F,H,Ret. This layout does have back edges, which is a negative, but it has a bigger compensating positive, which is that it handles the case where there are long strings of skipped blocks much better than the original layout. Both layouts handle runs of executed blocks equally well. Branch prediction also improves if there is any correlation between subsequent optional blocks. Here is the resulting concrete layout: straight_test: # @straight_test ; BB#0: # %entry ; A (merged with test1) mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_4 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_5 .LBB0_2: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_3: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 bne 0, .LBB0_7 b .LBB0_8 .LBB0_4: # %optional1 ; B (Copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_2 .LBB0_5: # %optional2 ; D (Copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_3 .LBB0_6: # %optional3 ; F (Copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit Differential Revision: https://reviews.llvm.org/D28522 llvm-svn: 295223
2017-02-16 03:49:14 +08:00
;CHECK-NEXT: beq 0, .[[TEST3LABEL]]
;CHECK-NEXT: .[[OPT2LABEL]]:
Codegen: Tail-duplicate during placement. The tail duplication pass uses an assumed layout when making duplication decisions. This is fine, but passes up duplication opportunities that may arise when blocks are outlined. Because we want the updated CFG to affect subsequent placement decisions, this change must occur during placement. In order to achieve this goal, TailDuplicationPass is split into a utility class, TailDuplicator, and the pass itself. The pass delegates nearly everything to the TailDuplicator object, except for looping over the blocks in a function. This allows the same code to be used for tail duplication in both places. This change, in concert with outlining optional branches, allows triangle shaped code to perform much better, esepecially when the taken/untaken branches are correlated, as it creates a second spine when the tests are small enough. Issue from previous rollback fixed, and a new test was added for that case as well. Issue was worklist/scheduling/taildup issue in layout. Issue from 2nd rollback fixed, with 2 additional tests. Issue was tail merging/loop info/tail-duplication causing issue with loops that share a header block. Issue with early tail-duplication of blocks that branch to a fallthrough predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll Differential revision: https://reviews.llvm.org/D18226 llvm-svn: 283934
2016-10-12 04:36:43 +08:00
;CHECK: rlwinm. {{[0-9]+}}, [[TAGREG]], 0, 29, 29
Codegen: Make chains from trellis-shaped CFGs Lay out trellis-shaped CFGs optimally. A trellis of the shape below: A B |\ /| | \ / | | X | | / \ | |/ \| C D would be laid out A; B->C ; D by the current layout algorithm. Now we identify trellises and lay them out either A->C; B->D or A->D; B->C. This scales with an increasing number of predecessors. A trellis is a a group of 2 or more predecessor blocks that all have the same successors. because of this we can tail duplicate to extend existing trellises. As an example consider the following CFG: B D F H / \ / \ / \ / \ A---C---E---G---Ret Where A,C,E,G are all small (Currently 2 instructions). The CFG preserving layout is then A,B,C,D,E,F,G,H,Ret. The current code will copy C into B, E into D and G into F and yield the layout A,C,B(C),E,D(E),F(G),G,H,ret define void @straight_test(i32 %tag) { entry: br label %test1 test1: ; A %tagbit1 = and i32 %tag, 1 %tagbit1eq0 = icmp eq i32 %tagbit1, 0 br i1 %tagbit1eq0, label %test2, label %optional1 optional1: ; B call void @a() br label %test2 test2: ; C %tagbit2 = and i32 %tag, 2 %tagbit2eq0 = icmp eq i32 %tagbit2, 0 br i1 %tagbit2eq0, label %test3, label %optional2 optional2: ; D call void @b() br label %test3 test3: ; E %tagbit3 = and i32 %tag, 4 %tagbit3eq0 = icmp eq i32 %tagbit3, 0 br i1 %tagbit3eq0, label %test4, label %optional3 optional3: ; F call void @c() br label %test4 test4: ; G %tagbit4 = and i32 %tag, 8 %tagbit4eq0 = icmp eq i32 %tagbit4, 0 br i1 %tagbit4eq0, label %exit, label %optional4 optional4: ; H call void @d() br label %exit exit: ret void } here is the layout after D27742: straight_test: # @straight_test ; ... Prologue elided ; BB#0: # %entry ; A (merged with test1) ; ... More prologue elided mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_2 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_3 b .LBB0_4 .LBB0_2: # %optional1 ; B (copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_4 .LBB0_3: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_5 b .LBB0_6 .LBB0_4: # %optional2 ; D (copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_5: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 b .LBB0_7 .LBB0_6: # %optional3 ; F (copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit ; Ret ld 30, 96(1) # 8-byte Folded Reload addi 1, 1, 112 ld 0, 16(1) mtlr 0 blr The tail-duplication has produced some benefit, but it has also produced a trellis which is not laid out optimally. With this patch, we improve the layouts of such trellises, and decrease the cost calculation for tail-duplication accordingly. This patch produces the layout A,C,E,G,B,D,F,H,Ret. This layout does have back edges, which is a negative, but it has a bigger compensating positive, which is that it handles the case where there are long strings of skipped blocks much better than the original layout. Both layouts handle runs of executed blocks equally well. Branch prediction also improves if there is any correlation between subsequent optional blocks. Here is the resulting concrete layout: straight_test: # @straight_test ; BB#0: # %entry ; A (merged with test1) mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_4 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_5 .LBB0_2: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_3: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 bne 0, .LBB0_7 b .LBB0_8 .LBB0_4: # %optional1 ; B (Copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_2 .LBB0_5: # %optional2 ; D (Copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_3 .LBB0_6: # %optional3 ; F (Copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit Differential Revision: https://reviews.llvm.org/D28522 llvm-svn: 295223
2017-02-16 03:49:14 +08:00
;CHECK-NEXT: beq 0, .[[TEST4LABEL]]
;CHECK-NEXT: .[[OPT3LABEL]]:
Codegen: Tail-duplicate during placement. The tail duplication pass uses an assumed layout when making duplication decisions. This is fine, but passes up duplication opportunities that may arise when blocks are outlined. Because we want the updated CFG to affect subsequent placement decisions, this change must occur during placement. In order to achieve this goal, TailDuplicationPass is split into a utility class, TailDuplicator, and the pass itself. The pass delegates nearly everything to the TailDuplicator object, except for looping over the blocks in a function. This allows the same code to be used for tail duplication in both places. This change, in concert with outlining optional branches, allows triangle shaped code to perform much better, esepecially when the taken/untaken branches are correlated, as it creates a second spine when the tests are small enough. Issue from previous rollback fixed, and a new test was added for that case as well. Issue was worklist/scheduling/taildup issue in layout. Issue from 2nd rollback fixed, with 2 additional tests. Issue was tail merging/loop info/tail-duplication causing issue with loops that share a header block. Issue with early tail-duplication of blocks that branch to a fallthrough predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll Differential revision: https://reviews.llvm.org/D18226 llvm-svn: 283934
2016-10-12 04:36:43 +08:00
;CHECK: rlwinm. {{[0-9]+}}, [[TAGREG]], 0, 28, 28
Codegen: Make chains from trellis-shaped CFGs Lay out trellis-shaped CFGs optimally. A trellis of the shape below: A B |\ /| | \ / | | X | | / \ | |/ \| C D would be laid out A; B->C ; D by the current layout algorithm. Now we identify trellises and lay them out either A->C; B->D or A->D; B->C. This scales with an increasing number of predecessors. A trellis is a a group of 2 or more predecessor blocks that all have the same successors. because of this we can tail duplicate to extend existing trellises. As an example consider the following CFG: B D F H / \ / \ / \ / \ A---C---E---G---Ret Where A,C,E,G are all small (Currently 2 instructions). The CFG preserving layout is then A,B,C,D,E,F,G,H,Ret. The current code will copy C into B, E into D and G into F and yield the layout A,C,B(C),E,D(E),F(G),G,H,ret define void @straight_test(i32 %tag) { entry: br label %test1 test1: ; A %tagbit1 = and i32 %tag, 1 %tagbit1eq0 = icmp eq i32 %tagbit1, 0 br i1 %tagbit1eq0, label %test2, label %optional1 optional1: ; B call void @a() br label %test2 test2: ; C %tagbit2 = and i32 %tag, 2 %tagbit2eq0 = icmp eq i32 %tagbit2, 0 br i1 %tagbit2eq0, label %test3, label %optional2 optional2: ; D call void @b() br label %test3 test3: ; E %tagbit3 = and i32 %tag, 4 %tagbit3eq0 = icmp eq i32 %tagbit3, 0 br i1 %tagbit3eq0, label %test4, label %optional3 optional3: ; F call void @c() br label %test4 test4: ; G %tagbit4 = and i32 %tag, 8 %tagbit4eq0 = icmp eq i32 %tagbit4, 0 br i1 %tagbit4eq0, label %exit, label %optional4 optional4: ; H call void @d() br label %exit exit: ret void } here is the layout after D27742: straight_test: # @straight_test ; ... Prologue elided ; BB#0: # %entry ; A (merged with test1) ; ... More prologue elided mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_2 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_3 b .LBB0_4 .LBB0_2: # %optional1 ; B (copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_4 .LBB0_3: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_5 b .LBB0_6 .LBB0_4: # %optional2 ; D (copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_5: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 b .LBB0_7 .LBB0_6: # %optional3 ; F (copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit ; Ret ld 30, 96(1) # 8-byte Folded Reload addi 1, 1, 112 ld 0, 16(1) mtlr 0 blr The tail-duplication has produced some benefit, but it has also produced a trellis which is not laid out optimally. With this patch, we improve the layouts of such trellises, and decrease the cost calculation for tail-duplication accordingly. This patch produces the layout A,C,E,G,B,D,F,H,Ret. This layout does have back edges, which is a negative, but it has a bigger compensating positive, which is that it handles the case where there are long strings of skipped blocks much better than the original layout. Both layouts handle runs of executed blocks equally well. Branch prediction also improves if there is any correlation between subsequent optional blocks. Here is the resulting concrete layout: straight_test: # @straight_test ; BB#0: # %entry ; A (merged with test1) mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_4 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_5 .LBB0_2: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_3: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 bne 0, .LBB0_7 b .LBB0_8 .LBB0_4: # %optional1 ; B (Copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_2 .LBB0_5: # %optional2 ; D (Copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_3 .LBB0_6: # %optional3 ; F (Copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit Differential Revision: https://reviews.llvm.org/D28522 llvm-svn: 295223
