maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/clang/test/OpenMP/target_map_codegen.cpp

4557 lines
248 KiB
C++
Raw Normal View History

[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// expected-no-diagnostics
#ifndef HEADER
#define HEADER
///
/// Implicit maps.
///
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK1 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK1 --check-prefix CK1-64
// RUN: %clang_cc1 -DCK1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK1 --check-prefix CK1-64
// RUN: %clang_cc1 -DCK1 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK1 --check-prefix CK1-32
// RUN: %clang_cc1 -DCK1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK1 --check-prefix CK1-32
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
#ifdef CK1
[OPENMP] Fix for PR34014: OpenMP 4.5: Target construct in static method of class fails to map class static variable. If the global variable is captured and it has several redeclarations, sometimes it may lead to a compiler crash. Patch fixes this by working only with canonical declarations. llvm-svn: 311479
2017-08-23 01:54:52 +08:00
class B {
public:
static double VAR;
B() {
}
static void modify(int &res) {
#pragma omp target map(tofrom \
: res)
{
res = B::VAR;
}
}
};
double B::VAR = 1.0;
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK1-DAG: [[SIZES:@.+]] = {{.+}}constant [1 x i[[sz:64|32]]] [i{{64|32}} 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// Map types: OMP_MAP_PRIVATE_VAL | OMP_MAP_IS_FIRST = 288
// CK1-DAG: [[TYPES:@.+]] = {{.+}}constant [1 x i32] [i32 288]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK1-LABEL: implicit_maps_integer
void implicit_maps_integer (int a){
[OPENMP] Fix for PR34014: OpenMP 4.5: Target construct in static method of class fails to map class static variable. If the global variable is captured and it has several redeclarations, sometimes it may lead to a compiler crash. Patch fixes this by working only with canonical declarations. llvm-svn: 311479
2017-08-23 01:54:52 +08:00
// CK1: call void{{.*}}modify
B::modify(a);
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
int i = a;
// CK1-DAG: call i32 @__tgt_target(i32 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}})
// CK1-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
// CK1-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
// CK1-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
// CK1-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK1-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz]]*
// CK1-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
// CK1-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
// CK1-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK1-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
// CK1-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to i32*
// CK1-64-DAG: store i32 {{.+}}, i32* [[CADDR]],
// CK1: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
#pragma omp target
{
++i;
}
}
// CK1: define internal void [[KERNEL]](i[[sz]] [[ARG:%.+]])
// CK1: [[ADDR:%.+]] = alloca i[[sz]],
// CK1: store i[[sz]] [[ARG]], i[[sz]]* [[ADDR]],
// CK1-64: [[CADDR:%.+]] = bitcast i64* [[ADDR]] to i32*
// CK1-64: {{.+}} = load i32, i32* [[CADDR]],
// CK1-32: {{.+}} = load i32, i32* [[ADDR]],
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK2 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK2 --check-prefix CK2-64
// RUN: %clang_cc1 -DCK2 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK2 --check-prefix CK2-64
// RUN: %clang_cc1 -DCK2 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK2 --check-prefix CK2-32
// RUN: %clang_cc1 -DCK2 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK2 --check-prefix CK2-32
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
#ifdef CK2
[OpenMP] Add support for mapping array sections through pointer references. Summary: This patch fixes a bug in the map of array sections whose base is a reference to a pointer. The existing mapping support was not prepared to deal with it, causing the compiler to crash. Mapping a reference to a pointer enjoys the same characteristics of a regular pointer, i.e., it is passed by value. Therefore, the reference has to be materialized in the target region. Reviewers: hfinkel, carlo.bertolli, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: https://reviews.llvm.org/D22690 llvm-svn: 276933
2016-07-28 06:49:49 +08:00
// CK2: [[SIZES:@.+]] = {{.+}}constant [1 x i[[sz:64|32]]] [i{{64|32}} 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// Map types: OMP_MAP_PRIVATE_VAL | OMP_MAP_IS_FIRST = 288
[OpenMP] Add support for mapping array sections through pointer references. Summary: This patch fixes a bug in the map of array sections whose base is a reference to a pointer. The existing mapping support was not prepared to deal with it, causing the compiler to crash. Mapping a reference to a pointer enjoys the same characteristics of a regular pointer, i.e., it is passed by value. Therefore, the reference has to be materialized in the target region. Reviewers: hfinkel, carlo.bertolli, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: https://reviews.llvm.org/D22690 llvm-svn: 276933
2016-07-28 06:49:49 +08:00
// CK2: [[TYPES:@.+]] = {{.+}}constant [1 x i32] [i32 288]
// CK2: [[SIZES2:@.+]] = {{.+}}constant [1 x i[[sz]]] zeroinitializer
// Map types: OMP_MAP_IS_PTR = 32
// CK2: [[TYPES2:@.+]] = {{.+}}constant [1 x i32] [i32 32]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
[OpenMP] Add support for mapping array sections through pointer references. Summary: This patch fixes a bug in the map of array sections whose base is a reference to a pointer. The existing mapping support was not prepared to deal with it, causing the compiler to crash. Mapping a reference to a pointer enjoys the same characteristics of a regular pointer, i.e., it is passed by value. Therefore, the reference has to be materialized in the target region. Reviewers: hfinkel, carlo.bertolli, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: https://reviews.llvm.org/D22690 llvm-svn: 276933
2016-07-28 06:49:49 +08:00
// CK2-LABEL: implicit_maps_reference
void implicit_maps_reference (int a, int *b){
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
int &i = a;
// CK2-DAG: call i32 @__tgt_target(i32 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}})
// CK2-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
// CK2-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
// CK2-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
// CK2-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK2-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz]]*
// CK2-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
// CK2-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
// CK2-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK2-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
// CK2-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to i32*
// CK2-64-DAG: store i32 {{.+}}, i32* [[CADDR]],
// CK2: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
#pragma omp target
{
++i;
}
[OpenMP] Add support for mapping array sections through pointer references. Summary: This patch fixes a bug in the map of array sections whose base is a reference to a pointer. The existing mapping support was not prepared to deal with it, causing the compiler to crash. Mapping a reference to a pointer enjoys the same characteristics of a regular pointer, i.e., it is passed by value. Therefore, the reference has to be materialized in the target region. Reviewers: hfinkel, carlo.bertolli, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: https://reviews.llvm.org/D22690 llvm-svn: 276933
2016-07-28 06:49:49 +08:00
int *&p = b;
// CK2-DAG: call i32 @__tgt_target(i32 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES2]]{{.+}}, {{.+}}[[TYPES2]]{{.+}})
// CK2-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
// CK2-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
// CK2-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
// CK2-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK2-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
// CK2-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
// CK2-DAG: store i32* [[VAL:%[^,]+]], i32** [[CBP1]]
// CK2-DAG: store i32* [[VAL]], i32** [[CP1]]
[OpenMP] Add support for mapping array sections through pointer references. Summary: This patch fixes a bug in the map of array sections whose base is a reference to a pointer. The existing mapping support was not prepared to deal with it, causing the compiler to crash. Mapping a reference to a pointer enjoys the same characteristics of a regular pointer, i.e., it is passed by value. Therefore, the reference has to be materialized in the target region. Reviewers: hfinkel, carlo.bertolli, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: https://reviews.llvm.org/D22690 llvm-svn: 276933
2016-07-28 06:49:49 +08:00
// CK2-DAG: [[VAL]] = load i32*, i32** [[ADDR:%.+]],
// CK2-DAG: [[ADDR]] = load i32**, i32*** [[ADDR2:%.+]],
// CK2: call void [[KERNEL2:@.+]](i32* [[VAL]])
#pragma omp target
{
++p;
}
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
}
// CK2: define internal void [[KERNEL]](i[[sz]] [[ARG:%.+]])
// CK2: [[ADDR:%.+]] = alloca i[[sz]],
// CK2: [[REF:%.+]] = alloca i32*,
// CK2: store i[[sz]] [[ARG]], i[[sz]]* [[ADDR]],
// CK2-64: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to i32*
// CK2-64: store i32* [[CADDR]], i32** [[REF]],
// CK2-64: [[RVAL:%.+]] = load i32*, i32** [[REF]],
// CK2-64: {{.+}} = load i32, i32* [[RVAL]],
// CK2-32: store i32* [[ADDR]], i32** [[REF]],
// CK2-32: [[RVAL:%.+]] = load i32*, i32** [[REF]],
// CK2-32: {{.+}} = load i32, i32* [[RVAL]],
[OpenMP] Add support for mapping array sections through pointer references. Summary: This patch fixes a bug in the map of array sections whose base is a reference to a pointer. The existing mapping support was not prepared to deal with it, causing the compiler to crash. Mapping a reference to a pointer enjoys the same characteristics of a regular pointer, i.e., it is passed by value. Therefore, the reference has to be materialized in the target region. Reviewers: hfinkel, carlo.bertolli, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: https://reviews.llvm.org/D22690 llvm-svn: 276933
2016-07-28 06:49:49 +08:00
// CK2: define internal void [[KERNEL2]](i32* [[ARG:%.+]])
// CK2: [[ADDR:%.+]] = alloca i32*,
// CK2: [[REF:%.+]] = alloca i32**,
// CK2: store i32* [[ARG]], i32** [[ADDR]],
// CK2: store i32** [[ADDR]], i32*** [[REF]],
// CK2: [[T:%.+]] = load i32**, i32*** [[REF]],
// CK2: [[TT:%.+]] = load i32*, i32** [[T]],
// CK2: getelementptr inbounds i32, i32* [[TT]], i32 1
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK3 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK3 --check-prefix CK3-64
// RUN: %clang_cc1 -DCK3 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK3 --check-prefix CK3-64
// RUN: %clang_cc1 -DCK3 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK3 --check-prefix CK3-32
// RUN: %clang_cc1 -DCK3 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK3 --check-prefix CK3-32
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
#ifdef CK3
// CK3-DAG: [[SIZES:@.+]] = {{.+}}constant [1 x i[[sz:64|32]]] [i{{64|32}} 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// Map types: OMP_MAP_PRIVATE_VAL | OMP_MAP_IS_FIRST = 288
// CK3-DAG: [[TYPES:@.+]] = {{.+}}constant [1 x i32] [i32 288]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK3-LABEL: implicit_maps_parameter
void implicit_maps_parameter (int a){
// CK3-DAG: call i32 @__tgt_target(i32 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}})
// CK3-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
// CK3-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
// CK3-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
// CK3-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK3-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz]]*
// CK3-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
// CK3-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
// CK3-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK3-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
// CK3-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to i32*
// CK3-64-DAG: store i32 {{.+}}, i32* [[CADDR]],
// CK3: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
#pragma omp target
{
++a;
}
}
// CK3: define internal void [[KERNEL]](i[[sz]] [[ARG:%.+]])
// CK3: [[ADDR:%.+]] = alloca i[[sz]],
// CK3: store i[[sz]] [[ARG]], i[[sz]]* [[ADDR]],
// CK3-64: [[CADDR:%.+]] = bitcast i64* [[ADDR]] to i32*
// CK3-64: {{.+}} = load i32, i32* [[CADDR]],
// CK3-32: {{.+}} = load i32, i32* [[ADDR]],
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK4 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK4 --check-prefix CK4-64
// RUN: %clang_cc1 -DCK4 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK4 --check-prefix CK4-64
// RUN: %clang_cc1 -DCK4 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK4 --check-prefix CK4-32
// RUN: %clang_cc1 -DCK4 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK4 --check-prefix CK4-32
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
#ifdef CK4
// CK4-DAG: [[SIZES:@.+]] = {{.+}}constant [1 x i[[sz:64|32]]] [i{{64|32}} 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// Map types: OMP_MAP_PRIVATE_VAL | OMP_MAP_IS_FIRST = 288
// CK4-DAG: [[TYPES:@.+]] = {{.+}}constant [1 x i32] [i32 288]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK4-LABEL: implicit_maps_nested_integer
void implicit_maps_nested_integer (int a){
int i = a;
// The captures in parallel are by reference. Only the capture in target is by
// copy.
// CK4: call void {{.+}}@__kmpc_fork_call({{.+}} [[KERNELP1:@.+]] to void (i32*, i32*, ...)*), i32* {{.+}})
// CK4: define internal void [[KERNELP1]](i32* {{[^,]+}}, i32* {{[^,]+}}, i32* {{[^,]+}})
#pragma omp parallel
{
// CK4-DAG: call i32 @__tgt_target(i32 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}})
// CK4-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
// CK4-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
// CK4-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
// CK4-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK4-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz]]*
// CK4-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
// CK4-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
// CK4-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK4-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
// CK4-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to i32*
// CK4-64-DAG: store i32 {{.+}}, i32* [[CADDR]],
// CK4: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
#pragma omp target
{
#pragma omp parallel
{
++i;
}
}
}
}
// CK4: define internal void [[KERNEL]](i[[sz]] [[ARG:%.+]])
// CK4: [[ADDR:%.+]] = alloca i[[sz]],
// CK4: store i[[sz]] [[ARG]], i[[sz]]* [[ADDR]],
// CK4-64: [[CADDR:%.+]] = bitcast i64* [[ADDR]] to i32*
// CK4-64: call void {{.+}}@__kmpc_fork_call({{.+}} [[KERNELP2:@.+]] to void (i32*, i32*, ...)*), i32* [[CADDR]])
// CK4-32: call void {{.+}}@__kmpc_fork_call({{.+}} [[KERNELP2:@.+]] to void (i32*, i32*, ...)*), i32* [[ADDR]])
// CK4: define internal void [[KERNELP2]](i32* {{[^,]+}}, i32* {{[^,]+}}, i32* {{[^,]+}})
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK5 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK5 --check-prefix CK5-64
// RUN: %clang_cc1 -DCK5 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK5 --check-prefix CK5-64
// RUN: %clang_cc1 -DCK5 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK5 --check-prefix CK5-32
// RUN: %clang_cc1 -DCK5 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK5 --check-prefix CK5-32
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
#ifdef CK5
// CK5-DAG: [[SIZES:@.+]] = {{.+}}constant [1 x i[[sz:64|32]]] [i{{64|32}} 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// Map types: OMP_MAP_PRIVATE_VAL | OMP_MAP_IS_FIRST = 288
// CK5-DAG: [[TYPES:@.+]] = {{.+}}constant [1 x i32] [i32 288]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK5-LABEL: implicit_maps_nested_integer_and_enum
void implicit_maps_nested_integer_and_enum (int a){
enum Bla {
SomeEnum = 0x09
};
// Using an enum should not change the mapping information.
int i = a;
// CK5-DAG: call i32 @__tgt_target(i32 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}})
// CK5-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
// CK5-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
// CK5-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
// CK5-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK5-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz]]*
// CK5-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
// CK5-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
// CK5-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK5-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
// CK5-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to i32*
// CK5-64-DAG: store i32 {{.+}}, i32* [[CADDR]],
// CK5: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
#pragma omp target
{
++i;
i += SomeEnum;
}
}
// CK5: define internal void [[KERNEL]](i[[sz]] [[ARG:%.+]])
// CK5: [[ADDR:%.+]] = alloca i[[sz]],
// CK5: store i[[sz]] [[ARG]], i[[sz]]* [[ADDR]],
// CK5-64: [[CADDR:%.+]] = bitcast i64* [[ADDR]] to i32*
// CK5-64: {{.+}} = load i32, i32* [[CADDR]],
// CK5-32: {{.+}} = load i32, i32* [[ADDR]],
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK6 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK6 --check-prefix CK6-64
// RUN: %clang_cc1 -DCK6 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK6 --check-prefix CK6-64
// RUN: %clang_cc1 -DCK6 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK6 --check-prefix CK6-32
// RUN: %clang_cc1 -DCK6 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK6 --check-prefix CK6-32
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
#ifdef CK6
// CK6-DAG: [[GBL:@Gi]] = global i32 0
// CK6-DAG: [[SIZES:@.+]] = {{.+}}constant [1 x i[[sz:64|32]]] [i{{64|32}} 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// Map types: OMP_MAP_PRIVATE_VAL | OMP_MAP_IS_FIRST = 288
// CK6-DAG: [[TYPES:@.+]] = {{.+}}constant [1 x i32] [i32 288]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK6-LABEL: implicit_maps_host_global
int Gi;
void implicit_maps_host_global (int a){
// CK6-DAG: call i32 @__tgt_target(i32 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}})
// CK6-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
// CK6-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
// CK6-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
// CK6-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK6-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz]]*
// CK6-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
// CK6-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
// CK6-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK6-64-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
// CK6-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to i32*
// CK6-64-DAG: store i32 [[GBLVAL:%.+]], i32* [[CADDR]],
// CK6-64-DAG: [[GBLVAL]] = load i32, i32* [[GBL]],
// CK6-32-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[GBLVAL:%.+]],
// CK6: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
#pragma omp target
{
++Gi;
}
}
// CK6: define internal void [[KERNEL]](i[[sz]] [[ARG:%.+]])
// CK6: [[ADDR:%.+]] = alloca i[[sz]],
// CK6: store i[[sz]] [[ARG]], i[[sz]]* [[ADDR]],
// CK6-64: [[CADDR:%.+]] = bitcast i64* [[ADDR]] to i32*
// CK6-64: {{.+}} = load i32, i32* [[CADDR]],
// CK6-32: {{.+}} = load i32, i32* [[ADDR]],
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK7 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK7 --check-prefix CK7-64
// RUN: %clang_cc1 -DCK7 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK7 --check-prefix CK7-64
// RUN: %clang_cc1 -DCK7 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK7 --check-prefix CK7-32
// RUN: %clang_cc1 -DCK7 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK7 --check-prefix CK7-32
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
#ifdef CK7
// For a 32-bit targets, the value doesn't fit the size of the pointer,
// therefore it is passed by reference with a map 'to' specification.
// CK7-DAG: [[SIZES:@.+]] = {{.+}}constant [1 x i[[sz:64|32]]] [i{{64|32}} 8]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// Map types: OMP_MAP_PRIVATE_VAL | OMP_MAP_IS_FIRST = 288
// CK7-64-DAG: [[TYPES:@.+]] = {{.+}}constant [1 x i32] [i32 288]
[OPENMP] Fix mapping|privatization of implicitly captured variables. If the `defaultmap(tofrom:scalar)` clause is specified, the scalars must be mapped with 'tofrom' modifiers, otherwise they must be captured as firstprivates. llvm-svn: 314995
2017-10-06 01:51:39 +08:00
// Map types: OMP_MAP_TO | OMP_MAP_PRIVATE_PTR | OMP_MAP_FIRST_REF = 161
// CK7-32-DAG: [[TYPES:@.+]] = {{.+}}constant [1 x i32] [i32 161]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK7-LABEL: implicit_maps_double
void implicit_maps_double (int a){
double d = (double)a;
// CK7-DAG: call i32 @__tgt_target(i32 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}})
// CK7-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
// CK7-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
// CK7-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
// CK7-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK7-64-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz]]*
// CK7-64-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
// CK7-64-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
// CK7-64-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK7-64-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
// CK7-64-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to double*
// CK7-64-64-DAG: store double {{.+}}, double* [[CADDR]],
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK7-32-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to double**
// CK7-32-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to double**
// CK7-32-DAG: store double* [[DECL:%[^,]+]], double** [[CBP1]]
// CK7-32-DAG: store double* [[DECL]], double** [[CP1]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK7-64: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
// CK7-32: call void [[KERNEL:@.+]](double* [[DECL]])
#pragma omp target
{
d += 1.0;
}
}
// CK7-64: define internal void [[KERNEL]](i[[sz]] [[ARG:%.+]])
// CK7-64: [[ADDR:%.+]] = alloca i[[sz]],
// CK7-64: store i[[sz]] [[ARG]], i[[sz]]* [[ADDR]],
// CK7-64: [[CADDR:%.+]] = bitcast i64* [[ADDR]] to double*
// CK7-64: {{.+}} = load double, double* [[CADDR]],
// CK7-32: define internal void [[KERNEL]](double* {{.+}}[[ARG:%.+]])
// CK7-32: [[ADDR:%.+]] = alloca double*,
// CK7-32: store double* [[ARG]], double** [[ADDR]],
// CK7-32: [[REF:%.+]] = load double*, double** [[ADDR]],
// CK7-32: {{.+}} = load double, double* [[REF]],
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK8 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK8
// RUN: %clang_cc1 -DCK8 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK8
// RUN: %clang_cc1 -DCK8 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK8
// RUN: %clang_cc1 -DCK8 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK8
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
#ifdef CK8
// CK8-DAG: [[SIZES:@.+]] = {{.+}}constant [1 x i[[sz:64|32]]] [i{{64|32}} 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// Map types: OMP_MAP_PRIVATE_VAL | OMP_MAP_IS_FIRST = 288
// CK8-DAG: [[TYPES:@.+]] = {{.+}}constant [1 x i32] [i32 288]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK8-LABEL: implicit_maps_float
void implicit_maps_float (int a){
float f = (float)a;
// CK8-DAG: call i32 @__tgt_target(i32 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}})
// CK8-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
// CK8-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
// CK8-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
// CK8-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK8-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz]]*
// CK8-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
// CK8-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
// CK8-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK8-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
// CK8-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to float*
// CK8-DAG: store float {{.+}}, float* [[CADDR]],
// CK8: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
#pragma omp target
{
f += 1.0;
}
}
// CK8: define internal void [[KERNEL]](i[[sz]] [[ARG:%.+]])
// CK8: [[ADDR:%.+]] = alloca i[[sz]],
// CK8: store i[[sz]] [[ARG]], i[[sz]]* [[ADDR]],
// CK8: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to float*
// CK8: {{.+}} = load float, float* [[CADDR]],
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK9 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK9
// RUN: %clang_cc1 -DCK9 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK9
// RUN: %clang_cc1 -DCK9 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK9
// RUN: %clang_cc1 -DCK9 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK9
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
#ifdef CK9
// CK9-DAG: [[SIZES:@.+]] = {{.+}}constant [1 x i[[sz:64|32]]] [i{{64|32}} 16]
[OPENMP] Generate implicit map|firstprivate clauses for target-based directives. If the variable is used in the target-based region but is not found in any private|mapping clause, then generate implicit firstprivate|map clauses for these implicitly mapped variables. llvm-svn: 314205
2017-09-26 21:47:31 +08:00
// Map types: OMP_MAP_TO + OMP_MAP_FROM + OMP_MAP_IS_FIRST + OMP_MAP_IMPLICIT = 547
// CK9-DAG: [[TYPES:@.+]] = {{.+}}constant [1 x i32] [i32 547]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK9-LABEL: implicit_maps_array
void implicit_maps_array (int a){
double darr[2] = {(double)a, (double)a};
// CK9-DAG: call i32 @__tgt_target(i32 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}})
// CK9-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
// CK9-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
// CK9-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
// CK9-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK9-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [2 x double]**
// CK9-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [2 x double]**
// CK9-DAG: store [2 x double]* [[DECL:%[^,]+]], [2 x double]** [[CBP1]]
// CK9-DAG: store [2 x double]* [[DECL]], [2 x double]** [[CP1]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK9: call void [[KERNEL:@.+]]([2 x double]* [[DECL]])
#pragma omp target
{
darr[0] += 1.0;
darr[1] += 1.0;
}
}
// CK9: define internal void [[KERNEL]]([2 x double]* {{.+}}[[ARG:%.+]])
// CK9: [[ADDR:%.+]] = alloca [2 x double]*,
// CK9: store [2 x double]* [[ARG]], [2 x double]** [[ADDR]],
// CK9: [[REF:%.+]] = load [2 x double]*, [2 x double]** [[ADDR]],
// CK9: {{.+}} = getelementptr inbounds [2 x double], [2 x double]* [[REF]], i[[sz]] 0, i[[sz]] 0
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK10 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK10
// RUN: %clang_cc1 -DCK10 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK10
// RUN: %clang_cc1 -DCK10 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK10
// RUN: %clang_cc1 -DCK10 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK10
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
#ifdef CK10
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK10-DAG: [[SIZES:@.+]] = {{.+}}constant [1 x i[[sz:64|32]]] zeroinitializer
// Map types: OMP_MAP_IS_FIRST = 32
// CK10-DAG: [[TYPES:@.+]] = {{.+}}constant [1 x i32] [i32 32]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK10-LABEL: implicit_maps_pointer
void implicit_maps_pointer (){
double *ddyn;
// CK10-DAG: call i32 @__tgt_target(i32 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}})
// CK10-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
// CK10-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
// CK10-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
// CK10-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK10-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to double**
// CK10-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to double**
// CK10-DAG: store double* [[PTR:%[^,]+]], double** [[CBP1]]
// CK10-DAG: store double* [[PTR]], double** [[CP1]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK10: call void [[KERNEL:@.+]](double* [[PTR]])
#pragma omp target
{
ddyn[0] += 1.0;
ddyn[1] += 1.0;
}
}
// CK10: define internal void [[KERNEL]](double* {{.*}}[[ARG:%.+]])
// CK10: [[ADDR:%.+]] = alloca double*,
// CK10: store double* [[ARG]], double** [[ADDR]],
// CK10: [[REF:%.+]] = load double*, double** [[ADDR]],
// CK10: {{.+}} = getelementptr inbounds double, double* [[REF]], i[[sz]] 0
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK11 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK11
// RUN: %clang_cc1 -DCK11 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK11
// RUN: %clang_cc1 -DCK11 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK11
// RUN: %clang_cc1 -DCK11 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK11
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
#ifdef CK11
// CK11-DAG: [[SIZES:@.+]] = {{.+}}constant [1 x i[[sz:64|32]]] [i{{64|32}} 16]
[OPENMP] Generate implicit map|firstprivate clauses for target-based directives. If the variable is used in the target-based region but is not found in any private|mapping clause, then generate implicit firstprivate|map clauses for these implicitly mapped variables. llvm-svn: 314205
2017-09-26 21:47:31 +08:00
// Map types: OMP_MAP_TO + OMP_MAP_FROM + OMP_MAP_IS_FIRST + OMP_MAP_IMPLICIT = 547
// CK11-DAG: [[TYPES:@.+]] = {{.+}}constant [1 x i32] [i32 547]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK11-LABEL: implicit_maps_double_complex
void implicit_maps_double_complex (int a){
double _Complex dc = (double)a;
// CK11-DAG: call i32 @__tgt_target(i32 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}})
// CK11-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
// CK11-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
// CK11-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
// CK11-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK11-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to { double, double }**
// CK11-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to { double, double }**
// CK11-DAG: store { double, double }* [[PTR:%[^,]+]], { double, double }** [[CBP1]]
// CK11-DAG: store { double, double }* [[PTR]], { double, double }** [[CP1]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK11: call void [[KERNEL:@.+]]({ double, double }* [[PTR]])
[OPENMP] Fix mapping|privatization of implicitly captured variables. If the `defaultmap(tofrom:scalar)` clause is specified, the scalars must be mapped with 'tofrom' modifiers, otherwise they must be captured as firstprivates. llvm-svn: 314995
2017-10-06 01:51:39 +08:00
#pragma omp target defaultmap(tofrom:scalar)
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
{
dc *= dc;
}
}
// CK11: define internal void [[KERNEL]]({ double, double }* {{.*}}[[ARG:%.+]])
// CK11: [[ADDR:%.+]] = alloca { double, double }*,
// CK11: store { double, double }* [[ARG]], { double, double }** [[ADDR]],
// CK11: [[REF:%.+]] = load { double, double }*, { double, double }** [[ADDR]],
// CK11: {{.+}} = getelementptr inbounds { double, double }, { double, double }* [[REF]], i32 0, i32 0
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK12 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK12 --check-prefix CK12-64
// RUN: %clang_cc1 -DCK12 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK12 --check-prefix CK12-64
// RUN: %clang_cc1 -DCK12 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK12 --check-prefix CK12-32
// RUN: %clang_cc1 -DCK12 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK12 --check-prefix CK12-32
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
#ifdef CK12
// For a 32-bit targets, the value doesn't fit the size of the pointer,
// therefore it is passed by reference with a map 'to' specification.
