2014-03-11 12:37:49 +08:00
|
|
|
// Test instrumentation of general constructs in C.
|
2014-01-07 06:27:43 +08:00
|
|
|
|
2016-02-05 02:39:09 +08:00
|
|
|
// RUN: %clang_cc1 -triple x86_64-apple-macosx10.9 -main-file-name c-general.c %s -o - -emit-llvm -fprofile-instrument=clang | FileCheck -check-prefix=PGOGEN %s
|
2014-04-18 06:49:06 +08:00
|
|
|
|
|
|
|
|
// RUN: llvm-profdata merge %S/Inputs/c-general.proftext -o %t.profdata
|
2016-03-03 04:59:36 +08:00
|
|
|
// RUN: %clang_cc1 -triple x86_64-apple-macosx10.9 -main-file-name c-general.c %s -o - -emit-llvm -fprofile-instrument-use-path=%t.profdata | FileCheck -check-prefix=PGOUSE %s
|
|
|
|
|
// RUN: %clang_cc1 -triple x86_64-apple-macosx10.9 -main-file-name c-general.c %s -o - -emit-llvm -fprofile-instrument-use-path=%S/Inputs/c-general.profdata.v3 | FileCheck -check-prefix=PGOUSE %s
|
2014-08-02 06:50:16 +08:00
|
|
|
// Also check compatibility with older profiles.
|
2016-03-03 04:59:36 +08:00
|
|
|
// RUN: %clang_cc1 -triple x86_64-apple-macosx10.9 -main-file-name c-general.c %s -o - -emit-llvm -fprofile-instrument-use-path=%S/Inputs/c-general.profdata.v1 | FileCheck -check-prefix=PGOUSE %s
|
2014-08-02 06:50:16 +08:00
|
|
|
|
2015-12-15 08:33:12 +08:00
|
|
|
// PGOGEN: @[[SLC:__profc_simple_loops]] = private global [4 x i64] zeroinitializer
|
|
|
|
|
// PGOGEN: @[[IFC:__profc_conditionals]] = private global [11 x i64] zeroinitializer
|
|
|
|
|
// PGOGEN: @[[EEC:__profc_early_exits]] = private global [9 x i64] zeroinitializer
|
|
|
|
|
// PGOGEN: @[[JMC:__profc_jumps]] = private global [22 x i64] zeroinitializer
|
|
|
|
|
// PGOGEN: @[[SWC:__profc_switches]] = private global [19 x i64] zeroinitializer
|
|
|
|
|
// PGOGEN: @[[BSC:__profc_big_switch]] = private global [17 x i64] zeroinitializer
|
|
|
|
|
// PGOGEN: @[[BOC:__profc_boolean_operators]] = private global [8 x i64] zeroinitializer
|
|
|
|
|
// PGOGEN: @[[BLC:__profc_boolop_loops]] = private global [9 x i64] zeroinitializer
|
|
|
|
|
// PGOGEN: @[[COC:__profc_conditional_operator]] = private global [3 x i64] zeroinitializer
|
|
|
|
|
// PGOGEN: @[[DFC:__profc_do_fallthrough]] = private global [4 x i64] zeroinitializer
|
|
|
|
|
// PGOGEN: @[[MAC:__profc_main]] = private global [1 x i64] zeroinitializer
|
|
|
|
|
// PGOGEN: @[[STC:__profc_c_general.c_static_func]] = private global [2 x i64] zeroinitializer
|
2014-01-07 06:27:43 +08:00
|
|
|
|
|
|
|
|
// PGOGEN-LABEL: @simple_loops()
|
|
|
|
|
// PGOUSE-LABEL: @simple_loops()
|
|
|
|
|
// PGOGEN: store {{.*}} @[[SLC]], i64 0, i64 0
|
|
|
|
|
void simple_loops() {
|
|
|
|
|
int i;
|
|
|
|
|
// PGOGEN: store {{.*}} @[[SLC]], i64 0, i64 1
|
|
|
|
|
// PGOUSE: br {{.*}} !prof ![[SL1:[0-9]+]]
|
|
|
|
|
for (i = 0; i < 100; ++i) {
|
|
|
|
|
}
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[SLC]], i64 0, i64 2
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[SL2:[0-9]+]]
|
|
|
|
|
while (i > 0)
|
|
|
|
|
i--;
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[SLC]], i64 0, i64 3
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[SL3:[0-9]+]]
|
|
|
|
|
do {} while (i++ < 75);
|
|
|
|
|
|
|
|
|
|
// PGOGEN-NOT: store {{.*}} @[[SLC]],
|
|
|
|
|
// PGOUSE-NOT: br {{.*}} !prof ![0-9]+
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// PGOGEN-LABEL: @conditionals()
|
|
|
|
|
// PGOUSE-LABEL: @conditionals()
|
|
|
|
|
// PGOGEN: store {{.*}} @[[IFC]], i64 0, i64 0
|
|
|
|
|
void conditionals() {
|
|
|
|
|
// PGOGEN: store {{.*}} @[[IFC]], i64 0, i64 1
|
|
|
|
|
// PGOUSE: br {{.*}} !prof ![[IF1:[0-9]+]]
|
|
|
|
|
for (int i = 0; i < 100; ++i) {
