Propagation of profile samples through the CFG.
This adds a propagation heuristic to convert instruction samples
into branch weights. It implements a similar heuristic to the one
implemented by Dehao Chen on GCC.
The propagation proceeds in 3 phases:
1- Assignment of block weights. All the basic blocks in the function
are initial assigned the same weight as their most frequently
executed instruction.
2- Creation of equivalence classes. Since samples may be missing from
blocks, we can fill in the gaps by setting the weights of all the
blocks in the same equivalence class to the same weight. To compute
the concept of equivalence, we use dominance and loop information.
Two blocks B1 and B2 are in the same equivalence class if B1
dominates B2, B2 post-dominates B1 and both are in the same loop.
3- Propagation of block weights into edges. This uses a simple
propagation heuristic. The following rules are applied to every
block B in the CFG:
- If B has a single predecessor/successor, then the weight
of that edge is the weight of the block.
- If all the edges are known except one, and the weight of the
block is already known, the weight of the unknown edge will
be the weight of the block minus the sum of all the known
edges. If the sum of all the known edges is larger than B's weight,
we set the unknown edge weight to zero.
- If there is a self-referential edge, and the weight of the block is
known, the weight for that edge is set to the weight of the block
minus the weight of the other incoming edges to that block (if
known).
Since this propagation is not guaranteed to finalize for every CFG, we
only allow it to proceed for a limited number of iterations (controlled
by -sample-profile-max-propagate-iterations). It currently uses the same
GCC default of 100.
Before propagation starts, the pass builds (for each block) a list of
unique predecessors and successors. This is necessary to handle
identical edges in multiway branches. Since we visit all blocks and all
edges of the CFG, it is cleaner to build these lists once at the start
of the pass.
Finally, the patch fixes the computation of relative line locations.
The profiler emits lines relative to the function header. To discover
it, we traverse the compilation unit looking for the subprogram
corresponding to the function. The line number of that subprogram is the
line where the function begins. That becomes line zero for all the
relative locations.
llvm-svn: 198972
2014-01-11 07:23:46 +08:00
; RUN: opt < %s -sample-profile -sample-profile-file=%S/Inputs/propagate.prof | opt -analyze -branch-prob | FileCheck %s
2016-05-28 07:20:16 +08:00
; RUN: opt < %s -passes=sample-profile -sample-profile-file=%S/Inputs/propagate.prof | opt -analyze -branch-prob | FileCheck %s
Propagation of profile samples through the CFG.
This adds a propagation heuristic to convert instruction samples
into branch weights. It implements a similar heuristic to the one
implemented by Dehao Chen on GCC.
The propagation proceeds in 3 phases:
1- Assignment of block weights. All the basic blocks in the function
are initial assigned the same weight as their most frequently
executed instruction.
2- Creation of equivalence classes. Since samples may be missing from
blocks, we can fill in the gaps by setting the weights of all the
blocks in the same equivalence class to the same weight. To compute
the concept of equivalence, we use dominance and loop information.
Two blocks B1 and B2 are in the same equivalence class if B1
dominates B2, B2 post-dominates B1 and both are in the same loop.
3- Propagation of block weights into edges. This uses a simple
propagation heuristic. The following rules are applied to every
block B in the CFG:
- If B has a single predecessor/successor, then the weight
of that edge is the weight of the block.
- If all the edges are known except one, and the weight of the
block is already known, the weight of the unknown edge will
be the weight of the block minus the sum of all the known
edges. If the sum of all the known edges is larger than B's weight,
we set the unknown edge weight to zero.
- If there is a self-referential edge, and the weight of the block is
known, the weight for that edge is set to the weight of the block
minus the weight of the other incoming edges to that block (if
known).
Since this propagation is not guaranteed to finalize for every CFG, we
only allow it to proceed for a limited number of iterations (controlled
by -sample-profile-max-propagate-iterations). It currently uses the same
GCC default of 100.
Before propagation starts, the pass builds (for each block) a list of
unique predecessors and successors. This is necessary to handle
identical edges in multiway branches. Since we visit all blocks and all
edges of the CFG, it is cleaner to build these lists once at the start
of the pass.
Finally, the patch fixes the computation of relative line locations.
The profiler emits lines relative to the function header. To discover
it, we traverse the compilation unit looking for the subprogram
corresponding to the function. The line number of that subprogram is the
line where the function begins. That becomes line zero for all the
relative locations.
llvm-svn: 198972
2014-01-11 07:23:46 +08:00
; Original C++ code for this test case:
;
; #include <stdio.h>
;
; long foo(int x, int y, long N) {
; if (x < y) {
; return y - x;
; } else {
; for (long i = 0; i < N; i++) {
; if (i > N / 3)
; x--;
; if (i > N / 4) {
; y++;
; x += 3;
; } else {
2016-04-26 12:59:11 +08:00
; for (unsigned j = 0; j < 100; j++) {
Propagation of profile samples through the CFG.
This adds a propagation heuristic to convert instruction samples
into branch weights. It implements a similar heuristic to the one
implemented by Dehao Chen on GCC.
The propagation proceeds in 3 phases:
1- Assignment of block weights. All the basic blocks in the function
are initial assigned the same weight as their most frequently
executed instruction.
2- Creation of equivalence classes. Since samples may be missing from
blocks, we can fill in the gaps by setting the weights of all the
blocks in the same equivalence class to the same weight. To compute
the concept of equivalence, we use dominance and loop information.
Two blocks B1 and B2 are in the same equivalence class if B1
dominates B2, B2 post-dominates B1 and both are in the same loop.
3- Propagation of block weights into edges. This uses a simple
propagation heuristic. The following rules are applied to every
block B in the CFG:
- If B has a single predecessor/successor, then the weight
of that edge is the weight of the block.
- If all the edges are known except one, and the weight of the
block is already known, the weight of the unknown edge will
be the weight of the block minus the sum of all the known
edges. If the sum of all the known edges is larger than B's weight,
we set the unknown edge weight to zero.
- If there is a self-referential edge, and the weight of the block is
known, the weight for that edge is set to the weight of the block
minus the weight of the other incoming edges to that block (if
known).
Since this propagation is not guaranteed to finalize for every CFG, we
only allow it to proceed for a limited number of iterations (controlled
by -sample-profile-max-propagate-iterations). It currently uses the same
GCC default of 100.
Before propagation starts, the pass builds (for each block) a list of
unique predecessors and successors. This is necessary to handle
identical edges in multiway branches. Since we visit all blocks and all
edges of the CFG, it is cleaner to build these lists once at the start
of the pass.
Finally, the patch fixes the computation of relative line locations.
The profiler emits lines relative to the function header. To discover
it, we traverse the compilation unit looking for the subprogram
corresponding to the function. The line number of that subprogram is the
line where the function begins. That becomes line zero for all the
relative locations.
llvm-svn: 198972
2014-01-11 07:23:46 +08:00
; x += j;
; y -= 3;
; }
; }
; }
; }
; return y * x;
; }
;
; int main() {
; int x = 5678;
; int y = 1234;
2016-04-26 12:59:11 +08:00
; long N = 9999999;
Propagation of profile samples through the CFG.
This adds a propagation heuristic to convert instruction samples
into branch weights. It implements a similar heuristic to the one
implemented by Dehao Chen on GCC.
The propagation proceeds in 3 phases:
1- Assignment of block weights. All the basic blocks in the function
are initial assigned the same weight as their most frequently
executed instruction.
2- Creation of equivalence classes. Since samples may be missing from
blocks, we can fill in the gaps by setting the weights of all the
blocks in the same equivalence class to the same weight. To compute
the concept of equivalence, we use dominance and loop information.
Two blocks B1 and B2 are in the same equivalence class if B1
dominates B2, B2 post-dominates B1 and both are in the same loop.
3- Propagation of block weights into edges. This uses a simple
propagation heuristic. The following rules are applied to every
block B in the CFG:
- If B has a single predecessor/successor, then the weight
of that edge is the weight of the block.
- If all the edges are known except one, and the weight of the
block is already known, the weight of the unknown edge will
be the weight of the block minus the sum of all the known
edges. If the sum of all the known edges is larger than B's weight,
we set the unknown edge weight to zero.
- If there is a self-referential edge, and the weight of the block is
known, the weight for that edge is set to the weight of the block
minus the weight of the other incoming edges to that block (if
known).
Since this propagation is not guaranteed to finalize for every CFG, we
only allow it to proceed for a limited number of iterations (controlled
by -sample-profile-max-propagate-iterations). It currently uses the same
GCC default of 100.
Before propagation starts, the pass builds (for each block) a list of
unique predecessors and successors. This is necessary to handle
identical edges in multiway branches. Since we visit all blocks and all
edges of the CFG, it is cleaner to build these lists once at the start
of the pass.
Finally, the patch fixes the computation of relative line locations.
The profiler emits lines relative to the function header. To discover
it, we traverse the compilation unit looking for the subprogram
corresponding to the function. The line number of that subprogram is the
line where the function begins. That becomes line zero for all the
relative locations.
llvm-svn: 198972
2014-01-11 07:23:46 +08:00
; printf("foo(%d, %d, %ld) = %ld\n", x, y, N, foo(x, y, N));
; return 0;
; }
; ModuleID = 'propagate.cc'
2016-04-26 12:59:11 +08:00
target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
Propagation of profile samples through the CFG.