2017-02-16 03:49:14 +08:00
;CHECK-NEXT: beq 0, .[[EXITLABEL]]
;CHECK-NEXT: .[[OPT4LABEL]]:
;CHECK: b .[[EXITLABEL]]
Codegen: Tail-duplicate during placement. The tail duplication pass uses an assumed layout when making duplication decisions. This is fine, but passes up duplication opportunities that may arise when blocks are outlined. Because we want the updated CFG to affect subsequent placement decisions, this change must occur during placement. In order to achieve this goal, TailDuplicationPass is split into a utility class, TailDuplicator, and the pass itself. The pass delegates nearly everything to the TailDuplicator object, except for looping over the blocks in a function. This allows the same code to be used for tail duplication in both places. This change, in concert with outlining optional branches, allows triangle shaped code to perform much better, esepecially when the taken/untaken branches are correlated, as it creates a second spine when the tests are small enough. Issue from previous rollback fixed, and a new test was added for that case as well. Issue was worklist/scheduling/taildup issue in layout. Issue from 2nd rollback fixed, with 2 additional tests. Issue was tail merging/loop info/tail-duplication causing issue with loops that share a header block. Issue with early tail-duplication of blocks that branch to a fallthrough predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll Differential revision: https://reviews.llvm.org/D18226 llvm-svn: 283934
2016-10-12 04:36:43 +08:00
Codegen: Make chains from trellis-shaped CFGs Lay out trellis-shaped CFGs optimally. A trellis of the shape below: A B |\ /| | \ / | | X | | / \ | |/ \| C D would be laid out A; B->C ; D by the current layout algorithm. Now we identify trellises and lay them out either A->C; B->D or A->D; B->C. This scales with an increasing number of predecessors. A trellis is a a group of 2 or more predecessor blocks that all have the same successors. because of this we can tail duplicate to extend existing trellises. As an example consider the following CFG: B D F H / \ / \ / \ / \ A---C---E---G---Ret Where A,C,E,G are all small (Currently 2 instructions). The CFG preserving layout is then A,B,C,D,E,F,G,H,Ret. The current code will copy C into B, E into D and G into F and yield the layout A,C,B(C),E,D(E),F(G),G,H,ret define void @straight_test(i32 %tag) { entry: br label %test1 test1: ; A %tagbit1 = and i32 %tag, 1 %tagbit1eq0 = icmp eq i32 %tagbit1, 0 br i1 %tagbit1eq0, label %test2, label %optional1 optional1: ; B call void @a() br label %test2 test2: ; C %tagbit2 = and i32 %tag, 2 %tagbit2eq0 = icmp eq i32 %tagbit2, 0 br i1 %tagbit2eq0, label %test3, label %optional2 optional2: ; D call void @b() br label %test3 test3: ; E %tagbit3 = and i32 %tag, 4 %tagbit3eq0 = icmp eq i32 %tagbit3, 0 br i1 %tagbit3eq0, label %test4, label %optional3 optional3: ; F call void @c() br label %test4 test4: ; G %tagbit4 = and i32 %tag, 8 %tagbit4eq0 = icmp eq i32 %tagbit4, 0 br i1 %tagbit4eq0, label %exit, label %optional4 optional4: ; H call void @d() br label %exit exit: ret void } here is the layout after D27742: straight_test: # @straight_test ; ... Prologue elided ; BB#0: # %entry ; A (merged with test1) ; ... More prologue elided mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_2 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_3 b .LBB0_4 .LBB0_2: # %optional1 ; B (copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_4 .LBB0_3: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_5 b .LBB0_6 .LBB0_4: # %optional2 ; D (copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_5: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 b .LBB0_7 .LBB0_6: # %optional3 ; F (copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit ; Ret ld 30, 96(1) # 8-byte Folded Reload addi 1, 1, 112 ld 0, 16(1) mtlr 0 blr The tail-duplication has produced some benefit, but it has also produced a trellis which is not laid out optimally. With this patch, we improve the layouts of such trellises, and decrease the cost calculation for tail-duplication accordingly. This patch produces the layout A,C,E,G,B,D,F,H,Ret. This layout does have back edges, which is a negative, but it has a bigger compensating positive, which is that it handles the case where there are long strings of skipped blocks much better than the original layout. Both layouts handle runs of executed blocks equally well. Branch prediction also improves if there is any correlation between subsequent optional blocks. Here is the resulting concrete layout: straight_test: # @straight_test ; BB#0: # %entry ; A (merged with test1) mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_4 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_5 .LBB0_2: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_3: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 bne 0, .LBB0_7 b .LBB0_8 .LBB0_4: # %optional1 ; B (Copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_2 .LBB0_5: # %optional2 ; D (Copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_3 .LBB0_6: # %optional3 ; F (Copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit Differential Revision: https://reviews.llvm.org/D28522 llvm-svn: 295223
2017-02-16 03:49:14 +08:00
define void @straight_test(i32 %tag) {
Codegen: Tail-duplicate during placement. The tail duplication pass uses an assumed layout when making duplication decisions. This is fine, but passes up duplication opportunities that may arise when blocks are outlined. Because we want the updated CFG to affect subsequent placement decisions, this change must occur during placement. In order to achieve this goal, TailDuplicationPass is split into a utility class, TailDuplicator, and the pass itself. The pass delegates nearly everything to the TailDuplicator object, except for looping over the blocks in a function. This allows the same code to be used for tail duplication in both places. This change, in concert with outlining optional branches, allows triangle shaped code to perform much better, esepecially when the taken/untaken branches are correlated, as it creates a second spine when the tests are small enough. Issue from previous rollback fixed, and a new test was added for that case as well. Issue was worklist/scheduling/taildup issue in layout. Issue from 2nd rollback fixed, with 2 additional tests. Issue was tail merging/loop info/tail-duplication causing issue with loops that share a header block. Issue with early tail-duplication of blocks that branch to a fallthrough predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll Differential revision: https://reviews.llvm.org/D18226 llvm-svn: 283934
2016-10-12 04:36:43 +08:00
entry:
br label %test1
test1:
%tagbit1 = and i32 %tag, 1
%tagbit1eq0 = icmp eq i32 %tagbit1, 0
Codegen: Make chains from trellis-shaped CFGs Lay out trellis-shaped CFGs optimally. A trellis of the shape below: A B |\ /| | \ / | | X | | / \ | |/ \| C D would be laid out A; B->C ; D by the current layout algorithm. Now we identify trellises and lay them out either A->C; B->D or A->D; B->C. This scales with an increasing number of predecessors. A trellis is a a group of 2 or more predecessor blocks that all have the same successors. because of this we can tail duplicate to extend existing trellises. As an example consider the following CFG: B D F H / \ / \ / \ / \ A---C---E---G---Ret Where A,C,E,G are all small (Currently 2 instructions). The CFG preserving layout is then A,B,C,D,E,F,G,H,Ret. The current code will copy C into B, E into D and G into F and yield the layout A,C,B(C),E,D(E),F(G),G,H,ret define void @straight_test(i32 %tag) { entry: br label %test1 test1: ; A %tagbit1 = and i32 %tag, 1 %tagbit1eq0 = icmp eq i32 %tagbit1, 0 br i1 %tagbit1eq0, label %test2, label %optional1 optional1: ; B call void @a() br label %test2 test2: ; C %tagbit2 = and i32 %tag, 2 %tagbit2eq0 = icmp eq i32 %tagbit2, 0 br i1 %tagbit2eq0, label %test3, label %optional2 optional2: ; D call void @b() br label %test3 test3: ; E %tagbit3 = and i32 %tag, 4 %tagbit3eq0 = icmp eq i32 %tagbit3, 0 br i1 %tagbit3eq0, label %test4, label %optional3 optional3: ; F call void @c() br label %test4 test4: ; G %tagbit4 = and i32 %tag, 8 %tagbit4eq0 = icmp eq i32 %tagbit4, 0 br i1 %tagbit4eq0, label %exit, label %optional4 optional4: ; H call void @d() br label %exit exit: ret void } here is the layout after D27742: straight_test: # @straight_test ; ... Prologue elided ; BB#0: # %entry ; A (merged with test1) ; ... More prologue elided mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_2 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_3 b .LBB0_4 .LBB0_2: # %optional1 ; B (copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_4 .LBB0_3: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_5 b .LBB0_6 .LBB0_4: # %optional2 ; D (copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_5: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 b .LBB0_7 .LBB0_6: # %optional3 ; F (copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit ; Ret ld 30, 96(1) # 8-byte Folded Reload addi 1, 1, 112 ld 0, 16(1) mtlr 0 blr The tail-duplication has produced some benefit, but it has also produced a trellis which is not laid out optimally. With this patch, we improve the layouts of such trellises, and decrease the cost calculation for tail-duplication accordingly. This patch produces the layout A,C,E,G,B,D,F,H,Ret. This layout does have back edges, which is a negative, but it has a bigger compensating positive, which is that it handles the case where there are long strings of skipped blocks much better than the original layout. Both layouts handle runs of executed blocks equally well. Branch prediction also improves if there is any correlation between subsequent optional blocks. Here is the resulting concrete layout: straight_test: # @straight_test ; BB#0: # %entry ; A (merged with test1) mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_4 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_5 .LBB0_2: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_3: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 bne 0, .LBB0_7 b .LBB0_8 .LBB0_4: # %optional1 ; B (Copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_2 .LBB0_5: # %optional2 ; D (Copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_3 .LBB0_6: # %optional3 ; F (Copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit Differential Revision: https://reviews.llvm.org/D28522 llvm-svn: 295223