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[IFC]], i64 0, i64 2
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[IF2:[0-9]+]]
|
|
|
|
|
if (i % 2) {
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[IFC]], i64 0, i64 3
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[IF3:[0-9]+]]
|
|
|
|
|
if (i) {}
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[IFC]], i64 0, i64 4
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[IF4:[0-9]+]]
|
|
|
|
|
} else if (i % 3) {
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[IFC]], i64 0, i64 5
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[IF5:[0-9]+]]
|
|
|
|
|
if (i) {}
|
|
|
|
|
} else {
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[IFC]], i64 0, i64 6
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[IF6:[0-9]+]]
|
|
|
|
|
if (i) {}
|
|
|
|
|
}
|
|
|
|
|
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[IFC]], i64 0, i64 8
|
|
|
|
|
// PGOGEN: store {{.*}} @[[IFC]], i64 0, i64 7
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[IF7:[0-9]+]]
|
|
|
|
|
if (1 && i) {}
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[IFC]], i64 0, i64 10
|
|
|
|
|
// PGOGEN: store {{.*}} @[[IFC]], i64 0, i64 9
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[IF8:[0-9]+]]
|
|
|
|
|
if (0 || i) {}
|
|
|
|
|
}
|
|
|
|
|
|
2014-01-24 01:34:24 +08:00
|
|
|
// PGOGEN-NOT: store {{.*}} @[[IFC]],
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE-NOT: br {{.*}} !prof ![0-9]+
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// PGOGEN-LABEL: @early_exits()
|
|
|
|
|
// PGOUSE-LABEL: @early_exits()
|
|
|
|
|
// PGOGEN: store {{.*}} @[[EEC]], i64 0, i64 0
|
|
|
|
|
void early_exits() {
|
|
|
|
|
int i = 0;
|
|
|
|
|
|
|
|
|
|
// PGOGEN: store {{.*}} @[[EEC]], i64 0, i64 1
|
|
|
|
|
// PGOUSE: br {{.*}} !prof ![[EE1:[0-9]+]]
|
|
|
|
|
if (i) {}
|
|
|
|
|
|
|
|
|
|
// PGOGEN: store {{.*}} @[[EEC]], i64 0, i64 2
|
|
|
|
|
// PGOUSE: br {{.*}} !prof ![[EE2:[0-9]+]]
|
|
|
|
|
while (i < 100) {
|
|
|
|
|
i++;
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[EEC]], i64 0, i64 3
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[EE3:[0-9]+]]
|
|
|
|
|
if (i > 50)
|
|
|
|
|
break;
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[EEC]], i64 0, i64 4
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[EE4:[0-9]+]]
|
|
|
|
|
if (i % 2)
|
|
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[EEC]], i64 0, i64 5
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[EE5:[0-9]+]]
|
|
|
|
|
if (i) {}
|
|
|
|
|
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[EEC]], i64 0, i64 6
|
2014-01-07 06:27:43 +08:00
|
|
|
do {
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[EEC]], i64 0, i64 7
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[EE6:[0-9]+]]
|
|
|
|
|
if (i > 75)
|
|
|
|
|
return;
|
|
|
|
|
else
|
|
|
|
|
i++;
|
|
|
|
|
// PGOUSE: br {{.*}} !prof ![[EE7:[0-9]+]]
|
|
|
|
|
} while (i < 100);
|
|
|
|
|
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[EEC]], i64 0, i64 8
|
2014-01-07 06:27:43 +08:00
|
|
|
// Never reached -> no weights
|
|
|
|
|
if (i) {}
|
|
|
|
|
|
|
|
|
|
// PGOGEN-NOT: store {{.*}} @[[EEC]],
|
|
|
|
|
// PGOUSE-NOT: br {{.*}} !prof ![0-9]+
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// PGOGEN-LABEL: @jumps()
|
|
|
|
|
// PGOUSE-LABEL: @jumps()
|
|
|
|
|
// PGOGEN: store {{.*}} @[[JMC]], i64 0, i64 0
|
|
|
|
|
void jumps() {
|
|
|
|
|
int i;
|
|
|
|
|
|
|
|
|
|
// PGOGEN: store {{.*}} @[[JMC]], i64 0, i64 1
|
|
|
|
|
// PGOUSE: br {{.*}} !prof ![[JM1:[0-9]+]]
|
|
|