This adds a propagation heuristic to convert instruction samples
into branch weights. It implements a similar heuristic to the one
implemented by Dehao Chen on GCC.
The propagation proceeds in 3 phases:
1- Assignment of block weights. All the basic blocks in the function
are initial assigned the same weight as their most frequently
executed instruction.
2- Creation of equivalence classes. Since samples may be missing from
blocks, we can fill in the gaps by setting the weights of all the
blocks in the same equivalence class to the same weight. To compute
the concept of equivalence, we use dominance and loop information.
Two blocks B1 and B2 are in the same equivalence class if B1
dominates B2, B2 post-dominates B1 and both are in the same loop.
3- Propagation of block weights into edges. This uses a simple
propagation heuristic. The following rules are applied to every
block B in the CFG:
- If B has a single predecessor/successor, then the weight
of that edge is the weight of the block.
- If all the edges are known except one, and the weight of the
block is already known, the weight of the unknown edge will
be the weight of the block minus the sum of all the known
edges. If the sum of all the known edges is larger than B's weight,
we set the unknown edge weight to zero.
- If there is a self-referential edge, and the weight of the block is
known, the weight for that edge is set to the weight of the block
minus the weight of the other incoming edges to that block (if
known).
Since this propagation is not guaranteed to finalize for every CFG, we
only allow it to proceed for a limited number of iterations (controlled
by -sample-profile-max-propagate-iterations). It currently uses the same
GCC default of 100.
Before propagation starts, the pass builds (for each block) a list of
unique predecessors and successors. This is necessary to handle
identical edges in multiway branches. Since we visit all blocks and all
edges of the CFG, it is cleaner to build these lists once at the start
of the pass.
Finally, the patch fixes the computation of relative line locations.
The profiler emits lines relative to the function header. To discover
it, we traverse the compilation unit looking for the subprogram
corresponding to the function. The line number of that subprogram is the
line where the function begins. That becomes line zero for all the
relative locations.
llvm-svn: 198972
2014-01-11 07:23:46 +08:00
target triple = "x86_64-unknown-linux-gnu"
@.str = private unnamed_addr constant [ 24 x i8 ] c "foo(%d, %d, %ld) = %ld\0A\00" , align 1
; Function Attrs: nounwind uwtable
2016-04-26 12:59:11 +08:00
define i64 @_Z3fooiil ( i32 %x , i32 %y , i64 %N ) #0 !dbg !6 {
Propagation of profile samples through the CFG.
This adds a propagation heuristic to convert instruction samples
into branch weights. It implements a similar heuristic to the one
implemented by Dehao Chen on GCC.
The propagation proceeds in 3 phases:
1- Assignment of block weights. All the basic blocks in the function
are initial assigned the same weight as their most frequently
executed instruction.
2- Creation of equivalence classes. Since samples may be missing from
blocks, we can fill in the gaps by setting the weights of all the
blocks in the same equivalence class to the same weight. To compute
the concept of equivalence, we use dominance and loop information.
Two blocks B1 and B2 are in the same equivalence class if B1
dominates B2, B2 post-dominates B1 and both are in the same loop.
3- Propagation of block weights into edges. This uses a simple
propagation heuristic. The following rules are applied to every
block B in the CFG:
- If B has a single predecessor/successor, then the weight
of that edge is the weight of the block.
- If all the edges are known except one, and the weight of the
block is already known, the weight of the unknown edge will
be the weight of the block minus the sum of all the known
edges. If the sum of all the known edges is larger than B's weight,
we set the unknown edge weight to zero.
- If there is a self-referential edge, and the weight of the block is
known, the weight for that edge is set to the weight of the block
minus the weight of the other incoming edges to that block (if
known).
Since this propagation is not guaranteed to finalize for every CFG, we
only allow it to proceed for a limited number of iterations (controlled
by -sample-profile-max-propagate-iterations). It currently uses the same
GCC default of 100.
Before propagation starts, the pass builds (for each block) a list of
unique predecessors and successors. This is necessary to handle
identical edges in multiway branches. Since we visit all blocks and all
edges of the CFG, it is cleaner to build these lists once at the start
of the pass.
Finally, the patch fixes the computation of relative line locations.
The profiler emits lines relative to the function header. To discover
it, we traverse the compilation unit looking for the subprogram
corresponding to the function. The line number of that subprogram is the
line where the function begins. That becomes line zero for all the
relative locations.
llvm-svn: 198972
2014-01-11 07:23:46 +08:00
entry:
%retval = alloca i64 , align 8
%x.addr = alloca i32 , align 4
%y.addr = alloca i32 , align 4
%N.addr = alloca i64 , align 8
%i = alloca i64 , align 8
2016-04-26 12:59:11 +08:00
%j = alloca i64 , align 8
Propagation of profile samples through the CFG.
This adds a propagation heuristic to convert instruction samples
into branch weights. It implements a similar heuristic to the one
implemented by Dehao Chen on GCC.
The propagation proceeds in 3 phases:
1- Assignment of block weights. All the basic blocks in the function
are initial assigned the same weight as their most frequently
executed instruction.
2- Creation of equivalence classes. Since samples may be missing from
blocks, we can fill in the gaps by setting the weights of all the
blocks in the same equivalence class to the same weight. To compute
the concept of equivalence, we use dominance and loop information.
Two blocks B1 and B2 are in the same equivalence class if B1
dominates B2, B2 post-dominates B1 and both are in the same loop.
3- Propagation of block weights into edges. This uses a simple
propagation heuristic. The following rules are applied to every
block B in the CFG:
- If B has a single predecessor/successor, then the weight
of that edge is the weight of the block.
- If all the edges are known except one, and the weight of the
block is already known, the weight of the unknown edge will
be the weight of the block minus the sum of all the known
edges. If the sum of all the known edges is larger than B's weight,
we set the unknown edge weight to zero.
- If there is a self-referential edge, and the weight of the block is
known, the weight for that edge is set to the weight of the block
minus the weight of the other incoming edges to that block (if
known).
Since this propagation is not guaranteed to finalize for every CFG, we
only allow it to proceed for a limited number of iterations (controlled
by -sample-profile-max-propagate-iterations). It currently uses the same
GCC default of 100.
Before propagation starts, the pass builds (for each block) a list of
unique predecessors and successors. This is necessary to handle
identical edges in multiway branches. Since we visit all blocks and all
edges of the CFG, it is cleaner to build these lists once at the start
of the pass.
Finally, the patch fixes the computation of relative line locations.
The profiler emits lines relative to the function header. To discover
it, we traverse the compilation unit looking for the subprogram
corresponding to the function. The line number of that subprogram is the
line where the function begins. That becomes line zero for all the
relative locations.
llvm-svn: 198972
2014-01-11 07:23:46 +08:00
store i32 %x , i32 * %x.addr , align 4
2016-04-26 12:59:11 +08:00
call void @llvm.dbg.declare ( metadata i32 * %x.addr , metadata !11 , metadata !12 ) , !dbg !13
Propagation of profile samples through the CFG.
This adds a propagation heuristic to convert instruction samples
into branch weights. It implements a similar heuristic to the one
implemented by Dehao Chen on GCC.
The propagation proceeds in 3 phases:
1- Assignment of block weights. All the basic blocks in the function
are initial assigned the same weight as their most frequently
executed instruction.
2- Creation of equivalence classes. Since samples may be missing from
blocks, we can fill in the gaps by setting the weights of all the
blocks in the same equivalence class to the same weight. To compute
the concept of equivalence, we use dominance and loop information.
Two blocks B1 and B2 are in the same equivalence class if B1
dominates B2, B2 post-dominates B1 and both are in the same loop.
3- Propagation of block weights into edges. This uses a simple
propagation heuristic. The following rules are applied to every
block B in the CFG:
- If B has a single predecessor/successor, then the weight
of that edge is the weight of the block.
- If all the edges are known except one, and the weight of the
block is already known, the weight of the unknown edge will
be the weight of the block minus the sum of all the known
edges. If the sum of all the known edges is larger than B's weight,
we set the unknown edge weight to zero.
- If there is a self-referential edge, and the weight of the block is
known, the weight for that edge is set to the weight of the block
minus the weight of the other incoming edges to that block (if
known).
Since this propagation is not guaranteed to finalize for every CFG, we
only allow it to proceed for a limited number of iterations (controlled
by -sample-profile-max-propagate-iterations). It currently uses the same
GCC default of 100.
Before propagation starts, the pass builds (for each block) a list of
unique predecessors and successors. This is necessary to handle
identical edges in multiway branches. Since we visit all blocks and all
edges of the CFG, it is cleaner to build these lists once at the start
of the pass.
Finally, the patch fixes the computation of relative line locations.
The profiler emits lines relative to the function header. To discover
it, we traverse the compilation unit looking for the subprogram
corresponding to the function. The line number of that subprogram is the
line where the function begins. That becomes line zero for all the
relative locations.
llvm-svn: 198972
2014-01-11 07:23:46 +08:00
store i32 %y , i32 * %y.addr , align 4
2016-04-26 12:59:11 +08:00
call void @llvm.dbg.declare ( metadata i32 * %y.addr , metadata !14 , metadata !12 ) , !dbg !15
Propagation of profile samples through the CFG.