2017-02-16 03:49:14 +08:00
br i1 %tagbit1eq0, label %test2, label %optional1, !prof !1
Codegen: Tail-duplicate during placement. The tail duplication pass uses an assumed layout when making duplication decisions. This is fine, but passes up duplication opportunities that may arise when blocks are outlined. Because we want the updated CFG to affect subsequent placement decisions, this change must occur during placement. In order to achieve this goal, TailDuplicationPass is split into a utility class, TailDuplicator, and the pass itself. The pass delegates nearly everything to the TailDuplicator object, except for looping over the blocks in a function. This allows the same code to be used for tail duplication in both places. This change, in concert with outlining optional branches, allows triangle shaped code to perform much better, esepecially when the taken/untaken branches are correlated, as it creates a second spine when the tests are small enough. Issue from previous rollback fixed, and a new test was added for that case as well. Issue was worklist/scheduling/taildup issue in layout. Issue from 2nd rollback fixed, with 2 additional tests. Issue was tail merging/loop info/tail-duplication causing issue with loops that share a header block. Issue with early tail-duplication of blocks that branch to a fallthrough predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll Differential revision: https://reviews.llvm.org/D18226 llvm-svn: 283934
2016-10-12 04:36:43 +08:00
optional1:
call void @a()
call void @a()
call void @a()
call void @a()
br label %test2
test2:
%tagbit2 = and i32 %tag, 2
%tagbit2eq0 = icmp eq i32 %tagbit2, 0
Codegen: Make chains from trellis-shaped CFGs Lay out trellis-shaped CFGs optimally. A trellis of the shape below: A B |\ /| | \ / | | X | | / \ | |/ \| C D would be laid out A; B->C ; D by the current layout algorithm. Now we identify trellises and lay them out either A->C; B->D or A->D; B->C. This scales with an increasing number of predecessors. A trellis is a a group of 2 or more predecessor blocks that all have the same successors. because of this we can tail duplicate to extend existing trellises. As an example consider the following CFG: B D F H / \ / \ / \ / \ A---C---E---G---Ret Where A,C,E,G are all small (Currently 2 instructions). The CFG preserving layout is then A,B,C,D,E,F,G,H,Ret. The current code will copy C into B, E into D and G into F and yield the layout A,C,B(C),E,D(E),F(G),G,H,ret define void @straight_test(i32 %tag) { entry: br label %test1 test1: ; A %tagbit1 = and i32 %tag, 1 %tagbit1eq0 = icmp eq i32 %tagbit1, 0 br i1 %tagbit1eq0, label %test2, label %optional1 optional1: ; B call void @a() br label %test2 test2: ; C %tagbit2 = and i32 %tag, 2 %tagbit2eq0 = icmp eq i32 %tagbit2, 0 br i1 %tagbit2eq0, label %test3, label %optional2 optional2: ; D call void @b() br label %test3 test3: ; E %tagbit3 = and i32 %tag, 4 %tagbit3eq0 = icmp eq i32 %tagbit3, 0 br i1 %tagbit3eq0, label %test4, label %optional3 optional3: ; F call void @c() br label %test4 test4: ; G %tagbit4 = and i32 %tag, 8 %tagbit4eq0 = icmp eq i32 %tagbit4, 0 br i1 %tagbit4eq0, label %exit, label %optional4 optional4: ; H call void @d() br label %exit exit: ret void } here is the layout after D27742: straight_test: # @straight_test ; ... Prologue elided ; BB#0: # %entry ; A (merged with test1) ; ... More prologue elided mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_2 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_3 b .LBB0_4 .LBB0_2: # %optional1 ; B (copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_4 .LBB0_3: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_5 b .LBB0_6 .LBB0_4: # %optional2 ; D (copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_5: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 b .LBB0_7 .LBB0_6: # %optional3 ; F (copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit ; Ret ld 30, 96(1) # 8-byte Folded Reload addi 1, 1, 112 ld 0, 16(1) mtlr 0 blr The tail-duplication has produced some benefit, but it has also produced a trellis which is not laid out optimally. With this patch, we improve the layouts of such trellises, and decrease the cost calculation for tail-duplication accordingly. This patch produces the layout A,C,E,G,B,D,F,H,Ret. This layout does have back edges, which is a negative, but it has a bigger compensating positive, which is that it handles the case where there are long strings of skipped blocks much better than the original layout. Both layouts handle runs of executed blocks equally well. Branch prediction also improves if there is any correlation between subsequent optional blocks. Here is the resulting concrete layout: straight_test: # @straight_test ; BB#0: # %entry ; A (merged with test1) mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_4 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_5 .LBB0_2: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_3: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 bne 0, .LBB0_7 b .LBB0_8 .LBB0_4: # %optional1 ; B (Copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_2 .LBB0_5: # %optional2 ; D (Copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_3 .LBB0_6: # %optional3 ; F (Copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit Differential Revision: https://reviews.llvm.org/D28522 llvm-svn: 295223
2017-02-16 03:49:14 +08:00
br i1 %tagbit2eq0, label %test3, label %optional2, !prof !1
Codegen: Tail-duplicate during placement. The tail duplication pass uses an assumed layout when making duplication decisions. This is fine, but passes up duplication opportunities that may arise when blocks are outlined. Because we want the updated CFG to affect subsequent placement decisions, this change must occur during placement. In order to achieve this goal, TailDuplicationPass is split into a utility class, TailDuplicator, and the pass itself. The pass delegates nearly everything to the TailDuplicator object, except for looping over the blocks in a function. This allows the same code to be used for tail duplication in both places. This change, in concert with outlining optional branches, allows triangle shaped code to perform much better, esepecially when the taken/untaken branches are correlated, as it creates a second spine when the tests are small enough. Issue from previous rollback fixed, and a new test was added for that case as well. Issue was worklist/scheduling/taildup issue in layout. Issue from 2nd rollback fixed, with 2 additional tests. Issue was tail merging/loop info/tail-duplication causing issue with loops that share a header block. Issue with early tail-duplication of blocks that branch to a fallthrough predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll Differential revision: https://reviews.llvm.org/D18226 llvm-svn: 283934
2016-10-12 04:36:43 +08:00
optional2:
call void @b()
call void @b()
call void @b()
call void @b()
br label %test3
test3:
%tagbit3 = and i32 %tag, 4
%tagbit3eq0 = icmp eq i32 %tagbit3, 0
Codegen: Make chains from trellis-shaped CFGs Lay out trellis-shaped CFGs optimally. A trellis of the shape below: A B |\ /| | \ / | | X | | / \ | |/ \| C D would be laid out A; B->C ; D by the current layout algorithm. Now we identify trellises and lay them out either A->C; B->D or A->D; B->C. This scales with an increasing number of predecessors. A trellis is a a group of 2 or more predecessor blocks that all have the same successors. because of this we can tail duplicate to extend existing trellises. As an example consider the following CFG: B D F H / \ / \ / \ / \ A---C---E---G---Ret Where A,C,E,G are all small (Currently 2 instructions). The CFG preserving layout is then A,B,C,D,E,F,G,H,Ret. The current code will copy C into B, E into D and G into F and yield the layout A,C,B(C),E,D(E),F(G),G,H,ret define void @straight_test(i32 %tag) { entry: br label %test1 test1: ; A %tagbit1 = and i32 %tag, 1 %tagbit1eq0 = icmp eq i32 %tagbit1, 0 br i1 %tagbit1eq0, label %test2, label %optional1 optional1: ; B call void @a() br label %test2 test2: ; C %tagbit2 = and i32 %tag, 2 %tagbit2eq0 = icmp eq i32 %tagbit2, 0 br i1 %tagbit2eq0, label %test3, label %optional2 optional2: ; D call void @b() br label %test3 test3: ; E %tagbit3 = and i32 %tag, 4 %tagbit3eq0 = icmp eq i32 %tagbit3, 0 br i1 %tagbit3eq0, label %test4, label %optional3 optional3: ; F call void @c() br label %test4 test4: ; G %tagbit4 = and i32 %tag, 8 %tagbit4eq0 = icmp eq i32 %tagbit4, 0 br i1 %tagbit4eq0, label %exit, label %optional4 optional4: ; H call void @d() br label %exit exit: ret void } here is the layout after D27742: straight_test: # @straight_test ; ... Prologue elided ; BB#0: # %entry ; A (merged with test1) ; ... More prologue elided mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_2 