// CK12-DAG: [[SIZES:@.+]] = {{.+}}constant [1 x i[[sz:64|32]]] [i{{64|32}} 8]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// Map types: OMP_MAP_PRIVATE_VAL + OMP_MAP_IS_FIRST = 288
// CK12-64-DAG: [[TYPES:@.+]] = {{.+}}constant [1 x i32] [i32 288]
[OPENMP] Fix mapping|privatization of implicitly captured variables. If the `defaultmap(tofrom:scalar)` clause is specified, the scalars must be mapped with 'tofrom' modifiers, otherwise they must be captured as firstprivates. llvm-svn: 314995
2017-10-06 01:51:39 +08:00
// Map types: OMP_MAP_TO | OMP_MAP_PRIVATE_PTR | OMP_MAP_FIRST_REF = 161
// CK12-32-DAG: [[TYPES:@.+]] = {{.+}}constant [1 x i32] [i32 161]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK12-LABEL: implicit_maps_float_complex
void implicit_maps_float_complex (int a){
float _Complex fc = (float)a;
// CK12-DAG: call i32 @__tgt_target(i32 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}})
// CK12-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
// CK12-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
// CK12-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
// CK12-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK12-64-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz]]*
// CK12-64-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
// CK12-64-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
// CK12-64-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK12-64-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
// CK12-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to { float, float }*
// CK12-64-DAG: store { float, float } {{.+}}, { float, float }* [[CADDR]],
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK12-32-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to { float, float }**
// CK12-32-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to { float, float }**
// CK12-32-DAG: store { float, float }* [[DECL:%[^,]+]], { float, float }** [[CBP1]]
// CK12-32-DAG: store { float, float }* [[DECL]], { float, float }** [[CP1]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK12-64: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
// CK12-32: call void [[KERNEL:@.+]]({ float, float }* [[DECL]])
#pragma omp target
{
fc *= fc;
}
}
// CK12-64: define internal void [[KERNEL]](i[[sz]] [[ARG:%.+]])
// CK12-64: [[ADDR:%.+]] = alloca i[[sz]],
// CK12-64: store i[[sz]] [[ARG]], i[[sz]]* [[ADDR]],
// CK12-64: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to { float, float }*
// CK12-64: {{.+}} = getelementptr inbounds { float, float }, { float, float }* [[CADDR]], i32 0, i32 0
// CK12-32: define internal void [[KERNEL]]({ float, float }* {{.+}}[[ARG:%.+]])
// CK12-32: [[ADDR:%.+]] = alloca { float, float }*,
// CK12-32: store { float, float }* [[ARG]], { float, float }** [[ADDR]],
// CK12-32: [[REF:%.+]] = load { float, float }*, { float, float }** [[ADDR]],
// CK12-32: {{.+}} = getelementptr inbounds { float, float }, { float, float }* [[REF]], i32 0, i32 0
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK13 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK13
// RUN: %clang_cc1 -DCK13 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK13
// RUN: %clang_cc1 -DCK13 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK13
// RUN: %clang_cc1 -DCK13 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK13
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
#ifdef CK13
// We don't have a constant map size for VLAs.
// Map types:
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// - OMP_MAP_PRIVATE_VAL + OMP_MAP_IS_FIRST = 288 (vla size)
// - OMP_MAP_PRIVATE_VAL + OMP_MAP_IS_FIRST = 288 (vla size)
[OPENMP] Generate implicit map|firstprivate clauses for target-based directives. If the variable is used in the target-based region but is not found in any private|mapping clause, then generate implicit firstprivate|map clauses for these implicitly mapped variables. llvm-svn: 314205
2017-09-26 21:47:31 +08:00
// - OMP_MAP_TO + OMP_MAP_FROM + OMP_MAP_IS_FIRST + OMP_IMPICIT_MAP = 547
// CK13-DAG: [[TYPES:@.+]] = {{.+}}constant [3 x i32] [i32 288, i32 288, i32 547]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK13-LABEL: implicit_maps_variable_length_array
void implicit_maps_variable_length_array (int a){
double vla[2][a];
// CK13-DAG: call i32 @__tgt_target(i32 {{.+}}, i8* {{.+}}, i32 3, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], i[[sz:64|32]]* [[SGEP:%[^,]+]], {{.+}}[[TYPES]]{{.+}})
// CK13-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
// CK13-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
// CK13-DAG: [[SGEP]] = getelementptr inbounds {{.+}}[[SS:%[^,]+]], i32 0, i32 0
// CK13-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
// CK13-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
// CK13-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[SS]], i32 0, i32 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK13-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i[[sz]]*
// CK13-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i[[sz]]*
// CK13-DAG: store i[[sz]] 2, i[[sz]]* [[CBP0]]
// CK13-DAG: store i[[sz]] 2, i[[sz]]* [[CP0]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK13-DAG: store i[[sz]] {{8|4}}, i[[sz]]* [[S0]],
// CK13-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 1
// CK13-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 1
// CK13-DAG: [[S1:%.+]] = getelementptr inbounds {{.+}}[[SS]], i32 0, i32 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK13-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz]]*
// CK13-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
// CK13-DAG: store i[[sz]] [[VAL:%.+]], i[[sz]]* [[CBP1]]
// CK13-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK13-DAG: store i[[sz]] {{8|4}}, i[[sz]]* [[S1]],
// CK13-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 2
// CK13-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 2
// CK13-DAG: [[S2:%.+]] = getelementptr inbounds {{.+}}[[SS]], i32 0, i32 2
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK13-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to double**
// CK13-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to double**
// CK13-DAG: store double* [[DECL:%.+]], double** [[CBP2]]
// CK13-DAG: store double* [[DECL]], double** [[CP2]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK13-DAG: store i[[sz]] [[VALS2:%.+]], i[[sz]]* [[S2]],
// CK13-DAG: [[VALS2]] = mul nuw i[[sz]] %{{.+}}, 8
// CK13: call void [[KERNEL:@.+]](i[[sz]] {{.+}}, i[[sz]] {{.+}}, double* [[DECL]])
#pragma omp target
{
vla[1][3] += 1.0;
}
}
Fix for PR30639: CGDebugInfo Null dereference with OpenMP array access, by Erich Keane OpenMP creates a variable array type with a a null size-expr. The Debug generation failed to due to this. This patch corrects the openmp implementation, updates the tests, and adds a new one for this condition. Differential Revision: https://reviews.llvm.org/D25373 llvm-svn: 284110
2016-10-13 17:52:46 +08:00
// CK13: define internal void [[KERNEL]](i[[sz]] [[VLA0:%.+]], i[[sz]] [[VLA1:%.+]], double* {{.*}}[[ARG:%.+]])
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK13: [[ADDR0:%.+]] = alloca i[[sz]],
// CK13: [[ADDR1:%.+]] = alloca i[[sz]],
// CK13: [[ADDR2:%.+]] = alloca double*,
// CK13: store i[[sz]] [[VLA0]], i[[sz]]* [[ADDR0]],
// CK13: store i[[sz]] [[VLA1]], i[[sz]]* [[ADDR1]],
// CK13: store double* [[ARG]], double** [[ADDR2]],
// CK13: {{.+}} = load i[[sz]], i[[sz]]* [[ADDR0]],
// CK13: {{.+}} = load i[[sz]], i[[sz]]* [[ADDR1]],
// CK13: [[REF:%.+]] = load double*, double** [[ADDR2]],
// CK13: {{.+}} = getelementptr inbounds double, double* [[REF]], i[[sz]] %{{.+}}
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK14 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK14 --check-prefix CK14-64
// RUN: %clang_cc1 -DCK14 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK14 --check-prefix CK14-64
// RUN: %clang_cc1 -DCK14 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK14 --check-prefix CK14-32
// RUN: %clang_cc1 -DCK14 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK14 --check-prefix CK14-32
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
#ifdef CK14
// CK14-DAG: [[ST:%.+]] = type { i32, double }
[OPENMP] Generate implicit map|firstprivate clauses for target-based directives. If the variable is used in the target-based region but is not found in any private|mapping clause, then generate implicit firstprivate|map clauses for these implicitly mapped variables. llvm-svn: 314205
2017-09-26 21:47:31 +08:00
// CK14-DAG: [[SIZES:@.+]] = {{.+}}constant [3 x i[[sz:64|32]]] [i{{64|32}} 4, i{{64|32}} 8, i{{64|32}} 4]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// Map types:
[OPENMP] Generate implicit map|firstprivate clauses for target-based directives. If the variable is used in the target-based region but is not found in any private|mapping clause, then generate implicit firstprivate|map clauses for these implicitly mapped variables. llvm-svn: 314205
2017-09-26 21:47:31 +08:00
// - OMP_MAP_TO + OMP_MAP_FROM + OMP_MAP_IS_FIRST + OMP_IMPLICIT_MAP = 547
// - OMP_MAP_TO + OMP_MAP_FROM + OMP_IMPLICIT_MAP = 515
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// - OMP_MAP_PRIVATE_VAL + OMP_MAP_IS_FIRST = 288
[OPENMP] Generate implicit map|firstprivate clauses for target-based directives. If the variable is used in the target-based region but is not found in any private|mapping clause, then generate implicit firstprivate|map clauses for these implicitly mapped variables. llvm-svn: 314205
2017-09-26 21:47:31 +08:00
// CK14-DAG: [[TYPES:@.+]] = {{.+}}constant [3 x i32] [i32 547, i32 515, i32 288]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
class SSS {
public:
int a;
double b;
void foo(int c) {
#pragma omp target
{
a += c;
b += (double)c;
}
}
SSS(int a, double b) : a(a), b(b) {}
};
// CK14-LABEL: implicit_maps_class
void implicit_maps_class (int a){
SSS sss(a, (double)a);
// CK14: define {{.*}}void @{{.+}}foo{{.+}}([[ST]]* {{[^,]+}}, i32 {{[^,]+}})
[OPENMP] Generate implicit map|firstprivate clauses for target-based directives. If the variable is used in the target-based region but is not found in any private|mapping clause, then generate implicit firstprivate|map clauses for these implicitly mapped variables. llvm-svn: 314205
2017-09-26 21:47:31 +08:00
// CK14-DAG: call i32 @__tgt_target(i32 {{.+}}, i8* {{.+}}, i32 3, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}})
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK14-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
// CK14-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
// CK14-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
// CK14-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK14-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
[OPENMP] Generate implicit map|firstprivate clauses for target-based directives. If the variable is used in the target-based region but is not found in any private|mapping clause, then generate implicit firstprivate|map clauses for these implicitly mapped variables. llvm-svn: 314205
2017-09-26 21:47:31 +08:00
// CK14-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK14-DAG: store [[ST]]* [[DECL:%.+]], [[ST]]** [[CBP0]]
[OPENMP] Generate implicit map|firstprivate clauses for target-based directives. If the variable is used in the target-based region but is not found in any private|mapping clause, then generate implicit firstprivate|map clauses for these implicitly mapped variables. llvm-svn: 314205
2017-09-26 21:47:31 +08:00
// CK14-DAG: store i32* %{{.+}}, i32** [[CP0]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK14-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 1
// CK14-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 1
[OPENMP] Generate implicit map|firstprivate clauses for target-based directives. If the variable is used in the target-based region but is not found in any private|mapping clause, then generate implicit firstprivate|map clauses for these implicitly mapped variables. llvm-svn: 314205
2017-09-26 21:47:31 +08:00
// CK14-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[ST]]**
// CK14-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to double**
// CK14-DAG: store [[ST]]* [[DECL]], [[ST]]** [[CBP1]]
// CK14-DAG: store double* %{{.+}}, double** [[CP1]]
// CK14-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 2
// CK14-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 2
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK14-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz]]*
// CK14-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
// CK14-DAG: store i[[sz]] [[VAL:%.+]], i[[sz]]* [[CBP1]]
// CK14-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK14-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
// CK14-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to i32*
// CK14-64-DAG: store i32 {{.+}}, i32* [[CADDR]],
// CK14: call void [[KERNEL:@.+]]([[ST]]* [[DECL]], i[[sz]] {{.+}})
sss.foo(123);
}
// CK14: define internal void [[KERNEL]]([[ST]]* [[THIS:%.+]], i[[sz]] [[ARG:%.+]])
// CK14: [[ADDR0:%.+]] = alloca [[ST]]*,
// CK14: [[ADDR1:%.+]] = alloca i[[sz]],
// CK14: store [[ST]]* [[THIS]], [[ST]]** [[ADDR0]],
// CK14: store i[[sz]] [[ARG]], i[[sz]]* [[ADDR1]],
// CK14: [[REF0:%.+]] = load [[ST]]*, [[ST]]** [[ADDR0]],
// CK14-64: [[CADDR1:%.+]] = bitcast i[[sz]]* [[ADDR1]] to i32*
// CK14-64: {{.+}} = load i32, i32* [[CADDR1]],
// CK14-32: {{.+}} = load i32, i32* [[ADDR1]],
// CK14: {{.+}} = getelementptr inbounds [[ST]], [[ST]]* [[REF0]], i32 0, i32 0
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK15 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK15 --check-prefix CK15-64
// RUN: %clang_cc1 -DCK15 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK15 --check-prefix CK15-64
// RUN: %clang_cc1 -DCK15 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK15 --check-prefix CK15-32
// RUN: %clang_cc1 -DCK15 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK15 --check-prefix CK15-32
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
#ifdef CK15
// CK15: [[ST:%.+]] = type { i32, double, i32* }
[OPENMP] Generate implicit map|firstprivate clauses for target-based directives. If the variable is used in the target-based region but is not found in any private|mapping clause, then generate implicit firstprivate|map clauses for these implicitly mapped variables. llvm-svn: 314205
2017-09-26 21:47:31 +08:00
// CK15: [[SIZES:@.+]] = {{.+}}constant [5 x i[[sz:64|32]]] [i{{64|32}} 4, i{{64|32}} 8, i{{64|32}} {{8|4}}, i{{64|32}} 4, i{{64|32}} 4]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// Map types:
[OPENMP] Generate implicit map|firstprivate clauses for target-based directives. If the variable is used in the target-based region but is not found in any private|mapping clause, then generate implicit firstprivate|map clauses for these implicitly mapped variables. llvm-svn: 314205
2017-09-26 21:47:31 +08:00
// - OMP_MAP_TO + OMP_MAP_FROM + OMP_MAP_IS_FIRST + OMP_MAP_IMPLICIT = 547
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// - OMP_MAP_PRIVATE_VAL + OMP_MAP_IS_FIRST = 288
[OPENMP] Generate implicit map|firstprivate clauses for target-based directives. If the variable is used in the target-based region but is not found in any private|mapping clause, then generate implicit firstprivate|map clauses for these implicitly mapped variables. llvm-svn: 314205
2017-09-26 21:47:31 +08:00
// CK15: [[TYPES:@.+]] = {{.+}}constant [5 x i32] [i32 547, i32 515, i32 512, i32 531, i32 288]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
[OPENMP] Generate implicit map|firstprivate clauses for target-based directives. If the variable is used in the target-based region but is not found in any private|mapping clause, then generate implicit firstprivate|map clauses for these implicitly mapped variables. llvm-svn: 314205
2017-09-26 21:47:31 +08:00
// CK15: [[SIZES2:@.+]] = {{.+}}constant [5 x i[[sz]]] [i{{64|32}} 4, i{{64|32}} 8, i{{64|32}} {{8|4}}, i{{64|32}} 4, i{{64|32}} 4]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// Map types:
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// - OMP_MAP_TO + OMP_MAP_FROM + OMP_MAP_IS_FIRST = 35
// - OMP_MAP_PRIVATE_VAL + OMP_MAP_IS_FIRST = 288
[OPENMP] Generate implicit map|firstprivate clauses for target-based directives. If the variable is used in the target-based region but is not found in any private|mapping clause, then generate implicit firstprivate|map clauses for these implicitly mapped variables. llvm-svn: 314205
2017-09-26 21:47:31 +08:00
// CK15: [[TYPES2:@.+]] = {{.+}}constant [5 x i32] [i32 547, i32 515, i32 512, i32 531, i32 288]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
template<int x>
class SSST {
public:
int a;
double b;
int &r;
void foo(int c) {
#pragma omp target
{
a += c + x;
b += (double)(c + x);
r += x;
}
}
template<int y>
void bar(int c) {
#pragma omp target
{
a += c + x + y;
b += (double)(c + x + y);
r += x + y;
}
}
SSST(int a, double b, int &r) : a(a), b(b), r(r) {}
};
// CK15-LABEL: implicit_maps_templated_class
void implicit_maps_templated_class (int a){
SSST<123> ssst(a, (double)a, a);
// CK15: define {{.*}}void @{{.+}}foo{{.+}}([[ST]]* {{[^,]+}}, i32 {{[^,]+}})
[OPENMP] Generate implicit map|firstprivate clauses for target-based directives. If the variable is used in the target-based region but is not found in any private|mapping clause, then generate implicit firstprivate|map clauses for these implicitly mapped variables. llvm-svn: 314205
2017-09-26 21:47:31 +08:00
// CK15-DAG: call i32 @__tgt_target(i32 {{.+}}, i8* {{.+}}, i32 5, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}})
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK15-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
// CK15-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
// CK15-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
// CK15-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK15-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
[OPENMP] Generate implicit map|firstprivate clauses for target-based directives. If the variable is used in the target-based region but is not found in any private|mapping clause, then generate implicit firstprivate|map clauses for these implicitly mapped variables. llvm-svn: 314205
2017-09-26 21:47:31 +08:00
// CK15-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK15-DAG: store [[ST]]* [[DECL:%.+]], [[ST]]** [[CBP0]]
[OPENMP] Generate implicit map|firstprivate clauses for target-based directives. If the variable is used in the target-based region but is not found in any private|mapping clause, then generate implicit firstprivate|map clauses for these implicitly mapped variables. llvm-svn: 314205
2017-09-26 21:47:31 +08:00
// CK15-DAG: store i32* %{{.+}}, i32** [[CP0]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK15-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 1
// CK15-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 1
[OPENMP] Generate implicit map|firstprivate clauses for target-based directives. If the variable is used in the target-based region but is not found in any private|mapping clause, then generate implicit firstprivate|map clauses for these implicitly mapped variables. llvm-svn: 314205
2017-09-26 21:47:31 +08:00
// CK15-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[ST]]**
// CK15-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to double**
// CK15-DAG: store [[ST]]* [[DECL]], [[ST]]** [[CBP1]]
// CK15-DAG: store double* %{{.+}}, double** [[CP1]]
// CK15-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 2
// CK15-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 2
// CK15-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to [[ST]]**
// CK15-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to i32***
// CK15-DAG: store [[ST]]* [[DECL]], [[ST]]** [[CBP2]]
// CK15-DAG: store i32** %{{.+}}, i32*** [[CP2]]
// CK15-DAG: [[BP3:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 3
// CK15-DAG: [[P3:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 3
// CK15-DAG: [[CBP3:%.+]] = bitcast i8** [[BP3]] to i32***
// CK15-DAG: [[CP3:%.+]] = bitcast i8** [[P3]] to i32**
// CK15-DAG: store i32** %{{.+}}, i32*** [[CBP3]]
// CK15-DAG: store i32* %{{.+}}, i32** [[CP3]]
// CK15-DAG: [[BP4:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 4
// CK15-DAG: [[P4:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 4
// CK15-DAG: [[CBP4:%.+]] = bitcast i8** [[BP4]] to i[[sz]]*
// CK15-DAG: [[CP4:%.+]] = bitcast i8** [[P4]] to i[[sz]]*
// CK15-DAG: store i[[sz]] [[VAL:%.+]], i[[sz]]* [[CBP4]]
// CK15-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP4]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK15-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
// CK15-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to i32*
// CK15-64-DAG: store i32 {{.+}}, i32* [[CADDR]],
// CK15: call void [[KERNEL:@.+]]([[ST]]* [[DECL]], i[[sz]] {{.+}})
ssst.foo(456);
// CK15: define {{.*}}void @{{.+}}bar{{.+}}([[ST]]* {{[^,]+}}, i32 {{[^,]+}})
[OPENMP] Generate implicit map|firstprivate clauses for target-based directives. If the variable is used in the target-based region but is not found in any private|mapping clause, then generate implicit firstprivate|map clauses for these implicitly mapped variables. llvm-svn: 314205
2017-09-26 21:47:31 +08:00
// CK15-DAG: call i32 @__tgt_target(i32 {{.+}}, i8* {{.+}}, i32 5, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES2]]{{.+}}, {{.+}}[[TYPES2]]{{.+}})
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK15-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
// CK15-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
// CK15-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
// CK15-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK15-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
[OPENMP] Generate implicit map|firstprivate clauses for target-based directives. If the variable is used in the target-based region but is not found in any private|mapping clause, then generate implicit firstprivate|map clauses for these implicitly mapped variables. llvm-svn: 314205
2017-09-26 21:47:31 +08:00
// CK15-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK15-DAG: store [[ST]]* [[DECL]], [[ST]]** [[CBP0]]
// CK15-DAG: store i32* %{{.+}}, i32** [[CP0]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK15-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 1
// CK15-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 1
[OPENMP] Generate implicit map|firstprivate clauses for target-based directives. If the variable is used in the target-based region but is not found in any private|mapping clause, then generate implicit firstprivate|map clauses for these implicitly mapped variables. llvm-svn: 314205
2017-09-26 21:47:31 +08:00
// CK15-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[ST]]**
// CK15-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to double**
// CK15-DAG: store [[ST]]* [[DECL]], [[ST]]** [[CBP1]]
// CK15-DAG: store double* %{{.+}}, double** [[CP1]]
// CK15-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 2
// CK15-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 2
// CK15-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to [[ST]]**
// CK15-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to i32***
// CK15-DAG: store [[ST]]* [[DECL]], [[ST]]** [[CBP2]]
// CK15-DAG: store i32** %{{.+}}, i32*** [[CP2]]
// CK15-DAG: [[BP3:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 3
// CK15-DAG: [[P3:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 3
// CK15-DAG: [[CBP3:%.+]] = bitcast i8** [[BP3]] to i32***
// CK15-DAG: [[CP3:%.+]] = bitcast i8** [[P3]] to i32**
// CK15-DAG: store i32** %{{.+}}, i32*** [[CBP3]]
// CK15-DAG: store i32* %{{.+}}, i32** [[CP3]]
// CK15-DAG: [[BP4:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 4
// CK15-DAG: [[P4:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 4
// CK15-DAG: [[CBP4:%.+]] = bitcast i8** [[BP4]] to i[[sz]]*
// CK15-DAG: [[CP4:%.+]] = bitcast i8** [[P4]] to i[[sz]]*
// CK15-DAG: store i[[sz]] [[VAL:%.+]], i[[sz]]* [[CBP4]]
// CK15-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP4]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK15-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
// CK15-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to i32*
// CK15-64-DAG: store i32 {{.+}}, i32* [[CADDR]],
// CK15: call void [[KERNEL2:@.+]]([[ST]]* [[DECL]], i[[sz]] {{.+}})
ssst.bar<210>(789);
}
// CK15: define internal void [[KERNEL]]([[ST]]* [[THIS:%.+]], i[[sz]] [[ARG:%.+]])
// CK15: [[ADDR0:%.+]] = alloca [[ST]]*,
// CK15: [[ADDR1:%.+]] = alloca i[[sz]],
// CK15: store [[ST]]* [[THIS]], [[ST]]** [[ADDR0]],
// CK15: store i[[sz]] [[ARG]], i[[sz]]* [[ADDR1]],
// CK15: [[REF0:%.+]] = load [[ST]]*, [[ST]]** [[ADDR0]],
// CK15-64: [[CADDR1:%.+]] = bitcast i[[sz]]* [[ADDR1]] to i32*
// CK15-64: {{.+}} = load i32, i32* [[CADDR1]],
// CK15-32: {{.+}} = load i32, i32* [[ADDR1]],
// CK15: {{.+}} = getelementptr inbounds [[ST]], [[ST]]* [[REF0]], i32 0, i32 0
// CK15: define internal void [[KERNEL2]]([[ST]]* [[THIS:%.+]], i[[sz]] [[ARG:%.+]])
// CK15: [[ADDR0:%.+]] = alloca [[ST]]*,
// CK15: [[ADDR1:%.+]] = alloca i[[sz]],
// CK15: store [[ST]]* [[THIS]], [[ST]]** [[ADDR0]],
// CK15: store i[[sz]] [[ARG]], i[[sz]]* [[ADDR1]],
// CK15: [[REF0:%.+]] = load [[ST]]*, [[ST]]** [[ADDR0]],
// CK15-64: [[CADDR1:%.+]] = bitcast i[[sz]]* [[ADDR1]] to i32*
// CK15-64: {{.+}} = load i32, i32* [[CADDR1]],
// CK15-32: {{.+}} = load i32, i32* [[ADDR1]],
// CK15: {{.+}} = getelementptr inbounds [[ST]], [[ST]]* [[REF0]], i32 0, i32 0
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK16 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK16 --check-prefix CK16-64
// RUN: %clang_cc1 -DCK16 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK16 --check-prefix CK16-64
// RUN: %clang_cc1 -DCK16 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK16 --check-prefix CK16-32
// RUN: %clang_cc1 -DCK16 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK16 --check-prefix CK16-32