This adds a propagation heuristic to convert instruction samples
into branch weights. It implements a similar heuristic to the one
implemented by Dehao Chen on GCC.
The propagation proceeds in 3 phases:
1- Assignment of block weights. All the basic blocks in the function
are initial assigned the same weight as their most frequently
executed instruction.
2- Creation of equivalence classes. Since samples may be missing from
blocks, we can fill in the gaps by setting the weights of all the
blocks in the same equivalence class to the same weight. To compute
the concept of equivalence, we use dominance and loop information.
Two blocks B1 and B2 are in the same equivalence class if B1
dominates B2, B2 post-dominates B1 and both are in the same loop.
3- Propagation of block weights into edges. This uses a simple
propagation heuristic. The following rules are applied to every
block B in the CFG:
- If B has a single predecessor/successor, then the weight
of that edge is the weight of the block.
- If all the edges are known except one, and the weight of the
block is already known, the weight of the unknown edge will
be the weight of the block minus the sum of all the known
edges. If the sum of all the known edges is larger than B's weight,
we set the unknown edge weight to zero.
- If there is a self-referential edge, and the weight of the block is
known, the weight for that edge is set to the weight of the block
minus the weight of the other incoming edges to that block (if
known).
Since this propagation is not guaranteed to finalize for every CFG, we
only allow it to proceed for a limited number of iterations (controlled
by -sample-profile-max-propagate-iterations). It currently uses the same
GCC default of 100.
Before propagation starts, the pass builds (for each block) a list of
unique predecessors and successors. This is necessary to handle
identical edges in multiway branches. Since we visit all blocks and all
edges of the CFG, it is cleaner to build these lists once at the start
of the pass.
Finally, the patch fixes the computation of relative line locations.
The profiler emits lines relative to the function header. To discover
it, we traverse the compilation unit looking for the subprogram
corresponding to the function. The line number of that subprogram is the
line where the function begins. That becomes line zero for all the
relative locations.
llvm-svn: 198972
2014-01-11 07:23:46 +08:00
store i64 %N , i64 * %N.addr , align 8
2016-04-26 12:59:11 +08:00
call void @llvm.dbg.declare ( metadata i64 * %N.addr , metadata !16 , metadata !12 ) , !dbg !17
%0 = load i32 , i32 * %x.addr , align 4 , !dbg !18
%1 = load i32 , i32 * %y.addr , align 4 , !dbg !20
%cmp = icmp slt i32 %0 , %1 , !dbg !21
br i1 %cmp , label %if.then , label %if.else , !dbg !22
Propagation of profile samples through the CFG.
This adds a propagation heuristic to convert instruction samples
into branch weights. It implements a similar heuristic to the one
implemented by Dehao Chen on GCC.
The propagation proceeds in 3 phases:
1- Assignment of block weights. All the basic blocks in the function
are initial assigned the same weight as their most frequently
executed instruction.
2- Creation of equivalence classes. Since samples may be missing from
blocks, we can fill in the gaps by setting the weights of all the
blocks in the same equivalence class to the same weight. To compute
the concept of equivalence, we use dominance and loop information.
Two blocks B1 and B2 are in the same equivalence class if B1
dominates B2, B2 post-dominates B1 and both are in the same loop.
3- Propagation of block weights into edges. This uses a simple
propagation heuristic. The following rules are applied to every
block B in the CFG:
- If B has a single predecessor/successor, then the weight
of that edge is the weight of the block.
- If all the edges are known except one, and the weight of the
block is already known, the weight of the unknown edge will
be the weight of the block minus the sum of all the known
edges. If the sum of all the known edges is larger than B's weight,
we set the unknown edge weight to zero.
- If there is a self-referential edge, and the weight of the block is
known, the weight for that edge is set to the weight of the block
minus the weight of the other incoming edges to that block (if
known).
Since this propagation is not guaranteed to finalize for every CFG, we
only allow it to proceed for a limited number of iterations (controlled
by -sample-profile-max-propagate-iterations). It currently uses the same
GCC default of 100.
Before propagation starts, the pass builds (for each block) a list of
unique predecessors and successors. This is necessary to handle
identical edges in multiway branches. Since we visit all blocks and all
edges of the CFG, it is cleaner to build these lists once at the start
of the pass.
Finally, the patch fixes the computation of relative line locations.
The profiler emits lines relative to the function header. To discover
it, we traverse the compilation unit looking for the subprogram
corresponding to the function. The line number of that subprogram is the
line where the function begins. That becomes line zero for all the
relative locations.
llvm-svn: 198972
2014-01-11 07:23:46 +08:00
if.then: ; preds = %entry
2016-04-26 12:59:11 +08:00
%2 = load i32 , i32 * %y.addr , align 4 , !dbg !23
%3 = load i32 , i32 * %x.addr , align 4 , !dbg !25
%sub = sub nsw i32 %2 , %3 , !dbg !26
%conv = sext i32 %sub to i64 , !dbg !23
store i64 %conv , i64 * %retval , align 8 , !dbg !27
br label %return , !dbg !27
Propagation of profile samples through the CFG.
This adds a propagation heuristic to convert instruction samples
into branch weights. It implements a similar heuristic to the one
implemented by Dehao Chen on GCC.
The propagation proceeds in 3 phases:
1- Assignment of block weights. All the basic blocks in the function
are initial assigned the same weight as their most frequently
executed instruction.
2- Creation of equivalence classes. Since samples may be missing from
blocks, we can fill in the gaps by setting the weights of all the
blocks in the same equivalence class to the same weight. To compute
the concept of equivalence, we use dominance and loop information.
Two blocks B1 and B2 are in the same equivalence class if B1
dominates B2, B2 post-dominates B1 and both are in the same loop.
3- Propagation of block weights into edges. This uses a simple
propagation heuristic. The following rules are applied to every
block B in the CFG:
- If B has a single predecessor/successor, then the weight
of that edge is the weight of the block.
- If all the edges are known except one, and the weight of the
block is already known, the weight of the unknown edge will
be the weight of the block minus the sum of all the known
edges. If the sum of all the known edges is larger than B's weight,
we set the unknown edge weight to zero.
- If there is a self-referential edge, and the weight of the block is
known, the weight for that edge is set to the weight of the block
minus the weight of the other incoming edges to that block (if
known).
Since this propagation is not guaranteed to finalize for every CFG, we
only allow it to proceed for a limited number of iterations (controlled
by -sample-profile-max-propagate-iterations). It currently uses the same
GCC default of 100.
Before propagation starts, the pass builds (for each block) a list of
unique predecessors and successors. This is necessary to handle
identical edges in multiway branches. Since we visit all blocks and all
edges of the CFG, it is cleaner to build these lists once at the start
of the pass.
Finally, the patch fixes the computation of relative line locations.
The profiler emits lines relative to the function header. To discover
it, we traverse the compilation unit looking for the subprogram
corresponding to the function. The line number of that subprogram is the
line where the function begins. That becomes line zero for all the
relative locations.
llvm-svn: 198972
2014-01-11 07:23:46 +08:00
if.else: ; preds = %entry
2016-04-26 12:59:11 +08:00
call void @llvm.dbg.declare ( metadata i64 * %i , metadata !28 , metadata !12 ) , !dbg !31
store i64 0 , i64 * %i , align 8 , !dbg !31
br label %for.cond , !dbg !32
Propagation of profile samples through the CFG.
This adds a propagation heuristic to convert instruction samples
into branch weights. It implements a similar heuristic to the one
implemented by Dehao Chen on GCC.
The propagation proceeds in 3 phases:
1- Assignment of block weights. All the basic blocks in the function
are initial assigned the same weight as their most frequently
executed instruction.
2- Creation of equivalence classes. Since samples may be missing from
blocks, we can fill in the gaps by setting the weights of all the
blocks in the same equivalence class to the same weight. To compute
the concept of equivalence, we use dominance and loop information.
Two blocks B1 and B2 are in the same equivalence class if B1
dominates B2, B2 post-dominates B1 and both are in the same loop.
3- Propagation of block weights into edges. This uses a simple
propagation heuristic. The following rules are applied to every
block B in the CFG:
- If B has a single predecessor/successor, then the weight
of that edge is the weight of the block.
- If all the edges are known except one, and the weight of the
block is already known, the weight of the unknown edge will
be the weight of the block minus the sum of all the known
edges. If the sum of all the known edges is larger than B's weight,
we set the unknown edge weight to zero.
- If there is a self-referential edge, and the weight of the block is
known, the weight for that edge is set to the weight of the block
minus the weight of the other incoming edges to that block (if
known).
Since this propagation is not guaranteed to finalize for every CFG, we
only allow it to proceed for a limited number of iterations (controlled
by -sample-profile-max-propagate-iterations). It currently uses the same
GCC default of 100.
Before propagation starts, the pass builds (for each block) a list of
unique predecessors and successors. This is necessary to handle
identical edges in multiway branches. Since we visit all blocks and all
edges of the CFG, it is cleaner to build these lists once at the start
of the pass.
Finally, the patch fixes the computation of relative line locations.