|
|
for (i = 0; i < 2; ++i) {
|
|
|
|
|
goto outofloop;
|
|
|
|
|
// Never reached -> no weights
|
|
|
|
|
if (i) {}
|
|
|
|
|
}
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[JMC]], i64 0, i64 3
|
2014-01-07 06:27:43 +08:00
|
|
|
outofloop:
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[JMC]], i64 0, i64 4
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[JM2:[0-9]+]]
|
|
|
|
|
if (i) {}
|
|
|
|
|
|
|
|
|
|
goto loop1;
|
|
|
|
|
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[JMC]], i64 0, i64 5
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[JM3:[0-9]+]]
|
|
|
|
|
while (i) {
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[JMC]], i64 0, i64 6
|
2014-01-07 06:27:43 +08:00
|
|
|
loop1:
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[JMC]], i64 0, i64 7
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[JM4:[0-9]+]]
|
|
|
|
|
if (i) {}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
goto loop2;
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[JMC]], i64 0, i64 8
|
2014-01-07 06:27:43 +08:00
|
|
|
first:
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[JMC]], i64 0, i64 9
|
2014-01-07 06:27:43 +08:00
|
|
|
second:
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[JMC]], i64 0, i64 10
|
2014-01-07 06:27:43 +08:00
|
|
|
third:
|
|
|
|
|
i++;
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[JMC]], i64 0, i64 11
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[JM5:[0-9]+]]
|
|
|
|
|
if (i < 3)
|
|
|
|
|
goto loop2;
|
|
|
|
|
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[JMC]], i64 0, i64 12
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[JM6:[0-9]+]]
|
|
|
|
|
while (i < 3) {
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[JMC]], i64 0, i64 13
|
2014-01-07 06:27:43 +08:00
|
|
|
loop2:
|
|
|
|
|
// PGOUSE: switch {{.*}} [
|
|
|
|
|
// PGOUSE: ], !prof ![[JM7:[0-9]+]]
|
|
|
|
|
switch (i) {
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[JMC]], i64 0, i64 15
|
2014-01-07 06:27:43 +08:00
|
|
|
case 0:
|
|
|
|
|
goto first;
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[JMC]], i64 0, i64 16
|
2014-01-07 06:27:43 +08:00
|
|
|
case 1:
|
|
|
|
|
goto second;
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[JMC]], i64 0, i64 17
|
2014-01-07 06:27:43 +08:00
|
|
|
case 2:
|
|
|
|
|
goto third;
|
|
|
|
|
}
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[JMC]], i64 0, i64 14
|
2014-01-07 06:27:43 +08:00
|
|
|
}
|
|
|
|
|
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[JMC]], i64 0, i64 18
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[JM8:[0-9]+]]
|
|
|
|
|
for (i = 0; i < 10; ++i) {
|
|
|
|
|
goto withinloop;
|
|
|
|
|
// never reached -> no weights
|
|
|
|
|
if (i) {}
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[JMC]], i64 0, i64 20
|
2014-01-07 06:27:43 +08:00
|
|
|
withinloop:
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[JMC]], i64 0, i64 21
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[JM9:[0-9]+]]
|
|
|
|
|
if (i) {}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// PGOGEN-NOT: store {{.*}} @[[JMC]],
|
|
|
|
|
// PGOUSE-NOT: br {{.*}} !prof ![0-9]+
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// PGOGEN-LABEL: @switches()
|
|
|
|
|
// PGOUSE-LABEL: @switches()
|
|
|
|
|
// PGOGEN: store {{.*}} @[[SWC]], i64 0, i64 0
|
|
|
|
|
void switches() {
|
|
|
|
|
static int weights[] = {1, 2, 2, 3, 3, 3, 4, 4, 4, 4, 5, 5, 5, 5, 5};
|
|
|
|
|
|
|
|
|
|
// No cases -> no weights
|
|
|
|
|
switch (weights[0]) {
|
|
|
|
|
// PGOGEN: store {{.*}} @[[SWC]], i64 0, i64 2