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_3 b .LBB0_4 .LBB0_2: # %optional1 ; B (copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_4 .LBB0_3: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_5 b .LBB0_6 .LBB0_4: # %optional2 ; D (copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_5: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 b .LBB0_7 .LBB0_6: # %optional3 ; F (copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit ; Ret ld 30, 96(1) # 8-byte Folded Reload addi 1, 1, 112 ld 0, 16(1) mtlr 0 blr The tail-duplication has produced some benefit, but it has also produced a trellis which is not laid out optimally. With this patch, we improve the layouts of such trellises, and decrease the cost calculation for tail-duplication accordingly. This patch produces the layout A,C,E,G,B,D,F,H,Ret. This layout does have back edges, which is a negative, but it has a bigger compensating positive, which is that it handles the case where there are long strings of skipped blocks much better than the original layout. Both layouts handle runs of executed blocks equally well. Branch prediction also improves if there is any correlation between subsequent optional blocks. Here is the resulting concrete layout: straight_test: # @straight_test ; BB#0: # %entry ; A (merged with test1) mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_4 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_5 .LBB0_2: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_3: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 bne 0, .LBB0_7 b .LBB0_8 .LBB0_4: # %optional1 ; B (Copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_2 .LBB0_5: # %optional2 ; D (Copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_3 .LBB0_6: # %optional3 ; F (Copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit Differential Revision: https://reviews.llvm.org/D28522 llvm-svn: 295223
2017-02-16 03:49:14 +08:00
br i1 %tagbit3eq0, label %test4, label %optional3, !prof !1
Codegen: Tail-duplicate during placement. The tail duplication pass uses an assumed layout when making duplication decisions. This is fine, but passes up duplication opportunities that may arise when blocks are outlined. Because we want the updated CFG to affect subsequent placement decisions, this change must occur during placement. In order to achieve this goal, TailDuplicationPass is split into a utility class, TailDuplicator, and the pass itself. The pass delegates nearly everything to the TailDuplicator object, except for looping over the blocks in a function. This allows the same code to be used for tail duplication in both places. This change, in concert with outlining optional branches, allows triangle shaped code to perform much better, esepecially when the taken/untaken branches are correlated, as it creates a second spine when the tests are small enough. Issue from previous rollback fixed, and a new test was added for that case as well. Issue was worklist/scheduling/taildup issue in layout. Issue from 2nd rollback fixed, with 2 additional tests. Issue was tail merging/loop info/tail-duplication causing issue with loops that share a header block. Issue with early tail-duplication of blocks that branch to a fallthrough predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll Differential revision: https://reviews.llvm.org/D18226 llvm-svn: 283934
2016-10-12 04:36:43 +08:00
optional3:
call void @c()
call void @c()
call void @c()
call void @c()
br label %test4
test4:
%tagbit4 = and i32 %tag, 8
%tagbit4eq0 = icmp eq i32 %tagbit4, 0
Codegen: Make chains from trellis-shaped CFGs Lay out trellis-shaped CFGs optimally. A trellis of the shape below: A B |\ /| | \ / | | X | | / \ | |/ \| C D would be laid out A; B->C ; D by the current layout algorithm. Now we identify trellises and lay them out either A->C; B->D or A->D; B->C. This scales with an increasing number of predecessors. A trellis is a a group of 2 or more predecessor blocks that all have the same successors. because of this we can tail duplicate to extend existing trellises. As an example consider the following CFG: B D F H / \ / \ / \ / \ A---C---E---G---Ret Where A,C,E,G are all small (Currently 2 instructions). The CFG preserving layout is then A,B,C,D,E,F,G,H,Ret. The current code will copy C into B, E into D and G into F and yield the layout A,C,B(C),E,D(E),F(G),G,H,ret define void @straight_test(i32 %tag) { entry: br label %test1 test1: ; A %tagbit1 = and i32 %tag, 1 %tagbit1eq0 = icmp eq i32 %tagbit1, 0 br i1 %tagbit1eq0, label %test2, label %optional1 optional1: ; B call void @a() br label %test2 test2: ; C %tagbit2 = and i32 %tag, 2 %tagbit2eq0 = icmp eq i32 %tagbit2, 0 br i1 %tagbit2eq0, label %test3, label %optional2 optional2: ; D call void @b() br label %test3 test3: ; E %tagbit3 = and i32 %tag, 4 %tagbit3eq0 = icmp eq i32 %tagbit3, 0 br i1 %tagbit3eq0, label %test4, label %optional3 optional3: ; F call void @c() br label %test4 test4: ; G %tagbit4 = and i32 %tag, 8 %tagbit4eq0 = icmp eq i32 %tagbit4, 0 br i1 %tagbit4eq0, label %exit, label %optional4 optional4: ; H call void @d() br label %exit exit: ret void } here is the layout after D27742: straight_test: # @straight_test ; ... Prologue elided ; BB#0: # %entry ; A (merged with test1) ; ... More prologue elided mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_2 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_3 b .LBB0_4 .LBB0_2: # %optional1 ; B (copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_4 .LBB0_3: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_5 b .LBB0_6 .LBB0_4: # %optional2 ; D (copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_5: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 b .LBB0_7 .LBB0_6: # %optional3 ; F (copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit ; Ret ld 30, 96(1) # 8-byte Folded Reload addi 1, 1, 112 ld 0, 16(1) mtlr 0 blr The tail-duplication has produced some benefit, but it has also produced a trellis which is not laid out optimally. With this patch, we improve the layouts of such trellises, and decrease the cost calculation for tail-duplication accordingly. This patch produces the layout A,C,E,G,B,D,F,H,Ret. This layout does have back edges, which is a negative, but it has a bigger compensating positive, which is that it handles the case where there are long strings of skipped blocks much better than the original layout. Both layouts handle runs of executed blocks equally well. Branch prediction also improves if there is any correlation between subsequent optional blocks. Here is the resulting concrete layout: straight_test: # @straight_test ; BB#0: # %entry ; A (merged with test1) mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_4 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_5 .LBB0_2: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_3: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 bne 0, .LBB0_7 b .LBB0_8 .LBB0_4: # %optional1 ; B (Copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_2 .LBB0_5: # %optional2 ; D (Copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_3 .LBB0_6: # %optional3 ; F (Copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit Differential Revision: https://reviews.llvm.org/D28522 llvm-svn: 295223
2017-02-16 03:49:14 +08:00
br i1 %tagbit4eq0, label %exit, label %optional4, !prof !1
Codegen: Tail-duplicate during placement. The tail duplication pass uses an assumed layout when making duplication decisions. This is fine, but passes up duplication opportunities that may arise when blocks are outlined. Because we want the updated CFG to affect subsequent placement decisions, this change must occur during placement. In order to achieve this goal, TailDuplicationPass is split into a utility class, TailDuplicator, and the pass itself. The pass delegates nearly everything to the TailDuplicator object, except for looping over the blocks in a function. This allows the same code to be used for tail duplication in both places. This change, in concert with outlining optional branches, allows triangle shaped code to perform much better, esepecially when the taken/untaken branches are correlated, as it creates a second spine when the tests are small enough. Issue from previous rollback fixed, and a new test was added for that case as well. Issue was worklist/scheduling/taildup issue in layout. Issue from 2nd rollback fixed, with 2 additional tests. Issue was tail merging/loop info/tail-duplication causing issue with loops that share a header block. Issue with early tail-duplication of blocks that branch to a fallthrough predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll Differential revision: https://reviews.llvm.org/D18226 llvm-svn: 283934
2016-10-12 04:36:43 +08:00
optional4:
call void @d()
call void @d()
call void @d()
call void @d()
br label %exit
exit:
ret void
}
CodeGen: BlockPlacement: Precompute layout for chains of triangles. For chains of triangles with small join blocks that can be tail duplicated, a simple calculation of probabilities is insufficient. Tail duplication can be profitable in 3 different ways for these cases: 1) The post-dominators marked 50% are actually taken 56% (This shrinks with longer chains) 2) The chains are statically correlated. Branch probabilities have a very U-shaped distribution. [http://nrs.harvard.edu/urn-3:HUL.InstRepos:24015805] If the branches in a chain are likely to be from the same side of the distribution as their predecessor, but are independent at runtime, this transformation is profitable. (Because the cost of being wrong is a small fixed cost, unlike the standard triangle layout where the cost of being wrong scales with the # of triangles.) 3) The chains are dynamically correlated. If the probability that a previous branch was taken positively influences whether the next branch will be taken We believe that 2 and 3 are common enough to justify the small margin in 1. The code pre-scans a function's CFG to identify this pattern and marks the edges so that the standard layout algorithm can use the computed results. llvm-svn: 296845
2017-03-03 09:00:22 +08:00
; Intended layout:
; The chain-of-triangles based duplicating produces the layout
; test1
; test2
; test3
; optional1
; optional2
; optional3
; exit
; even for 50/50 branches.