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
#ifdef CK16
// CK16-DAG: [[SIZES:@.+]] = {{.+}}constant [1 x i[[sz:64|32]]] [i{{64|32}} 4]
// Map types:
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// - OMP_MAP_PRIVATE_VAL + OMP_MAP_IS_FIRST = 288
// CK16-DAG: [[TYPES:@.+]] = {{.+}}constant [1 x i32] [i32 288]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
template<int y>
int foo(int d) {
int res = d;
#pragma omp target
{
res += y;
}
return res;
}
// CK16-LABEL: implicit_maps_templated_function
void implicit_maps_templated_function (int a){
int i = a;
// CK16: define {{.*}}i32 @{{.+}}foo{{.+}}(i32 {{[^,]+}})
// CK16-DAG: call i32 @__tgt_target(i32 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}})
// CK16-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
// CK16-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
// CK16-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
// CK16-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK16-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz]]*
// CK16-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
// CK16-DAG: store i[[sz]] [[VAL:%.+]], i[[sz]]* [[CBP1]]
// CK16-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK16-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
// CK16-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to i32*
// CK16-64-DAG: store i32 {{.+}}, i32* [[CADDR]],
// CK16: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
i = foo<543>(i);
}
// CK16: define internal void [[KERNEL]](i[[sz]] [[ARG:%.+]])
// CK16: [[ADDR:%.+]] = alloca i[[sz]],
// CK16: store i[[sz]] [[ARG]], i[[sz]]* [[ADDR]],
// CK16-64: [[CADDR:%.+]] = bitcast i64* [[ADDR]] to i32*
// CK16-64: {{.+}} = load i32, i32* [[CADDR]],
// CK16-32: {{.+}} = load i32, i32* [[ADDR]],
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK17 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK17
// RUN: %clang_cc1 -DCK17 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK17
// RUN: %clang_cc1 -DCK17 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK17
// RUN: %clang_cc1 -DCK17 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK17
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
#ifdef CK17
// CK17-DAG: [[ST:%.+]] = type { i32, double }
// CK17-DAG: [[SIZES:@.+]] = {{.+}}constant [1 x i[[sz:64|32]]] [i{{64|32}} {{16|12}}]
[OPENMP] Generate implicit map|firstprivate clauses for target-based directives. If the variable is used in the target-based region but is not found in any private|mapping clause, then generate implicit firstprivate|map clauses for these implicitly mapped variables. llvm-svn: 314205
2017-09-26 21:47:31 +08:00
// Map types: OMP_MAP_TO + OMP_MAP_FROM + OMP_MAP_IS_FIRST + OMP_MAP_IMPLICIT = 547
// CK17-DAG: [[TYPES:@.+]] = {{.+}}constant [1 x i32] [i32 547]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
class SSS {
public:
int a;
double b;
};
// CK17-LABEL: implicit_maps_struct
void implicit_maps_struct (int a){
SSS s = {a, (double)a};
// CK17-DAG: call i32 @__tgt_target(i32 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}})
// CK17-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
// CK17-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
// CK17-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
// CK17-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK17-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[ST]]**
// CK17-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [[ST]]**
// CK17-DAG: store [[ST]]* [[DECL:%.+]], [[ST]]** [[CBP1]]
// CK17-DAG: store [[ST]]* [[DECL]], [[ST]]** [[CP1]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK17: call void [[KERNEL:@.+]]([[ST]]* [[DECL]])
#pragma omp target
{
s.a += 1;
s.b += 1.0;
}
}
// CK17: define internal void [[KERNEL]]([[ST]]* {{.+}}[[ARG:%.+]])
// CK17: [[ADDR:%.+]] = alloca [[ST]]*,
// CK17: store [[ST]]* [[ARG]], [[ST]]** [[ADDR]],
// CK17: [[REF:%.+]] = load [[ST]]*, [[ST]]** [[ADDR]],
// CK17: {{.+}} = getelementptr inbounds [[ST]], [[ST]]* [[REF]], i32 0, i32 0
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK18 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK18 --check-prefix CK18-64
// RUN: %clang_cc1 -DCK18 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK18 --check-prefix CK18-64
// RUN: %clang_cc1 -DCK18 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK18 --check-prefix CK18-32
// RUN: %clang_cc1 -DCK18 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK18 --check-prefix CK18-32
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
#ifdef CK18
// CK18-DAG: [[SIZES:@.+]] = {{.+}}constant [1 x i[[sz:64|32]]] [i{{64|32}} 4]
// Map types:
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// - OMP_MAP_PRIVATE_VAL + OMP_MAP_IS_FIRST = 288
// CK18-DAG: [[TYPES:@.+]] = {{.+}}constant [1 x i32] [i32 288]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
template<typename T>
int foo(T d) {
#pragma omp target
{
d += (T)1;
}
return d;
}
// CK18-LABEL: implicit_maps_template_type_capture
void implicit_maps_template_type_capture (int a){
int i = a;
// CK18: define {{.*}}i32 @{{.+}}foo{{.+}}(i32 {{[^,]+}})
// CK18-DAG: call i32 @__tgt_target(i32 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}})
// CK18-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
// CK18-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
// CK18-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
// CK18-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK18-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz]]*
// CK18-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
// CK18-DAG: store i[[sz]] [[VAL:%.+]], i[[sz]]* [[CBP1]]
// CK18-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
// CK18-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
// CK18-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to i32*
// CK18-64-DAG: store i32 {{.+}}, i32* [[CADDR]],
// CK18: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
i = foo(i);
}
// CK18: define internal void [[KERNEL]](i[[sz]] [[ARG:%.+]])
// CK18: [[ADDR:%.+]] = alloca i[[sz]],
// CK18: store i[[sz]] [[ARG]], i[[sz]]* [[ADDR]],
// CK18-64: [[CADDR:%.+]] = bitcast i64* [[ADDR]] to i32*
// CK18-64: {{.+}} = load i32, i32* [[CADDR]],
// CK18-32: {{.+}} = load i32, i32* [[ADDR]],
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK19 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK19 --check-prefix CK19-64
// RUN: %clang_cc1 -DCK19 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK19 --check-prefix CK19-64
// RUN: %clang_cc1 -DCK19 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK19 --check-prefix CK19-32
// RUN: %clang_cc1 -DCK19 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK19 --check-prefix CK19-32
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
#ifdef CK19
// CK19: [[SIZE00:@.+]] = private {{.*}}constant [1 x i[[Z:64|32]]] [i[[Z:64|32]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE00:@.+]] = private {{.*}}constant [1 x i32] [i32 32]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
[OpenMP] Fix mapping of scalars for combined directives Combined directives like 'target parallel' have two captured statements. Sema has to check the right one from the right direction. Previously, Sema::IsOpenMPCapturedByRef would return false for mapped scalars on combined directives. This results in a wrong signature of the outlined function which triggers an assertion: void llvm::CallInst::init(llvm::FunctionType *, llvm::Value *, ArrayRef<llvm::Value *>, ArrayRef<OperandBundleDef>, const llvm::Twine &): Assertion `(i >= FTy->getNumParams() || FTy->getParamType(i) == Args[i]->getType()) && "Calling a function with a bad signature!"' failed. Fixes PR30975 (and PR31985). New function was taken from clang-ykt. Differential Revision: https://reviews.llvm.org/D34888 llvm-svn: 306956
2017-07-01 18:40:50 +08:00
// CK19: [[SIZE00n:@.+]] = private {{.*}}constant [1 x i[[Z:64|32]]] [i[[Z:64|32]] 4]
// CK19: [[MTYPE00n:@.+]] = private {{.*}}constant [1 x i32] [i32 32]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE01:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 400]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE01:@.+]] = private {{.*}}constant [1 x i32] [i32 33]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE02:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 240]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE02:@.+]] = private {{.*}}constant [1 x i32] [i32 34]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE03:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 240]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE03:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE04:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 400]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE04:@.+]] = private {{.*}}constant [1 x i32] [i32 32]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE05:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE05:@.+]] = private {{.*}}constant [1 x i32] [i32 33]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE06:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE07:@.+]] = private {{.*}}constant [1 x i32] [i32 32]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE08:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE08:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE09:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] {{8|4}}]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE09:@.+]] = private {{.*}}constant [1 x i32] [i32 34]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE10:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 240]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE10:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE11:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 240]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE11:@.+]] = private {{.*}}constant [1 x i32] [i32 32]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE12:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE12:@.+]] = private {{.*}}constant [1 x i32] [i32 33]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE13:@.+]] = private {{.*}}constant [1 x i32] [i32 32]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE14:@.+]] = private {{.*}}constant [1 x i32] [i32 33]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE15:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE15:@.+]] = private {{.*}}constant [1 x i32] [i32 34]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE16:@.+]] = private {{.*}}constant [2 x i32] [i32 288, i32 33]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE17:@.+]] = private {{.*}}constant [2 x i[[Z]]] [i[[Z]] {{8|4}}, i[[Z]] 240]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE17:@.+]] = private {{.*}}constant [2 x i32] [i32 288, i32 34]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE18:@.+]] = private {{.*}}constant [2 x i[[Z]]] [i[[Z]] {{8|4}}, i[[Z]] 240]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE18:@.+]] = private {{.*}}constant [2 x i32] [i32 288, i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE19:@.+]] = private {{.*}}constant [2 x i32] [i32 288, i32 32]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE20:@.+]] = private {{.*}}constant [2 x i[[Z]]] [i[[Z]] {{8|4}}, i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE20:@.+]] = private {{.*}}constant [2 x i32] [i32 288, i32 33]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE21:@.+]] = private {{.*}}constant [2 x i32] [i32 288, i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE22:@.+]] = private {{.*}}constant [2 x i[[Z]]] [i[[Z]] {{8|4}}, i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE22:@.+]] = private {{.*}}constant [2 x i32] [i32 288, i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE23:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE23:@.+]] = private {{.*}}constant [1 x i32] [i32 39]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE24:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 480]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE24:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE25:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 16]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE25:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE26:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 24]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE26:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE27:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE27:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE28:@.+]] = private {{.*}}constant [3 x i[[Z]]] [i[[Z]] {{8|4}}, i[[Z]] {{8|4}}, i[[Z]] 16]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE28:@.+]] = private {{.*}}constant [3 x i32] [i32 35, i32 19, i32 19]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE29:@.+]] = private {{.*}}constant [3 x i[[Z]]] [i[[Z]] {{8|4}}, i[[Z]] {{8|4}}, i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE29:@.+]] = private {{.*}}constant [3 x i32] [i32 35, i32 19, i32 19]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE30:@.+]] = private {{.*}}constant [4 x i32] [i32 288, i32 288, i32 288, i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE31:@.+]] = private {{.*}}constant [4 x i[[Z]]] [i[[Z]] {{8|4}}, i[[Z]] {{8|4}}, i[[Z]] {{8|4}}, i[[Z]] 40]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE31:@.+]] = private {{.*}}constant [4 x i32] [i32 288, i32 288, i32 288, i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE32:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 13728]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE32:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE33:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 13728]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE33:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE34:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 13728]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE34:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE35:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE36:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 208]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE36:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE37:@.+]] = private {{.*}}constant [3 x i32] [i32 288, i32 288, i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE38:@.+]] = private {{.*}}constant [3 x i32] [i32 288, i32 288, i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE39:@.+]] = private {{.*}}constant [3 x i32] [i32 288, i32 288, i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE40:@.+]] = private {{.*}}constant [3 x i32] [i32 288, i32 288, i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE41:@.+]] = private {{.*}}constant [3 x i[[Z]]] [i[[Z]] {{8|4}}, i[[Z]] {{8|4}}, i[[Z]] 208]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE41:@.+]] = private {{.*}}constant [3 x i32] [i32 288, i32 288, i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: [[SIZE42:@.+]] = private {{.*}}constant [3 x i[[Z]]] [i[[Z]] {{8|4}}, i[[Z]] {{8|4}}, i[[Z]] 104]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE42:@.+]] = private {{.*}}constant [3 x i32] [i32 35, i32 19, i32 19]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK19: [[MTYPE43:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-LABEL: explicit_maps_single
void explicit_maps_single (int ii){
// Map of a scalar.
int a = ii;
// Region 00
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE00]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE00]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK19-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
// CK19-DAG: store i32* [[VAR0]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: call void [[CALL00:@.+]](i32* {{[^,]+}})
#pragma omp target map(alloc:a)
{
++a;
}
[OpenMP] Fix mapping of scalars for combined directives Combined directives like 'target parallel' have two captured statements. Sema has to check the right one from the right direction. Previously, Sema::IsOpenMPCapturedByRef would return false for mapped scalars on combined directives. This results in a wrong signature of the outlined function which triggers an assertion: void llvm::CallInst::init(llvm::FunctionType *, llvm::Value *, ArrayRef<llvm::Value *>, ArrayRef<OperandBundleDef>, const llvm::Twine &): Assertion `(i >= FTy->getNumParams() || FTy->getParamType(i) == Args[i]->getType()) && "Calling a function with a bad signature!"' failed. Fixes PR30975 (and PR31985). New function was taken from clang-ykt. Differential Revision: https://reviews.llvm.org/D34888 llvm-svn: 306956
2017-07-01 18:40:50 +08:00
// Map of a scalar in nested region.
int b = a;
// Region 00n
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE00n]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE00n]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK19-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
// CK19-DAG: store i32* [[VAR0]], i32** [[CP0]]
// CK19: call void [[CALL00n:@.+]](i32* {{[^,]+}})
#pragma omp target map(alloc:b)
#pragma omp parallel
{
++b;
}
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// Map of an array.
int arra[100];
// Region 01
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE01]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE01]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x i32]**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [100 x i32]**
// CK19-DAG: store [100 x i32]* [[VAR0:%.+]], [100 x i32]** [[CBP0]]
// CK19-DAG: store [100 x i32]* [[VAR0]], [100 x i32]** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: call void [[CALL01:@.+]]([100 x i32]* {{[^,]+}})
#pragma omp target map(to:arra)
{
arra[50]++;
}
// Region 02
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE02]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE02]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x i32]**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK19-DAG: store [100 x i32]* [[VAR0:%.+]], [100 x i32]** [[CBP0]]
// CK19-DAG: store i32* [[SEC0:%[^,]+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[SEC0]] = getelementptr {{.*}}[100 x i32]* [[VAR0]], i{{.+}} 0, i{{.+}} 20
// CK19: call void [[CALL02:@.+]]([100 x i32]* {{[^,]+}})
#pragma omp target map(from:arra[20:60])
{
arra[50]++;
}
// Region 03
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE03]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE03]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x i32]**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK19-DAG: store [100 x i32]* [[VAR0:%.+]], [100 x i32]** [[CBP0]]
// CK19-DAG: store i32* [[SEC0:%[^,]+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[SEC0]] = getelementptr {{.*}}[100 x i32]* [[VAR0]], i{{.+}} 0, i{{.+}} 0
// CK19: call void [[CALL03:@.+]]([100 x i32]* {{[^,]+}})
#pragma omp target map(tofrom:arra[:60])
{
arra[50]++;
}
// Region 04
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE04]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE04]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x i32]**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK19-DAG: store [100 x i32]* [[VAR0:%.+]], [100 x i32]** [[CBP0]]
// CK19-DAG: store i32* [[SEC0:%[^,]+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[SEC0]] = getelementptr {{.*}}[100 x i32]* [[VAR0]], i{{.+}} 0, i{{.+}} 0
// CK19: call void [[CALL04:@.+]]([100 x i32]* {{[^,]+}})
#pragma omp target map(alloc:arra[:])
{
arra[50]++;
}
// Region 05
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE05]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE05]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x i32]**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK19-DAG: store [100 x i32]* [[VAR0:%.+]], [100 x i32]** [[CBP0]]
// CK19-DAG: store i32* [[SEC0:%[^,]+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[SEC0]] = getelementptr {{.*}}[100 x i32]* [[VAR0]], i{{.+}} 0, i{{.+}} 15
// CK19: call void [[CALL05:@.+]]([100 x i32]* {{[^,]+}})
#pragma omp target map(to:arra[15])
{
arra[15]++;
}
// Region 06
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i[[Z]]* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE06]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x i32]**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK19-DAG: store [100 x i32]* [[VAR0:%.+]], [100 x i32]** [[CBP0]]
// CK19-DAG: store i32* [[SEC0:%[^,]+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i{{.+}} [[CSVAL0:%[^,]+]], i{{.+}}* [[S0]]
// CK19-DAG: [[CSVAL0]] = mul nuw i{{.+}} %{{.*}}, 4
// CK19-DAG: [[SEC0]] = getelementptr {{.*}}[100 x i32]* [[VAR0]], i{{.+}} 0, i{{.+}} %{{.*}}
// CK19: call void [[CALL06:@.+]]([100 x i32]* {{[^,]+}})
#pragma omp target map(tofrom:arra[ii:ii+23])
{
arra[50]++;
}
// Region 07
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i[[Z]]* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE07]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x i32]**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK19-DAG: store [100 x i32]* [[VAR0:%.+]], [100 x i32]** [[CBP0]]
// CK19-DAG: store i32* [[SEC0:%[^,]+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i{{.+}} [[CSVAL0:%[^,]+]], i{{.+}}* [[S0]]
// CK19-DAG: [[CSVAL0]] = mul nuw i{{.+}} %{{.*}}, 4
// CK19-DAG: [[SEC0]] = getelementptr {{.*}}[100 x i32]* [[VAR0]], i{{.+}} 0, i{{.+}} 0
// CK19: call void [[CALL07:@.+]]([100 x i32]* {{[^,]+}})
#pragma omp target map(alloc:arra[:ii])
{
arra[50]++;
}
// Region 08
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE08]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE08]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x i32]**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK19-DAG: store [100 x i32]* [[VAR0:%.+]], [100 x i32]** [[CBP0]]
// CK19-DAG: store i32* [[SEC0:%[^,]+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[SEC0]] = getelementptr {{.*}}[100 x i32]* [[VAR0]], i{{.+}} 0, i{{.+}} %{{.*}}
// CK19: call void [[CALL08:@.+]]([100 x i32]* {{[^,]+}})
#pragma omp target map(tofrom:arra[ii])
{
arra[15]++;
}
// Map of a pointer.
int *pa;
// Region 09
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE09]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE09]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32***
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32***
// CK19-DAG: store i32** [[VAR0:%.+]], i32*** [[CBP0]]
// CK19-DAG: store i32** [[VAR0]], i32*** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: call void [[CALL09:@.+]](i32** {{[^,]+}})
#pragma omp target map(from:pa)
{
pa[50]++;
}
// Region 10
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE10]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE10]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK19-DAG: store i32* [[RVAR0:%.+]], i32** [[CBP0]]
// CK19-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[RVAR0]] = load i32*, i32** [[VAR0:%[^,]+]]
// CK19-DAG: [[SEC0]] = getelementptr {{.*}}i32* [[RVAR00:%.+]], i{{.+}} 20
// CK19-DAG: [[RVAR00]] = load i32*, i32** [[VAR0]]
// CK19: call void [[CALL10:@.+]](i32* {{[^,]+}})
#pragma omp target map(tofrom:pa[20:60])
{
pa[50]++;
}
// Region 11
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE11]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE11]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK19-DAG: store i32* [[RVAR0:%.+]], i32** [[CBP0]]
// CK19-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[RVAR0]] = load i32*, i32** [[VAR0:%[^,]+]]
// CK19-DAG: [[SEC0]] = getelementptr {{.*}}i32* [[RVAR00:%.+]], i{{.+}} 0
// CK19-DAG: [[RVAR00]] = load i32*, i32** [[VAR0]]
// CK19: call void [[CALL11:@.+]](i32* {{[^,]+}})
#pragma omp target map(alloc:pa[:60])
{
pa[50]++;
}
// Region 12
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE12]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE12]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK19-DAG: store i32* [[RVAR0:%.+]], i32** [[CBP0]]
// CK19-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[RVAR0]] = load i32*, i32** [[VAR0:%[^,]+]]
// CK19-DAG: [[SEC0]] = getelementptr {{.*}}i32* [[RVAR00:%.+]], i{{.+}} 15
// CK19-DAG: [[RVAR00]] = load i32*, i32** [[VAR0]]
// CK19: call void [[CALL12:@.+]](i32* {{[^,]+}})
#pragma omp target map(to:pa[15])
{
pa[15]++;
}
// Region 13
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i[[Z]]* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE13]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK19-DAG: store i32* [[RVAR0:%.+]], i32** [[CBP0]]
// CK19-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i{{.+}} [[CSVAL0:%[^,]+]], i{{.+}}* [[S0]]
// CK19-DAG: [[CSVAL0]] = mul nuw i{{.+}} %{{.*}}, 4
// CK19-DAG: [[RVAR0]] = load i32*, i32** [[VAR0:%[^,]+]]
// CK19-DAG: [[SEC0]] = getelementptr {{.*}}i32* [[RVAR00:%.+]], i{{.+}} %{{.*}}
// CK19-DAG: [[RVAR00]] = load i32*, i32** [[VAR0]]
// CK19: call void [[CALL13:@.+]](i32* {{[^,]+}})
#pragma omp target map(alloc:pa[ii-23:ii])
{
pa[50]++;
}
// Region 14
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i[[Z]]* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE14]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK19-DAG: store i32* [[RVAR0:%.+]], i32** [[CBP0]]
// CK19-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i{{.+}} [[CSVAL0:%[^,]+]], i{{.+}}* [[S0]]
// CK19-DAG: [[CSVAL0]] = mul nuw i{{.+}} %{{.*}}, 4
// CK19-DAG: [[RVAR0]] = load i32*, i32** [[VAR0:%[^,]+]]
// CK19-DAG: [[SEC0]] = getelementptr {{.*}}i32* [[RVAR00:%.+]], i{{.+}} 0
// CK19-DAG: [[RVAR00]] = load i32*, i32** [[VAR0]]
// CK19: call void [[CALL14:@.+]](i32* {{[^,]+}})
#pragma omp target map(to:pa[:ii])
{
pa[50]++;
}
// Region 15
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE15]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE15]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK19-DAG: store i32* [[RVAR0:%.+]], i32** [[CBP0]]
// CK19-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[RVAR0]] = load i32*, i32** [[VAR0:%[^,]+]]
// CK19-DAG: [[SEC0]] = getelementptr {{.*}}i32* [[RVAR00:%.+]], i{{.+}} %{{.*}}
// CK19-DAG: [[RVAR00]] = load i32*, i32** [[VAR0]]
// CK19: call void [[CALL15:@.+]](i32* {{[^,]+}})
#pragma omp target map(from:pa[ii+12])
{
pa[15]++;
}
// Map of a variable-size array.