|
|
|
|
|
default:
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
// PGOGEN: store {{.*}} @[[SWC]], i64 0, i64 1
|
|
|
|
|
|
|
|
|
|
// PGOGEN: store {{.*}} @[[SWC]], i64 0, i64 3
|
|
|
|
|
// PGOUSE: br {{.*}} !prof ![[SW1:[0-9]+]]
|
|
|
|
|
for (int i = 0, len = sizeof(weights) / sizeof(weights[0]); i < len; ++i) {
|
|
|
|
|
// PGOUSE: switch {{.*}} [
|
|
|
|
|
// PGOUSE: ], !prof ![[SW2:[0-9]+]]
|
|
|
|
|
switch (i[weights]) {
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[SWC]], i64 0, i64 5
|
2014-01-07 06:27:43 +08:00
|
|
|
case 1:
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[SWC]], i64 0, i64 6
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[SW3:[0-9]+]]
|
|
|
|
|
if (i) {}
|
|
|
|
|
// fallthrough
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[SWC]], i64 0, i64 7
|
2014-01-07 06:27:43 +08:00
|
|
|
case 2:
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[SWC]], i64 0, i64 8
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[SW4:[0-9]+]]
|
|
|
|
|
if (i) {}
|
|
|
|
|
break;
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[SWC]], i64 0, i64 9
|
2014-01-07 06:27:43 +08:00
|
|
|
case 3:
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[SWC]], i64 0, i64 10
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[SW5:[0-9]+]]
|
|
|
|
|
if (i) {}
|
|
|
|
|
continue;
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[SWC]], i64 0, i64 11
|
2014-01-07 06:27:43 +08:00
|
|
|
case 4:
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[SWC]], i64 0, i64 12
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[SW6:[0-9]+]]
|
|
|
|
|
if (i) {}
|
|
|
|
|
// PGOUSE: switch {{.*}} [
|
|
|
|
|
// PGOUSE: ], !prof ![[SW7:[0-9]+]]
|
|
|
|
|
switch (i) {
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[SWC]], i64 0, i64 14
|
2014-01-07 06:27:43 +08:00
|
|
|
case 6 ... 9:
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[SWC]], i64 0, i64 15
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[SW8:[0-9]+]]
|
|
|
|
|
if (i) {}
|
|
|
|
|
continue;
|
|
|
|
|
}
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[SWC]], i64 0, i64 13
|
2014-01-07 06:27:43 +08:00
|
|
|
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[SWC]], i64 0, i64 16
|
2014-01-07 06:27:43 +08:00
|
|
|
default:
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[SWC]], i64 0, i64 17
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[SW9:[0-9]+]]
|
|
|
|
|
if (i == len - 1)
|
|
|
|
|
return;
|
|
|
|
|
}
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[SWC]], i64 0, i64 4
|
2014-01-07 06:27:43 +08:00
|
|
|
}
|
|
|
|
|
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[SWC]], i64 0, i64 18
|
2014-01-07 06:27:43 +08:00
|
|
|
// Never reached -> no weights
|
|
|
|
|
if (weights[0]) {}
|
|
|
|
|
|
|
|
|
|
// PGOGEN-NOT: store {{.*}} @[[SWC]],
|
|
|
|
|
// PGOUSE-NOT: br {{.*}} !prof ![0-9]+
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// PGOGEN-LABEL: @big_switch()
|
|
|
|
|
// PGOUSE-LABEL: @big_switch()
|
|
|
|
|
// PGOGEN: store {{.*}} @[[BSC]], i64 0, i64 0
|
|
|
|
|
void big_switch() {
|
|
|
|
|
// PGOGEN: store {{.*}} @[[BSC]], i64 0, i64 1
|
|
|
|
|
// PGOUSE: br {{.*}} !prof ![[BS1:[0-9]+]]
|
|
|
|
|
for (int i = 0; i < 32; ++i) {
|
|
|
|
|
// PGOUSE: switch {{.*}} [
|
|
|
|
|
// PGOUSE: ], !prof ![[BS2:[0-9]+]]
|
|
|
|
|
switch (1 << i) {
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[BSC]], i64 0, i64 3
|
2014-01-07 06:27:43 +08:00
|
|
|
case (1 << 0):
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[BSC]], i64 0, i64 4