; Tail duplication puts test n+1 at the end of optional n
; so optional1 includes a copy of test2 at the end, and branches
; to test3 (at the top) or falls through to optional 2.
; The CHECK statements check for the whole string of tests
; and then check that the correct test has been duplicated into the end of
; the optional blocks and that the optional blocks are in the correct order.
;CHECK-LABEL: straight_test_50:
; test1 may have been merged with entry
;CHECK: mr [[TAGREG:[0-9]+]], 3
;CHECK: andi. {{[0-9]+}}, [[TAGREG]], 1
;CHECK-NEXT: bc 12, 1, .[[OPT1LABEL:[_0-9A-Za-z]+]]
;CHECK-NEXT: # %test2
;CHECK-NEXT: rlwinm. {{[0-9]+}}, [[TAGREG]], 0, 30, 30
;CHECK-NEXT: bne 0, .[[OPT2LABEL:[_0-9A-Za-z]+]]
;CHECK-NEXT: .[[TEST3LABEL:[_0-9A-Za-z]+]]: # %test3
;CHECK-NEXT: rlwinm. {{[0-9]+}}, [[TAGREG]], 0, 29, 29
;CHECK-NEXT: bne 0, .[[OPT3LABEL:[_0-9A-Za-z]+]]
;CHECK-NEXT: .[[EXITLABEL:[_0-9A-Za-z]+]]: # %exit
;CHECK: blr
;CHECK-NEXT: .[[OPT1LABEL]]:
;CHECK: rlwinm. {{[0-9]+}}, [[TAGREG]], 0, 30, 30
;CHECK-NEXT: beq 0, .[[TEST3LABEL]]
;CHECK-NEXT: .[[OPT2LABEL]]:
;CHECK: rlwinm. {{[0-9]+}}, [[TAGREG]], 0, 29, 29
;CHECK-NEXT: beq 0, .[[EXITLABEL]]
CodeGen: BlockPlacement: Adjust test case so it covers rL297925. NFC I had ajusted the test case before when testing a chain of length 2, and then reverted it with rL296845 when I switched to 3 triangles. After running benchmarks and examining generated code at length 2 I forgot to put the test back. llvm-svn: 298000
2017-03-17 05:33:29 +08:00
;CHECK-NEXT: .[[OPT3LABEL]]:
CodeGen: BlockPlacement: Precompute layout for chains of triangles. For chains of triangles with small join blocks that can be tail duplicated, a simple calculation of probabilities is insufficient. Tail duplication can be profitable in 3 different ways for these cases: 1) The post-dominators marked 50% are actually taken 56% (This shrinks with longer chains) 2) The chains are statically correlated. Branch probabilities have a very U-shaped distribution. [http://nrs.harvard.edu/urn-3:HUL.InstRepos:24015805] If the branches in a chain are likely to be from the same side of the distribution as their predecessor, but are independent at runtime, this transformation is profitable. (Because the cost of being wrong is a small fixed cost, unlike the standard triangle layout where the cost of being wrong scales with the # of triangles.) 3) The chains are dynamically correlated. If the probability that a previous branch was taken positively influences whether the next branch will be taken We believe that 2 and 3 are common enough to justify the small margin in 1. The code pre-scans a function's CFG to identify this pattern and marks the edges so that the standard layout algorithm can use the computed results. llvm-svn: 296845
2017-03-03 09:00:22 +08:00
;CHECK: b .[[EXITLABEL]]
define void @straight_test_50(i32 %tag) {
entry:
br label %test1
test1:
%tagbit1 = and i32 %tag, 1
%tagbit1eq0 = icmp eq i32 %tagbit1, 0
br i1 %tagbit1eq0, label %test2, label %optional1, !prof !2
optional1:
call void @a()
br label %test2
test2:
%tagbit2 = and i32 %tag, 2
%tagbit2eq0 = icmp eq i32 %tagbit2, 0
br i1 %tagbit2eq0, label %test3, label %optional2, !prof !2
optional2:
call void @b()
br label %test3
test3:
%tagbit3 = and i32 %tag, 4
%tagbit3eq0 = icmp eq i32 %tagbit3, 0
CodeGen: BlockPlacement: Adjust test case so it covers rL297925. NFC I had ajusted the test case before when testing a chain of length 2, and then reverted it with rL296845 when I switched to 3 triangles. After running benchmarks and examining generated code at length 2 I forgot to put the test back. llvm-svn: 298000
2017-03-17 05:33:29 +08:00
br i1 %tagbit3eq0, label %exit, label %optional3, !prof !1
CodeGen: BlockPlacement: Precompute layout for chains of triangles. For chains of triangles with small join blocks that can be tail duplicated, a simple calculation of probabilities is insufficient. Tail duplication can be profitable in 3 different ways for these cases: 1) The post-dominators marked 50% are actually taken 56% (This shrinks with longer chains) 2) The chains are statically correlated. Branch probabilities have a very U-shaped distribution. [http://nrs.harvard.edu/urn-3:HUL.InstRepos:24015805] If the branches in a chain are likely to be from the same side of the distribution as their predecessor, but are independent at runtime, this transformation is profitable. (Because the cost of being wrong is a small fixed cost, unlike the standard triangle layout where the cost of being wrong scales with the # of triangles.) 3) The chains are dynamically correlated. If the probability that a previous branch was taken positively influences whether the next branch will be taken We believe that 2 and 3 are common enough to justify the small margin in 1. The code pre-scans a function's CFG to identify this pattern and marks the edges so that the standard layout algorithm can use the computed results. llvm-svn: 296845
2017-03-03 09:00:22 +08:00
optional3:
call void @c()
br label %exit
exit:
ret void
}
CodeGen: BlockPlacement: Increase tail duplication size for O3. At O3 we are more willing to increase size if we believe it will improve performance. The current threshold for tail-duplication of 2 instructions is conservative, and can be relaxed at O3. Benchmark results: llvm test-suite: 6% improvement in aha, due to duplication of loop latch 3% improvement in hexxagon 2% slowdown in lpbench. Seems related, but couldn't completely diagnose. Internal google benchmark: Produces 4% improvement on internal google protocol buffer serialization benchmarks. Differential-Revision: https://reviews.llvm.org/D32324 llvm-svn: 303084
2017-05-16 01:30:47 +08:00
; Intended layout:
; The chain-of-triangles based duplicating produces the layout when 3
; instructions are allowed for tail-duplication.