int va[ii];
// Region 16
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i[[Z]]* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE16]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i[[Z]]*
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i[[Z]]*
// CK19-DAG: store i[[Z]] {{%.+}}, i[[Z]]* [[CBP0]]
// CK19-DAG: store i[[Z]] {{%.+}}, i[[Z]]* [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i{{.+}} {{8|4}}, i{{.+}}* [[S0]]
// CK19-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK19-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
// CK19-DAG: [[S1:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
// CK19-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
// CK19-DAG: store i32* [[VAR1:%.+]], i32** [[CBP1]]
// CK19-DAG: store i32* [[VAR1]], i32** [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i{{.+}} [[CSVAL1:%[^,]+]], i{{.+}}* [[S1]]
// CK19-DAG: [[CSVAL1]] = mul nuw i{{.+}} %{{.*}}, 4
// CK19: call void [[CALL16:@.+]](i{{.+}} {{[^,]+}}, i32* {{[^,]+}})
#pragma omp target map(to:va)
{
va[50]++;
}
// Region 17
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[SIZE17]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE17]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i[[Z]]*
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i[[Z]]*
// CK19-DAG: store i[[Z]] {{%.+}}, i[[Z]]* [[CBP0]]
// CK19-DAG: store i[[Z]] {{%.+}}, i[[Z]]* [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK19-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
// CK19-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
// CK19-DAG: store i32* [[VAR1:%.+]], i32** [[CBP1]]
// CK19-DAG: store i32* [[SEC1:%.+]], i32** [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[SEC1]] = getelementptr {{.*}}i32* [[VAR1]], i{{.+}} 20
// CK19: call void [[CALL17:@.+]](i{{.+}} {{[^,]+}}, i32* {{[^,]+}})
#pragma omp target map(from:va[20:60])
{
va[50]++;
}
// Region 18
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[SIZE18]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE18]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i[[Z]]*
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i[[Z]]*
// CK19-DAG: store i[[Z]] {{%.+}}, i[[Z]]* [[CBP0]]
// CK19-DAG: store i[[Z]] {{%.+}}, i[[Z]]* [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK19-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
// CK19-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
// CK19-DAG: store i32* [[VAR1:%.+]], i32** [[CBP1]]
// CK19-DAG: store i32* [[SEC1:%.+]], i32** [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[SEC1]] = getelementptr {{.*}}i32* [[VAR1]], i{{.+}} 0
// CK19: call void [[CALL18:@.+]](i{{.+}} {{[^,]+}}, i32* {{[^,]+}})
#pragma omp target map(tofrom:va[:60])
{
va[50]++;
}
// Region 19
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i[[Z]]* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE19]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i[[Z]]*
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i[[Z]]*
// CK19-DAG: store i[[Z]] {{%.+}}, i[[Z]]* [[CBP0]]
// CK19-DAG: store i[[Z]] {{%.+}}, i[[Z]]* [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i{{.+}} {{8|4}}, i{{.+}}* [[S0]]
// CK19-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK19-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
// CK19-DAG: [[S1:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
// CK19-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
// CK19-DAG: store i32* [[VAR1:%.+]], i32** [[CBP1]]
// CK19-DAG: store i32* [[SEC1:%.+]], i32** [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i{{.+}} [[CSVAL1:%[^,]+]], i{{.+}}* [[S1]]
// CK19-DAG: [[CSVAL1]] = mul nuw i{{.+}} %{{.*}}, 4
// CK19-DAG: [[SEC1]] = getelementptr {{.*}}i32* [[VAR1]], i{{.+}} 0
// CK19: call void [[CALL19:@.+]](i{{.+}} {{[^,]+}}, i32* {{[^,]+}})
#pragma omp target map(alloc:va[:])
{
va[50]++;
}
// Region 20
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[SIZE20]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE20]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i[[Z]]*
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i[[Z]]*
// CK19-DAG: store i[[Z]] {{%.+}}, i[[Z]]* [[CBP0]]
// CK19-DAG: store i[[Z]] {{%.+}}, i[[Z]]* [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK19-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
// CK19-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
// CK19-DAG: store i32* [[VAR1:%.+]], i32** [[CBP1]]
// CK19-DAG: store i32* [[SEC1:%.+]], i32** [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[SEC1]] = getelementptr {{.*}}i32* [[VAR1]], i{{.+}} 15
// CK19: call void [[CALL20:@.+]](i{{.+}} {{[^,]+}}, i32* {{[^,]+}})
#pragma omp target map(to:va[15])
{
va[15]++;
}
// Region 21
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i[[Z]]* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE21]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i[[Z]]*
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i[[Z]]*
// CK19-DAG: store i[[Z]] {{%.+}}, i[[Z]]* [[CBP0]]
// CK19-DAG: store i[[Z]] {{%.+}}, i[[Z]]* [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i{{.+}} {{8|4}}, i{{.+}}* [[S0]]
// CK19-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK19-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
// CK19-DAG: [[S1:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
// CK19-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
// CK19-DAG: store i32* [[VAR1:%.+]], i32** [[CBP1]]
// CK19-DAG: store i32* [[SEC1:%.+]], i32** [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i{{.+}} [[CSVAL1:%[^,]+]], i{{.+}}* [[S1]]
// CK19-DAG: [[CSVAL1]] = mul nuw i{{.+}} %{{.*}}, 4
// CK19-DAG: [[SEC1]] = getelementptr {{.*}}i32* [[VAR1]], i{{.+}} %{{.+}}
// CK19: call void [[CALL21:@.+]](i{{.+}} {{[^,]+}}, i32* {{[^,]+}})
#pragma omp target map(tofrom:va[ii:ii+23])
{
va[50]++;
}
// Region 22
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[SIZE22]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE22]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i[[Z]]*
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i[[Z]]*
// CK19-DAG: store i[[Z]] {{%.+}}, i[[Z]]* [[CBP0]]
// CK19-DAG: store i[[Z]] {{%.+}}, i[[Z]]* [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK19-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
// CK19-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
// CK19-DAG: store i32* [[VAR1:%.+]], i32** [[CBP1]]
// CK19-DAG: store i32* [[SEC1:%.+]], i32** [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[SEC1]] = getelementptr {{.*}}i32* [[VAR1]], i{{.+}} %{{.+}}
// CK19: call void [[CALL22:@.+]](i{{.+}} {{[^,]+}}, i32* {{[^,]+}})
#pragma omp target map(tofrom:va[ii])
{
va[15]++;
}
// Always.
// Region 23
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE23]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE23]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK19-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
// CK19-DAG: store i32* [[VAR0]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: call void [[CALL23:@.+]](i32* {{[^,]+}})
#pragma omp target map(always, tofrom: a)
{
a++;
}
// Multidimensional arrays.
int marr[4][5][6];
int ***mptr;
// Region 24
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE24]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE24]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [4 x [5 x [6 x i32]]]**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [4 x [5 x [6 x i32]]]**
// CK19-DAG: store [4 x [5 x [6 x i32]]]* [[VAR0:%.+]], [4 x [5 x [6 x i32]]]** [[CBP0]]
// CK19-DAG: store [4 x [5 x [6 x i32]]]* [[VAR0]], [4 x [5 x [6 x i32]]]** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: call void [[CALL24:@.+]]([4 x [5 x [6 x i32]]]* {{[^,]+}})
#pragma omp target map(tofrom: marr)
{
marr[1][2][3]++;
}
// Region 25
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE25]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE25]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [4 x [5 x [6 x i32]]]**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK19-DAG: store [4 x [5 x [6 x i32]]]* [[VAR0:%.+]], [4 x [5 x [6 x i32]]]** [[CBP0]]
// CK19-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[SEC0]] = getelementptr {{.*}}[6 x i32]* [[SEC00:[^,]+]], i{{.+}} 0, i{{.+}} 2
// CK19-DAG: [[SEC00]] = getelementptr {{.*}}[5 x [6 x i32]]* [[SEC000:[^,]+]], i{{.+}} 0, i{{.+}} 2
// CK19-DAG: [[SEC000]] = getelementptr {{.*}}[4 x [5 x [6 x i32]]]* [[VAR0]], i{{.+}} 0, i{{.+}} 1
// CK19: call void [[CALL25:@.+]]([4 x [5 x [6 x i32]]]* {{[^,]+}})
#pragma omp target map(tofrom: marr[1][2][2:4])
{
marr[1][2][3]++;
}
// Region 26
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE26]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE26]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [4 x [5 x [6 x i32]]]**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK19-DAG: store [4 x [5 x [6 x i32]]]* [[VAR0:%.+]], [4 x [5 x [6 x i32]]]** [[CBP0]]
// CK19-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[SEC0]] = getelementptr {{.*}}[6 x i32]* [[SEC00:[^,]+]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[SEC00]] = getelementptr {{.*}}[5 x [6 x i32]]* [[SEC000:[^,]+]], i{{.+}} 0, i{{.+}} 2
// CK19-DAG: [[SEC000]] = getelementptr {{.*}}[4 x [5 x [6 x i32]]]* [[VAR0]], i{{.+}} 0, i{{.+}} 1
// CK19: call void [[CALL26:@.+]]([4 x [5 x [6 x i32]]]* {{[^,]+}})
#pragma omp target map(tofrom: marr[1][2][:])
{
marr[1][2][3]++;
}
// Region 27
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE27]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE27]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [4 x [5 x [6 x i32]]]**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK19-DAG: store [4 x [5 x [6 x i32]]]* [[VAR0:%.+]], [4 x [5 x [6 x i32]]]** [[CBP0]]
// CK19-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[SEC0]] = getelementptr {{.*}}[6 x i32]* [[SEC00:[^,]+]], i{{.+}} 0, i{{.+}} 3
// CK19-DAG: [[SEC00]] = getelementptr {{.*}}[5 x [6 x i32]]* [[SEC000:[^,]+]], i{{.+}} 0, i{{.+}} 2
// CK19-DAG: [[SEC000]] = getelementptr {{.*}}[4 x [5 x [6 x i32]]]* [[VAR0]], i{{.+}} 0, i{{.+}} 1
// CK19: call void [[CALL27:@.+]]([4 x [5 x [6 x i32]]]* {{[^,]+}})
#pragma omp target map(tofrom: marr[1][2][3])
{
marr[1][2][3]++;
}
// Region 28
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 3, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[3 x i{{.+}}]* [[SIZE28]], {{.+}}getelementptr {{.+}}[3 x i{{.+}}]* [[MTYPE28]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32****
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32****
// CK19-DAG: store i32*** [[VAR0:%.+]], i32**** [[CBP0]]
// CK19-DAG: store i32*** [[SEC0:%.+]], i32**** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[VAR0]] = load i32***, i32**** [[PTR:%[^,]+]],
// CK19-DAG: [[SEC0]] = getelementptr {{.*}}i32*** [[SEC00:[^,]+]], i{{.+}} 1
// CK19-DAG: [[SEC00]] = load i32***, i32**** [[PTR]],
// CK19-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK19-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32****
// CK19-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32***
// CK19-DAG: store i32*** [[SEC0]], i32**** [[CBP1]]
// CK19-DAG: store i32** [[SEC1:%.+]], i32*** [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[SEC1]] = getelementptr {{.*}}i32** [[SEC11:[^,]+]], i{{.+}} 2
// CK19-DAG: [[SEC11]] = load i32**, i32*** [[SEC111:%[^,]+]],
// CK19-DAG: [[SEC111]] = getelementptr {{.*}}i32*** [[SEC1111:[^,]+]], i{{.+}} 1
// CK19-DAG: [[SEC1111]] = load i32***, i32**** [[PTR]],
// CK19-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 2
// CK19-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 2
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to i32***
// CK19-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to i32**
// CK19-DAG: store i32** [[SEC1]], i32*** [[CBP2]]
// CK19-DAG: store i32* [[SEC2:%.+]], i32** [[CP2]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[SEC2]] = getelementptr {{.*}}i32* [[SEC22:[^,]+]], i{{.+}} 2
// CK19-DAG: [[SEC22]] = load i32*, i32** [[SEC222:%[^,]+]],
// CK19-DAG: [[SEC222]] = getelementptr {{.*}}i32** [[SEC2222:[^,]+]], i{{.+}} 2
// CK19-DAG: [[SEC2222]] = load i32**, i32*** [[SEC22222:%[^,]+]],
// CK19-DAG: [[SEC22222]] = getelementptr {{.*}}i32*** [[SEC222222:[^,]+]], i{{.+}} 1
// CK19-DAG: [[SEC222222]] = load i32***, i32**** [[PTR]],
// CK19: call void [[CALL28:@.+]](i32*** {{[^,]+}})
#pragma omp target map(tofrom: mptr[1][2][2:4])
{
mptr[1][2][3]++;
}
// Region 29
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 3, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[3 x i{{.+}}]* [[SIZE29]], {{.+}}getelementptr {{.+}}[3 x i{{.+}}]* [[MTYPE29]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32****
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32****
// CK19-DAG: store i32*** [[VAR0:%.+]], i32**** [[CBP0]]
// CK19-DAG: store i32*** [[SEC0:%.+]], i32**** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[VAR0]] = load i32***, i32**** [[PTR:%[^,]+]],
// CK19-DAG: [[SEC0]] = getelementptr {{.*}}i32*** [[SEC00:[^,]+]], i{{.+}} 1
// CK19-DAG: [[SEC00]] = load i32***, i32**** [[PTR]],
// CK19-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK19-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32****
// CK19-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32***
// CK19-DAG: store i32*** [[SEC0]], i32**** [[CBP1]]
// CK19-DAG: store i32** [[SEC1:%.+]], i32*** [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[SEC1]] = getelementptr {{.*}}i32** [[SEC11:[^,]+]], i{{.+}} 2
// CK19-DAG: [[SEC11]] = load i32**, i32*** [[SEC111:%[^,]+]],
// CK19-DAG: [[SEC111]] = getelementptr {{.*}}i32*** [[SEC1111:[^,]+]], i{{.+}} 1
// CK19-DAG: [[SEC1111]] = load i32***, i32**** [[PTR]],
// CK19-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 2
// CK19-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 2
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to i32***
// CK19-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to i32**
// CK19-DAG: store i32** [[SEC1]], i32*** [[CBP2]]
// CK19-DAG: store i32* [[SEC2:%.+]], i32** [[CP2]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[SEC2]] = getelementptr {{.*}}i32* [[SEC22:[^,]+]], i{{.+}} 3
// CK19-DAG: [[SEC22]] = load i32*, i32** [[SEC222:%[^,]+]],
// CK19-DAG: [[SEC222]] = getelementptr {{.*}}i32** [[SEC2222:[^,]+]], i{{.+}} 2
// CK19-DAG: [[SEC2222]] = load i32**, i32*** [[SEC22222:%[^,]+]],
// CK19-DAG: [[SEC22222]] = getelementptr {{.*}}i32*** [[SEC222222:[^,]+]], i{{.+}} 1
// CK19-DAG: [[SEC222222]] = load i32***, i32**** [[PTR]],
// CK19: call void [[CALL29:@.+]](i32*** {{[^,]+}})
#pragma omp target map(tofrom: mptr[1][2][3])
{
mptr[1][2][3]++;
}
// Multidimensional VLA.
double mva[23][ii][ii+5];
// Region 30
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 4, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i[[Z]]* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[4 x i{{.+}}]* [[MTYPE30]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
//
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i[[Z]]*
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i[[Z]]*
// CK19-DAG: store i[[Z]] 23, i[[Z]]* [[CBP0]]
// CK19-DAG: store i[[Z]] 23, i[[Z]]* [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i[[Z]] {{8|4}}, i[[Z]]* [[S0]]
//
// CK19-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK19-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
// CK19-DAG: [[S1:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[Z]]*
// CK19-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[Z]]*
// CK19-DAG: store i[[Z]] [[VAR1:%.+]], i[[Z]]* [[CBP1]]
// CK19-DAG: store i[[Z]] [[VAR11:%.+]], i[[Z]]* [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i[[Z]] {{8|4}}, i[[Z]]* [[S1]]
// CK19-64-DAG: [[VAR1]] = zext i32 %{{[^,]+}} to i64
// CK19-64-DAG: [[VAR11]] = zext i32 %{{[^,]+}} to i64
//
// CK19-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 2
// CK19-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 2
// CK19-DAG: [[S2:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 2
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to i[[Z]]*
// CK19-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to i[[Z]]*
// CK19-DAG: store i[[Z]] [[VAR2:%.+]], i[[Z]]* [[CBP2]]
// CK19-DAG: store i[[Z]] [[VAR22:%.+]], i[[Z]]* [[CP2]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i[[Z]] {{8|4}}, i[[Z]]* [[S2]]
// CK19-64-DAG: [[VAR2]] = zext i32 %{{[^,]+}} to i64
// CK19-64-DAG: [[VAR22]] = zext i32 %{{[^,]+}} to i64
//
// CK19-DAG: [[BP3:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 3
// CK19-DAG: [[P3:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 3
// CK19-DAG: [[S3:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 3
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP3:%.+]] = bitcast i8** [[BP3]] to double**
// CK19-DAG: [[CP3:%.+]] = bitcast i8** [[P3]] to double**
// CK19-DAG: store double* [[VAR3:%.+]], double** [[CBP3]]
// CK19-DAG: store double* [[VAR3]], double** [[CP3]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i[[Z]] [[CSVAL3:%[^,]+]], i[[Z]]* [[S3]]
// CK19-DAG: [[CSVAL3]] = mul nuw i[[Z]] %{{[^,]+}}, {{8|4}}
// CK19: call void [[CALL30:@.+]](i[[Z]] 23, i[[Z]] %{{[^,]+}}, i[[Z]] %{{[^,]+}}, double* %{{[^,]+}})
#pragma omp target map(tofrom: mva)
{
mva[1][2][3]++;
}
// Region 31
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 4, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[4 x i{{.+}}]* [[SIZE31]], {{.+}}getelementptr {{.+}}[4 x i{{.+}}]* [[MTYPE31]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
//
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i[[Z]]*
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i[[Z]]*
// CK19-DAG: store i[[Z]] 23, i[[Z]]* [[CBP0]]
// CK19-DAG: store i[[Z]] 23, i[[Z]]* [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
//
// CK19-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK19-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[Z]]*
// CK19-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[Z]]*
// CK19-DAG: store i[[Z]] [[VAR1:%.+]], i[[Z]]* [[CBP1]]
// CK19-DAG: store i[[Z]] [[VAR11:%.+]], i[[Z]]* [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
//
// CK19-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 2
// CK19-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 2
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to i[[Z]]*
// CK19-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to i[[Z]]*
// CK19-DAG: store i[[Z]] [[VAR2:%.+]], i[[Z]]* [[CBP2]]
// CK19-DAG: store i[[Z]] [[VAR22:%.+]], i[[Z]]* [[CP2]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
//
// CK19-DAG: [[BP3:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 3
// CK19-DAG: [[P3:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 3
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP3:%.+]] = bitcast i8** [[BP3]] to double**
// CK19-DAG: [[CP3:%.+]] = bitcast i8** [[P3]] to double**
// CK19-DAG: store double* [[VAR3:%.+]], double** [[CBP3]]
// CK19-DAG: store double* [[SEC3:%.+]], double** [[CP3]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[SEC3]] = getelementptr {{.*}}double* [[SEC33:%.+]], i[[Z]] 0
// CK19-DAG: [[SEC33]] = getelementptr {{.*}}double* [[SEC333:%.+]], i[[Z]] [[IDX3:%.+]]
// CK19-DAG: [[IDX3]] = mul nsw i[[Z]] %{{[^,]+}}, %{{[^,]+}}
// CK19-DAG: [[SEC333]] = getelementptr {{.*}}double* [[VAR3]], i[[Z]] [[IDX33:%.+]]
// CK19-DAG: [[IDX33]] = mul nsw i[[Z]] 1, %{{[^,]+}}
// CK19: call void [[CALL31:@.+]](i[[Z]] 23, i[[Z]] %{{[^,]+}}, i[[Z]] %{{[^,]+}}, double* %{{[^,]+}})
#pragma omp target map(tofrom: mva[1][ii-2][:5])
{
mva[1][2][3]++;
}
// Multidimensional array sections.