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[BS3:[0-9]+]]
|
|
|
|
|
if (i) {}
|
|
|
|
|
// fallthrough
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[BSC]], i64 0, i64 5
|
2014-01-07 06:27:43 +08:00
|
|
|
case (1 << 1):
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[BSC]], i64 0, i64 6
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[BS4:[0-9]+]]
|
|
|
|
|
if (i) {}
|
|
|
|
|
break;
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[BSC]], i64 0, i64 7
|
2014-01-07 06:27:43 +08:00
|
|
|
case (1 << 2) ... (1 << 12):
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[BSC]], i64 0, i64 8
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[BS5:[0-9]+]]
|
|
|
|
|
if (i) {}
|
|
|
|
|
break;
|
|
|
|
|
// The branch for the large case range above appears after the case body
|
|
|
|
|
// PGOUSE: br {{.*}} !prof ![[BS6:[0-9]+]]
|
|
|
|
|
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[BSC]], i64 0, i64 9
|
2014-01-07 06:27:43 +08:00
|
|
|
case (1 << 13):
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[BSC]], i64 0, i64 10
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[BS7:[0-9]+]]
|
|
|
|
|
if (i) {}
|
|
|
|
|
break;
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[BSC]], i64 0, i64 11
|
2014-01-07 06:27:43 +08:00
|
|
|
case (1 << 14) ... (1 << 28):
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[BSC]], i64 0, i64 12
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[BS8:[0-9]+]]
|
|
|
|
|
if (i) {}
|
|
|
|
|
break;
|
|
|
|
|
// The branch for the large case range above appears after the case body
|
|
|
|
|
// PGOUSE: br {{.*}} !prof ![[BS9:[0-9]+]]
|
|
|
|
|
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[BSC]], i64 0, i64 13
|
2014-01-07 06:27:43 +08:00
|
|
|
case (1 << 29) ... ((1 << 29) + 1):
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[BSC]], i64 0, i64 14
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[BS10:[0-9]+]]
|
|
|
|
|
if (i) {}
|
|
|
|
|
break;
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[BSC]], i64 0, i64 15
|
2014-01-07 06:27:43 +08:00
|
|
|
default:
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[BSC]], i64 0, i64 16
|
2014-01-07 06:27:43 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[BS11:[0-9]+]]
|
|
|
|
|
if (i) {}
|
|
|
|
|
break;
|
|
|
|
|
}
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[BSC]], i64 0, i64 2
|
2014-01-07 06:27:43 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// PGOGEN-NOT: store {{.*}} @[[BSC]],
|
|
|
|
|
// PGOUSE-NOT: br {{.*}} !prof ![0-9]+
|
|
|
|
|
// PGOUSE: ret void
|
|
|
|
|
}
|
|
|
|
|
|
2014-01-23 10:54:30 +08:00
|
|
|
// PGOGEN-LABEL: @boolean_operators()
|
|
|
|
|
// PGOUSE-LABEL: @boolean_operators()
|
|
|
|
|
// PGOGEN: store {{.*}} @[[BOC]], i64 0, i64 0
|
|
|
|
|
void boolean_operators() {
|
|
|
|
|
int v;
|
|
|
|
|
// PGOGEN: store {{.*}} @[[BOC]], i64 0, i64 1
|
|
|
|
|
// PGOUSE: br {{.*}} !prof ![[BO1:[0-9]+]]
|
|
|
|
|
for (int i = 0; i < 100; ++i) {
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[BOC]], i64 0, i64 2
|
2014-01-23 10:54:30 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[BO2:[0-9]+]]
|
|
|
|
|
v = i % 3 || i;
|
|
|
|
|
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[BOC]], i64 0, i64 3
|
2014-01-23 10:54:30 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[BO3:[0-9]+]]
|
|
|
|
|
v = i % 3 && i;
|
|
|
|
|
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[BOC]], i64 0, i64 5
|
|
|
|
|
// PGOGEN: store {{.*}} @[[BOC]], i64 0, i64 4
|
2014-01-23 10:54:30 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[BO4:[0-9]+]]
|
|
|
|
|