The profiler emits lines relative to the function header. To discover
it, we traverse the compilation unit looking for the subprogram
corresponding to the function. The line number of that subprogram is the
line where the function begins. That becomes line zero for all the
relative locations.
llvm-svn: 198972
2014-01-11 07:23:46 +08:00
2016-04-26 12:59:11 +08:00
for.cond: ; preds = %for.inc17, %if.else
%4 = load i64 , i64 * %i , align 8 , !dbg !33
%5 = load i64 , i64 * %N.addr , align 8 , !dbg !36
%cmp1 = icmp slt i64 %4 , %5 , !dbg !37
br i1 %cmp1 , label %for.body , label %for.end19 , !dbg !38
Propagation of profile samples through the CFG.
This adds a propagation heuristic to convert instruction samples
into branch weights. It implements a similar heuristic to the one
implemented by Dehao Chen on GCC.
The propagation proceeds in 3 phases:
1- Assignment of block weights. All the basic blocks in the function
are initial assigned the same weight as their most frequently
executed instruction.
2- Creation of equivalence classes. Since samples may be missing from
blocks, we can fill in the gaps by setting the weights of all the
blocks in the same equivalence class to the same weight. To compute
the concept of equivalence, we use dominance and loop information.
Two blocks B1 and B2 are in the same equivalence class if B1
dominates B2, B2 post-dominates B1 and both are in the same loop.
3- Propagation of block weights into edges. This uses a simple
propagation heuristic. The following rules are applied to every
block B in the CFG:
- If B has a single predecessor/successor, then the weight
of that edge is the weight of the block.
- If all the edges are known except one, and the weight of the
block is already known, the weight of the unknown edge will
be the weight of the block minus the sum of all the known
edges. If the sum of all the known edges is larger than B's weight,
we set the unknown edge weight to zero.
- If there is a self-referential edge, and the weight of the block is
known, the weight for that edge is set to the weight of the block
minus the weight of the other incoming edges to that block (if
known).
Since this propagation is not guaranteed to finalize for every CFG, we
only allow it to proceed for a limited number of iterations (controlled
by -sample-profile-max-propagate-iterations). It currently uses the same
GCC default of 100.
Before propagation starts, the pass builds (for each block) a list of
unique predecessors and successors. This is necessary to handle
identical edges in multiway branches. Since we visit all blocks and all
edges of the CFG, it is cleaner to build these lists once at the start
of the pass.
Finally, the patch fixes the computation of relative line locations.
The profiler emits lines relative to the function header. To discover
it, we traverse the compilation unit looking for the subprogram
corresponding to the function. The line number of that subprogram is the
line where the function begins. That becomes line zero for all the
relative locations.
llvm-svn: 198972
2014-01-11 07:23:46 +08:00
for.body: ; preds = %for.cond
2016-04-26 12:59:11 +08:00
%6 = load i64 , i64 * %i , align 8 , !dbg !39
%7 = load i64 , i64 * %N.addr , align 8 , !dbg !42
%div = sdiv i64 %7 , 3 , !dbg !43
%cmp2 = icmp sgt i64 %6 , %div , !dbg !44
br i1 %cmp2 , label %if.then3 , label %if.end , !dbg !45
2016-08-13 00:22:12 +08:00
; CHECK: edge for.body -> if.then3 probability is 0x51292fa6 / 0x80000000 = 63.41%
; CHECK: edge for.body -> if.end probability is 0x2ed6d05a / 0x80000000 = 36.59%
Propagation of profile samples through the CFG.
This adds a propagation heuristic to convert instruction samples
into branch weights. It implements a similar heuristic to the one
implemented by Dehao Chen on GCC.
The propagation proceeds in 3 phases:
1- Assignment of block weights. All the basic blocks in the function
are initial assigned the same weight as their most frequently
executed instruction.
2- Creation of equivalence classes. Since samples may be missing from
blocks, we can fill in the gaps by setting the weights of all the
blocks in the same equivalence class to the same weight. To compute
the concept of equivalence, we use dominance and loop information.
Two blocks B1 and B2 are in the same equivalence class if B1
dominates B2, B2 post-dominates B1 and both are in the same loop.
3- Propagation of block weights into edges. This uses a simple
propagation heuristic. The following rules are applied to every
block B in the CFG:
- If B has a single predecessor/successor, then the weight
of that edge is the weight of the block.
- If all the edges are known except one, and the weight of the
block is already known, the weight of the unknown edge will
be the weight of the block minus the sum of all the known
edges. If the sum of all the known edges is larger than B's weight,
we set the unknown edge weight to zero.
- If there is a self-referential edge, and the weight of the block is
known, the weight for that edge is set to the weight of the block
minus the weight of the other incoming edges to that block (if
known).
Since this propagation is not guaranteed to finalize for every CFG, we
only allow it to proceed for a limited number of iterations (controlled
by -sample-profile-max-propagate-iterations). It currently uses the same
GCC default of 100.
Before propagation starts, the pass builds (for each block) a list of
unique predecessors and successors. This is necessary to handle
identical edges in multiway branches. Since we visit all blocks and all
edges of the CFG, it is cleaner to build these lists once at the start
of the pass.
Finally, the patch fixes the computation of relative line locations.
The profiler emits lines relative to the function header. To discover
it, we traverse the compilation unit looking for the subprogram
corresponding to the function. The line number of that subprogram is the
line where the function begins. That becomes line zero for all the
relative locations.
llvm-svn: 198972
2014-01-11 07:23:46 +08:00
if.then3: ; preds = %for.body
2016-04-26 12:59:11 +08:00
%8 = load i32 , i32 * %x.addr , align 4 , !dbg !46
%dec = add nsw i32 %8 , -1 , !dbg !46
store i32 %dec , i32 * %x.addr , align 4 , !dbg !46
br label %if.end , !dbg !47
Propagation of profile samples through the CFG.
This adds a propagation heuristic to convert instruction samples
into branch weights. It implements a similar heuristic to the one
implemented by Dehao Chen on GCC.
The propagation proceeds in 3 phases:
1- Assignment of block weights. All the basic blocks in the function
are initial assigned the same weight as their most frequently
executed instruction.
2- Creation of equivalence classes. Since samples may be missing from
blocks, we can fill in the gaps by setting the weights of all the
blocks in the same equivalence class to the same weight. To compute
the concept of equivalence, we use dominance and loop information.
Two blocks B1 and B2 are in the same equivalence class if B1
dominates B2, B2 post-dominates B1 and both are in the same loop.
3- Propagation of block weights into edges. This uses a simple
propagation heuristic. The following rules are applied to every
block B in the CFG:
- If B has a single predecessor/successor, then the weight
of that edge is the weight of the block.
- If all the edges are known except one, and the weight of the
block is already known, the weight of the unknown edge will
be the weight of the block minus the sum of all the known
edges. If the sum of all the known edges is larger than B's weight,
we set the unknown edge weight to zero.
- If there is a self-referential edge, and the weight of the block is
known, the weight for that edge is set to the weight of the block
minus the weight of the other incoming edges to that block (if
known).
Since this propagation is not guaranteed to finalize for every CFG, we
only allow it to proceed for a limited number of iterations (controlled
by -sample-profile-max-propagate-iterations). It currently uses the same
GCC default of 100.
Before propagation starts, the pass builds (for each block) a list of
unique predecessors and successors. This is necessary to handle
identical edges in multiway branches. Since we visit all blocks and all
edges of the CFG, it is cleaner to build these lists once at the start
of the pass.
Finally, the patch fixes the computation of relative line locations.
The profiler emits lines relative to the function header. To discover
it, we traverse the compilation unit looking for the subprogram
corresponding to the function. The line number of that subprogram is the
line where the function begins. That becomes line zero for all the
relative locations.
llvm-svn: 198972
2014-01-11 07:23:46 +08:00
if.end: ; preds = %if.then3, %for.body
2016-04-26 12:59:11 +08:00
%9 = load i64 , i64 * %i , align 8 , !dbg !48
%10 = load i64 , i64 * %N.addr , align 8 , !dbg !50
%div4 = sdiv i64 %10 , 4 , !dbg !51
%cmp5 = icmp sgt i64 %9 , %div4 , !dbg !52
br i1 %cmp5 , label %if.then6 , label %if.else7 , !dbg !53
2016-08-13 00:22:12 +08:00
; CHECK: edge if.end -> if.then6 probability is 0x5d89d89e / 0x80000000 = 73.08%
; CHECK: edge if.end -> if.else7 probability is 0x22762762 / 0x80000000 = 26.92%
Propagation of profile samples through the CFG.
This adds a propagation heuristic to convert instruction samples
into branch weights. It implements a similar heuristic to the one
implemented by Dehao Chen on GCC.
The propagation proceeds in 3 phases:
1- Assignment of block weights. All the basic blocks in the function
are initial assigned the same weight as their most frequently
executed instruction.
2- Creation of equivalence classes. Since samples may be missing from
blocks, we can fill in the gaps by setting the weights of all the
blocks in the same equivalence class to the same weight. To compute
the concept of equivalence, we use dominance and loop information.
Two blocks B1 and B2 are in the same equivalence class if B1
dominates B2, B2 post-dominates B1 and both are in the same loop.