double marras[11][12][13];
double mvlaas[11][ii][13];
double ***mptras;
// Region 32
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE32]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE32]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [11 x [12 x [13 x double]]]**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [11 x [12 x [13 x double]]]**
// CK19-DAG: store [11 x [12 x [13 x double]]]* [[VAR0:%.+]], [11 x [12 x [13 x double]]]** [[CBP0]]
// CK19-DAG: store [11 x [12 x [13 x double]]]* [[VAR0]], [11 x [12 x [13 x double]]]** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: call void [[CALL32:@.+]]([11 x [12 x [13 x double]]]* {{[^,]+}})
#pragma omp target map(marras)
{
marras[1][2][3]++;
}
// Region 33
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE33]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE33]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [11 x [12 x [13 x double]]]**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [12 x [13 x double]]**
// CK19-DAG: store [11 x [12 x [13 x double]]]* [[VAR0:%.+]], [11 x [12 x [13 x double]]]** [[CBP0]]
// CK19-DAG: store [12 x [13 x double]]* [[SEC0:%.+]], [12 x [13 x double]]** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[SEC0]] = getelementptr {{.+}}[11 x [12 x [13 x double]]]* [[VAR0]], i[[Z]] 0, i[[Z]] 0
// CK19: call void [[CALL33:@.+]]([11 x [12 x [13 x double]]]* {{[^,]+}})
#pragma omp target map(marras[:])
{
marras[1][2][3]++;
}
// Region 34
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE34]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE34]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [11 x [12 x [13 x double]]]**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [12 x [13 x double]]**
// CK19-DAG: store [11 x [12 x [13 x double]]]* [[VAR0:%.+]], [11 x [12 x [13 x double]]]** [[CBP0]]
// CK19-DAG: store [12 x [13 x double]]* [[SEC0:%.+]], [12 x [13 x double]]** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[SEC0]] = getelementptr {{.+}}[11 x [12 x [13 x double]]]* [[VAR0]], i[[Z]] 0, i[[Z]] 0
// CK19: call void [[CALL34:@.+]]([11 x [12 x [13 x double]]]* {{[^,]+}})
#pragma omp target map(marras[:][:][:])
{
marras[1][2][3]++;
}
// Region 35
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i[[Z]]* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE35]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
//
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [11 x [12 x [13 x double]]]**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [13 x double]**
// CK19-DAG: store [11 x [12 x [13 x double]]]* [[VAR0:%.+]], [11 x [12 x [13 x double]]]** [[CBP0]]
// CK19-DAG: store [13 x double]* [[SEC0:%.+]], [13 x double]** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i[[Z]] [[CSVAL0:%[^,]+]], i[[Z]]* [[S0]]
// CK19-DAG: [[SEC0]] = getelementptr {{.+}}[12 x [13 x double]]* [[SEC00:%[^,]+]], i[[Z]] 0, i[[Z]] 0
// CK19-DAG: [[SEC00]] = getelementptr {{.+}}[11 x [12 x [13 x double]]]* [[VAR0]], i[[Z]] 0, i[[Z]] 1
// CK19-DAG: [[CSVAL0]] = mul nuw i[[Z]] %{{[^,]+}}, 104
// CK19: call void [[CALL35:@.+]]([11 x [12 x [13 x double]]]* {{[^,]+}})
#pragma omp target map(marras[1][:ii][:])
{
marras[1][2][3]++;
}
// Region 36
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE36]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE36]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [11 x [12 x [13 x double]]]**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [13 x double]**
// CK19-DAG: store [11 x [12 x [13 x double]]]* [[VAR0:%.+]], [11 x [12 x [13 x double]]]** [[CBP0]]
// CK19-DAG: store [13 x double]* [[SEC0:%.+]], [13 x double]** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[SEC0]] = getelementptr {{.+}}[13 x double]* [[SEC00:%[^,]+]], i{{.+}} 0
// CK19-DAG: [[SEC00]] = getelementptr {{.+}}[12 x [13 x double]]* [[SEC000:%[^,]+]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[SEC000]] = getelementptr {{.+}}[11 x [12 x [13 x double]]]* [[VAR0]], i{{.+}} 0, i{{.+}} 0
// CK19: call void [[CALL36:@.+]]([11 x [12 x [13 x double]]]* {{[^,]+}})
#pragma omp target map(marras[:1][:2][:13])
{
marras[1][2][3]++;
}
// Region 37
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 3, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i[[Z]]* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[3 x i{{.+}}]* [[MTYPE37]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
//
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i[[Z]]*
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i[[Z]]*
// CK19-DAG: store i[[Z]] 11, i[[Z]]* [[CBP0]]
// CK19-DAG: store i[[Z]] 11, i[[Z]]* [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i[[Z]] {{8|4}}, i[[Z]]* [[S0]]
//
// CK19-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK19-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
// CK19-DAG: [[S1:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[Z]]*
// CK19-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[Z]]*
// CK19-DAG: store i[[Z]] [[VAR1:%.+]], i[[Z]]* [[CBP1]]
// CK19-DAG: store i[[Z]] [[VAR11:%.+]], i[[Z]]* [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i[[Z]] {{8|4}}, i[[Z]]* [[S1]]
//
// CK19-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 2
// CK19-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 2
// CK19-DAG: [[S2:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 2
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to [13 x double]**
// CK19-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to [13 x double]**
// CK19-DAG: store [13 x double]* [[VAR2:%.+]], [13 x double]** [[CBP2]]
// CK19-DAG: store [13 x double]* [[VAR2]], [13 x double]** [[CP2]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i[[Z]] [[CSVAL2:%[^,]+]], i[[Z]]* [[S2]]
// CK19-DAG: [[CSVAL2]] = mul nuw i[[Z]] %{{[^,]+}}, 104
// CK19: call void [[CALL37:@.+]](i[[Z]] 11, i[[Z]] %{{[^,]+}}, [13 x double]* %{{[^,]+}})
#pragma omp target map(mvlaas)
{
mvlaas[1][2][3]++;
}
// Region 38
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 3, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i[[Z]]* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[3 x i{{.+}}]* [[MTYPE38]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
//
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i[[Z]]*
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i[[Z]]*
// CK19-DAG: store i[[Z]] 11, i[[Z]]* [[CBP0]]
// CK19-DAG: store i[[Z]] 11, i[[Z]]* [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i[[Z]] {{8|4}}, i[[Z]]* [[S0]]
//
// CK19-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK19-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
// CK19-DAG: [[S1:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[Z]]*
// CK19-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[Z]]*
// CK19-DAG: store i[[Z]] [[VAR1:%.+]], i[[Z]]* [[CBP1]]
// CK19-DAG: store i[[Z]] [[VAR11:%.+]], i[[Z]]* [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i[[Z]] {{8|4}}, i[[Z]]* [[S1]]
//
// CK19-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 2
// CK19-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 2
// CK19-DAG: [[S2:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 2
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to [13 x double]**
// CK19-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to [13 x double]**
// CK19-DAG: store [13 x double]* [[VAR2:%.+]], [13 x double]** [[CBP2]]
// CK19-DAG: store [13 x double]* [[SEC2:%.+]], [13 x double]** [[CP2]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i[[Z]] [[CSVAL2:%[^,]+]], i[[Z]]* [[S2]]
// CK19-DAG: [[SEC2]] = getelementptr {{.+}}[13 x double]* [[VAR2]], i[[Z]] [[SEC22:%[^,]+]]
// CK19-DAG: [[SEC22]] = mul nsw i[[Z]] 0, %{{[^,]+}}
// CK19-DAG: [[CSVAL2]] = mul nuw i[[Z]] %{{[^,]+}}, 104
// CK19: call void [[CALL38:@.+]](i[[Z]] 11, i[[Z]] %{{[^,]+}}, [13 x double]* %{{[^,]+}})
#pragma omp target map(mvlaas[:])
{
mvlaas[1][2][3]++;
}
// Region 39
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 3, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i[[Z]]* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[3 x i{{.+}}]* [[MTYPE39]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
//
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i[[Z]]*
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i[[Z]]*
// CK19-DAG: store i[[Z]] 11, i[[Z]]* [[CBP0]]
// CK19-DAG: store i[[Z]] 11, i[[Z]]* [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i[[Z]] {{8|4}}, i[[Z]]* [[S0]]
//
// CK19-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK19-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
// CK19-DAG: [[S1:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[Z]]*
// CK19-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[Z]]*
// CK19-DAG: store i[[Z]] [[VAR1:%.+]], i[[Z]]* [[CBP1]]
// CK19-DAG: store i[[Z]] [[VAR11:%.+]], i[[Z]]* [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i[[Z]] {{8|4}}, i[[Z]]* [[S1]]
//
// CK19-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 2
// CK19-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 2
// CK19-DAG: [[S2:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 2
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to [13 x double]**
// CK19-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to [13 x double]**
// CK19-DAG: store [13 x double]* [[VAR2:%.+]], [13 x double]** [[CBP2]]
// CK19-DAG: store [13 x double]* [[SEC2:%.+]], [13 x double]** [[CP2]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i[[Z]] [[CSVAL2:%[^,]+]], i[[Z]]* [[S2]]
// CK19-DAG: [[SEC2]] = getelementptr {{.+}}[13 x double]* [[VAR2]], i[[Z]] [[SEC22:%[^,]+]]
// CK19-DAG: [[SEC22]] = mul nsw i[[Z]] 0, %{{[^,]+}}
// CK19-DAG: [[CSVAL2]] = mul nuw i[[Z]] %{{[^,]+}}, 104
// CK19: call void [[CALL39:@.+]](i[[Z]] 11, i[[Z]] %{{[^,]+}}, [13 x double]* %{{[^,]+}})
#pragma omp target map(mvlaas[:][:][:])
{
mvlaas[1][2][3]++;
}
// Region 40
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 3, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i[[Z]]* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[3 x i{{.+}}]* [[MTYPE40]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
//
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i[[Z]]*
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i[[Z]]*
// CK19-DAG: store i[[Z]] 11, i[[Z]]* [[CBP0]]
// CK19-DAG: store i[[Z]] 11, i[[Z]]* [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i[[Z]] {{8|4}}, i[[Z]]* [[S0]]
//
// CK19-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK19-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
// CK19-DAG: [[S1:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[Z]]*
// CK19-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[Z]]*
// CK19-DAG: store i[[Z]] [[VAR1:%.+]], i[[Z]]* [[CBP1]]
// CK19-DAG: store i[[Z]] [[VAR11:%.+]], i[[Z]]* [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i[[Z]] {{8|4}}, i[[Z]]* [[S1]]
//
// CK19-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 2
// CK19-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 2
// CK19-DAG: [[S2:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 2
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to [13 x double]**
// CK19-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to [13 x double]**
// CK19-DAG: store [13 x double]* [[VAR2:%.+]], [13 x double]** [[CBP2]]
// CK19-DAG: store [13 x double]* [[SEC2:%.+]], [13 x double]** [[CP2]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i[[Z]] [[CSVAL2:%[^,]+]], i[[Z]]* [[S2]]
// CK19-DAG: [[SEC2]] = getelementptr {{.+}}[13 x double]* [[SEC22:%[^,]+]], i[[Z]] 0
// CK19-DAG: [[SEC22]] = getelementptr {{.+}}[13 x double]* [[VAR2]], i[[Z]] [[SEC222:%[^,]+]]
// CK19-DAG: [[SEC222]] = mul nsw i[[Z]] 1, %{{[^,]+}}
// CK19: call void [[CALL40:@.+]](i[[Z]] 11, i[[Z]] %{{[^,]+}}, [13 x double]* %{{[^,]+}})
#pragma omp target map(mvlaas[1][:ii][:])
{
mvlaas[1][2][3]++;
}
// Region 41
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 3, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[3 x i{{.+}}]* [[SIZE41]], {{.+}}getelementptr {{.+}}[3 x i{{.+}}]* [[MTYPE41]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
//
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i[[Z]]*
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i[[Z]]*
// CK19-DAG: store i[[Z]] 11, i[[Z]]* [[CBP0]]
// CK19-DAG: store i[[Z]] 11, i[[Z]]* [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
//
// CK19-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK19-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[Z]]*
// CK19-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[Z]]*
// CK19-DAG: store i[[Z]] [[VAR1:%.+]], i[[Z]]* [[CBP1]]
// CK19-DAG: store i[[Z]] [[VAR11:%.+]], i[[Z]]* [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
//
// CK19-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 2
// CK19-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 2
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to [13 x double]**
// CK19-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to [13 x double]**
// CK19-DAG: store [13 x double]* [[VAR2:%.+]], [13 x double]** [[CBP2]]
// CK19-DAG: store [13 x double]* [[SEC2:%.+]], [13 x double]** [[CP2]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[SEC2]] = getelementptr {{.+}}[13 x double]* [[SEC22:%[^,]+]], i[[Z]] 0
// CK19-DAG: [[SEC22]] = getelementptr {{.+}}[13 x double]* [[VAR2]], i[[Z]] [[SEC222:%[^,]+]]
// CK19-DAG: [[SEC222]] = mul nsw i[[Z]] 0, %{{[^,]+}}
// CK19: call void [[CALL41:@.+]](i[[Z]] 11, i[[Z]] %{{[^,]+}}, [13 x double]* %{{[^,]+}})
#pragma omp target map(mvlaas[:1][:2][:13])
{
mvlaas[1][2][3]++;
}
// Region 42
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 3, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[3 x i{{.+}}]* [[SIZE42]], {{.+}}getelementptr {{.+}}[3 x i{{.+}}]* [[MTYPE42]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to double****
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to double****
// CK19-DAG: store double*** [[VAR0:%.+]], double**** [[CBP0]]
// CK19-DAG: store double*** [[SEC0:%.+]], double**** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[VAR0]] = load double***, double**** [[PTR:%[^,]+]],
// CK19-DAG: [[SEC0]] = getelementptr {{.*}}double*** [[SEC00:[^,]+]], i{{.+}} 0
// CK19-DAG: [[SEC00]] = load double***, double**** [[PTR]],
// CK19-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK19-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to double****
// CK19-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to double***
// CK19-DAG: store double*** [[SEC0]], double**** [[CBP1]]
// CK19-DAG: store double** [[SEC1:%.+]], double*** [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[SEC1]] = getelementptr {{.*}}double** [[SEC11:[^,]+]], i{{.+}} 2
// CK19-DAG: [[SEC11]] = load double**, double*** [[SEC111:%[^,]+]],
// CK19-DAG: [[SEC111]] = getelementptr {{.*}}double*** [[SEC1111:[^,]+]], i{{.+}} 0
// CK19-DAG: [[SEC1111]] = load double***, double**** [[PTR]],
// CK19-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 2
// CK19-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 2
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to double***
// CK19-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to double**
// CK19-DAG: store double** [[SEC1]], double*** [[CBP2]]
// CK19-DAG: store double* [[SEC2:%.+]], double** [[CP2]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: [[SEC2]] = getelementptr {{.*}}double* [[SEC22:[^,]+]], i{{.+}} 0
// CK19-DAG: [[SEC22]] = load double*, double** [[SEC222:%[^,]+]],
// CK19-DAG: [[SEC222]] = getelementptr {{.*}}double** [[SEC2222:[^,]+]], i{{.+}} 2
// CK19-DAG: [[SEC2222]] = load double**, double*** [[SEC22222:%[^,]+]],
// CK19-DAG: [[SEC22222]] = getelementptr {{.*}}double*** [[SEC222222:[^,]+]], i{{.+}} 0
// CK19-DAG: [[SEC222222]] = load double***, double**** [[PTR]],
// CK19: call void [[CALL42:@.+]](double*** {{[^,]+}})
#pragma omp target map(mptras[:1][2][:13])
{
mptras[1][2][3]++;
}
// Region 43 - the memory is not contiguous for this map - will map the whole last dimension.
// CK19-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i[[Z]]* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE43]]{{.+}})
// CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
//
// CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
// CK19-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK19-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [11 x [12 x [13 x double]]]**
// CK19-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [13 x double]**
// CK19-DAG: store [11 x [12 x [13 x double]]]* [[VAR0:%.+]], [11 x [12 x [13 x double]]]** [[CBP0]]
// CK19-DAG: store [13 x double]* [[SEC0:%.+]], [13 x double]** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19-DAG: store i[[Z]] [[CSVAL0:%[^,]+]], i[[Z]]* [[S0]]
// CK19-DAG: [[SEC0]] = getelementptr {{.+}}[12 x [13 x double]]* [[SEC00:%[^,]+]], i[[Z]] 0, i[[Z]] 0
// CK19-DAG: [[SEC00]] = getelementptr {{.+}}[11 x [12 x [13 x double]]]* [[VAR0]], i[[Z]] 0, i[[Z]] 1
// CK19-DAG: [[CSVAL0]] = mul nuw i[[Z]] %{{[^,]+}}, 104
// CK19: call void [[CALL43:@.+]]([11 x [12 x [13 x double]]]* {{[^,]+}})
#pragma omp target map(marras[1][:ii][1:])
{
marras[1][2][3]++;
}
}
// CK19: define {{.+}}[[CALL00]]
[OpenMP] Fix mapping of scalars for combined directives Combined directives like 'target parallel' have two captured statements. Sema has to check the right one from the right direction. Previously, Sema::IsOpenMPCapturedByRef would return false for mapped scalars on combined directives. This results in a wrong signature of the outlined function which triggers an assertion: void llvm::CallInst::init(llvm::FunctionType *, llvm::Value *, ArrayRef<llvm::Value *>, ArrayRef<OperandBundleDef>, const llvm::Twine &): Assertion `(i >= FTy->getNumParams() || FTy->getParamType(i) == Args[i]->getType()) && "Calling a function with a bad signature!"' failed. Fixes PR30975 (and PR31985). New function was taken from clang-ykt. Differential Revision: https://reviews.llvm.org/D34888 llvm-svn: 306956
2017-07-01 18:40:50 +08:00
// CK19: define {{.+}}[[CALL00n]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK19: define {{.+}}[[CALL01]]
// CK19: define {{.+}}[[CALL02]]
// CK19: define {{.+}}[[CALL03]]
// CK19: define {{.+}}[[CALL04]]
// CK19: define {{.+}}[[CALL05]]
// CK19: define {{.+}}[[CALL06]]
// CK19: define {{.+}}[[CALL07]]
// CK19: define {{.+}}[[CALL08]]
// CK19: define {{.+}}[[CALL09]]
// CK19: define {{.+}}[[CALL10]]
// CK19: define {{.+}}[[CALL11]]
// CK19: define {{.+}}[[CALL12]]
// CK19: define {{.+}}[[CALL13]]
// CK19: define {{.+}}[[CALL14]]
// CK19: define {{.+}}[[CALL15]]
// CK19: define {{.+}}[[CALL16]]
// CK19: define {{.+}}[[CALL17]]
// CK19: define {{.+}}[[CALL18]]
// CK19: define {{.+}}[[CALL19]]
// CK19: define {{.+}}[[CALL20]]
// CK19: define {{.+}}[[CALL21]]
// CK19: define {{.+}}[[CALL22]]
// CK19: define {{.+}}[[CALL23]]
// CK19: define {{.+}}[[CALL24]]
// CK19: define {{.+}}[[CALL25]]
// CK19: define {{.+}}[[CALL26]]
// CK19: define {{.+}}[[CALL27]]
// CK19: define {{.+}}[[CALL28]]
// CK19: define {{.+}}[[CALL29]]
// CK19: define {{.+}}[[CALL30]]
// CK19: define {{.+}}[[CALL31]]
// CK19: define {{.+}}[[CALL32]]
// CK19: define {{.+}}[[CALL33]]
// CK19: define {{.+}}[[CALL34]]
// CK19: define {{.+}}[[CALL35]]
// CK19: define {{.+}}[[CALL36]]
// CK19: define {{.+}}[[CALL37]]
// CK19: define {{.+}}[[CALL38]]
// CK19: define {{.+}}[[CALL39]]
// CK19: define {{.+}}[[CALL40]]
// CK19: define {{.+}}[[CALL41]]
// CK19: define {{.+}}[[CALL42]]
// CK19: define {{.+}}[[CALL43]]
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK20 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK20 --check-prefix CK20-64
// RUN: %clang_cc1 -DCK20 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK20 --check-prefix CK20-64
// RUN: %clang_cc1 -DCK20 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK20 --check-prefix CK20-32
// RUN: %clang_cc1 -DCK20 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK20 --check-prefix CK20-32
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
#ifdef CK20
// CK20: [[SIZE00:@.+]] = private {{.*}}constant [1 x i[[Z:64|32]]] [i[[Z:64|32]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK20: [[MTYPE00:@.+]] = private {{.*}}constant [1 x i32] [i32 33]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK20: [[SIZE01:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 20]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK20: [[MTYPE01:@.+]] = private {{.*}}constant [1 x i32] [i32 33]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK20: [[SIZE02:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK20: [[MTYPE02:@.+]] = private {{.*}}constant [1 x i32] [i32 34]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK20: [[SIZE03:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 12]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK20: [[MTYPE03:@.+]] = private {{.*}}constant [1 x i32] [i32 34]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK20-LABEL: explicit_maps_references_and_function_args
void explicit_maps_references_and_function_args (int a, float b, int (&c)[10], float *d){
int &aa = a;
float &bb = b;
int (&cc)[10] = c;
float *&dd = d;
// Region 00
// CK20-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE00]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE00]]{{.+}})
// CK20-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK20-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK20-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK20-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK20-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
// CK20-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK20-DAG: store i32* [[RVAR0:%.+]], i32** [[CBP0]]
// CK20-DAG: store i32* [[RVAR00:%.+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK20-DAG: [[RVAR0]] = load i32*, i32** [[VAR0:%[^,]+]]
// CK20-DAG: [[RVAR00]] = load i32*, i32** [[VAR0]]
// CK20: call void [[CALL00:@.+]](i32* {{[^,]+}})
#pragma omp target map(to:aa)
{
aa += 1;
}
// Region 01
// CK20-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE01]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE01]]{{.+}})
// CK20-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK20-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK20-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK20-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK20-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [10 x i32]**
// CK20-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK20-DAG: store [10 x i32]* [[RVAR0:%.+]], [10 x i32]** [[CBP0]]
// CK20-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK20-DAG: [[SEC0]] = getelementptr {{.*}}[10 x i32]* [[RVAR00:%.+]], i{{.+}} 0, i{{.+}} 0
// CK20-DAG: [[RVAR0]] = load [10 x i32]*, [10 x i32]** [[VAR0:%[^,]+]]
// CK20-DAG: [[RVAR00]] = load [10 x i32]*, [10 x i32]** [[VAR0]]
// CK20: call void [[CALL01:@.+]]([10 x i32]* {{[^,]+}})
#pragma omp target map(to:cc[:5])
{
cc[3] += 1;
}
// Region 02
// CK20-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE02]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE02]]{{.+}})
// CK20-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK20-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK20-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK20-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK20-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to float**
// CK20-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to float**
// CK20-DAG: store float* [[VAR0:%.+]], float** [[CBP0]]
// CK20-DAG: store float* [[VAR0]], float** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK20: call void [[CALL02:@.+]](float* {{[^,]+}})
#pragma omp target map(from:b)
{
b += 1.0f;
}
// Region 03
// CK20-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE03]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE03]]{{.+}})
// CK20-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK20-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK20-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK20-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK20-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to float**
// CK20-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to float**
// CK20-DAG: store float* [[RVAR0:%.+]], float** [[CBP0]]
// CK20-DAG: store float* [[SEC0:%.+]], float** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK20-DAG: [[RVAR0]] = load float*, float** [[VAR0:%[^,]+]]
// CK20-DAG: [[SEC0]] = getelementptr {{.*}}float* [[RVAR00:%.+]], i{{.+}} 2
// CK20-DAG: [[RVAR00]] = load float*, float** [[VAR0]]
// CK20: call void [[CALL03:@.+]](float* {{[^,]+}})
#pragma omp target map(from:d[2:3])
{
d[2] += 1.0f;
}
}
// CK20: define {{.+}}[[CALL00]]
// CK20: define {{.+}}[[CALL01]]
// CK20: define {{.+}}[[CALL02]]
// CK20: define {{.+}}[[CALL03]]
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK21 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK21 --check-prefix CK21-64
// RUN: %clang_cc1 -DCK21 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK21 --check-prefix CK21-64
// RUN: %clang_cc1 -DCK21 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK21 --check-prefix CK21-32
// RUN: %clang_cc1 -DCK21 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK21 --check-prefix CK21-32
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
#ifdef CK21
// CK21: [[ST:%.+]] = type { i32, i32, float* }
// CK21: [[SIZE00:@.+]] = private {{.*}}constant [1 x i[[Z:64|32]]] [i[[Z:64|32]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK21: [[MTYPE00:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK21: [[SIZE01:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 492]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK21: [[MTYPE01:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK21: [[SIZE02:@.+]] = private {{.*}}constant [2 x i[[Z]]] [i[[Z]] {{8|4}}, i[[Z]] 500]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK21: [[MTYPE02:@.+]] = private {{.*}}constant [2 x i32] [i32 34, i32 18]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK21: [[SIZE03:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 492]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK21: [[MTYPE03:@.+]] = private {{.*}}constant [1 x i32] [i32 34]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK21: [[SIZE04:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK21: [[MTYPE04:@.+]] = private {{.*}}constant [1 x i32] [i32 34]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK21: [[SIZE05:@.+]] = private {{.*}}constant [2 x i[[Z]]] [i[[Z]] 4, i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK21: [[MTYPE05:@.+]] = private {{.*}}constant [2 x i32] [i32 35, i32 3]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK21-LABEL: explicit_maps_template_args_and_members
template <int X, typename T>
struct CC {
T A;
int A2;
float *B;
int foo(T arg) {
float la[X];
T *lb;
// Region 00
// CK21-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE00]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE00]]{{.+}})
// CK21-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK21-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK21-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK21-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK21-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
// CK21-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK21-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
// CK21-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK21-DAG: [[SEC0]] = getelementptr {{.*}}[[ST]]* [[VAR0:%.+]], i{{.+}} 0, i{{.+}} 0
// CK21: call void [[CALL00:@.+]]([[ST]]* {{[^,]+}})
#pragma omp target map(A)
{
A += 1;
}
// Region 01
// CK21-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE01]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE01]]{{.+}})
// CK21-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK21-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK21-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK21-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK21-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
// CK21-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK21-DAG: store i32* [[RVAR0:%.+]], i32** [[CBP0]]
// CK21-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK21-DAG: [[RVAR0]] = load i32*, i32** [[VAR0:%[^,]+]]
// CK21-DAG: [[SEC0]] = getelementptr {{.*}}i32* [[RVAR00:%.+]], i{{.+}} 0
// CK21-DAG: [[RVAR00]] = load i32*, i32** [[VAR0]]
// CK21: call void [[CALL01:@.+]](i32* {{[^,]+}})
#pragma omp target map(lb[:X])
{
lb[4] += 1;
}
// Region 02
// CK21-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[SIZE02]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE02]]{{.+}})
// CK21-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK21-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK21-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK21-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK21-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
// CK21-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to float***
// CK21-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
// CK21-DAG: store float** [[SEC0:%.+]], float*** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK21-DAG: [[SEC0]] = getelementptr {{.*}}[[ST]]* [[VAR0]], i{{.+}} 0, i{{.+}} 2
// CK21-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK21-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK21-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to float***
// CK21-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to float**
// CK21-DAG: store float** [[SEC0]], float*** [[CBP1]]
// CK21-DAG: store float* [[SEC1:%.+]], float** [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK21-DAG: [[SEC1]] = getelementptr {{.*}}float* [[RVAR1:%[^,]+]], i{{.+}} 123
// CK21-DAG: [[RVAR1]] = load float*, float** [[SEC1_:%[^,]+]]
// CK21-DAG: [[SEC1_]] = getelementptr {{.*}}[[ST]]* [[VAR0]], i{{.+}} 0, i{{.+}} 2
// CK21: call void [[CALL02:@.+]]([[ST]]* {{[^,]+}})
#pragma omp target map(from:B[X:X+2])
{
B[2] += 1.0f;
}
// Region 03
// CK21-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE03]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE03]]{{.+}})
// CK21-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK21-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK21-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK21-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK21-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [123 x float]**
// CK21-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [123 x float]**
// CK21-DAG: store [123 x float]* [[VAR0:%.+]], [123 x float]** [[CBP0]]
// CK21-DAG: store [123 x float]* [[VAR0]], [123 x float]** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK21: call void [[CALL03:@.+]]([123 x float]* {{[^,]+}})
#pragma omp target map(from:la)
{
la[3] += 1.0f;
}
// Region 04
// CK21-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE04]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE04]]{{.+}})
// CK21-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK21-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK21-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK21-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK21-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
// CK21-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK21-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
// CK21-DAG: store i32* [[VAR0]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK21: call void [[CALL04:@.+]](i32* {{[^,]+}})
#pragma omp target map(from:arg)
{
arg +=1;
}
// Make sure the extra flag is passed to the second map.