// PGOUSE: br {{.*}} !prof ![[BO5:[0-9]+]]
|
|
|
|
|
v = i % 3 || i % 2 || i;
|
|
|
|
|
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[BOC]], i64 0, i64 7
|
|
|
|
|
// PGOGEN: store {{.*}} @[[BOC]], i64 0, i64 6
|
2014-01-23 10:54:30 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[BO6:[0-9]+]]
|
|
|
|
|
// PGOUSE: br {{.*}} !prof ![[BO7:[0-9]+]]
|
|
|
|
|
v = i % 2 && i % 3 && i;
|
|
|
|
|
}
|
|
|
|
|
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
// PGOGEN-NOT: store {{.*}} @[[BOC]],
|
|
|
|
|
// PGOUSE-NOT: br {{.*}} !prof ![0-9]+
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// PGOGEN-LABEL: @boolop_loops()
|
|
|
|
|
// PGOUSE-LABEL: @boolop_loops()
|
|
|
|
|
// PGOGEN: store {{.*}} @[[BLC]], i64 0, i64 0
|
|
|
|
|
void boolop_loops() {
|
|
|
|
|
int i = 100;
|
|
|
|
|
|
|
|
|
|
// PGOGEN: store {{.*}} @[[BLC]], i64 0, i64 2
|
|
|
|
|
// PGOGEN: store {{.*}} @[[BLC]], i64 0, i64 1
|
|
|
|
|
// PGOUSE: br {{.*}} !prof ![[BL1:[0-9]+]]
|
|
|
|
|
// PGOUSE: br {{.*}} !prof ![[BL2:[0-9]+]]
|
|
|
|
|
while (i && i > 50)
|
|
|
|
|
i--;
|
|
|
|
|
|
|
|
|
|
// PGOGEN: store {{.*}} @[[BLC]], i64 0, i64 4
|
|
|
|
|
// PGOGEN: store {{.*}} @[[BLC]], i64 0, i64 3
|
|
|
|
|
// PGOUSE: br {{.*}} !prof ![[BL3:[0-9]+]]
|
|
|
|
|
// PGOUSE: br {{.*}} !prof ![[BL4:[0-9]+]]
|
|
|
|
|
while ((i % 2) || (i > 0))
|
|
|
|
|
i--;
|
|
|
|
|
|
|
|
|
|
// PGOGEN: store {{.*}} @[[BLC]], i64 0, i64 6
|
|
|
|
|
// PGOGEN: store {{.*}} @[[BLC]], i64 0, i64 5
|
|
|
|
|
// PGOUSE: br {{.*}} !prof ![[BL5:[0-9]+]]
|
|
|
|
|
// PGOUSE: br {{.*}} !prof ![[BL6:[0-9]+]]
|
|
|
|
|
for (i = 100; i && i > 50; --i);
|
|
|
|
|
|
|
|
|
|
// PGOGEN: store {{.*}} @[[BLC]], i64 0, i64 8
|
|
|
|
|
// PGOGEN: store {{.*}} @[[BLC]], i64 0, i64 7
|
|
|
|
|
// PGOUSE: br {{.*}} !prof ![[BL7:[0-9]+]]
|
|
|
|
|
// PGOUSE: br {{.*}} !prof ![[BL8:[0-9]+]]
|
|
|
|
|
for (; (i % 2) || (i > 0); --i);
|
|
|
|
|
|
|
|
|
|
// PGOGEN-NOT: store {{.*}} @[[BLC]],
|
2014-01-23 10:54:30 +08:00
|
|
|
// PGOUSE-NOT: br {{.*}} !prof ![0-9]+
|
|
|
|
|
}
|
|
|
|
|
|
2014-04-11 14:10:10 +08:00
|
|
|
// PGOGEN-LABEL: @conditional_operator()
|
|
|
|
|
// PGOUSE-LABEL: @conditional_operator()
|
|
|
|
|
// PGOGEN: store {{.*}} @[[COC]], i64 0, i64 0
|
|
|
|
|
void conditional_operator() {
|
|
|
|
|
int i = 100;
|
|
|
|
|
|
|
|
|
|
// PGOGEN: store {{.*}} @[[COC]], i64 0, i64 1
|
|
|
|
|
// PGOUSE: br {{.*}} !prof ![[CO1:[0-9]+]]
|
|
|
|
|
int j = i < 50 ? i : 1;
|
|
|
|
|
|
|
|
|
|
// PGOGEN: store {{.*}} @[[COC]], i64 0, i64 2
|
|
|
|
|
// PGOUSE: br {{.*}} !prof ![[CO2:[0-9]+]]
|
|
|
|
|
int k = i ?: 0;
|
|
|
|
|
|
|
|
|
|
// PGOGEN-NOT: store {{.*}} @[[COC]],
|
|
|
|
|
// PGOUSE-NOT: br {{.*}} !prof ![0-9]+
|
|
|
|
|
}
|
|
|
|
|
|
2015-05-02 06:52:48 +08:00
|
|
|
// PGOGEN-LABEL: @do_fallthrough()
|
|
|
|
|
// PGOUSE-LABEL: @do_fallthrough()
|
|
|
|
|
// PGOGEN: store {{.*}} @[[DFC]], i64 0, i64 0
|
2014-02-28 13:57:14 +08:00
|
|
|
void do_fallthrough() {
|
2015-05-02 06:52:48 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[DFC]], i64 0, i64 1
|
|
|
|
|
// PGOUSE: br {{.*}} !prof ![[DF1:[0-9]+]]
|
2014-02-28 13:57:14 +08:00
|
|
|
for (int i = 0; i < 10; ++i) {
|
|
|
|
|
int j = 0;
|
2015-05-02 06:52:48 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[DFC]], i64 0, i64 2
|
2014-02-28 13:57:14 +08:00
|
|
|
do {
|
|
|
|
|
// The number of exits out of this do-loop via the break statement
|
|
|
|
|
// exceeds the counter value for the loop (which does not include the
|
|
|
|
|
// fallthrough count). Make sure that does not violate any assertions.