3- Propagation of block weights into edges. This uses a simple
propagation heuristic. The following rules are applied to every
block B in the CFG:
- If B has a single predecessor/successor, then the weight
of that edge is the weight of the block.
- If all the edges are known except one, and the weight of the
block is already known, the weight of the unknown edge will
be the weight of the block minus the sum of all the known
edges. If the sum of all the known edges is larger than B's weight,
we set the unknown edge weight to zero.
- If there is a self-referential edge, and the weight of the block is
known, the weight for that edge is set to the weight of the block
minus the weight of the other incoming edges to that block (if
known).
Since this propagation is not guaranteed to finalize for every CFG, we
only allow it to proceed for a limited number of iterations (controlled
by -sample-profile-max-propagate-iterations). It currently uses the same
GCC default of 100.
Before propagation starts, the pass builds (for each block) a list of
unique predecessors and successors. This is necessary to handle
identical edges in multiway branches. Since we visit all blocks and all
edges of the CFG, it is cleaner to build these lists once at the start
of the pass.
Finally, the patch fixes the computation of relative line locations.
The profiler emits lines relative to the function header. To discover
it, we traverse the compilation unit looking for the subprogram
corresponding to the function. The line number of that subprogram is the
line where the function begins. That becomes line zero for all the
relative locations.
llvm-svn: 198972
2014-01-11 07:23:46 +08:00
if.then6: ; preds = %if.end
2016-04-26 12:59:11 +08:00
%11 = load i32 , i32 * %y.addr , align 4 , !dbg !54
%inc = add nsw i32 %11 , 1 , !dbg !54
store i32 %inc , i32 * %y.addr , align 4 , !dbg !54
%12 = load i32 , i32 * %x.addr , align 4 , !dbg !56
%add = add nsw i32 %12 , 3 , !dbg !56
store i32 %add , i32 * %x.addr , align 4 , !dbg !56
br label %if.end16 , !dbg !57
Propagation of profile samples through the CFG.
This adds a propagation heuristic to convert instruction samples
into branch weights. It implements a similar heuristic to the one
implemented by Dehao Chen on GCC.
The propagation proceeds in 3 phases:
1- Assignment of block weights. All the basic blocks in the function
are initial assigned the same weight as their most frequently
executed instruction.
2- Creation of equivalence classes. Since samples may be missing from
blocks, we can fill in the gaps by setting the weights of all the
blocks in the same equivalence class to the same weight. To compute
the concept of equivalence, we use dominance and loop information.
Two blocks B1 and B2 are in the same equivalence class if B1
dominates B2, B2 post-dominates B1 and both are in the same loop.
3- Propagation of block weights into edges. This uses a simple
propagation heuristic. The following rules are applied to every
block B in the CFG:
- If B has a single predecessor/successor, then the weight
of that edge is the weight of the block.
- If all the edges are known except one, and the weight of the
block is already known, the weight of the unknown edge will
be the weight of the block minus the sum of all the known
edges. If the sum of all the known edges is larger than B's weight,
we set the unknown edge weight to zero.
- If there is a self-referential edge, and the weight of the block is
known, the weight for that edge is set to the weight of the block
minus the weight of the other incoming edges to that block (if
known).
Since this propagation is not guaranteed to finalize for every CFG, we
only allow it to proceed for a limited number of iterations (controlled
by -sample-profile-max-propagate-iterations). It currently uses the same
GCC default of 100.
Before propagation starts, the pass builds (for each block) a list of
unique predecessors and successors. This is necessary to handle
identical edges in multiway branches. Since we visit all blocks and all
edges of the CFG, it is cleaner to build these lists once at the start
of the pass.
Finally, the patch fixes the computation of relative line locations.
The profiler emits lines relative to the function header. To discover
it, we traverse the compilation unit looking for the subprogram
corresponding to the function. The line number of that subprogram is the
line where the function begins. That becomes line zero for all the
relative locations.
llvm-svn: 198972
2014-01-11 07:23:46 +08:00
if.else7: ; preds = %if.end
2016-04-26 12:59:11 +08:00
call void @llvm.dbg.declare ( metadata i64 * %j , metadata !58 , metadata !12 ) , !dbg !62
store i64 0 , i64 * %j , align 8 , !dbg !62
br label %for.cond8 , !dbg !63
Propagation of profile samples through the CFG.
This adds a propagation heuristic to convert instruction samples
into branch weights. It implements a similar heuristic to the one
implemented by Dehao Chen on GCC.
The propagation proceeds in 3 phases:
1- Assignment of block weights. All the basic blocks in the function
are initial assigned the same weight as their most frequently
executed instruction.
2- Creation of equivalence classes. Since samples may be missing from
blocks, we can fill in the gaps by setting the weights of all the
blocks in the same equivalence class to the same weight. To compute
the concept of equivalence, we use dominance and loop information.
Two blocks B1 and B2 are in the same equivalence class if B1
dominates B2, B2 post-dominates B1 and both are in the same loop.
3- Propagation of block weights into edges. This uses a simple
propagation heuristic. The following rules are applied to every
block B in the CFG:
- If B has a single predecessor/successor, then the weight
of that edge is the weight of the block.
- If all the edges are known except one, and the weight of the
block is already known, the weight of the unknown edge will
be the weight of the block minus the sum of all the known
edges. If the sum of all the known edges is larger than B's weight,
we set the unknown edge weight to zero.
- If there is a self-referential edge, and the weight of the block is
known, the weight for that edge is set to the weight of the block
minus the weight of the other incoming edges to that block (if
known).
Since this propagation is not guaranteed to finalize for every CFG, we
only allow it to proceed for a limited number of iterations (controlled
by -sample-profile-max-propagate-iterations). It currently uses the same
GCC default of 100.
Before propagation starts, the pass builds (for each block) a list of
unique predecessors and successors. This is necessary to handle
identical edges in multiway branches. Since we visit all blocks and all
edges of the CFG, it is cleaner to build these lists once at the start
of the pass.
Finally, the patch fixes the computation of relative line locations.
The profiler emits lines relative to the function header. To discover
it, we traverse the compilation unit looking for the subprogram
corresponding to the function. The line number of that subprogram is the
line where the function begins. That becomes line zero for all the
relative locations.
llvm-svn: 198972
2014-01-11 07:23:46 +08:00
for.cond8: ; preds = %for.inc, %if.else7
2016-04-26 12:59:11 +08:00
%13 = load i64 , i64 * %j , align 8 , !dbg !64
%cmp9 = icmp slt i64 %13 , 100 , !dbg !67
br i1 %cmp9 , label %for.body10 , label %for.end , !dbg !68
2016-08-13 00:22:12 +08:00
; CHECK: edge for.cond8 -> for.body10 probability is 0x7e941a89 / 0x80000000 = 98.89% [HOT edge]
; CHECK: edge for.cond8 -> for.end probability is 0x016be577 / 0x80000000 = 1.11%
2016-04-26 12:59:11 +08:00
for.body10: ; preds = %for.cond8
%14 = load i64 , i64 * %j , align 8 , !dbg !69
%15 = load i32 , i32 * %x.addr , align 4 , !dbg !71
%conv11 = sext i32 %15 to i64 , !dbg !71
%add12 = add nsw i64 %conv11 , %14 , !dbg !71
%conv13 = trunc i64 %add12 to i32 , !dbg !71
store i32 %conv13 , i32 * %x.addr , align 4 , !dbg !71
%16 = load i32 , i32 * %y.addr , align 4 , !dbg !72
%sub14 = sub nsw i32 %16 , 3 , !dbg !72
store i32 %sub14 , i32 * %y.addr , align 4 , !dbg !72
br label %for.inc , !dbg !73
for.inc: ; preds = %for.body10
%17 = load i64 , i64 * %j , align 8 , !dbg !74
%inc15 = add nsw i64 %17 , 1 , !dbg !74
store i64 %inc15 , i64 * %j , align 8 , !dbg !74
br label %for.cond8 , !dbg !76
Propagation of profile samples through the CFG.
This adds a propagation heuristic to convert instruction samples
into branch weights. It implements a similar heuristic to the one
implemented by Dehao Chen on GCC.
The propagation proceeds in 3 phases:
1- Assignment of block weights. All the basic blocks in the function
are initial assigned the same weight as their most frequently
executed instruction.
2- Creation of equivalence classes. Since samples may be missing from
blocks, we can fill in the gaps by setting the weights of all the
blocks in the same equivalence class to the same weight. To compute
the concept of equivalence, we use dominance and loop information.
Two blocks B1 and B2 are in the same equivalence class if B1
dominates B2, B2 post-dominates B1 and both are in the same loop.
3- Propagation of block weights into edges. This uses a simple
propagation heuristic. The following rules are applied to every
block B in the CFG:
- If B has a single predecessor/successor, then the weight
of that edge is the weight of the block.
- If all the edges are known except one, and the weight of the
block is already known, the weight of the unknown edge will
be the weight of the block minus the sum of all the known
edges. If the sum of all the known edges is larger than B's weight,
we set the unknown edge weight to zero.
- If there is a self-referential edge, and the weight of the block is
known, the weight for that edge is set to the weight of the block
minus the weight of the other incoming edges to that block (if
known).