; test1
; test2
; test3
; optional1
; optional2
; optional3
; exit
;
; Otherwise it produces the layout:
; test1
; optional1
; test2
; optional2
; test3
; optional3
; exit
;CHECK-LABEL: straight_test_3_instr_test:
; test1 may have been merged with entry
;CHECK: mr [[TAGREG:[0-9]+]], 3
;CHECK: clrlwi {{[0-9]+}}, [[TAGREG]], 30
;CHECK-NEXT: cmplwi {{[0-9]+}}, 2
;CHECK-O3-NEXT: bne 0, .[[OPT1LABEL:[_0-9A-Za-z]+]]
;CHECK-O3-NEXT: # %test2
;CHECK-O3-NEXT: rlwinm {{[0-9]+}}, [[TAGREG]], 0, 28, 29
;CHECK-O3-NEXT: cmplwi {{[0-9]+}}, 8
;CHECK-O3-NEXT: bne 0, .[[OPT2LABEL:[_0-9A-Za-z]+]]
;CHECK-O3-NEXT: .[[TEST3LABEL:[_0-9A-Za-z]+]]: # %test3
;CHECK-O3-NEXT: rlwinm {{[0-9]+}}, [[TAGREG]], 0, 26, 27
;CHECK-O3-NEXT: cmplwi {{[0-9]+}}, 32
;CHECK-O3-NEXT: bne 0, .[[OPT3LABEL:[_0-9A-Za-z]+]]
;CHECK-O3-NEXT: .[[EXITLABEL:[_0-9A-Za-z]+]]: # %exit
;CHECK-O3: blr
;CHECK-O3-NEXT: .[[OPT1LABEL]]:
;CHECK-O3: rlwinm {{[0-9]+}}, [[TAGREG]], 0, 28, 29
;CHECK-O3-NEXT: cmplwi {{[0-9]+}}, 8
;CHECK-O3-NEXT: beq 0, .[[TEST3LABEL]]
;CHECK-O3-NEXT: .[[OPT2LABEL]]:
;CHECK-O3: rlwinm {{[0-9]+}}, [[TAGREG]], 0, 26, 27
;CHECK-O3-NEXT: cmplwi {{[0-9]+}}, 32
;CHECK-O3-NEXT: beq 0, .[[EXITLABEL]]
;CHECK-O3-NEXT: .[[OPT3LABEL]]:
;CHECK-O3: b .[[EXITLABEL]]
;CHECK-O2-NEXT: beq 0, .[[TEST2LABEL:[_0-9A-Za-z]+]]
;CHECK-O2-NEXT: # %optional1
;CHECK-O2: .[[TEST2LABEL]]: # %test2
;CHECK-O2-NEXT: rlwinm {{[0-9]+}}, [[TAGREG]], 0, 28, 29
;CHECK-O2-NEXT: cmplwi {{[0-9]+}}, 8
;CHECK-O2-NEXT: beq 0, .[[TEST3LABEL:[_0-9A-Za-z]+]]
;CHECK-O2-NEXT: # %optional2
;CHECK-O2: .[[TEST3LABEL]]: # %test3
;CHECK-O2-NEXT: rlwinm {{[0-9]+}}, [[TAGREG]], 0, 26, 27
;CHECK-O2-NEXT: cmplwi {{[0-9]+}}, 32
;CHECK-O2-NEXT: beq 0, .[[EXITLABEL:[_0-9A-Za-z]+]]
;CHECK-O2-NEXT: # %optional3
;CHECK-O2: .[[EXITLABEL:[_0-9A-Za-z]+]]: # %exit
;CHECK-O2: blr
define void @straight_test_3_instr_test(i32 %tag) {
entry:
br label %test1
test1:
%tagbit1 = and i32 %tag, 3
%tagbit1eq0 = icmp eq i32 %tagbit1, 2
br i1 %tagbit1eq0, label %test2, label %optional1, !prof !2
optional1:
call void @a()
br label %test2
test2:
%tagbit2 = and i32 %tag, 12
%tagbit2eq0 = icmp eq i32 %tagbit2, 8
br i1 %tagbit2eq0, label %test3, label %optional2, !prof !2
optional2:
call void @b()
br label %test3
test3:
%tagbit3 = and i32 %tag, 48
%tagbit3eq0 = icmp eq i32 %tagbit3, 32
br i1 %tagbit3eq0, label %exit, label %optional3, !prof !1
optional3:
call void @c()
br label %exit
exit:
ret void
}
Codegen: Make chains from trellis-shaped CFGs Lay out trellis-shaped CFGs optimally. A trellis of the shape below: A B |\ /| | \ / | | X | | / \ | |/ \| C D would be laid out A; B->C ; D by the current layout algorithm. Now we identify trellises and lay them out either A->C; B->D or A->D; B->C. This scales with an increasing number of predecessors. A trellis is a a group of 2 or more predecessor blocks that all have the same successors. because of this we can tail duplicate to extend existing trellises. As an example consider the following CFG: B D F H / \ / \ / \ / \ A---C---E---G---Ret Where A,C,E,G are all small (Currently 2 instructions). The CFG preserving layout is then A,B,C,D,E,F,G,H,Ret. The current code will copy C into B, E into D and G into F and yield the layout A,C,B(C),E,D(E),F(G),G,H,ret define void @straight_test(i32 %tag) { entry: br label %test1 test1: ; A %tagbit1 = and i32 %tag, 1 %tagbit1eq0 = icmp eq i32 %tagbit1, 0 br i1 %tagbit1eq0, label %test2, label %optional1 optional1: ; B call void @a() br label %test2 test2: ; C %tagbit2 = and i32 %tag, 2 %tagbit2eq0 = icmp eq i32 %tagbit2, 0 br i1 %tagbit2eq0, label %test3, label %optional2 optional2: ; D call void @b() br label %test3 test3: ; E %tagbit3 = and i32 %tag, 4 %tagbit3eq0 = icmp eq i32 %tagbit3, 0 br i1 %tagbit3eq0, label %test4, label %optional3 optional3: ; F call void @c() br label %test4 test4: ; G %tagbit4 = and i32 %tag, 8 %tagbit4eq0 = icmp eq i32 %tagbit4, 0 br i1 %tagbit4eq0, label %exit, label %optional4 optional4: ; H call void @d() br label %exit exit: ret void } here is the layout after D27742: straight_test: # @straight_test ; ... Prologue elided ; BB#0: # %entry ; A (merged with test1) ; ... More prologue elided mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_2 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_3 b .LBB0_4 .LBB0_2: # %optional1 ; B (copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_4 .LBB0_3: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_5 b .LBB0_6 .LBB0_4: # %optional2 ; D (copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_5: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 b .LBB0_7 .LBB0_6: # %optional3 ; F (copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit ; Ret ld 30, 96(1) # 8-byte Folded Reload addi 1, 1, 112 ld 0, 16(1) mtlr 0 blr The tail-duplication has produced some benefit, but it has also produced a trellis which is not laid out optimally. With this patch, we improve the layouts of such trellises, and decrease the cost calculation for tail-duplication accordingly. This patch produces the layout A,C,E,G,B,D,F,H,Ret. This layout does have back edges, which is a negative, but it has a bigger compensating positive, which is that it handles the case where there are long strings of skipped blocks much better than the original layout. Both layouts handle runs of executed blocks equally well. Branch prediction also improves if there is any correlation between subsequent optional blocks. Here is the resulting concrete layout: straight_test: # @straight_test ; BB#0: # %entry ; A (merged with test1) mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_4 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_5 .LBB0_2: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_3: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 bne 0, .LBB0_7 b .LBB0_8 .LBB0_4: # %optional1 ; B (Copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_2 .LBB0_5: # %optional2 ; D (Copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_3 .LBB0_6: # %optional3 ; F (Copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit Differential Revision: https://reviews.llvm.org/D28522 llvm-svn: 295223
2017-02-16 03:49:14 +08:00
; Intended layout:
; The chain-based outlining produces the layout
; entry
; --- Begin loop ---
; for.latch
; for.check
; test1
; test2
; test3
; test4
; optional1
; optional2
; optional3
; optional4
; --- End loop ---
; exit
; The CHECK statements check for the whole string of tests and exit block,
; and then check that the correct test has been duplicated into the end of
; the optional blocks and that the optional blocks are in the correct order.
;CHECK-LABEL: loop_test:
;CHECK: add [[TAGPTRREG:[0-9]+]], 3, 4
;CHECK: .[[LATCHLABEL:[._0-9A-Za-z]+]]: # %for.latch
;CHECK: addi
;CHECK: .[[CHECKLABEL:[._0-9A-Za-z]+]]: # %for.check
;CHECK: lwz [[TAGREG:[0-9]+]], 0([[TAGPTRREG]])
;CHECK: # %test1
;CHECK: andi. {{[0-9]+}}, [[TAGREG]], 1
;CHECK-NEXT: bc 12, 1, .[[OPT1LABEL:[._0-9A-Za-z]+]]
;CHECK-NEXT: # %test2
;CHECK: rlwinm. {{[0-9]+}}, [[TAGREG]], 0, 30, 30
;CHECK-NEXT: bne 0, .[[OPT2LABEL:[._0-9A-Za-z]+]]
;CHECK-NEXT: .[[TEST3LABEL:[._0-9A-Za-z]+]]: # %test3
;CHECK: rlwinm. {{[0-9]+}}, [[TAGREG]], 0, 29, 29
;CHECK-NEXT: bne 0, .[[OPT3LABEL:[._0-9A-Za-z]+]]
;CHECK-NEXT: .[[TEST4LABEL:[._0-9A-Za-z]+]]: # %{{(test4|optional3)}}
;CHECK: rlwinm. {{[0-9]+}}, [[TAGREG]], 0, 28, 28
;CHECK-NEXT: beq 0, .[[LATCHLABEL]]
;CHECK-NEXT: b .[[OPT4LABEL:[._0-9A-Za-z]+]]
;CHECK: [[OPT1LABEL]]
;CHECK: rlwinm. {{[0-9]+}}, [[TAGREG]], 0, 30, 30
;CHECK-NEXT: beq 0, .[[TEST3LABEL]]
;CHECK-NEXT: .[[OPT2LABEL]]
;CHECK: rlwinm. {{[0-9]+}}, [[TAGREG]], 0, 29, 29
;CHECK-NEXT: beq 0, .[[TEST4LABEL]]
;CHECK-NEXT: .[[OPT3LABEL]]
;CHECK: rlwinm. {{[0-9]+}}, [[TAGREG]], 0, 28, 28
;CHECK-NEXT: beq 0, .[[LATCHLABEL]]
;CHECK: [[OPT4LABEL]]:
;CHECK: b .[[LATCHLABEL]]
define void @loop_test(i32* %tags, i32 %count) {
entry:
br label %for.check
for.check:
%count.loop = phi i32 [%count, %entry], [%count.sub, %for.latch]
%done.count = icmp ugt i32 %count.loop, 0
%tag_ptr = getelementptr inbounds i32, i32* %tags, i32 %count
%tag = load i32, i32* %tag_ptr
%done.tag = icmp eq i32 %tag, 0
%done = and i1 %done.count, %done.tag
br i1 %done, label %test1, label %exit, !prof !1
test1:
%tagbit1 = and i32 %tag, 1
%tagbit1eq0 = icmp eq i32 %tagbit1, 0
br i1 %tagbit1eq0, label %test2, label %optional1, !prof !1
optional1:
call void @a()
call void @a()
call void @a()
call void @a()
br label %test2
test2:
%tagbit2 = and i32 %tag, 2
%tagbit2eq0 = icmp eq i32 %tagbit2, 0
br i1 %tagbit2eq0, label %test3, label %optional2, !prof !1
optional2:
call void @b()
call void @b()
call void @b()
call void @b()
br label %test3
test3:
%tagbit3 = and i32 %tag, 4
%tagbit3eq0 = icmp eq i32 %tagbit3, 0
br i1 %tagbit3eq0, label %test4, label %optional3, !prof !1
optional3:
call void @c()
call void @c()
call void @c()
call void @c()
br label %test4
test4:
%tagbit4 = and i32 %tag, 8
%tagbit4eq0 = icmp eq i32 %tagbit4, 0
br i1 %tagbit4eq0, label %for.latch, label %optional4, !prof !1
optional4:
call void @d()
call void @d()
call void @d()
call void @d()
br label %for.latch
for.latch:
%count.sub = sub i32 %count.loop, 1
br label %for.check
exit:
ret void
}
; The block then2 is not unavoidable, meaning it does not dominate the exit.