|
2015-05-02 06:52:48 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[DFC]], i64 0, i64 3
|
|
|
|
|
// PGOUSE: br {{.*}} !prof ![[DF3:[0-9]+]]
|
2014-02-28 13:57:14 +08:00
|
|
|
if (i < 8) break;
|
|
|
|
|
j++;
|
2015-05-02 06:52:48 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[DF2:[0-9]+]]
|
2014-02-28 13:57:14 +08:00
|
|
|
} while (j < 2);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2014-03-06 12:55:41 +08:00
|
|
|
// PGOGEN-LABEL: @static_func()
|
|
|
|
|
// PGOUSE-LABEL: @static_func()
|
2014-03-07 05:35:59 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[STC]], i64 0, i64 0
|
2014-03-06 12:55:41 +08:00
|
|
|
static void static_func() {
|
2014-03-07 05:35:59 +08:00
|
|
|
// PGOGEN: store {{.*}} @[[STC]], i64 0, i64 1
|
2014-03-06 12:55:41 +08:00
|
|
|
// PGOUSE: br {{.*}} !prof ![[ST1:[0-9]+]]
|
|
|
|
|
for (int i = 0; i < 10; ++i) {
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2014-12-16 03:10:08 +08:00
|
|
|
// PGOUSE-DAG: ![[SL1]] = !{!"branch_weights", i32 101, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[SL2]] = !{!"branch_weights", i32 101, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[SL3]] = !{!"branch_weights", i32 76, i32 2}
|
|
|
|
|
|
|
|
|
|
// PGOUSE-DAG: ![[EE1]] = !{!"branch_weights", i32 1, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[EE2]] = !{!"branch_weights", i32 52, i32 1}
|
|
|
|
|
// PGOUSE-DAG: ![[EE3]] = !{!"branch_weights", i32 2, i32 51}
|
|
|
|
|
// PGOUSE-DAG: ![[EE4]] = !{!"branch_weights", i32 26, i32 26}
|
|
|
|
|
// PGOUSE-DAG: ![[EE5]] = !{!"branch_weights", i32 2, i32 1}
|
|
|
|
|
// PGOUSE-DAG: ![[EE6]] = !{!"branch_weights", i32 2, i32 26}
|
|
|
|
|
// PGOUSE-DAG: ![[EE7]] = !{!"branch_weights", i32 26, i32 1}
|
|
|
|
|
|
|
|
|
|
// PGOUSE-DAG: ![[IF1]] = !{!"branch_weights", i32 101, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[IF2]] = !{!"branch_weights", i32 51, i32 51}
|
|
|
|
|
// PGOUSE-DAG: ![[IF3]] = !{!"branch_weights", i32 51, i32 1}
|
|
|
|
|
// PGOUSE-DAG: ![[IF4]] = !{!"branch_weights", i32 34, i32 18}
|
|
|
|
|
// PGOUSE-DAG: ![[IF5]] = !{!"branch_weights", i32 34, i32 1}
|
|
|
|
|
// PGOUSE-DAG: ![[IF6]] = !{!"branch_weights", i32 17, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[IF7]] = !{!"branch_weights", i32 100, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[IF8]] = !{!"branch_weights", i32 100, i32 2}
|
|
|
|
|
|
|
|
|
|
// PGOUSE-DAG: ![[JM1]] = !{!"branch_weights", i32 2, i32 1}
|
|
|
|
|
// PGOUSE-DAG: ![[JM2]] = !{!"branch_weights", i32 1, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[JM3]] = !{!"branch_weights", i32 1, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[JM4]] = !{!"branch_weights", i32 1, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[JM5]] = !{!"branch_weights", i32 3, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[JM6]] = !{!"branch_weights", i32 1, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[JM7]] = !{!"branch_weights", i32 1, i32 2, i32 2, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[JM8]] = !{!"branch_weights", i32 11, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[JM9]] = !{!"branch_weights", i32 10, i32 2}
|
|
|
|
|
|
|
|
|
|
// PGOUSE-DAG: ![[SW1]] = !{!"branch_weights", i32 16, i32 1}
|
|
|
|
|
// PGOUSE-DAG: ![[SW2]] = !{!"branch_weights", i32 6, i32 2, i32 3, i32 4, i32 5}
|
|
|
|
|
// PGOUSE-DAG: ![[SW3]] = !{!"branch_weights", i32 1, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[SW4]] = !{!"branch_weights", i32 3, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[SW5]] = !{!"branch_weights", i32 4, i32 1}
|
|
|
|
|