Since this propagation is not guaranteed to finalize for every CFG, we
only allow it to proceed for a limited number of iterations (controlled
by -sample-profile-max-propagate-iterations). It currently uses the same
GCC default of 100.
Before propagation starts, the pass builds (for each block) a list of
unique predecessors and successors. This is necessary to handle
identical edges in multiway branches. Since we visit all blocks and all
edges of the CFG, it is cleaner to build these lists once at the start
of the pass.
Finally, the patch fixes the computation of relative line locations.
The profiler emits lines relative to the function header. To discover
it, we traverse the compilation unit looking for the subprogram
corresponding to the function. The line number of that subprogram is the
line where the function begins. That becomes line zero for all the
relative locations.
llvm-svn: 198972
2014-01-11 07:23:46 +08:00
for.end: ; preds = %for.cond8
2016-04-26 12:59:11 +08:00
br label %if.end16
if.end16: ; preds = %for.end, %if.then6
br label %for.inc17 , !dbg !77
for.inc17: ; preds = %if.end16
%18 = load i64 , i64 * %i , align 8 , !dbg !78
%inc18 = add nsw i64 %18 , 1 , !dbg !78
store i64 %inc18 , i64 * %i , align 8 , !dbg !78
br label %for.cond , !dbg !80
for.end19: ; preds = %for.cond
br label %if.end20
if.end20: ; preds = %for.end19
%19 = load i32 , i32 * %y.addr , align 4 , !dbg !81
%20 = load i32 , i32 * %x.addr , align 4 , !dbg !82
%mul = mul nsw i32 %19 , %20 , !dbg !83
%conv21 = sext i32 %mul to i64 , !dbg !81
store i64 %conv21 , i64 * %retval , align 8 , !dbg !84
br label %return , !dbg !84
return: ; preds = %if.end20, %if.then
%21 = load i64 , i64 * %retval , align 8 , !dbg !85
ret i64 %21 , !dbg !85
Propagation of profile samples through the CFG.
This adds a propagation heuristic to convert instruction samples
into branch weights. It implements a similar heuristic to the one
implemented by Dehao Chen on GCC.
The propagation proceeds in 3 phases:
1- Assignment of block weights. All the basic blocks in the function
are initial assigned the same weight as their most frequently
executed instruction.
2- Creation of equivalence classes. Since samples may be missing from
blocks, we can fill in the gaps by setting the weights of all the
blocks in the same equivalence class to the same weight. To compute
the concept of equivalence, we use dominance and loop information.
Two blocks B1 and B2 are in the same equivalence class if B1
dominates B2, B2 post-dominates B1 and both are in the same loop.
3- Propagation of block weights into edges. This uses a simple
propagation heuristic. The following rules are applied to every
block B in the CFG:
- If B has a single predecessor/successor, then the weight
of that edge is the weight of the block.
- If all the edges are known except one, and the weight of the
block is already known, the weight of the unknown edge will
be the weight of the block minus the sum of all the known
edges. If the sum of all the known edges is larger than B's weight,
we set the unknown edge weight to zero.
- If there is a self-referential edge, and the weight of the block is
known, the weight for that edge is set to the weight of the block
minus the weight of the other incoming edges to that block (if
known).
Since this propagation is not guaranteed to finalize for every CFG, we
only allow it to proceed for a limited number of iterations (controlled
by -sample-profile-max-propagate-iterations). It currently uses the same
GCC default of 100.
Before propagation starts, the pass builds (for each block) a list of
unique predecessors and successors. This is necessary to handle
identical edges in multiway branches. Since we visit all blocks and all
edges of the CFG, it is cleaner to build these lists once at the start
of the pass.
Finally, the patch fixes the computation of relative line locations.
The profiler emits lines relative to the function header. To discover
it, we traverse the compilation unit looking for the subprogram
corresponding to the function. The line number of that subprogram is the
line where the function begins. That becomes line zero for all the
relative locations.
llvm-svn: 198972
2014-01-11 07:23:46 +08:00
}
2016-04-26 12:59:11 +08:00
; Function Attrs: nounwind readnone
declare void @llvm.dbg.declare ( metadata , metadata , metadata ) #1
; Function Attrs: norecurse uwtable
define i32 @main ( ) #2 !dbg !86 {
Propagation of profile samples through the CFG.
This adds a propagation heuristic to convert instruction samples
into branch weights. It implements a similar heuristic to the one
implemented by Dehao Chen on GCC.
The propagation proceeds in 3 phases:
1- Assignment of block weights. All the basic blocks in the function
are initial assigned the same weight as their most frequently
executed instruction.
2- Creation of equivalence classes. Since samples may be missing from
blocks, we can fill in the gaps by setting the weights of all the
blocks in the same equivalence class to the same weight. To compute
the concept of equivalence, we use dominance and loop information.
Two blocks B1 and B2 are in the same equivalence class if B1
dominates B2, B2 post-dominates B1 and both are in the same loop.
3- Propagation of block weights into edges. This uses a simple
propagation heuristic. The following rules are applied to every
block B in the CFG:
- If B has a single predecessor/successor, then the weight
of that edge is the weight of the block.
- If all the edges are known except one, and the weight of the
block is already known, the weight of the unknown edge will
be the weight of the block minus the sum of all the known
edges. If the sum of all the known edges is larger than B's weight,
we set the unknown edge weight to zero.
- If there is a self-referential edge, and the weight of the block is
known, the weight for that edge is set to the weight of the block
minus the weight of the other incoming edges to that block (if
known).
Since this propagation is not guaranteed to finalize for every CFG, we
only allow it to proceed for a limited number of iterations (controlled
by -sample-profile-max-propagate-iterations). It currently uses the same
GCC default of 100.
Before propagation starts, the pass builds (for each block) a list of
unique predecessors and successors. This is necessary to handle
identical edges in multiway branches. Since we visit all blocks and all
edges of the CFG, it is cleaner to build these lists once at the start
of the pass.
Finally, the patch fixes the computation of relative line locations.
The profiler emits lines relative to the function header. To discover
it, we traverse the compilation unit looking for the subprogram
corresponding to the function. The line number of that subprogram is the
line where the function begins. That becomes line zero for all the
relative locations.
llvm-svn: 198972
2014-01-11 07:23:46 +08:00
entry:
%retval = alloca i32 , align 4
%x = alloca i32 , align 4
%y = alloca i32 , align 4
%N = alloca i64 , align 8
2016-04-26 12:59:11 +08:00
store i32 0 , i32 * %retval , align 4
call void @llvm.dbg.declare ( metadata i32 * %x , metadata !89 , metadata !12 ) , !dbg !90
store i32 5678 , i32 * %x , align 4 , !dbg !90
call void @llvm.dbg.declare ( metadata i32 * %y , metadata !91 , metadata !12 ) , !dbg !92
store i32 1234 , i32 * %y , align 4 , !dbg !92
call void @llvm.dbg.declare ( metadata i64 * %N , metadata !93 , metadata !12 ) , !dbg !94
store i64 9999999 , i64 * %N , align 8 , !dbg !94
%0 = load i32 , i32 * %x , align 4 , !dbg !95
%1 = load i32 , i32 * %y , align 4 , !dbg !96
%2 = load i64 , i64 * %N , align 8 , !dbg !97
%3 = load i32 , i32 * %x , align 4 , !dbg !98
%4 = load i32 , i32 * %y , align 4 , !dbg !99
%5 = load i64 , i64 * %N , align 8 , !dbg !100
%call = call i64 @_Z3fooiil ( i32 %3 , i32 %4 , i64 %5 ) , !dbg !101
%call1 = call i32 ( i8 * , . . . ) @printf ( i8 * getelementptr inbounds ( [ 24 x i8 ] , [ 24 x i8 ] * @.str , i32 0 , i32 0 ) , i32 %0 , i32 %1 , i64 %2 , i64 %call ) , !dbg !102
ret i32 0 , !dbg !104
Propagation of profile samples through the CFG.
This adds a propagation heuristic to convert instruction samples
into branch weights. It implements a similar heuristic to the one
implemented by Dehao Chen on GCC.
The propagation proceeds in 3 phases:
1- Assignment of block weights. All the basic blocks in the function
are initial assigned the same weight as their most frequently
executed instruction.
2- Creation of equivalence classes. Since samples may be missing from
blocks, we can fill in the gaps by setting the weights of all the
blocks in the same equivalence class to the same weight. To compute
the concept of equivalence, we use dominance and loop information.
Two blocks B1 and B2 are in the same equivalence class if B1
dominates B2, B2 post-dominates B1 and both are in the same loop.
3- Propagation of block weights into edges. This uses a simple
propagation heuristic. The following rules are applied to every
block B in the CFG:
- If B has a single predecessor/successor, then the weight
of that edge is the weight of the block.
- If all the edges are known except one, and the weight of the
block is already known, the weight of the unknown edge will
be the weight of the block minus the sum of all the known
edges. If the sum of all the known edges is larger than B's weight,
we set the unknown edge weight to zero.
- If there is a self-referential edge, and the weight of the block is
known, the weight for that edge is set to the weight of the block
minus the weight of the other incoming edges to that block (if
known).
Since this propagation is not guaranteed to finalize for every CFG, we
only allow it to proceed for a limited number of iterations (controlled
by -sample-profile-max-propagate-iterations). It currently uses the same
GCC default of 100.