; But since it can be tail-duplicated, it should be placed as a fallthrough from
; test2 and copied. The purpose here is to make sure that the tail-duplication
; code is independent of the outlining code, which works by choosing the
; "unavoidable" blocks.
; CHECK-LABEL: avoidable_test:
; CHECK: # %entry
; CHECK: andi.
; CHECK: # %test2
; Make sure then2 falls through from test2
; CHECK-NOT: # %{{[-_a-zA-Z0-9]+}}
; CHECK: # %then2
; CHECK: rlwinm. {{[0-9]+}}, {{[0-9]+}}, 0, 29, 29
; CHECK: # %else1
; CHECK: bl a
; CHECK: bl a
; Make sure then2 was copied into else1
; CHECK: rlwinm. {{[0-9]+}}, {{[0-9]+}}, 0, 29, 29
; CHECK: # %end1
; CHECK: bl d
; CHECK: # %else2
; CHECK: bl c
; CHECK: # %end2
define void @avoidable_test(i32 %tag) {
entry:
br label %test1
test1:
%tagbit1 = and i32 %tag, 1
%tagbit1eq0 = icmp eq i32 %tagbit1, 0
br i1 %tagbit1eq0, label %test2, label %else1, !prof !1 ; %test2 more likely
else1:
call void @a()
call void @a()
br label %then2
test2:
%tagbit2 = and i32 %tag, 2
%tagbit2eq0 = icmp eq i32 %tagbit2, 0
br i1 %tagbit2eq0, label %then2, label %else2, !prof !1 ; %then2 more likely
then2:
%tagbit3 = and i32 %tag, 4
%tagbit3eq0 = icmp eq i32 %tagbit3, 0
br i1 %tagbit3eq0, label %end2, label %end1, !prof !1 ; %end2 more likely
else2:
call void @c()
br label %end2
end2:
ret void
end1:
call void @d()
ret void
}
; CHECK-LABEL: trellis_test
; The number in the block labels is the expected block frequency given the
; probabilities annotated. There is a conflict in the b;c->d;e trellis that
; should be resolved as c->e;b->d.
; The d;e->f;g trellis should be resolved as e->g;d->f.
; The f;g->h;i trellis should be resolved as f->i;g->h.
; The h;i->j;ret trellis contains a triangle edge, and should be resolved as
; h->j->ret
; CHECK: # %entry
; CHECK: # %c10
; CHECK: # %e9
; CHECK: # %g10
; CHECK: # %h10
; CHECK: # %j8
; CHECK: # %ret
; CHECK: # %b6
; CHECK: # %d7
; CHECK: # %f6
; CHECK: # %i6
define void @trellis_test(i32 %tag) {
entry:
br label %a16
a16:
call void @a()
call void @a()
%tagbits.a = and i32 %tag, 3
%tagbits.a.eq0 = icmp eq i32 %tagbits.a, 0
br i1 %tagbits.a.eq0, label %c10, label %b6, !prof !1 ; 10 to 6
c10:
call void @c()
call void @c()
%tagbits.c = and i32 %tag, 12
%tagbits.c.eq0 = icmp eq i32 %tagbits.c, 0
; Both of these edges should be hotter than the other incoming edge
; for e9 or d7
br i1 %tagbits.c.eq0, label %e9, label %d7, !prof !3 ; 6 to 4
e9:
call void @e()
call void @e()
%tagbits.e = and i32 %tag, 48
%tagbits.e.eq0 = icmp eq i32 %tagbits.e, 0
br i1 %tagbits.e.eq0, label %g10, label %f6, !prof !4 ; 7 to 2
g10:
call void @g()
call void @g()
%tagbits.g = and i32 %tag, 192
%tagbits.g.eq0 = icmp eq i32 %tagbits.g, 0
br i1 %tagbits.g.eq0, label %i6, label %h10, !prof !5 ; 2 to 8
i6:
call void @i()
call void @i()
%tagbits.i = and i32 %tag, 768
%tagbits.i.eq0 = icmp eq i32 %tagbits.i, 0
br i1 %tagbits.i.eq0, label %ret, label %j8, !prof !2 ; balanced (3 to 3)
b6:
call void @b()
call void @b()
%tagbits.b = and i32 %tag, 12
%tagbits.b.eq1 = icmp eq i32 %tagbits.b, 8
br i1 %tagbits.b.eq1, label %e9, label %d7, !prof !2 ; balanced (3 to 3)
d7:
call void @d()
call void @d()
%tagbits.d = and i32 %tag, 48
%tagbits.d.eq1 = icmp eq i32 %tagbits.d, 32
br i1 %tagbits.d.eq1, label %g10, label %f6, !prof !6 ; 3 to 4
f6:
call void @f()
call void @f()
%tagbits.f = and i32 %tag, 192
%tagbits.f.eq1 = icmp eq i32 %tagbits.f, 128
br i1 %tagbits.f.eq1, label %i6, label %h10, !prof !7 ; 4 to 2
h10:
call void @h()
call void @h()
%tagbits.h = and i32 %tag, 768
%tagbits.h.eq1 = icmp eq i32 %tagbits.h, 512
br i1 %tagbits.h.eq1, label %ret, label %j8, !prof !2 ; balanced (5 to 5)
j8:
call void @j()
call void @j()
br label %ret
ret:
ret void
}
; Verify that we still consider tail-duplication opportunities if we find a
; triangle trellis. Here D->F->G is the triangle, and D;E are both predecessors
; of both F and G. The basic trellis algorithm picks the F->G edge, but after
; checking, it's profitable to duplicate G into F. The weights here are not
; really important. They are there to help make the test stable.
; CHECK-LABEL: trellis_then_dup_test
; CHECK: # %entry
; CHECK: # %b
; CHECK: # %d
; CHECK: # %g
; CHECK: # %ret1
; CHECK: # %c
; CHECK: # %e
; CHECK: # %f
; CHECK: # %ret2
; CHECK: # %ret
define void @trellis_then_dup_test(i32 %tag) {
entry:
br label %a
a:
call void @a()
call void @a()
%tagbits.a = and i32 %tag, 3
%tagbits.a.eq0 = icmp eq i32 %tagbits.a, 0
br i1 %tagbits.a.eq0, label %b, label %c, !prof !1 ; 5 to 3
b:
call void @b()
call void @b()
%tagbits.b = and i32 %tag, 12
%tagbits.b.eq1 = icmp eq i32 %tagbits.b, 8
br i1 %tagbits.b.eq1, label %d, label %e, !prof !1 ; 5 to 3
d:
call void @d()
call void @d()
%tagbits.d = and i32 %tag, 48
%tagbits.d.eq1 = icmp eq i32 %tagbits.d, 32
br i1 %tagbits.d.eq1, label %g, label %f, !prof !1 ; 5 to 3
f:
call void @f()
call void @f()
br label %g
g:
%tagbits.g = and i32 %tag, 192
%tagbits.g.eq0 = icmp eq i32 %tagbits.g, 0
br i1 %tagbits.g.eq0, label %ret1, label %ret2, !prof !2 ; balanced
c:
call void @c()
call void @c()
%tagbits.c = and i32 %tag, 12
%tagbits.c.eq0 = icmp eq i32 %tagbits.c, 0
br i1 %tagbits.c.eq0, label %d, label %e, !prof !1 ; 5 to 3
e:
call void @e()
call void @e()
%tagbits.e = and i32 %tag, 48
%tagbits.e.eq0 = icmp eq i32 %tagbits.e, 0
br i1 %tagbits.e.eq0, label %g, label %f, !prof !1 ; 5 to 3
ret1:
call void @a()
br label %ret
ret2:
call void @b()
br label %ret
ret:
ret void
}
Fix trellis layout to avoid mis-identify triangle. Summary: For the following CFG: A->B B->C A->C If there is another edge B->D, then ABC should not be considered as triangle. Reviewers: davidxl, iteratee Reviewed By: iteratee Subscribers: nemanjai, llvm-commits Differential Revision: https://reviews.llvm.org/D31310 llvm-svn: 298661
2017-03-24 07:28:09 +08:00
; Verify that we did not mis-identify triangle trellises if it is not
; really a triangle.