// Region 05
// CK21-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[SIZE05]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE05]]{{.+}})
// CK21-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK21-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK21-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK21-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK21-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
// CK21-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK21-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
// CK21-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK21-DAG: [[SEC0]] = getelementptr {{.*}}[[ST]]* [[VAR0]], i{{.+}} 0, i{{.+}} 0
// CK21-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK21-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK21-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[ST]]**
// CK21-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
// CK21-DAG: store [[ST]]* [[VAR1:%.+]], [[ST]]** [[CBP1]]
// CK21-DAG: store i32* [[SEC1:%.+]], i32** [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK21-DAG: [[SEC1]] = getelementptr {{.*}}[[ST]]* [[VAR0]], i{{.+}} 0, i{{.+}} 1
// CK21: call void [[CALL05:@.+]]([[ST]]* {{[^,]+}})
#pragma omp target map(A, A2)
{
A += 1;
A2 += 1;
}
return A;
}
};
int explicit_maps_template_args_and_members(int a){
CC<123,int> c;
return c.foo(a);
}
// CK21: define {{.+}}[[CALL00]]
// CK21: define {{.+}}[[CALL01]]
// CK21: define {{.+}}[[CALL02]]
// CK21: define {{.+}}[[CALL03]]
// CK21: define {{.+}}[[CALL04]]
// CK21: define {{.+}}[[CALL05]]
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK22 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK22 --check-prefix CK22-64
// RUN: %clang_cc1 -DCK22 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK22 --check-prefix CK22-64
// RUN: %clang_cc1 -DCK22 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK22 --check-prefix CK22-32
// RUN: %clang_cc1 -DCK22 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK22 --check-prefix CK22-32
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
#ifdef CK22
// CK22-DAG: [[ST:%.+]] = type { float }
// CK22-DAG: [[STT:%.+]] = type { i32 }
// CK22: [[SIZE00:@.+]] = private {{.*}}constant [1 x i[[Z:64|32]]] [i[[Z:64|32]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK22: [[MTYPE00:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: [[SIZE01:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 400]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK22: [[MTYPE01:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: [[SIZE02:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] {{8|4}}]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK22: [[MTYPE02:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: [[SIZE03:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 16]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK22: [[MTYPE03:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: [[SIZE04:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 20]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK22: [[MTYPE04:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: [[SIZE05:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK22: [[MTYPE05:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: [[SIZE06:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 400]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK22: [[MTYPE06:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: [[SIZE07:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] {{8|4}}]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK22: [[MTYPE07:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: [[SIZE08:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 16]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK22: [[MTYPE08:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: [[SIZE09:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 20]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK22: [[MTYPE09:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: [[SIZE10:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK22: [[MTYPE10:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: [[SIZE11:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 400]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK22: [[MTYPE11:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: [[SIZE12:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] {{8|4}}]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK22: [[MTYPE12:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: [[SIZE13:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 16]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK22: [[MTYPE13:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: [[SIZE14:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 20]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK22: [[MTYPE14:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
int a;
int c[100];
int *d;
struct ST {
float fa;
};
ST sa ;
ST sc[100];
ST *sd;
template<typename T>
struct STT {
T fa;
};
STT<int> sta ;
STT<int> stc[100];
STT<int> *std;
// CK22-LABEL: explicit_maps_globals
int explicit_maps_globals(void){
// Region 00
// CK22-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE00]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE00]]{{.+}})
// CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK22-DAG: store i32* @a, i32** [[CBP0]]
// CK22-DAG: store i32* @a, i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: call void [[CALL00:@.+]](i32* {{[^,]+}})
#pragma omp target map(a)
{ a+=1; }
// Region 01
// CK22-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE01]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE01]]{{.+}})
// CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x i32]**
// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [100 x i32]**
// CK22-DAG: store [100 x i32]* @c, [100 x i32]** [[CBP0]]
// CK22-DAG: store [100 x i32]* @c, [100 x i32]** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: call void [[CALL01:@.+]]([100 x i32]* {{[^,]+}})
#pragma omp target map(c)
{ c[3]+=1; }
// Region 02
// CK22-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE02]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE02]]{{.+}})
// CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32***
// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32***
// CK22-DAG: store i32** @d, i32*** [[CBP0]]
// CK22-DAG: store i32** @d, i32*** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: call void [[CALL02:@.+]](i32** {{[^,]+}})
#pragma omp target map(d)
{ d[3]+=1; }
// Region 03
// CK22-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE03]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE03]]{{.+}})
// CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x i32]**
// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK22-DAG: store [100 x i32]* @c, [100 x i32]** [[CBP0]]
// CK22-DAG: store i32* getelementptr inbounds ([100 x i32], [100 x i32]* @c, i{{.+}} 0, i{{.+}} 1), i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: call void [[CALL03:@.+]]([100 x i32]* {{[^,]+}})
#pragma omp target map(c[1:4])
{ c[3]+=1; }
// Region 04
// CK22-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE04]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE04]]{{.+}})
// CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK22-DAG: store i32* [[RVAR0:%.+]], i32** [[CBP0]]
// CK22-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22-DAG: [[RVAR0]] = load i32*, i32** @d
// CK22-DAG: [[SEC0]] = getelementptr {{.*}}i32* [[RVAR00:%.+]], i{{.+}} 2
// CK22-DAG: [[RVAR00]] = load i32*, i32** @d
// CK22: call void [[CALL04:@.+]](i32* {{[^,]+}})
#pragma omp target map(d[2:5])
{ d[3]+=1; }
// Region 05
// CK22-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE05]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE05]]{{.+}})
// CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
// CK22-DAG: store [[ST]]* @sa, [[ST]]** [[CBP0]]
// CK22-DAG: store [[ST]]* @sa, [[ST]]** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: call void [[CALL05:@.+]]([[ST]]* {{[^,]+}})
#pragma omp target map(sa)
{ sa.fa+=1; }
// Region 06
// CK22-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE06]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE06]]{{.+}})
// CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x [[ST]]]**
// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [100 x [[ST]]]**
// CK22-DAG: store [100 x [[ST]]]* @sc, [100 x [[ST]]]** [[CBP0]]
// CK22-DAG: store [100 x [[ST]]]* @sc, [100 x [[ST]]]** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: call void [[CALL06:@.+]]([100 x [[ST]]]* {{[^,]+}})
#pragma omp target map(sc)
{ sc[3].fa+=1; }
// Region 07
// CK22-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE07]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE07]]{{.+}})
// CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]***
// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]***
// CK22-DAG: store [[ST]]** @sd, [[ST]]*** [[CBP0]]
// CK22-DAG: store [[ST]]** @sd, [[ST]]*** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: call void [[CALL07:@.+]]([[ST]]** {{[^,]+}})
#pragma omp target map(sd)
{ sd[3].fa+=1; }
// Region 08
// CK22-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE08]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE08]]{{.+}})
// CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x [[ST]]]**
// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
// CK22-DAG: store [100 x [[ST]]]* @sc, [100 x [[ST]]]** [[CBP0]]
// CK22-DAG: store [[ST]]* getelementptr inbounds ([100 x [[ST]]], [100 x [[ST]]]* @sc, i{{.+}} 0, i{{.+}} 1), [[ST]]** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: call void [[CALL08:@.+]]([100 x [[ST]]]* {{[^,]+}})
#pragma omp target map(sc[1:4])
{ sc[3].fa+=1; }
// Region 09
// CK22-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE09]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE09]]{{.+}})
// CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
// CK22-DAG: store [[ST]]* [[RVAR0:%.+]], [[ST]]** [[CBP0]]
// CK22-DAG: store [[ST]]* [[SEC0:%.+]], [[ST]]** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22-DAG: [[RVAR0]] = load [[ST]]*, [[ST]]** @sd
// CK22-DAG: [[SEC0]] = getelementptr {{.*}}[[ST]]* [[RVAR00:%.+]], i{{.+}} 2
// CK22-DAG: [[RVAR00]] = load [[ST]]*, [[ST]]** @sd
// CK22: call void [[CALL09:@.+]]([[ST]]* {{[^,]+}})
#pragma omp target map(sd[2:5])
{ sd[3].fa+=1; }
// Region 10
// CK22-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE10]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE10]]{{.+}})
// CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[STT]]**
// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[STT]]**
// CK22-DAG: store [[STT]]* @sta, [[STT]]** [[CBP0]]
// CK22-DAG: store [[STT]]* @sta, [[STT]]** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: call void [[CALL10:@.+]]([[STT]]* {{[^,]+}})
#pragma omp target map(sta)
{ sta.fa+=1; }
// Region 11
// CK22-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE11]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE11]]{{.+}})
// CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x [[STT]]]**
// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [100 x [[STT]]]**
// CK22-DAG: store [100 x [[STT]]]* @stc, [100 x [[STT]]]** [[CBP0]]
// CK22-DAG: store [100 x [[STT]]]* @stc, [100 x [[STT]]]** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: call void [[CALL11:@.+]]([100 x [[STT]]]* {{[^,]+}})
#pragma omp target map(stc)
{ stc[3].fa+=1; }
// Region 12
// CK22-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE12]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE12]]{{.+}})
// CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[STT]]***
// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[STT]]***
// CK22-DAG: store [[STT]]** @std, [[STT]]*** [[CBP0]]
// CK22-DAG: store [[STT]]** @std, [[STT]]*** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: call void [[CALL12:@.+]]([[STT]]** {{[^,]+}})
#pragma omp target map(std)
{ std[3].fa+=1; }
// Region 13
// CK22-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE13]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE13]]{{.+}})
// CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x [[STT]]]**
// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[STT]]**
// CK22-DAG: store [100 x [[STT]]]* @stc, [100 x [[STT]]]** [[CBP0]]
// CK22-DAG: store [[STT]]* getelementptr inbounds ([100 x [[STT]]], [100 x [[STT]]]* @stc, i{{.+}} 0, i{{.+}} 1), [[STT]]** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22: call void [[CALL13:@.+]]([100 x [[STT]]]* {{[^,]+}})
#pragma omp target map(stc[1:4])
{ stc[3].fa+=1; }
// Region 14
// CK22-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE14]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE14]]{{.+}})
// CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[STT]]**
// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[STT]]**
// CK22-DAG: store [[STT]]* [[RVAR0:%.+]], [[STT]]** [[CBP0]]
// CK22-DAG: store [[STT]]* [[SEC0:%.+]], [[STT]]** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK22-DAG: [[RVAR0]] = load [[STT]]*, [[STT]]** @std
// CK22-DAG: [[SEC0]] = getelementptr {{.*}}[[STT]]* [[RVAR00:%.+]], i{{.+}} 2
// CK22-DAG: [[RVAR00]] = load [[STT]]*, [[STT]]** @std
// CK22: call void [[CALL14:@.+]]([[STT]]* {{[^,]+}})
#pragma omp target map(std[2:5])
{ std[3].fa+=1; }
return 0;
}
// CK22: define {{.+}}[[CALL00]]
// CK22: define {{.+}}[[CALL01]]
// CK22: define {{.+}}[[CALL02]]
// CK22: define {{.+}}[[CALL03]]
// CK22: define {{.+}}[[CALL04]]
// CK22: define {{.+}}[[CALL05]]
// CK22: define {{.+}}[[CALL06]]
// CK22: define {{.+}}[[CALL07]]
// CK22: define {{.+}}[[CALL08]]
// CK22: define {{.+}}[[CALL09]]
// CK22: define {{.+}}[[CALL10]]
// CK22: define {{.+}}[[CALL11]]
// CK22: define {{.+}}[[CALL12]]
// CK22: define {{.+}}[[CALL13]]
// CK22: define {{.+}}[[CALL14]]
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -std=c++11 -DCK23 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK23 --check-prefix CK23-64
// RUN: %clang_cc1 -std=c++11 -DCK23 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -std=c++11 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK23 --check-prefix CK23-64
// RUN: %clang_cc1 -std=c++11 -DCK23 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK23 --check-prefix CK23-32
// RUN: %clang_cc1 -std=c++11 -DCK23 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -std=c++11 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK23 --check-prefix CK23-32
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
#ifdef CK23
// CK23: [[SIZE00:@.+]] = private {{.*}}constant [1 x i[[Z:64|32]]] [i[[Z:64|32]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK23: [[MTYPE00:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK23: [[SIZE01:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK23: [[MTYPE01:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK23: [[SIZE02:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 400]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK23: [[MTYPE02:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK23: [[SIZE03:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] {{8|4}}]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK23: [[MTYPE03:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK23: [[SIZE04:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 16]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK23: [[MTYPE04:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK23: [[SIZE05:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 16]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK23: [[MTYPE05:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK23-LABEL: explicit_maps_inside_captured
int explicit_maps_inside_captured(int a){
float b;
float c[100];
float *d;
// CK23: call void @{{.*}}explicit_maps_inside_captured{{.*}}([[SA:%.+]]* {{.*}})
// CK23: define {{.*}}explicit_maps_inside_captured{{.*}}
[&](void){
// Region 00
// CK23-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE00]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE00]]{{.+}})
// CK23-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK23-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK23-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK23-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK23-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
// CK23-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK23-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
// CK23-DAG: store i32* [[VAR00:%.+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK23-DAG: [[VAR0]] = load i32*, i32** [[CAP0:%[^,]+]]
// CK23-DAG: [[CAP0]] = getelementptr inbounds [[SA]], [[SA]]
// CK23-DAG: [[VAR00]] = load i32*, i32** [[CAP00:%[^,]+]]
// CK23-DAG: [[CAP00]] = getelementptr inbounds [[SA]], [[SA]]
// CK23: call void [[CALL00:@.+]](i32* {{[^,]+}})
#pragma omp target map(a)
{ a+=1; }
// Region 01
// CK23-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE01]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE01]]{{.+}})
// CK23-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK23-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK23-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK23-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK23-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to float**
// CK23-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to float**
// CK23-DAG: store float* [[VAR0:%.+]], float** [[CBP0]]
// CK23-DAG: store float* [[VAR00:%.+]], float** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK23-DAG: [[VAR0]] = load float*, float** [[CAP0:%[^,]+]]
// CK23-DAG: [[CAP0]] = getelementptr inbounds [[SA]], [[SA]]
// CK23-DAG: [[VAR00]] = load float*, float** [[CAP00:%[^,]+]]
// CK23-DAG: [[CAP00]] = getelementptr inbounds [[SA]], [[SA]]
// CK23: call void [[CALL01:@.+]](float* {{[^,]+}})
#pragma omp target map(b)
{ b+=1; }
// Region 02
// CK23-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE02]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE02]]{{.+}})
// CK23-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK23-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK23-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK23-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK23-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x float]**
// CK23-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [100 x float]**
// CK23-DAG: store [100 x float]* [[VAR0:%.+]], [100 x float]** [[CBP0]]
// CK23-DAG: store [100 x float]* [[VAR00:%.+]], [100 x float]** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK23-DAG: [[VAR0]] = load [100 x float]*, [100 x float]** [[CAP0:%[^,]+]]
// CK23-DAG: [[CAP0]] = getelementptr inbounds [[SA]], [[SA]]
// CK23-DAG: [[VAR00]] = load [100 x float]*, [100 x float]** [[CAP00:%[^,]+]]
// CK23-DAG: [[CAP00]] = getelementptr inbounds [[SA]], [[SA]]
// CK23: call void [[CALL02:@.+]]([100 x float]* {{[^,]+}})
#pragma omp target map(c)
{ c[3]+=1; }
// Region 03
// CK23-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE03]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE03]]{{.+}})
// CK23-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK23-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK23-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK23-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK23-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to float***
// CK23-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to float***
// CK23-DAG: store float** [[VAR0:%.+]], float*** [[CBP0]]
// CK23-DAG: store float** [[VAR00:%.+]], float*** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK23-DAG: [[VAR0]] = load float**, float*** [[CAP0:%[^,]+]]
// CK23-DAG: [[CAP0]] = getelementptr inbounds [[SA]], [[SA]]
// CK23-DAG: [[VAR00]] = load float**, float*** [[CAP00:%[^,]+]]
// CK23-DAG: [[CAP00]] = getelementptr inbounds [[SA]], [[SA]]
// CK23: call void [[CALL03:@.+]](float** {{[^,]+}})
#pragma omp target map(d)
{ d[3]+=1; }
// Region 04
// CK23-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE04]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE04]]{{.+}})
// CK23-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK23-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK23-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK23-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK23-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x float]**
// CK23-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to float**
// CK23-DAG: store [100 x float]* [[VAR0:%.+]], [100 x float]** [[CBP0]]
// CK23-DAG: store float* [[SEC0:%.+]], float** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK23-DAG: [[SEC0]] = getelementptr {{.*}}[100 x float]* [[VAR00:%.+]], i{{.+}} 0, i{{.+}} 2
// CK23-DAG: [[VAR0]] = load [100 x float]*, [100 x float]** [[CAP0:%[^,]+]]
// CK23-DAG: [[CAP0]] = getelementptr inbounds [[SA]], [[SA]]
// CK23-DAG: [[VAR00]] = load [100 x float]*, [100 x float]** [[CAP00:%[^,]+]]
// CK23-DAG: [[CAP00]] = getelementptr inbounds [[SA]], [[SA]]
// CK23: call void [[CALL04:@.+]]([100 x float]* {{[^,]+}})
#pragma omp target map(c[2:4])
{ c[3]+=1; }
// Region 05
// CK23-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE05]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE05]]{{.+}})
// CK23-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK23-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK23-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK23-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK23-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to float**
// CK23-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to float**
// CK23-DAG: store float* [[RVAR0:%.+]], float** [[CBP0]]
// CK23-DAG: store float* [[SEC0:%.+]], float** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK23-DAG: [[RVAR0]] = load float*, float** [[VAR0:%[^,]+]]
// CK23-DAG: [[SEC0]] = getelementptr {{.*}}float* [[RVAR00:%.+]], i{{.+}} 2
// CK23-DAG: [[RVAR00]] = load float*, float** [[VAR00:%[^,]+]]
// CK23-DAG: [[VAR0]] = load float**, float*** [[CAP0:%[^,]+]]
// CK23-DAG: [[CAP0]] = getelementptr inbounds [[SA]], [[SA]]
// CK23-DAG: [[VAR00]] = load float**, float*** [[CAP00:%[^,]+]]
// CK23-DAG: [[CAP00]] = getelementptr inbounds [[SA]], [[SA]]
// CK23: call void [[CALL05:@.+]](float* {{[^,]+}})
#pragma omp target map(d[2:4])
{ d[3]+=1; }
}();
return b;
}
// CK23: define {{.+}}[[CALL00]]
// CK23: define {{.+}}[[CALL01]]
// CK23: define {{.+}}[[CALL02]]
// CK23: define {{.+}}[[CALL03]]
// CK23: define {{.+}}[[CALL04]]
// CK23: define {{.+}}[[CALL05]]
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK24 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK24 --check-prefix CK24-64
// RUN: %clang_cc1 -DCK24 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK24 --check-prefix CK24-64
// RUN: %clang_cc1 -DCK24 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK24 --check-prefix CK24-32
// RUN: %clang_cc1 -DCK24 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK24 --check-prefix CK24-32
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
#ifdef CK24
// CK24-DAG: [[SC:%.+]] = type { i32, [[SB:%.+]], [[SB:%.+]]*, [10 x i32] }
// CK24-DAG: [[SB]] = type { i32, [[SA:%.+]], [10 x [[SA:%.+]]], [10 x [[SA:%.+]]*], [[SA:%.+]]* }
// CK24-DAG: [[SA]] = type { i32, [[SA]]*, [10 x i32] }
struct SA{
int a;
struct SA *p;
int b[10];
};
struct SB{
int a;
struct SA s;
struct SA sa[10];
struct SA *sp[10];
struct SA *p;
};
struct SC{
int a;
struct SB s;
struct SB *p;
int b[10];
};
// CK24: [[SIZE01:@.+]] = private {{.*}}constant [1 x i[[Z:64|32]]] [i[[Z:64|32]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK24: [[MTYPE01:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24: [[SIZE13:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK24: [[MTYPE13:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24: [[SIZE14:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] {{56|48}}]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK24: [[MTYPE14:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24: [[SIZE15:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK24: [[MTYPE15:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24: [[SIZE16:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 20]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK24: [[MTYPE16:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24: [[SIZE17:@.+]] = private {{.*}}constant [2 x i[[Z]]] [i[[Z]] {{8|4}}, i[[Z]] {{3560|2880}}]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK24: [[MTYPE17:@.+]] = private {{.*}}constant [2 x i32] [i32 35, i32 19]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24: [[SIZE18:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK24: [[MTYPE18:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24: [[SIZE19:@.+]] = private {{.*}}constant [2 x i[[Z]]] [i[[Z]] {{8|4}}, i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK24: [[MTYPE19:@.+]] = private {{.*}}constant [2 x i32] [i32 35, i32 19]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24: [[SIZE20:@.+]] = private {{.*}}constant [2 x i[[Z]]] [i[[Z]] {{8|4}}, i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK24: [[MTYPE20:@.+]] = private {{.*}}constant [2 x i32] [i32 35, i32 19]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24: [[SIZE21:@.+]] = private {{.*}}constant [2 x i[[Z]]] [i[[Z]] {{8|4}}, i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK24: [[MTYPE21:@.+]] = private {{.*}}constant [2 x i32] [i32 35, i32 19]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24: [[SIZE22:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] {{8|4}}]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK24: [[MTYPE22:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24: [[SIZE23:@.+]] = private {{.*}}constant [2 x i[[Z]]] [i[[Z]] {{8|4}}, i[[Z]] {{8|4}}]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK24: [[MTYPE23:@.+]] = private {{.*}}constant [2 x i32] [i32 35, i32 19]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24: [[SIZE24:@.+]] = private {{.*}}constant [4 x i[[Z]]] [i[[Z]] {{8|4}}, i[[Z]] {{8|4}}, i[[Z]] {{8|4}}, i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK24: [[MTYPE24:@.+]] = private {{.*}}constant [4 x i32] [i32 35, i32 19, i32 19, i32 19]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24-LABEL: explicit_maps_struct_fields
int explicit_maps_struct_fields(int a){
SC s;
SC *p;
// Region 01
// CK24-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE01]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE01]]{{.+}})
// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK24-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[SC]]**
// CK24-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK24-DAG: store [[SC]]* [[VAR0:%.+]], [[SC]]** [[CBP0]]
// CK24-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24-DAG: [[SEC0]] = getelementptr {{.*}}[[SC]]* [[VAR0]], i{{.+}} 0, i{{.+}} 0
// CK24: call void [[CALL01:@.+]]([[SC]]* {{[^,]+}})
#pragma omp target map(s.a)
{ s.a++; }
//
// Same thing but starting from a pointer.