Before propagation starts, the pass builds (for each block) a list of
unique predecessors and successors. This is necessary to handle
identical edges in multiway branches. Since we visit all blocks and all
edges of the CFG, it is cleaner to build these lists once at the start
of the pass.
Finally, the patch fixes the computation of relative line locations.
The profiler emits lines relative to the function header. To discover
it, we traverse the compilation unit looking for the subprogram
corresponding to the function. The line number of that subprogram is the
line where the function begins. That becomes line zero for all the
relative locations.
llvm-svn: 198972
2014-01-11 07:23:46 +08:00
}
2016-04-26 12:59:11 +08:00
declare i32 @printf ( i8 * , . . . ) #3
Propagation of profile samples through the CFG.
This adds a propagation heuristic to convert instruction samples
into branch weights. It implements a similar heuristic to the one
implemented by Dehao Chen on GCC.
The propagation proceeds in 3 phases:
1- Assignment of block weights. All the basic blocks in the function
are initial assigned the same weight as their most frequently
executed instruction.
2- Creation of equivalence classes. Since samples may be missing from
blocks, we can fill in the gaps by setting the weights of all the
blocks in the same equivalence class to the same weight. To compute
the concept of equivalence, we use dominance and loop information.
Two blocks B1 and B2 are in the same equivalence class if B1
dominates B2, B2 post-dominates B1 and both are in the same loop.
3- Propagation of block weights into edges. This uses a simple
propagation heuristic. The following rules are applied to every
block B in the CFG:
- If B has a single predecessor/successor, then the weight
of that edge is the weight of the block.
- If all the edges are known except one, and the weight of the
block is already known, the weight of the unknown edge will
be the weight of the block minus the sum of all the known
edges. If the sum of all the known edges is larger than B's weight,
we set the unknown edge weight to zero.
- If there is a self-referential edge, and the weight of the block is
known, the weight for that edge is set to the weight of the block
minus the weight of the other incoming edges to that block (if
known).
Since this propagation is not guaranteed to finalize for every CFG, we
only allow it to proceed for a limited number of iterations (controlled
by -sample-profile-max-propagate-iterations). It currently uses the same
GCC default of 100.
Before propagation starts, the pass builds (for each block) a list of
unique predecessors and successors. This is necessary to handle
identical edges in multiway branches. Since we visit all blocks and all
edges of the CFG, it is cleaner to build these lists once at the start
of the pass.
Finally, the patch fixes the computation of relative line locations.
The profiler emits lines relative to the function header. To discover
it, we traverse the compilation unit looking for the subprogram
corresponding to the function. The line number of that subprogram is the
line where the function begins. That becomes line zero for all the
relative locations.
llvm-svn: 198972
2014-01-11 07:23:46 +08:00
2016-04-26 12:59:11 +08:00
attributes #0 = { nounwind uwtable "disable-tail-calls" = "false" "less-precise-fpmad" = "false" "no-frame-pointer-elim" = "true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math" = "false" "no-jump-tables" = "false" "no-nans-fp-math" = "false" "stack-protector-buffer-size" = "8" "target-cpu" = "x86-64" "target-features" = "+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math" = "false" "use-soft-float" = "false" }
attributes #1 = { nounwind readnone }
attributes #2 = { norecurse uwtable "disable-tail-calls" = "false" "less-precise-fpmad" = "false" "no-frame-pointer-elim" = "true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math" = "false" "no-jump-tables" = "false" "no-nans-fp-math" = "false" "stack-protector-buffer-size" = "8" "target-cpu" = "x86-64" "target-features" = "+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math" = "false" "use-soft-float" = "false" }
attributes #3 = { "disable-tail-calls" = "false" "less-precise-fpmad" = "false" "no-frame-pointer-elim" = "true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math" = "false" "no-nans-fp-math" = "false" "stack-protector-buffer-size" = "8" "target-cpu" = "x86-64" "target-features" = "+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math" = "false" "use-soft-float" = "false" }
Propagation of profile samples through the CFG.
This adds a propagation heuristic to convert instruction samples
into branch weights. It implements a similar heuristic to the one
implemented by Dehao Chen on GCC.
The propagation proceeds in 3 phases:
1- Assignment of block weights. All the basic blocks in the function
are initial assigned the same weight as their most frequently
executed instruction.
2- Creation of equivalence classes. Since samples may be missing from
blocks, we can fill in the gaps by setting the weights of all the
blocks in the same equivalence class to the same weight. To compute
the concept of equivalence, we use dominance and loop information.
Two blocks B1 and B2 are in the same equivalence class if B1
dominates B2, B2 post-dominates B1 and both are in the same loop.
3- Propagation of block weights into edges. This uses a simple
propagation heuristic. The following rules are applied to every
block B in the CFG:
- If B has a single predecessor/successor, then the weight
of that edge is the weight of the block.
- If all the edges are known except one, and the weight of the
block is already known, the weight of the unknown edge will
be the weight of the block minus the sum of all the known
edges. If the sum of all the known edges is larger than B's weight,
we set the unknown edge weight to zero.
- If there is a self-referential edge, and the weight of the block is
known, the weight for that edge is set to the weight of the block
minus the weight of the other incoming edges to that block (if
known).
Since this propagation is not guaranteed to finalize for every CFG, we
only allow it to proceed for a limited number of iterations (controlled
by -sample-profile-max-propagate-iterations). It currently uses the same
GCC default of 100.
Before propagation starts, the pass builds (for each block) a list of
unique predecessors and successors. This is necessary to handle
identical edges in multiway branches. Since we visit all blocks and all
edges of the CFG, it is cleaner to build these lists once at the start
of the pass.
Finally, the patch fixes the computation of relative line locations.
The profiler emits lines relative to the function header. To discover
it, we traverse the compilation unit looking for the subprogram
corresponding to the function. The line number of that subprogram is the
line where the function begins. That becomes line zero for all the
relative locations.
llvm-svn: 198972
2014-01-11 07:23:46 +08:00
!llvm.dbg.cu = ! { !0 }
2016-04-26 12:59:11 +08:00
!llvm.module.flags = ! { !3 , !4 }
!llvm.ident = ! { !5 }
Propagation of profile samples through the CFG.
This adds a propagation heuristic to convert instruction samples
into branch weights. It implements a similar heuristic to the one
implemented by Dehao Chen on GCC.
The propagation proceeds in 3 phases:
1- Assignment of block weights. All the basic blocks in the function
are initial assigned the same weight as their most frequently
executed instruction.
2- Creation of equivalence classes. Since samples may be missing from
blocks, we can fill in the gaps by setting the weights of all the
blocks in the same equivalence class to the same weight. To compute
the concept of equivalence, we use dominance and loop information.
Two blocks B1 and B2 are in the same equivalence class if B1
dominates B2, B2 post-dominates B1 and both are in the same loop.
3- Propagation of block weights into edges. This uses a simple
propagation heuristic. The following rules are applied to every
block B in the CFG:
- If B has a single predecessor/successor, then the weight
of that edge is the weight of the block.
- If all the edges are known except one, and the weight of the
block is already known, the weight of the unknown edge will
be the weight of the block minus the sum of all the known
edges. If the sum of all the known edges is larger than B's weight,
we set the unknown edge weight to zero.
- If there is a self-referential edge, and the weight of the block is
known, the weight for that edge is set to the weight of the block
minus the weight of the other incoming edges to that block (if
known).
Since this propagation is not guaranteed to finalize for every CFG, we
only allow it to proceed for a limited number of iterations (controlled
by -sample-profile-max-propagate-iterations). It currently uses the same
GCC default of 100.
Before propagation starts, the pass builds (for each block) a list of
unique predecessors and successors. This is necessary to handle
identical edges in multiway branches. Since we visit all blocks and all
edges of the CFG, it is cleaner to build these lists once at the start
of the pass.
Finally, the patch fixes the computation of relative line locations.