; CHECK-LABEL: trellis_no_triangle
; CHECK: # %entry
; CHECK: # %b
; CHECK: # %d
; CHECK: # %ret
; CHECK: # %c
; CHECK: # %e
define void @trellis_no_triangle(i32 %tag) {
entry:
br label %a
a:
call void @a()
call void @a()
%tagbits.a = and i32 %tag, 3
%tagbits.a.eq0 = icmp eq i32 %tagbits.a, 0
br i1 %tagbits.a.eq0, label %b, label %c, !prof !8 ; 98 to 2
b:
call void @b()
call void @b()
%tagbits.b = and i32 %tag, 12
%tagbits.b.eq1 = icmp eq i32 %tagbits.b, 8
br i1 %tagbits.b.eq1, label %d, label %e, !prof !9 ; 97 to 1
d:
call void @d()
call void @d()
%tagbits.d = and i32 %tag, 48
%tagbits.d.eq1 = icmp eq i32 %tagbits.d, 32
br i1 %tagbits.d.eq1, label %ret, label %e, !prof !10 ; 96 to 2
c:
call void @c()
call void @c()
%tagbits.c = and i32 %tag, 12
%tagbits.c.eq0 = icmp eq i32 %tagbits.c, 0
br i1 %tagbits.c.eq0, label %d, label %e, !prof !2 ; 1 to 1
e:
call void @e()
call void @e()
br label %ret
ret:
call void @f()
ret void
}
Codegen: Tail-duplicate during placement. The tail duplication pass uses an assumed layout when making duplication decisions. This is fine, but passes up duplication opportunities that may arise when blocks are outlined. Because we want the updated CFG to affect subsequent placement decisions, this change must occur during placement. In order to achieve this goal, TailDuplicationPass is split into a utility class, TailDuplicator, and the pass itself. The pass delegates nearly everything to the TailDuplicator object, except for looping over the blocks in a function. This allows the same code to be used for tail duplication in both places. This change, in concert with outlining optional branches, allows triangle shaped code to perform much better, esepecially when the taken/untaken branches are correlated, as it creates a second spine when the tests are small enough. Issue from previous rollback fixed, and a new test was added for that case as well. Issue was worklist/scheduling/taildup issue in layout. Issue from 2nd rollback fixed, with 2 additional tests. Issue was tail merging/loop info/tail-duplication causing issue with loops that share a header block. Issue with early tail-duplication of blocks that branch to a fallthrough predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll Differential revision: https://reviews.llvm.org/D18226 llvm-svn: 283934
2016-10-12 04:36:43 +08:00
declare void @a()
declare void @b()
declare void @c()
declare void @d()
Codegen: Make chains from trellis-shaped CFGs Lay out trellis-shaped CFGs optimally. A trellis of the shape below: A B |\ /| | \ / | | X | | / \ | |/ \| C D would be laid out A; B->C ; D by the current layout algorithm. Now we identify trellises and lay them out either A->C; B->D or A->D; B->C. This scales with an increasing number of predecessors. A trellis is a a group of 2 or more predecessor blocks that all have the same successors. because of this we can tail duplicate to extend existing trellises. As an example consider the following CFG: B D F H / \ / \ / \ / \ A---C---E---G---Ret Where A,C,E,G are all small (Currently 2 instructions). The CFG preserving layout is then A,B,C,D,E,F,G,H,Ret. The current code will copy C into B, E into D and G into F and yield the layout A,C,B(C),E,D(E),F(G),G,H,ret define void @straight_test(i32 %tag) { entry: br label %test1 test1: ; A %tagbit1 = and i32 %tag, 1 %tagbit1eq0 = icmp eq i32 %tagbit1, 0 br i1 %tagbit1eq0, label %test2, label %optional1 optional1: ; B call void @a() br label %test2 test2: ; C %tagbit2 = and i32 %tag, 2 %tagbit2eq0 = icmp eq i32 %tagbit2, 0 br i1 %tagbit2eq0, label %test3, label %optional2 optional2: ; D call void @b() br label %test3 test3: ; E %tagbit3 = and i32 %tag, 4 %tagbit3eq0 = icmp eq i32 %tagbit3, 0 br i1 %tagbit3eq0, label %test4, label %optional3 optional3: ; F call void @c() br label %test4 test4: ; G %tagbit4 = and i32 %tag, 8 %tagbit4eq0 = icmp eq i32 %tagbit4, 0 br i1 %tagbit4eq0, label %exit, label %optional4 optional4: ; H call void @d() br label %exit exit: ret void } here is the layout after D27742: straight_test: # @straight_test ; ... Prologue elided ; BB#0: # %entry ; A (merged with test1) ; ... More prologue elided mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_2 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_3 b .LBB0_4 .LBB0_2: # %optional1 ; B (copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_4 .LBB0_3: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_5 b .LBB0_6 .LBB0_4: # %optional2 ; D (copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_5: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 b .LBB0_7 .LBB0_6: # %optional3 ; F (copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit ; Ret ld 30, 96(1) # 8-byte Folded Reload addi 1, 1, 112 ld 0, 16(1) mtlr 0 blr The tail-duplication has produced some benefit, but it has also produced a trellis which is not laid out optimally. With this patch, we improve the layouts of such trellises, and decrease the cost calculation for tail-duplication accordingly. This patch produces the layout A,C,E,G,B,D,F,H,Ret. This layout does have back edges, which is a negative, but it has a bigger compensating positive, which is that it handles the case where there are long strings of skipped blocks much better than the original layout. Both layouts handle runs of executed blocks equally well. Branch prediction also improves if there is any correlation between subsequent optional blocks. Here is the resulting concrete layout: straight_test: # @straight_test ; BB#0: # %entry ; A (merged with test1) mr 30, 3 andi. 3, 30, 1 bc 12, 1, .LBB0_4 ; BB#1: # %test2 ; C rlwinm. 3, 30, 0, 30, 30 bne 0, .LBB0_5 .LBB0_2: # %test3 ; E rlwinm. 3, 30, 0, 29, 29 bne 0, .LBB0_6 .LBB0_3: # %test4 ; G rlwinm. 3, 30, 0, 28, 28 bne 0, .LBB0_7 b .LBB0_8 .LBB0_4: # %optional1 ; B (Copy of C) bl a nop rlwinm. 3, 30, 0, 30, 30 beq 0, .LBB0_2 .LBB0_5: # %optional2 ; D (Copy of E) bl b nop rlwinm. 3, 30, 0, 29, 29 beq 0, .LBB0_3 .LBB0_6: # %optional3 ; F (Copy of G) bl c nop rlwinm. 3, 30, 0, 28, 28 beq 0, .LBB0_8 .LBB0_7: # %optional4 ; H bl d nop .LBB0_8: # %exit Differential Revision: https://reviews.llvm.org/D28522 llvm-svn: 295223
2017-02-16 03:49:14 +08:00
declare void @e()
declare void @f()
declare void @g()
declare void @h()
declare void @i()
declare void @j()
!1 = !{!"branch_weights", i32 5, i32 3}
!2 = !{!"branch_weights", i32 50, i32 50}
!3 = !{!"branch_weights", i32 6, i32 4}
!4 = !{!"branch_weights", i32 7, i32 2}
!5 = !{!"branch_weights", i32 2, i32 8}
!6 = !{!"branch_weights", i32 3, i32 4}
!7 = !{!"branch_weights", i32 4, i32 2}
Fix trellis layout to avoid mis-identify triangle. Summary: For the following CFG: A->B B->C A->C If there is another edge B->D, then ABC should not be considered as triangle. Reviewers: davidxl, iteratee Reviewed By: iteratee Subscribers: nemanjai, llvm-commits Differential Revision: https://reviews.llvm.org/D31310 llvm-svn: 298661
2017-03-24 07:28:09 +08:00
!8 = !{!"branch_weights", i32 98, i32 2}
!9 = !{!"branch_weights", i32 97, i32 1}
!10 = !{!"branch_weights", i32 96, i32 2}