//
// Region 13
// CK24-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE13]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE13]]{{.+}})
// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK24-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[SC]]**
// CK24-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK24-DAG: store [[SC]]* [[VAR0:%.+]], [[SC]]** [[CBP0]]
// CK24-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24-DAG: [[SEC0]] = getelementptr {{.*}}[[SC]]* [[VAR00:%.+]], i{{.+}} 0, i{{.+}} 0
// CK24-DAG: [[VAR0]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24-DAG: [[VAR00]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24: call void [[CALL13:@.+]]([[SC]]* {{[^,]+}})
#pragma omp target map(p->a)
{ p->a++; }
// Region 14
// CK24-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE14]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE14]]{{.+}})
// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK24-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[SC]]**
// CK24-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[SA]]**
// CK24-DAG: store [[SC]]* [[VAR0:%.+]], [[SC]]** [[CBP0]]
// CK24-DAG: store [[SA]]* [[SEC0:%.+]], [[SA]]** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24-DAG: [[SEC0]] = getelementptr {{.*}}[[SB]]* [[SEC00:%[^,]+]], i{{.+}} 0, i{{.+}} 1
// CK24-DAG: [[SEC00]] = getelementptr {{.*}}[[SC]]* [[VAR00:%.+]], i{{.+}} 0, i{{.+}} 1
// CK24-DAG: [[VAR0]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24-DAG: [[VAR00]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24: call void [[CALL14:@.+]]([[SC]]* {{[^,]+}})
#pragma omp target map(p->s.s)
{ p->a++; }
// Region 15
// CK24-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE15]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE15]]{{.+}})
// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK24-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[SC]]**
// CK24-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK24-DAG: store [[SC]]* [[VAR0:%.+]], [[SC]]** [[CBP0]]
// CK24-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24-DAG: [[SEC0]] = getelementptr {{.*}}[[SA]]* [[SEC00:%[^,]+]], i{{.+}} 0, i{{.+}} 0
// CK24-DAG: [[SEC00]] = getelementptr {{.*}}[[SB]]* [[SEC000:%[^,]+]], i{{.+}} 0, i{{.+}} 1
// CK24-DAG: [[SEC000]] = getelementptr {{.*}}[[SC]]* [[VAR00:%.+]], i{{.+}} 0, i{{.+}} 1
// CK24-DAG: [[VAR0]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24-DAG: [[VAR00]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24: call void [[CALL15:@.+]]([[SC]]* {{[^,]+}})
#pragma omp target map(p->s.s.a)
{ p->a++; }
// Region 16
// CK24-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE16]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE16]]{{.+}})
// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK24-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[SC]]**
// CK24-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK24-DAG: store [[SC]]* [[VAR0:%.+]], [[SC]]** [[CBP0]]
// CK24-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24-DAG: [[SEC0]] = getelementptr {{.*}}[10 x i32]* [[SEC00:%[^,]+]], i{{.+}} 0, i{{.+}} 0
// CK24-DAG: [[SEC00]] = getelementptr {{.*}}[[SC]]* [[VAR00:%.+]], i{{.+}} 0, i{{.+}} 3
// CK24-DAG: [[VAR0]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24-DAG: [[VAR00]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24: call void [[CALL16:@.+]]([[SC]]* {{[^,]+}})
#pragma omp target map(p->b[:5])
{ p->a++; }
// Region 17
// CK24-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[SIZE17]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE17]]{{.+}})
// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK24-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[SC]]**
// CK24-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[SB]]***
// CK24-DAG: store [[SC]]* [[VAR0:%.+]], [[SC]]** [[CBP0]]
// CK24-DAG: store [[SB]]** [[SEC0:%.+]], [[SB]]*** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24-DAG: [[SEC0]] = getelementptr {{.*}}[[SC]]* [[VAR00:%.+]], i{{.+}} 0, i{{.+}} 2
// CK24-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK24-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK24-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[SB]]***
// CK24-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [[SB]]**
// CK24-DAG: store [[SB]]** [[SEC0]], [[SB]]*** [[CBP1]]
// CK24-DAG: store [[SB]]* [[SEC1:%.+]], [[SB]]** [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24-DAG: [[SEC1]] = getelementptr {{.*}}[[SB]]* [[SEC11:%[^,]+]], i{{.+}} 0
// CK24-DAG: [[SEC11]] = load [[SB]]*, [[SB]]** [[SEC111:%[^,]+]],
// CK24-DAG: [[SEC111]] = getelementptr {{.*}}[[SC]]* [[VAR000:%.+]], i{{.+}} 0, i{{.+}} 2
// CK24-DAG: [[VAR0]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24-DAG: [[VAR00]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24-DAG: [[VAR000]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24: call void [[CALL17:@.+]]([[SC]]* {{[^,]+}})
#pragma omp target map(p->p[:5])
{ p->a++; }
// Region 18
// CK24-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE18]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE18]]{{.+}})
// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK24-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[SC]]**
// CK24-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK24-DAG: store [[SC]]* [[VAR0:%.+]], [[SC]]** [[CBP0]]
// CK24-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24-DAG: [[SEC0]] = getelementptr {{.*}}[[SA]]* [[SEC00:%[^,]+]], i{{.+}} 0, i{{.+}} 0
// CK24-DAG: [[SEC00]] = getelementptr {{.*}}[10 x [[SA]]]* [[SEC000:%[^,]+]], i{{.+}} 0, i{{.+}} 3
// CK24-DAG: [[SEC000]] = getelementptr {{.*}}[[SB]]* [[SEC0000:%[^,]+]], i{{.+}} 0, i{{.+}} 2
// CK24-DAG: [[SEC0000]] = getelementptr {{.*}}[[SC]]* [[VAR00:%.+]], i{{.+}} 0, i{{.+}} 1
// CK24-DAG: [[VAR0]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24-DAG: [[VAR00]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24: call void [[CALL18:@.+]]([[SC]]* {{[^,]+}})
#pragma omp target map(p->s.sa[3].a)
{ p->a++; }
// Region 19
// CK24-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[SIZE19]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE19]]{{.+}})
// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK24-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[SC]]**
// CK24-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[SA]]***
// CK24-DAG: store [[SC]]* [[VAR0:%.+]], [[SC]]** [[CBP0]]
// CK24-DAG: store [[SA]]** [[SEC0:%.+]], [[SA]]*** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24-DAG: [[SEC0]] = getelementptr {{.*}}[10 x [[SA]]*]* [[SEC00:%[^,]+]], i{{.+}} 0, i{{.+}} 3
// CK24-DAG: [[SEC00]] = getelementptr {{.*}}[[SB]]* [[SEC000:%[^,]+]], i{{.+}} 0, i{{.+}} 3
// CK24-DAG: [[SEC000]] = getelementptr {{.*}}[[SC]]* [[VAR00:%.+]], i{{.+}} 0, i{{.+}} 1
// CK24-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK24-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK24-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[SA]]***
// CK24-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
// CK24-DAG: store [[SA]]** [[SEC0]], [[SA]]*** [[CBP1]]
// CK24-DAG: store i32* [[SEC1:%.+]], i32** [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24-DAG: [[SEC1]] = getelementptr {{.*}}[[SA]]* [[SEC11:%[^,]+]], i{{.+}} 0, i{{.+}} 0
// CK24-DAG: [[SEC11]] = load [[SA]]*, [[SA]]** [[SEC111:%[^,]+]],
// CK24-DAG: [[SEC111]] = getelementptr {{.*}}[10 x [[SA]]*]* [[SEC1111:%[^,]+]], i{{.+}} 0, i{{.+}} 3
// CK24-DAG: [[SEC1111]] = getelementptr {{.*}}[[SB]]* [[SEC11111:%[^,]+]], i{{.+}} 0, i{{.+}} 3
// CK24-DAG: [[SEC11111]] = getelementptr {{.*}}[[SC]]* [[VAR000:%.+]], i{{.+}} 0, i{{.+}} 1
// CK24-DAG: [[VAR0]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24-DAG: [[VAR00]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24-DAG: [[VAR000]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24: call void [[CALL19:@.+]]([[SC]]* {{[^,]+}})
#pragma omp target map(p->s.sp[3]->a)
{ p->a++; }
// Region 20
// CK24-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[SIZE20]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE20]]{{.+}})
// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK24-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[SC]]**
// CK24-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[SB]]***
// CK24-DAG: store [[SC]]* [[VAR0:%.+]], [[SC]]** [[CBP0]]
// CK24-DAG: store [[SB]]** [[SEC0:%.+]], [[SB]]*** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24-DAG: [[SEC0]] = getelementptr {{.*}}[[SC]]* [[VAR00:%.+]], i{{.+}} 0, i{{.+}} 2
// CK24-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK24-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK24-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[SB]]***
// CK24-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
// CK24-DAG: store [[SB]]** [[SEC0]], [[SB]]*** [[CBP1]]
// CK24-DAG: store i32* [[SEC1:%.+]], i32** [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24-DAG: [[SEC1]] = getelementptr {{.*}}[[SB]]* [[SEC11:%[^,]+]], i{{.+}} 0
// CK24-DAG: [[SEC11]] = load [[SB]]*, [[SB]]** [[SEC111:%[^,]+]],
// CK24-DAG: [[SEC111]] = getelementptr {{.*}}[[SC]]* [[VAR000:%.+]], i{{.+}} 0, i{{.+}} 2
// CK24-DAG: [[VAR0]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24-DAG: [[VAR00]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24-DAG: [[VAR000]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24: call void [[CALL20:@.+]]([[SC]]* {{[^,]+}})
#pragma omp target map(p->p->a)
{ p->a++; }
// Region 21
// CK24-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[SIZE21]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE21]]{{.+}})
// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK24-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[SC]]**
// CK24-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[SA]]***
// CK24-DAG: store [[SC]]* [[VAR0:%.+]], [[SC]]** [[CBP0]]
// CK24-DAG: store [[SA]]** [[SEC0:%.+]], [[SA]]*** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24-DAG: [[SEC0]] = getelementptr {{.*}}[[SB]]* [[SEC00:[^,]+]], i{{.+}} 0, i{{.+}} 4
// CK24-DAG: [[SEC00]] = getelementptr {{.*}}[[SC]]* [[VAR00:%.+]], i{{.+}} 0, i{{.+}} 1
// CK24-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK24-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK24-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[SA]]***
// CK24-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
// CK24-DAG: store [[SA]]** [[SEC0]], [[SA]]*** [[CBP1]]
// CK24-DAG: store i32* [[SEC1:%.+]], i32** [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24-DAG: [[SEC1]] = getelementptr {{.*}}[[SA]]* [[SEC11:%[^,]+]], i{{.+}} 0
// CK24-DAG: [[SEC11]] = load [[SA]]*, [[SA]]** [[SEC111:%[^,]+]],
// CK24-DAG: [[SEC111]] = getelementptr {{.*}}[[SB]]* [[SEC1111:[^,]+]], i{{.+}} 0, i{{.+}} 4
// CK24-DAG: [[SEC1111]] = getelementptr {{.*}}[[SC]]* [[VAR000:%.+]], i{{.+}} 0, i{{.+}} 1
// CK24-DAG: [[VAR0]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24-DAG: [[VAR00]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24-DAG: [[VAR000]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24: call void [[CALL21:@.+]]([[SC]]* {{[^,]+}})
#pragma omp target map(p->s.p->a)
{ p->a++; }
// Region 22
// CK24-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE22]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE22]]{{.+}})
// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK24-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[SC]]**
// CK24-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK24-DAG: store [[SC]]* [[VAR0:%.+]], [[SC]]** [[CBP0]]
// CK24-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24-DAG: [[SEC0]] = getelementptr {{.*}}[10 x i32]* [[SEC00:%[^,]+]], i{{.+}} 0, i{{.+}} 0
// CK24-DAG: [[SEC00]] = getelementptr {{.*}}[[SA]]* [[SEC000:%[^,]+]], i{{.+}} 0, i{{.+}} 2
// CK24-DAG: [[SEC000]] = getelementptr {{.*}}[[SB]]* [[SEC0000:%[^,]+]], i{{.+}} 0, i{{.+}} 1
// CK24-DAG: [[SEC0000]] = getelementptr {{.*}}[[SC]]* [[VAR00:%.+]], i{{.+}} 0, i{{.+}} 1
// CK24-DAG: [[VAR0]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24-DAG: [[VAR00]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24: call void [[CALL22:@.+]]([[SC]]* {{[^,]+}})
#pragma omp target map(p->s.s.b[:2])
{ p->a++; }
// Region 23
// CK24-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[SIZE23]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE23]]{{.+}})
// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK24-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[SC]]**
// CK24-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[SA]]***
// CK24-DAG: store [[SC]]* [[VAR0:%.+]], [[SC]]** [[CBP0]]
// CK24-DAG: store [[SA]]** [[SEC0:%.+]], [[SA]]*** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24-DAG: [[SEC0]] = getelementptr {{.*}}[[SB]]* [[SEC00:%[^,]+]], i{{.+}} 0, i{{.+}} 4
// CK24-DAG: [[SEC00]] = getelementptr {{.*}}[[SC]]* [[VAR00:%.+]], i{{.+}} 0, i{{.+}} 1
// CK24-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK24-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK24-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[SA]]***
// CK24-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
// CK24-DAG: store [[SA]]** [[SEC0]], [[SA]]*** [[CBP1]]
// CK24-DAG: store i32* [[SEC1:%.+]], i32** [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24-DAG: [[SEC1]] = getelementptr {{.*}}[10 x i32]* [[SEC11:%[^,]+]], i{{.+}} 0, i{{.+}} 0
// CK24-DAG: [[SEC11]] = getelementptr {{.*}}[[SA]]* [[SEC111:%[^,]+]], i{{.+}} 0, i{{.+}} 2
// CK24-DAG: [[SEC111]] = load [[SA]]*, [[SA]]** [[SEC1111:%[^,]+]],
// CK24-DAG: [[SEC1111]] = getelementptr {{.*}}[[SB]]* [[SEC11111:%[^,]+]], i{{.+}} 0, i{{.+}} 4
// CK24-DAG: [[SEC11111]] = getelementptr {{.*}}[[SC]]* [[VAR000:%.+]], i{{.+}} 0, i{{.+}} 1
// CK24-DAG: [[VAR0]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24-DAG: [[VAR00]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24-DAG: [[VAR000]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24: call void [[CALL23:@.+]]([[SC]]* {{[^,]+}})
#pragma omp target map(p->s.p->b[:2])
{ p->a++; }
// Region 24
// CK24-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 4, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[4 x i{{.+}}]* [[SIZE24]], {{.+}}getelementptr {{.+}}[4 x i{{.+}}]* [[MTYPE24]]{{.+}})
// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK24-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[SC]]**
// CK24-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[SB]]***
// CK24-DAG: store [[SC]]* [[VAR0:%.+]], [[SC]]** [[CBP0]]
// CK24-DAG: store [[SB]]** [[SEC0:%.+]], [[SB]]*** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24-DAG: [[SEC0]] = getelementptr {{.*}}[[SC]]* [[VAR00:%.+]], i{{.+}} 0, i{{.+}} 2
// CK24-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK24-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK24-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[SB]]***
// CK24-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [[SA]]***
// CK24-DAG: store [[SB]]** [[SEC0]], [[SB]]*** [[CBP1]]
// CK24-DAG: store [[SA]]** [[SEC1:%.+]], [[SA]]*** [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24-DAG: [[SEC1]] = getelementptr {{.*}}[[SB]]* [[SEC11:%[^,]+]], i{{.+}} 0, i{{.+}} 4
// CK24-DAG: [[SEC11]] = load [[SB]]*, [[SB]]** [[SEC111:%[^,]+]],
// CK24-DAG: [[SEC111]] = getelementptr {{.*}}[[SC]]* [[VAR000:%.+]], i{{.+}} 0, i{{.+}} 2
// CK24-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 2
// CK24-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 2
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK24-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to [[SA]]***
// CK24-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to [[SA]]***
// CK24-DAG: store [[SA]]** [[SEC1]], [[SA]]*** [[CBP2]]
// CK24-DAG: store [[SA]]** [[SEC2:%.+]], [[SA]]*** [[CP2]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24-DAG: [[SEC2]] = getelementptr {{.*}}[[SA]]* [[SEC22:%[^,]+]], i{{.+}} 0, i{{.+}} 1
// CK24-DAG: [[SEC22]] = load [[SA]]*, [[SA]]** [[SEC222:%[^,]+]],
// CK24-DAG: [[SEC222]] = getelementptr {{.*}}[[SB]]* [[SEC2222:%[^,]+]], i{{.+}} 0, i{{.+}} 4
// CK24-DAG: [[SEC2222]] = load [[SB]]*, [[SB]]** [[SEC22222:%[^,]+]],
// CK24-DAG: [[SEC22222]] = getelementptr {{.*}}[[SC]]* [[VAR0000:%.+]], i{{.+}} 0, i{{.+}} 2
// CK24-DAG: [[BP3:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 3
// CK24-DAG: [[P3:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 3
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK24-DAG: [[CBP3:%.+]] = bitcast i8** [[BP3]] to [[SA]]***
// CK24-DAG: [[CP3:%.+]] = bitcast i8** [[P3]] to i32**
// CK24-DAG: store [[SA]]** [[SEC2]], [[SA]]*** [[CBP3]]
// CK24-DAG: store i32* [[SEC3:%.+]], i32** [[CP3]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK24-DAG: [[SEC3]] = getelementptr {{.*}}[[SA]]* [[SEC33:%[^,]+]], i{{.+}} 0, i{{.+}} 0
// CK24-DAG: [[SEC33]] = load [[SA]]*, [[SA]]** [[SEC333:%[^,]+]],
// CK24-DAG: [[SEC333]] = getelementptr {{.*}}[[SA]]* [[SEC3333:%[^,]+]], i{{.+}} 0, i{{.+}} 1
// CK24-DAG: [[SEC3333]] = load [[SA]]*, [[SA]]** [[SEC33333:%[^,]+]],
// CK24-DAG: [[SEC33333]] = getelementptr {{.*}}[[SB]]* [[SEC333333:%[^,]+]], i{{.+}} 0, i{{.+}} 4
// CK24-DAG: [[SEC333333]] = load [[SB]]*, [[SB]]** [[SEC3333333:%[^,]+]],
// CK24-DAG: [[SEC3333333]] = getelementptr {{.*}}[[SC]]* [[VAR00000:%.+]], i{{.+}} 0, i{{.+}} 2
// CK24-DAG: [[VAR0]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24-DAG: [[VAR00]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24-DAG: [[VAR000]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24-DAG: [[VAR0000]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24-DAG: [[VAR00000]] = load [[SC]]*, [[SC]]** %{{.+}}
// CK24: call void [[CALL24:@.+]]([[SC]]* {{[^,]+}})
#pragma omp target map(p->p->p->p->a)
{ p->a++; }
return s.a;
}
// CK24: define {{.+}}[[CALL01]]
// CK24: define {{.+}}[[CALL13]]
// CK24: define {{.+}}[[CALL14]]
// CK24: define {{.+}}[[CALL15]]
// CK24: define {{.+}}[[CALL16]]
// CK24: define {{.+}}[[CALL17]]
// CK24: define {{.+}}[[CALL18]]
// CK24: define {{.+}}[[CALL19]]
// CK24: define {{.+}}[[CALL20]]
// CK24: define {{.+}}[[CALL21]]
// CK24: define {{.+}}[[CALL22]]
// CK24: define {{.+}}[[CALL23]]
// CK24: define {{.+}}[[CALL24]]
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK25 -std=c++11 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK25 --check-prefix CK25-64
// RUN: %clang_cc1 -DCK25 -std=c++11 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -std=c++11 -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK25 --check-prefix CK25-64
// RUN: %clang_cc1 -DCK25 -std=c++11 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK25 --check-prefix CK25-32
// RUN: %clang_cc1 -DCK25 -std=c++11 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -std=c++11 -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK25 --check-prefix CK25-32
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
#ifdef CK25
// CK25: [[ST:%.+]] = type { i32, float }
// CK25: [[CA00:%.+]] = type { [[ST]]* }
// CK25: [[CA01:%.+]] = type { i32* }
// CK25: [[SIZE00:@.+]] = private {{.*}}constant [1 x i[[Z:64|32]]] [i[[Z:64|32]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK25: [[MTYPE00:@.+]] = private {{.*}}constant [1 x i32] [i32 33]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK25: [[SIZE01:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK25: [[MTYPE01:@.+]] = private {{.*}}constant [1 x i32] [i32 33]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK25-LABEL: explicit_maps_with_inner_lambda
template <int X, typename T>
struct CC {
T A;
float B;
int foo(T arg) {
// Region 00
// CK25-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE00]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE00]]{{.+}})
// CK25-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK25-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK25-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK25-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK25-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
// CK25-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK25-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
// CK25-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK25-DAG: [[SEC0]] = getelementptr {{.*}}[[ST]]* [[VAR0:%.+]], i{{.+}} 0, i{{.+}} 0
// CK25: call void [[CALL00:@.+]]([[ST]]* {{[^,]+}})
#pragma omp target map(to:A)
{
[&]() {
A += 1;
}();
}
// Region 01
// CK25-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE01]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE01]]{{.+}})
// CK25-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK25-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK25-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK25-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK25-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
// CK25-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK25-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
// CK25-DAG: store i32* [[VAR0]], i32** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK25: call void [[CALL01:@.+]](i32* {{[^,]+}})
#pragma omp target map(to:arg)
{
[&]() {
arg += 1;
}();
}
return A+arg;
}
};
int explicit_maps_with_inner_lambda(int a){
CC<123,int> c;
return c.foo(a);
}
// CK25: define {{.+}}[[CALL00]]([[ST]]* [[VAL:%.+]])
// CK25: store [[ST]]* [[VAL]], [[ST]]** [[VALADDR:%[^,]+]],
// CK25: [[VAL1:%.+]] = load [[ST]]*, [[ST]]** [[VALADDR]],
// CK25: [[VALADDR1:%.+]] = getelementptr inbounds [[CA00]], [[CA00]]* [[CA:%[^,]+]], i32 0, i32 0
// CK25: store [[ST]]* [[VAL1]], [[ST]]** [[VALADDR1]],
// CK25: call void {{.*}}[[LAMBDA:@.+]]{{.*}}([[CA00]]* [[CA]])
// CK25: define {{.+}}[[LAMBDA]]
// CK25: define {{.+}}[[CALL01]](i32* {{.*}}[[VAL:%.+]])
// CK25: store i32* [[VAL]], i32** [[VALADDR:%[^,]+]],
// CK25: [[VAL1:%.+]] = load i32*, i32** [[VALADDR]],
// CK25: [[VALADDR1:%.+]] = getelementptr inbounds [[CA01]], [[CA01]]* [[CA:%[^,]+]], i32 0, i32 0
// CK25: store i32* [[VAL1]], i32** [[VALADDR1]],
Fix for PR30632: Name mangling issue. There was a bug in the implementation of captured statements. If it has a lambda expression in it and the same lambda expression is used outside the captured region, clang produced an error: ``` error: definition with same mangled name as another definition ``` Here is an example: ``` struct A { template <typename L> void g(const L&) { } }; template<typename T> void f() { { A().g([](){}); } A().g([](){}); } int main() { f<void>(); } ``` Error report: ``` main.cpp:3:10: error: definition with same mangled name as another definition void g(const L&) { } ^ main.cpp:3:10: note: previous definition is here ``` Patch fixes this bug. llvm-svn: 284229
2016-10-14 20:43:59 +08:00
// CK25: call void {{.*}}[[LAMBDA2:@.+]]{{.*}}([[CA01]]* [[CA]])
// CK25: define {{.+}}[[LAMBDA2]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK26 -std=c++11 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK26 --check-prefix CK26-64
// RUN: %clang_cc1 -DCK26 -std=c++11 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -std=c++11 -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK26 --check-prefix CK26-64
// RUN: %clang_cc1 -DCK26 -std=c++11 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK26 --check-prefix CK26-32
// RUN: %clang_cc1 -DCK26 -std=c++11 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -std=c++11 -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK26 --check-prefix CK26-32
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
#ifdef CK26
// CK26: [[ST:%.+]] = type { i32, float*, i32, float* }
// CK26: [[SIZE00:@.+]] = private {{.*}}constant [2 x i[[Z:64|32]]] [i[[Z:64|32]] {{32|16}}, i[[Z:64|32]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK26: [[MTYPE00:@.+]] = private {{.*}}constant [2 x i32] [i32 35, i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK26: [[SIZE01:@.+]] = private {{.*}}constant [2 x i[[Z]]] [i[[Z]] {{32|16}}, i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK26: [[MTYPE01:@.+]] = private {{.*}}constant [2 x i32] [i32 35, i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK26: [[SIZE02:@.+]] = private {{.*}}constant [2 x i[[Z]]] [i[[Z]] {{32|16}}, i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK26: [[MTYPE02:@.+]] = private {{.*}}constant [2 x i32] [i32 35, i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK26: [[SIZE03:@.+]] = private {{.*}}constant [2 x i[[Z]]] [i[[Z]] {{32|16}}, i[[Z]] 4]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK26: [[MTYPE03:@.+]] = private {{.*}}constant [2 x i32] [i32 35, i32 35]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK26-LABEL: explicit_maps_with_private_class_members
struct CC {
int fA;
float &fB;
int pA;
float &pB;
CC(float &B) : fB(B), pB(B) {
// CK26: call {{.*}}@__kmpc_fork_call{{.*}} [[OUTCALL:@.+]] to void (i32*, i32*, ...)*
// define {{.*}}void [[OUTCALL]]
#pragma omp parallel firstprivate(fA,fB) private(pA,pB)
{
// Region 00
// CK26-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[SIZE00]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE00]]{{.+}})
// CK26-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK26-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK26-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK26-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK26-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
// CK26-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
// CK26-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
// CK26-DAG: store [[ST]]* [[VAR0]], [[ST]]** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK26-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK26-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK26-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
// CK26-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
// CK26-DAG: store i32* [[VAR1:%.+]], i32** [[CBP1]]
// CK26-DAG: store i32* [[SEC1:%.+]], i32** [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK26-DAG: [[VAR1]] = load i32*, i32** [[PVT:%.+]],
// CK26-DAG: [[SEC1]] = load i32*, i32** [[PVT]],
// CK26: call void [[CALL00:@.+]]([[ST]]* {{[^,]+}}, i32* {{[^,]+}})
#pragma omp target map(fA)
{
++fA;
}
// Region 01
// CK26-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[SIZE01]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE01]]{{.+}})
// CK26-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK26-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK26-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK26-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK26-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
// CK26-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
// CK26-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
// CK26-DAG: store [[ST]]* [[VAR0]], [[ST]]** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK26-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK26-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK26-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to float**
// CK26-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to float**
// CK26-DAG: store float* [[VAR1:%.+]], float** [[CBP1]]
// CK26-DAG: store float* [[SEC1:%.+]], float** [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK26-DAG: [[VAR1]] = load float*, float** [[PVT:%.+]],
// CK26-DAG: [[SEC1]] = load float*, float** [[PVT]],
// CK26: call void [[CALL01:@.+]]([[ST]]* {{[^,]+}}, float* {{[^,]+}})
#pragma omp target map(fB)
{
fB += 1.0;
}
// Region 02
// CK26-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[SIZE02]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE02]]{{.+}})
// CK26-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK26-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK26-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK26-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK26-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
// CK26-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
// CK26-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
// CK26-DAG: store [[ST]]* [[VAR0]], [[ST]]** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK26-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK26-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK26-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
// CK26-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
// CK26-DAG: store i32* [[VAR1:%.+]], i32** [[CBP1]]
// CK26-DAG: store i32* [[SEC1:%.+]], i32** [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK26-DAG: [[VAR1]] = load i32*, i32** [[PVT:%.+]],
// CK26-DAG: [[SEC1]] = load i32*, i32** [[PVT]],
// CK26: call void [[CALL02:@.+]]([[ST]]* {{[^,]+}}, i32* {{[^,]+}})
#pragma omp target map(pA)
{
++pA;
}
// Region 01
// CK26-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[SIZE03]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE03]]{{.+}})
// CK26-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK26-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK26-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK26-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK26-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
// CK26-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
// CK26-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
// CK26-DAG: store [[ST]]* [[VAR0]], [[ST]]** [[CP0]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK26-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK26-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK26-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to float**
// CK26-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to float**
// CK26-DAG: store float* [[VAR1:%.+]], float** [[CBP1]]
// CK26-DAG: store float* [[SEC1:%.+]], float** [[CP1]]
[OpenMP] Map clause codegeneration. Summary: Implement codegen for the map clause. All the new list items in 4.5 specification are supported. Fix bug in the generation of array sections that was exposed by some of the map clause tests: for pointer types the offsets have to be calculated from the pointee not the pointer. Reviewers: hfinkel, kkwli0, carlo.bertolli, arpith-jacob, ABataev Subscribers: ABataev, cfe-commits, caomhin, fraggamuffin Differential Revision: http://reviews.llvm.org/D16749 llvm-svn: 267808
2016-04-28 06:40:57 +08:00
// CK26-DAG: [[VAR1]] = load float*, float** [[PVT:%.+]],
// CK26-DAG: [[SEC1]] = load float*, float** [[PVT]],
// CK26: call void [[CALL03:@.+]]([[ST]]* {{[^,]+}}, float* {{[^,]+}})
#pragma omp target map(pB)
{
pB += 1.0;
}
}
}
int foo() {
return fA + pA;
}
};
// Make sure the private instance is used in all target regions.