The profiler emits lines relative to the function header. To discover
it, we traverse the compilation unit looking for the subprogram
corresponding to the function. The line number of that subprogram is the
line where the function begins. That becomes line zero for all the
relative locations.
llvm-svn: 198972
2014-01-11 07:23:46 +08:00
2016-04-26 12:59:11 +08:00
!0 = distinct !DICompileUnit ( language: D W _ L A N G _ C _ p l u s _ p l u s , file: !1 , producer: "clang version 3.9.0 (trunk 266819)" , isOptimized: false , runtimeVersion: 0 , emissionKind: F u l l D e b u g , enums: !2 )
2015-04-30 00:38:44 +08:00
!1 = !DIFile ( filename: "propagate.cc" , directory: "." )
2015-03-17 05:10:12 +08:00
!2 = ! { }
2016-04-26 12:59:11 +08:00
!3 = ! { i32 2 , !"Dwarf Version" , i32 4 }
!4 = ! { i32 2 , !"Debug Info Version" , i32 3 }
!5 = ! { !"clang version 3.9.0 (trunk 266819)" }
!6 = distinct !DISubprogram ( name: "foo" , linkageName: "_Z3fooiil" , scope: !1 , file: !1 , line: 3 , type: !7 , isLocal: false , isDefinition: true , scopeLine: 3 , flags: D I F l a g P r o t o t y p e d , isOptimized: false , unit: !0 , variables: !2 )
!7 = !DISubroutineType ( types: !8 )
!8 = ! { !9 , !10 , !10 , !9 }
!9 = !DIBasicType ( name: "long int" , size: 64 , align: 64 , encoding: D W _ A T E _ s i g n e d )
!10 = !DIBasicType ( name: "int" , size: 32 , align: 32 , encoding: D W _ A T E _ s i g n e d )
!11 = !DILocalVariable ( name: "x" , arg: 1 , scope: !6 , file: !1 , line: 3 , type: !10 )
!12 = !DIExpression ( )
!13 = !DILocation ( line: 3 , column: 14 , scope: !6 )
!14 = !DILocalVariable ( name: "y" , arg: 2 , scope: !6 , file: !1 , line: 3 , type: !10 )
!15 = !DILocation ( line: 3 , column: 21 , scope: !6 )
!16 = !DILocalVariable ( name: "N" , arg: 3 , scope: !6 , file: !1 , line: 3 , type: !9 )
!17 = !DILocation ( line: 3 , column: 29 , scope: !6 )
!18 = !DILocation ( line: 4 , column: 7 , scope: !19 )
!19 = distinct !DILexicalBlock ( scope: !6 , file: !1 , line: 4 , column: 7 )
!20 = !DILocation ( line: 4 , column: 11 , scope: !19 )
!21 = !DILocation ( line: 4 , column: 9 , scope: !19 )
!22 = !DILocation ( line: 4 , column: 7 , scope: !6 )
!23 = !DILocation ( line: 5 , column: 12 , scope: !24 )
!24 = distinct !DILexicalBlock ( scope: !19 , file: !1 , line: 4 , column: 14 )
!25 = !DILocation ( line: 5 , column: 16 , scope: !24 )
!26 = !DILocation ( line: 5 , column: 14 , scope: !24 )
!27 = !DILocation ( line: 5 , column: 5 , scope: !24 )
!28 = !DILocalVariable ( name: "i" , scope: !29 , file: !1 , line: 7 , type: !9 )
!29 = distinct !DILexicalBlock ( scope: !30 , file: !1 , line: 7 , column: 5 )
!30 = distinct !DILexicalBlock ( scope: !19 , file: !1 , line: 6 , column: 10 )
!31 = !DILocation ( line: 7 , column: 15 , scope: !29 )
!32 = !DILocation ( line: 7 , column: 10 , scope: !29 )
!33 = !DILocation ( line: 7 , column: 22 , scope: !34 )
2017-02-24 02:27:45 +08:00
!34 = !DILexicalBlockFile ( scope: !35 , file: !1 , discriminator: 2 )
2016-04-26 12:59:11 +08:00
!35 = distinct !DILexicalBlock ( scope: !29 , file: !1 , line: 7 , column: 5 )
!36 = !DILocation ( line: 7 , column: 26 , scope: !34 )
!37 = !DILocation ( line: 7 , column: 24 , scope: !34 )
!38 = !DILocation ( line: 7 , column: 5 , scope: !34 )
!39 = !DILocation ( line: 8 , column: 11 , scope: !40 )
!40 = distinct !DILexicalBlock ( scope: !41 , file: !1 , line: 8 , column: 11 )
!41 = distinct !DILexicalBlock ( scope: !35 , file: !1 , line: 7 , column: 34 )
!42 = !DILocation ( line: 8 , column: 15 , scope: !40 )
!43 = !DILocation ( line: 8 , column: 17 , scope: !40 )
!44 = !DILocation ( line: 8 , column: 13 , scope: !40 )
!45 = !DILocation ( line: 8 , column: 11 , scope: !41 )
!46 = !DILocation ( line: 9 , column: 10 , scope: !40 )
!47 = !DILocation ( line: 9 , column: 9 , scope: !40 )
!48 = !DILocation ( line: 10 , column: 11 , scope: !49 )
!49 = distinct !DILexicalBlock ( scope: !41 , file: !1 , line: 10 , column: 11 )
!50 = !DILocation ( line: 10 , column: 15 , scope: !49 )
!51 = !DILocation ( line: 10 , column: 17 , scope: !49 )
!52 = !DILocation ( line: 10 , column: 13 , scope: !49 )
!53 = !DILocation ( line: 10 , column: 11 , scope: !41 )
!54 = !DILocation ( line: 11 , column: 10 , scope: !55 )
!55 = distinct !DILexicalBlock ( scope: !49 , file: !1 , line: 10 , column: 22 )
!56 = !DILocation ( line: 12 , column: 11 , scope: !55 )
!57 = !DILocation ( line: 13 , column: 7 , scope: !55 )
!58 = !DILocalVariable ( name: "j" , scope: !59 , file: !1 , line: 14 , type: !61 )
!59 = distinct !DILexicalBlock ( scope: !60 , file: !1 , line: 14 , column: 9 )
!60 = distinct !DILexicalBlock ( scope: !49 , file: !1 , line: 13 , column: 14 )
!61 = !DIBasicType ( name: "long long int" , size: 64 , align: 64 , encoding: D W _ A T E _ s i g n e d )
!62 = !DILocation ( line: 14 , column: 24 , scope: !59 )
!63 = !DILocation ( line: 14 , column: 14 , scope: !59 )
!64 = !DILocation ( line: 14 , column: 31 , scope: !65 )
2017-02-24 02:27:45 +08:00
!65 = !DILexicalBlockFile ( scope: !66 , file: !1 , discriminator: 2 )
2016-04-26 12:59:11 +08:00
!66 = distinct !DILexicalBlock ( scope: !59 , file: !1 , line: 14 , column: 9 )
!67 = !DILocation ( line: 14 , column: 33 , scope: !65 )
!68 = !DILocation ( line: 14 , column: 9 , scope: !65 )
!69 = !DILocation ( line: 15 , column: 16 , scope: !70 )
!70 = distinct !DILexicalBlock ( scope: !66 , file: !1 , line: 14 , column: 45 )
!71 = !DILocation ( line: 15 , column: 13 , scope: !70 )
!72 = !DILocation ( line: 16 , column: 13 , scope: !70 )
!73 = !DILocation ( line: 17 , column: 9 , scope: !70 )
!74 = !DILocation ( line: 14 , column: 41 , scope: !75 )
2017-02-24 02:27:45 +08:00
!75 = !DILexicalBlockFile ( scope: !66 , file: !1 , discriminator: 4 )
2016-04-26 12:59:11 +08:00
!76 = !DILocation ( line: 14 , column: 9 , scope: !75 )
!77 = !DILocation ( line: 19 , column: 5 , scope: !41 )
!78 = !DILocation ( line: 7 , column: 30 , scope: !79 )
2017-02-24 02:27:45 +08:00
!79 = !DILexicalBlockFile ( scope: !35 , file: !1 , discriminator: 4 )
2016-04-26 12:59:11 +08:00
!80 = !DILocation ( line: 7 , column: 5 , scope: !79 )
!81 = !DILocation ( line: 21 , column: 10 , scope: !6 )
!82 = !DILocation ( line: 21 , column: 14 , scope: !6 )
!83 = !DILocation ( line: 21 , column: 12 , scope: !6 )
!84 = !DILocation ( line: 21 , column: 3 , scope: !6 )
!85 = !DILocation ( line: 22 , column: 1 , scope: !6 )
!86 = distinct !DISubprogram ( name: "main" , scope: !1 , file: !1 , line: 24 , type: !87 , isLocal: false , isDefinition: true , scopeLine: 24 , flags: D I F l a g P r o t o t y p e d , isOptimized: false , unit: !0 , variables: !2 )
!87 = !DISubroutineType ( types: !88 )
!88 = ! { !10 }
!89 = !DILocalVariable ( name: "x" , scope: !86 , file: !1 , line: 25 , type: !10 )
!90 = !DILocation ( line: 25 , column: 7 , scope: !86 )
!91 = !DILocalVariable ( name: "y" , scope: !86 , file: !1 , line: 26 , type: !10 )
!92 = !DILocation ( line: 26 , column: 7 , scope: !86 )
!93 = !DILocalVariable ( name: "N" , scope: !86 , file: !1 , line: 27 , type: !9 )
!94 = !DILocation ( line: 27 , column: 8 , scope: !86 )
!95 = !DILocation ( line: 28 , column: 38 , scope: !86 )
!96 = !DILocation ( line: 28 , column: 41 , scope: !86 )
!97 = !DILocation ( line: 28 , column: 44 , scope: !86 )
!98 = !DILocation ( line: 28 , column: 51 , scope: !86 )
!99 = !DILocation ( line: 28 , column: 54 , scope: !86 )
!100 = !DILocation ( line: 28 , column: 57 , scope: !86 )
!101 = !DILocation ( line: 28 , column: 47 , scope: !86 )
!102 = !DILocation ( line: 28 , column: 3 , scope: !103 )
2017-02-24 02:27:45 +08:00
!103 = !DILexicalBlockFile ( scope: !86 , file: !1 , discriminator: 2 )
2016-04-26 12:59:11 +08:00
!104 = !DILocation ( line: 29 , column: 3 , scope: !86 )