// PGOUSE-DAG: ![[SW6]] = !{!"branch_weights", i32 5, i32 1}
|
|
|
|
|
// PGOUSE-DAG: ![[SW7]] = !{!"branch_weights", i32 1, i32 2, i32 2, i32 2, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[SW8]] = !{!"branch_weights", i32 5, i32 1}
|
|
|
|
|
// PGOUSE-DAG: ![[SW9]] = !{!"branch_weights", i32 2, i32 5}
|
|
|
|
|
|
|
|
|
|
// PGOUSE-DAG: ![[BS1]] = !{!"branch_weights", i32 33, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[BS2]] = !{!"branch_weights", i32 29, i32 2, i32 2, i32 2, i32 2, i32 1}
|
|
|
|
|
// PGOUSE-DAG: ![[BS3]] = !{!"branch_weights", i32 1, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[BS4]] = !{!"branch_weights", i32 2, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[BS5]] = !{!"branch_weights", i32 12, i32 1}
|
|
|
|
|
// PGOUSE-DAG: ![[BS6]] = !{!"branch_weights", i32 12, i32 3}
|
|
|
|
|
// PGOUSE-DAG: ![[BS7]] = !{!"branch_weights", i32 2, i32 1}
|
|
|
|
|
// PGOUSE-DAG: ![[BS8]] = !{!"branch_weights", i32 16, i32 1}
|
|
|
|
|
// PGOUSE-DAG: ![[BS9]] = !{!"branch_weights", i32 16, i32 14}
|
|
|
|
|
// PGOUSE-DAG: ![[BS10]] = !{!"branch_weights", i32 2, i32 1}
|
|
|
|
|
// PGOUSE-DAG: ![[BS11]] = !{!"branch_weights", i32 3, i32 1}
|
|
|
|
|
|
|
|
|
|
// PGOUSE-DAG: ![[BO1]] = !{!"branch_weights", i32 101, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[BO2]] = !{!"branch_weights", i32 67, i32 35}
|
|
|
|
|
// PGOUSE-DAG: ![[BO3]] = !{!"branch_weights", i32 67, i32 35}
|
|
|
|
|
// PGOUSE-DAG: ![[BO4]] = !{!"branch_weights", i32 67, i32 35}
|
|
|
|
|
// PGOUSE-DAG: ![[BO5]] = !{!"branch_weights", i32 18, i32 18}
|
|
|
|
|
// PGOUSE-DAG: ![[BO6]] = !{!"branch_weights", i32 51, i32 51}
|
|
|
|
|
// PGOUSE-DAG: ![[BO7]] = !{!"branch_weights", i32 34, i32 18}
|
|
|
|
|
// PGOUSE-DAG: ![[BL1]] = !{!"branch_weights", i32 52, i32 1}
|
|
|
|
|
// PGOUSE-DAG: ![[BL2]] = !{!"branch_weights", i32 51, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[BL3]] = !{!"branch_weights", i32 26, i32 27}
|
|
|
|
|
// PGOUSE-DAG: ![[BL4]] = !{!"branch_weights", i32 51, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[BL5]] = !{!"branch_weights", i32 52, i32 1}
|
|
|
|
|
// PGOUSE-DAG: ![[BL6]] = !{!"branch_weights", i32 51, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[BL7]] = !{!"branch_weights", i32 26, i32 27}
|
|
|
|
|
// PGOUSE-DAG: ![[BL8]] = !{!"branch_weights", i32 51, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[CO1]] = !{!"branch_weights", i32 1, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[CO2]] = !{!"branch_weights", i32 2, i32 1}
|
2015-05-02 06:52:48 +08:00
|
|
|
|
|
|
|
|
// PGOUSE-DAG: ![[DF1]] = !{!"branch_weights", i32 11, i32 2}
|
|
|
|
|
// PGOUSE-DAG: ![[DF2]] = !{!"branch_weights", i32 3, i32 3}
|
|
|
|
|
// PGOUSE-DAG: ![[DF3]] = !{!"branch_weights", i32 9, i32 5}
|
|
|
|
|
|
2014-12-16 03:10:08 +08:00
|
|
|
// PGOUSE-DAG: ![[ST1]] = !{!"branch_weights", i32 11, i32 2}
|
2014-01-23 10:54:30 +08:00
|
|
|
|
2014-01-07 06:27:43 +08:00
|
|
|
int main(int argc, const char *argv[]) {
|
|
|
|
|
simple_loops();
|
|
|
|
|
conditionals();
|
|
|
|
|
early_exits();
|
|
|
|
|
jumps();
|
|
|
|
|
switches();
|
|
|
|
|
big_switch();
|
2014-01-23 10:54:30 +08:00
|
|
|
boolean_operators();
|
Change PGO instrumentation to compute counts in a separate AST traversal.
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
2014-02-18 03:21:09 +08:00
|
|
|
boolop_loops();
|
2014-04-11 14:10:10 +08:00
|
|
|
conditional_operator();
|
2014-02-28 13:57:14 +08:00
|
|
|
do_fallthrough();
|
2014-03-06 12:55:41 +08:00
|
|
|
static_func();
|
2014-01-07 06:27:43 +08:00
|
|
|
return 0;
|
|
|
|
|
}
|