// CK26: define {{.+}}[[CALL00]]({{.*}}i32*{{.*}}[[PVTARG:%.+]])
// CK26: store i32* [[PVTARG]], i32** [[PVTADDR:%.+]],
// CK26: [[ADDR:%.+]] = load i32*, i32** [[PVTADDR]],
// CK26: [[VAL:%.+]] = load i32, i32* [[ADDR]],
// CK26: add nsw i32 [[VAL]], 1
// CK26: define {{.+}}[[CALL01]]({{.*}}float*{{.*}}[[PVTARG:%.+]])
// CK26: store float* [[PVTARG]], float** [[PVTADDR:%.+]],
// CK26: [[ADDR:%.+]] = load float*, float** [[PVTADDR]],
// CK26: [[VAL:%.+]] = load float, float* [[ADDR]],
// CK26: [[EXT:%.+]] = fpext float [[VAL]] to double
// CK26: fadd double [[EXT]], 1.000000e+00
// CK26: define {{.+}}[[CALL02]]({{.*}}i32*{{.*}}[[PVTARG:%.+]])
// CK26: store i32* [[PVTARG]], i32** [[PVTADDR:%.+]],
// CK26: [[ADDR:%.+]] = load i32*, i32** [[PVTADDR]],
// CK26: [[VAL:%.+]] = load i32, i32* [[ADDR]],
// CK26: add nsw i32 [[VAL]], 1
// CK26: define {{.+}}[[CALL03]]({{.*}}float*{{.*}}[[PVTARG:%.+]])
// CK26: store float* [[PVTARG]], float** [[PVTADDR:%.+]],
// CK26: [[ADDR:%.+]] = load float*, float** [[PVTADDR]],
// CK26: [[VAL:%.+]] = load float, float* [[ADDR]],
// CK26: [[EXT:%.+]] = fpext float [[VAL]] to double
// CK26: fadd double [[EXT]], 1.000000e+00
int explicit_maps_with_private_class_members(){
float B;
CC c(B);
return c.foo();
}
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
#endif
///==========================================================================///
[OpenMP] Use fopenmp prefix for all options introduced by the offloading implementation. Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev Subscribers: kkwli0, cfe-commits, caomhin Differential Revision: http://reviews.llvm.org/D21841 llvm-svn: 274283
2016-07-01 05:22:08 +08:00
// RUN: %clang_cc1 -DCK27 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK27 --check-prefix CK27-64
// RUN: %clang_cc1 -DCK27 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK27 --check-prefix CK27-64
// RUN: %clang_cc1 -DCK27 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK27 --check-prefix CK27-32
// RUN: %clang_cc1 -DCK27 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK27 --check-prefix CK27-32
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
#ifdef CK27
// CK27: [[SIZE00:@.+]] = private {{.*}}constant [1 x i[[Z:64|32]]] zeroinitializer
// CK27: [[MTYPE00:@.+]] = private {{.*}}constant [1 x i32] [i32 32]
// CK27: [[SIZE01:@.+]] = private {{.*}}constant [1 x i[[Z]]] zeroinitializer
// CK27: [[MTYPE01:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
// CK27: [[SIZE02:@.+]] = private {{.*}}constant [1 x i[[Z]]] zeroinitializer
// CK27: [[MTYPE02:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
// CK27: [[SIZE03:@.+]] = private {{.*}}constant [1 x i[[Z]]] zeroinitializer
// CK27: [[MTYPE03:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
[OpenMP] Add support for the 'private pointer' flag to signal variables captured in target regions and used in first-private clauses. Summary: If a variable is implicitly mapped (doesn't show in a map clause), the runtime library has to be informed if the corresponding capture shows up in first-private clause, so that the storage previously allocated in the device is used. This patch adds the support for that. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20112 llvm-svn: 270870
2016-05-27 00:53:38 +08:00
// CK27: [[SIZE05:@.+]] = private {{.*}}constant [1 x i[[Z]]] zeroinitializer
// CK27: [[MTYPE05:@.+]] = private {{.*}}constant [1 x i32] [i32 32]
// CK27: [[SIZE07:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 4]
// CK27: [[MTYPE07:@.+]] = private {{.*}}constant [1 x i32] [i32 288]
// CK27: [[SIZE09:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 40]
// CK27: [[MTYPE09:@.+]] = private {{.*}}constant [1 x i32] [i32 161]
// CK27-LABEL: zero_size_section_and_private_maps
void zero_size_section_and_private_maps (int ii){
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// Map of a pointer.
int *pa;
// Region 00
// CK27-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE00]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE00]]{{.+}})
// CK27-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK27-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK27-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK27-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK27-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
// CK27-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK27-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
// CK27-DAG: store i32* [[VAR0]], i32** [[CP0]]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK27: call void [[CALL00:@.+]](i32* {{[^,]+}})
#pragma omp target
{
pa[50]++;
}
// Region 01
// CK27-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE01]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE01]]{{.+}})
// CK27-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK27-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK27-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK27-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK27-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
// CK27-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK27-DAG: store i32* [[RVAR0:%.+]], i32** [[CBP0]]
// CK27-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK27-DAG: [[RVAR0]] = load i32*, i32** [[VAR0:%[^,]+]]
// CK27-DAG: [[SEC0]] = getelementptr {{.*}}i32* [[RVAR00:%.+]], i{{.+}} 0
// CK27-DAG: [[RVAR00]] = load i32*, i32** [[VAR0]]
// CK27: call void [[CALL01:@.+]](i32* {{[^,]+}})
#pragma omp target map(pa[:0])
{
pa[50]++;
}
// Region 02
// CK27-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE02]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE02]]{{.+}})
// CK27-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK27-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK27-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK27-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK27-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
// CK27-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK27-DAG: store i32* [[RVAR0:%.+]], i32** [[CBP0]]
// CK27-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK27-DAG: [[RVAR0]] = load i32*, i32** [[VAR0:%[^,]+]]
// CK27-DAG: [[SEC0]] = getelementptr {{.*}}i32* [[RVAR00:%.+]], i{{.+}} 0
// CK27-DAG: [[RVAR00]] = load i32*, i32** [[VAR0]]
// CK27: call void [[CALL02:@.+]](i32* {{[^,]+}})
#pragma omp target map(pa[0:0])
{
pa[50]++;
}
// Region 03
// CK27-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE03]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE03]]{{.+}})
// CK27-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK27-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK27-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK27-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK27-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
// CK27-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK27-DAG: store i32* [[RVAR0:%.+]], i32** [[CBP0]]
// CK27-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
// CK27-DAG: [[RVAR0]] = load i32*, i32** [[VAR0:%[^,]+]]
// CK27-DAG: [[SEC0]] = getelementptr {{.*}}i32* [[RVAR00:%.+]], i{{.+}} %{{.+}}
// CK27-DAG: [[RVAR00]] = load i32*, i32** [[VAR0]]
// CK27: call void [[CALL03:@.+]](i32* {{[^,]+}})
#pragma omp target map(pa[ii:0])
{
pa[50]++;
}
[OpenMP] Add support for the 'private pointer' flag to signal variables captured in target regions and used in first-private clauses. Summary: If a variable is implicitly mapped (doesn't show in a map clause), the runtime library has to be informed if the corresponding capture shows up in first-private clause, so that the storage previously allocated in the device is used. This patch adds the support for that. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20112 llvm-svn: 270870
2016-05-27 00:53:38 +08:00
int *pvtPtr;
int pvtScl;
int pvtArr[10];
// Region 04
// CK27: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 0, i8** null, i8** null, i{{64|32}}* null, i32* null)
// CK27: call void [[CALL04:@.+]]()
#pragma omp target private(pvtPtr)
{
pvtPtr[5]++;
}
// Region 05
// CK27-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE05]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE05]]{{.+}})
// CK27-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK27-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK27-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK27-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK27-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
// CK27-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK27-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
// CK27-DAG: store i32* [[VAR0]], i32** [[CP0]]
[OpenMP] Add support for the 'private pointer' flag to signal variables captured in target regions and used in first-private clauses. Summary: If a variable is implicitly mapped (doesn't show in a map clause), the runtime library has to be informed if the corresponding capture shows up in first-private clause, so that the storage previously allocated in the device is used. This patch adds the support for that. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20112 llvm-svn: 270870
2016-05-27 00:53:38 +08:00
// CK27: call void [[CALL05:@.+]](i32* {{[^,]+}})
#pragma omp target firstprivate(pvtPtr)
{
pvtPtr[5]++;
}
// Region 06
// CK27: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 0, i8** null, i8** null, i{{64|32}}* null, i32* null)
// CK27: call void [[CALL06:@.+]]()
#pragma omp target private(pvtScl)
{
pvtScl++;
}
// Region 07
// CK27-DAG: call i32 @__tgt_target(i32 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZE07]]{{.+}}, {{.+}}[[MTYPE07]]{{.+}})
// CK27-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
// CK27-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
// CK27-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
// CK27-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK27-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[Z]]*
// CK27-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[Z]]*
// CK27-DAG: store i[[Z]] [[VAL:%.+]], i[[Z]]* [[CBP1]]
// CK27-DAG: store i[[Z]] [[VAL]], i[[Z]]* [[CP1]]
[OpenMP] Add support for the 'private pointer' flag to signal variables captured in target regions and used in first-private clauses. Summary: If a variable is implicitly mapped (doesn't show in a map clause), the runtime library has to be informed if the corresponding capture shows up in first-private clause, so that the storage previously allocated in the device is used. This patch adds the support for that. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20112 llvm-svn: 270870
2016-05-27 00:53:38 +08:00
// CK27-DAG: [[VAL]] = load i[[Z]], i[[Z]]* [[ADDR:%.+]],
// CK27-64-DAG: [[CADDR:%.+]] = bitcast i[[Z]]* [[ADDR]] to i32*
// CK27-64-DAG: store i32 {{.+}}, i32* [[CADDR]],
// CK27: call void [[CALL07:@.+]](i[[Z]] [[VAL]])
#pragma omp target firstprivate(pvtScl)
{
pvtScl++;
}
// Region 08
// CK27: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 0, i8** null, i8** null, i{{64|32}}* null, i32* null)
// CK27: call void [[CALL08:@.+]]()
#pragma omp target private(pvtArr)
{
pvtArr[5]++;
}
// Region 09
// CK27-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE09]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE09]]{{.+}})
// CK27-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK27-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK27-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK27-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK27-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [10 x i32]**
// CK27-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [10 x i32]**
// CK27-DAG: store [10 x i32]* [[VAR0:%.+]], [10 x i32]** [[CBP0]]
// CK27-DAG: store [10 x i32]* [[VAR0]], [10 x i32]** [[CP0]]
[OpenMP] Add support for the 'private pointer' flag to signal variables captured in target regions and used in first-private clauses. Summary: If a variable is implicitly mapped (doesn't show in a map clause), the runtime library has to be informed if the corresponding capture shows up in first-private clause, so that the storage previously allocated in the device is used. This patch adds the support for that. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20112 llvm-svn: 270870
2016-05-27 00:53:38 +08:00
// CK27: call void [[CALL09:@.+]]([10 x i32]* {{[^,]+}})
#pragma omp target firstprivate(pvtArr)
{
pvtArr[5]++;
}
[OpenMP] Adjust map type bits according to latest spec and use zero size array sections for pointers. Summary: This patch changes the bits used to specify the map types according to the latest version of the libomptarget document and add the support for zero size array section when pointers are being implicitly mapped. This completes the missing new 4.5 map semantics. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20111 llvm-svn: 270868
2016-05-27 00:48:10 +08:00
}
// CK27: define {{.+}}[[CALL00]]
// CK27: define {{.+}}[[CALL01]]
// CK27: define {{.+}}[[CALL02]]
// CK27: define {{.+}}[[CALL03]]
[OpenMP] Add support for the 'private pointer' flag to signal variables captured in target regions and used in first-private clauses. Summary: If a variable is implicitly mapped (doesn't show in a map clause), the runtime library has to be informed if the corresponding capture shows up in first-private clause, so that the storage previously allocated in the device is used. This patch adds the support for that. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: http://reviews.llvm.org/D20112 llvm-svn: 270870
2016-05-27 00:53:38 +08:00
// CK27: define {{.+}}[[CALL04]]
// CK27: define {{.+}}[[CALL05]]
// CK27: define {{.+}}[[CALL06]]
// CK27: define {{.+}}[[CALL07]]
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
#endif
[OpenMP] Add support for mapping array sections through pointer references. Summary: This patch fixes a bug in the map of array sections whose base is a reference to a pointer. The existing mapping support was not prepared to deal with it, causing the compiler to crash. Mapping a reference to a pointer enjoys the same characteristics of a regular pointer, i.e., it is passed by value. Therefore, the reference has to be materialized in the target region. Reviewers: hfinkel, carlo.bertolli, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: https://reviews.llvm.org/D22690 llvm-svn: 276933
2016-07-28 06:49:49 +08:00
///==========================================================================///
// RUN: %clang_cc1 -DCK28 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK28 --check-prefix CK28-64
// RUN: %clang_cc1 -DCK28 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK28 --check-prefix CK28-64
// RUN: %clang_cc1 -DCK28 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK28 --check-prefix CK28-32
// RUN: %clang_cc1 -DCK28 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK28 --check-prefix CK28-32
#ifdef CK28
// CK28: [[SIZE00:@.+]] = private {{.*}}constant [1 x i[[Z:64|32]]] [i[[Z:64|32]] {{8|4}}]
// CK28: [[MTYPE00:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
// CK28: [[SIZE01:@.+]] = private {{.*}}constant [1 x i[[Z]]] [i[[Z]] 400]
// CK28: [[MTYPE01:@.+]] = private {{.*}}constant [1 x i32] [i32 35]
// CK28-LABEL: explicit_maps_pointer_references
void explicit_maps_pointer_references (int *p){
int *&a = p;
// Region 00
// CK28-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE00]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE00]]{{.+}})
// CK28-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK28-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK28-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK28-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK28-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32***
// CK28-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32***
// CK28-DAG: store i32** [[VAR0:%.+]], i32*** [[CBP0]]
// CK28-DAG: store i32** [[VAR1:%.+]], i32*** [[CP0]]
[OpenMP] Add support for mapping array sections through pointer references. Summary: This patch fixes a bug in the map of array sections whose base is a reference to a pointer. The existing mapping support was not prepared to deal with it, causing the compiler to crash. Mapping a reference to a pointer enjoys the same characteristics of a regular pointer, i.e., it is passed by value. Therefore, the reference has to be materialized in the target region. Reviewers: hfinkel, carlo.bertolli, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: https://reviews.llvm.org/D22690 llvm-svn: 276933
2016-07-28 06:49:49 +08:00
// CK28-DAG: [[VAR0]] = load i32**, i32*** [[VAR00:%.+]],
// CK28-DAG: [[VAR1]] = load i32**, i32*** [[VAR11:%.+]],
// CK28: call void [[CALL00:@.+]](i32** {{[^,]+}})
#pragma omp target map(a)
{
++a;
}
// Region 01
// CK28-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE01]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE01]]{{.+}})
// CK28-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK28-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK28-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK28-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK28-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
// CK28-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
// CK28-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
// CK28-DAG: store i32* [[VAR1:%.+]], i32** [[CP0]]
[OpenMP] Add support for mapping array sections through pointer references. Summary: This patch fixes a bug in the map of array sections whose base is a reference to a pointer. The existing mapping support was not prepared to deal with it, causing the compiler to crash. Mapping a reference to a pointer enjoys the same characteristics of a regular pointer, i.e., it is passed by value. Therefore, the reference has to be materialized in the target region. Reviewers: hfinkel, carlo.bertolli, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: https://reviews.llvm.org/D22690 llvm-svn: 276933
2016-07-28 06:49:49 +08:00
// CK28-DAG: [[VAR0]] = load i32*, i32** [[VAR00:%.+]],
// CK28-DAG: [[VAR00]] = load i32**, i32*** [[VAR000:%.+]],
// CK28-DAG: [[VAR1]] = getelementptr inbounds i32, i32* [[VAR11:%.+]], i{{64|32}} 2
// CK28-DAG: [[VAR11]] = load i32*, i32** [[VAR111:%.+]],
// CK28-DAG: [[VAR111]] = load i32**, i32*** [[VAR1111:%.+]],
// CK28: call void [[CALL01:@.+]](i32* {{[^,]+}})
#pragma omp target map(a[2:100])
{
++a;
}
}
#endif
[OpenMP] Add support to map member expressions with references to pointers. Summary: This patch add support to map pointers through references in class members. Although a reference does not have storage that a user can access, it still has to be mapped in order to get the deep copy right and the dereferencing code work properly. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: https://reviews.llvm.org/D22787 llvm-svn: 276934
2016-07-28 06:52:16 +08:00
///==========================================================================///
// RUN: %clang_cc1 -DCK29 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK29 --check-prefix CK29-64
// RUN: %clang_cc1 -DCK29 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK29 --check-prefix CK29-64
// RUN: %clang_cc1 -DCK29 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix CK29 --check-prefix CK29-32
// RUN: %clang_cc1 -DCK29 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CK29 --check-prefix CK29-32
#ifdef CK29
// CK29: [[SSA:%.+]] = type { double*, double** }
// CK29: [[SSB:%.+]] = type { [[SSA]]*, [[SSA]]** }
// CK29: [[SIZE00:@.+]] = private {{.*}}constant [4 x i[[Z:64|32]]] [i[[Z:64|32]] {{8|4}}, i[[Z:64|32]] {{8|4}}, i[[Z:64|32]] {{8|4}}, i[[Z:64|32]] 80]
// CK29: [[MTYPE00:@.+]] = private {{.*}}constant [4 x i32] [i32 35, i32 16, i32 19, i32 19]
// CK29: [[SIZE01:@.+]] = private {{.*}}constant [4 x i[[Z]]] [i[[Z]] {{8|4}}, i[[Z]] {{8|4}}, i[[Z]] {{8|4}}, i[[Z]] 80]
// CK29: [[MTYPE01:@.+]] = private {{.*}}constant [4 x i32] [i32 32, i32 19, i32 19, i32 19]
// CK29: [[SIZE02:@.+]] = private {{.*}}constant [5 x i[[Z]]] [i[[Z]] {{8|4}}, i[[Z]] {{8|4}}, i[[Z]] {{8|4}}, i[[Z]] {{8|4}}, i[[Z]] 80]
// CK29: [[MTYPE02:@.+]] = private {{.*}}constant [5 x i32] [i32 32, i32 19, i32 16, i32 19, i32 19]
struct SSA{
double *p;
double *&pr;
SSA(double *&pr) : pr(pr) {}
};
struct SSB{
SSA *p;
SSA *&pr;
SSB(SSA *&pr) : pr(pr) {}
// CK29-LABEL: define {{.+}}foo
void foo() {
// Region 00
// CK29-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 4, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[4 x i{{.+}}]* [[SIZE00]], {{.+}}getelementptr {{.+}}[4 x i{{.+}}]* [[MTYPE00]]{{.+}})
// CK29-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK29-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK29-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK29-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK29-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[SSB]]**
// CK29-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[SSA]]**
// CK29-DAG: store [[SSB]]* [[VAR0:%.+]], [[SSB]]** [[CBP0]]
// CK29-DAG: store [[SSA]]** [[VAR00:%.+]], [[SSA]]*** [[CP0]]
[OpenMP] Add support to map member expressions with references to pointers. Summary: This patch add support to map pointers through references in class members. Although a reference does not have storage that a user can access, it still has to be mapped in order to get the deep copy right and the dereferencing code work properly. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: https://reviews.llvm.org/D22787 llvm-svn: 276934
2016-07-28 06:52:16 +08:00
// CK29-DAG: [[VAR0]] = load [[SSB]]*, [[SSB]]** %
// CK29-DAG: [[VAR00]] = getelementptr inbounds [[SSB]], [[SSB]]* [[VAR0]], i32 0, i32 0
// CK29-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK29-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK29-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[SSA]]***
// CK29-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to double****
// CK29-DAG: store [[SSA]]** [[VAR00]], [[SSA]]*** [[CBP1]]
// CK29-DAG: store double*** [[VAR1:%.+]], double**** [[CP1]]
[OpenMP] Add support to map member expressions with references to pointers. Summary: This patch add support to map pointers through references in class members. Although a reference does not have storage that a user can access, it still has to be mapped in order to get the deep copy right and the dereferencing code work properly. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: https://reviews.llvm.org/D22787 llvm-svn: 276934
2016-07-28 06:52:16 +08:00
// CK29-DAG: [[VAR1]] = getelementptr inbounds [[SSA]], [[SSA]]* %{{.+}}, i32 0, i32 1
// CK29-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 2
// CK29-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 2
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK29-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to double****
// CK29-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to double***
// CK29-DAG: store double*** [[VAR1]], double**** [[CBP2]]
// CK29-DAG: store double** [[VAR2:%.+]], double*** [[CP2]]
[OpenMP] Add support to map member expressions with references to pointers. Summary: This patch add support to map pointers through references in class members. Although a reference does not have storage that a user can access, it still has to be mapped in order to get the deep copy right and the dereferencing code work properly. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: https://reviews.llvm.org/D22787 llvm-svn: 276934
2016-07-28 06:52:16 +08:00
// CK29-DAG: [[VAR2]] = load double**, double*** [[VAR22:%.+]],
// CK29-DAG: [[VAR22]] = getelementptr inbounds [[SSA]], [[SSA]]* %{{.+}}, i32 0, i32 1
// CK29-DAG: [[BP3:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 3
// CK29-DAG: [[P3:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 3
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK29-DAG: [[CBP3:%.+]] = bitcast i8** [[BP3]] to double***
// CK29-DAG: [[CP3:%.+]] = bitcast i8** [[P3]] to double**
// CK29-DAG: store double** [[VAR2]], double*** [[CBP3]]
// CK29-DAG: store double* [[VAR3:%.+]], double** [[CP3]]
[OpenMP] Add support to map member expressions with references to pointers. Summary: This patch add support to map pointers through references in class members. Although a reference does not have storage that a user can access, it still has to be mapped in order to get the deep copy right and the dereferencing code work properly. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: https://reviews.llvm.org/D22787 llvm-svn: 276934
2016-07-28 06:52:16 +08:00
// CK29-DAG: [[VAR3]] = getelementptr inbounds double, double* [[VAR33:%.+]], i{{.+}} 0
// CK29-DAG: [[VAR33]] = load double*, double** %{{.+}},
// CK29: call void [[CALL00:@.+]]([[SSB]]* {{[^,]+}})
#pragma omp target map(p->pr[:10])
{
p->pr++;
}
// Region 01
// CK29-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 4, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[4 x i{{.+}}]* [[SIZE01]], {{.+}}getelementptr {{.+}}[4 x i{{.+}}]* [[MTYPE01]]{{.+}})
// CK29-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK29-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK29-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK29-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK29-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[SSB]]**
// CK29-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[SSA]]****
// CK29-DAG: store [[SSB]]* [[VAR0:%.+]], [[SSB]]** [[CBP0]]
// CK29-DAG: store [[SSA]]*** [[VAR00:%.+]], [[SSA]]**** [[CP0]]
[OpenMP] Add support to map member expressions with references to pointers. Summary: This patch add support to map pointers through references in class members. Although a reference does not have storage that a user can access, it still has to be mapped in order to get the deep copy right and the dereferencing code work properly. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: https://reviews.llvm.org/D22787 llvm-svn: 276934
2016-07-28 06:52:16 +08:00
// CK29-DAG: [[VAR00]] = getelementptr inbounds [[SSB]], [[SSB]]* [[VAR0]], i32 0, i32 1
// CK29-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK29-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK29-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[SSA]]****
// CK29-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [[SSA]]***
// CK29-DAG: store [[SSA]]*** [[VAR00]], [[SSA]]**** [[CBP1]]
// CK29-DAG: store [[SSA]]** [[VAR1:%.+]], [[SSA]]*** [[CP1]]
[OpenMP] Add support to map member expressions with references to pointers. Summary: This patch add support to map pointers through references in class members. Although a reference does not have storage that a user can access, it still has to be mapped in order to get the deep copy right and the dereferencing code work properly. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: https://reviews.llvm.org/D22787 llvm-svn: 276934
2016-07-28 06:52:16 +08:00
// CK29-DAG: [[VAR1]] = load [[SSA]]**, [[SSA]]*** [[VAR00]],
// CK29-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 2
// CK29-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 2
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK29-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to [[SSA]]***
// CK29-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to double***
// CK29-DAG: store [[SSA]]** [[VAR1]], [[SSA]]*** [[CBP2]]
// CK29-DAG: store double** [[VAR2:%.+]], double*** [[CP2]]
[OpenMP] Add support to map member expressions with references to pointers. Summary: This patch add support to map pointers through references in class members. Although a reference does not have storage that a user can access, it still has to be mapped in order to get the deep copy right and the dereferencing code work properly. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: https://reviews.llvm.org/D22787 llvm-svn: 276934
2016-07-28 06:52:16 +08:00
// CK29-DAG: [[VAR2]] = getelementptr inbounds [[SSA]], [[SSA]]* %{{.+}}, i32 0, i32 0
// CK29-DAG: [[BP3:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 3
// CK29-DAG: [[P3:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 3
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK29-DAG: [[CBP3:%.+]] = bitcast i8** [[BP3]] to double***
// CK29-DAG: [[CP3:%.+]] = bitcast i8** [[P3]] to double**
// CK29-DAG: store double** [[VAR2]], double*** [[CBP3]]
// CK29-DAG: store double* [[VAR3:%.+]], double** [[CP3]]
[OpenMP] Add support to map member expressions with references to pointers. Summary: This patch add support to map pointers through references in class members. Although a reference does not have storage that a user can access, it still has to be mapped in order to get the deep copy right and the dereferencing code work properly. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: https://reviews.llvm.org/D22787 llvm-svn: 276934
2016-07-28 06:52:16 +08:00
// CK29-DAG: [[VAR3]] = getelementptr inbounds double, double* [[VAR33:%.+]], i{{.+}} 0
// CK29-DAG: [[VAR33]] = load double*, double** %{{.+}},
// CK29: call void [[CALL00:@.+]]([[SSB]]* {{[^,]+}})
#pragma omp target map(pr->p[:10])
{
pr->p++;
}
// Region 02
// CK29-DAG: call i32 @__tgt_target(i32 {{[^,]+}}, i8* {{[^,]+}}, i32 5, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[5 x i{{.+}}]* [[SIZE02]], {{.+}}getelementptr {{.+}}[5 x i{{.+}}]* [[MTYPE02]]{{.+}})
// CK29-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
// CK29-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
// CK29-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
// CK29-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK29-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[SSB]]**
// CK29-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[SSA]]****
// CK29-DAG: store [[SSB]]* [[VAR0:%.+]], [[SSB]]** [[CBP0]]
// CK29-DAG: store [[SSA]]*** [[VAR00:%.+]], [[SSA]]**** [[CP0]]
[OpenMP] Add support to map member expressions with references to pointers. Summary: This patch add support to map pointers through references in class members. Although a reference does not have storage that a user can access, it still has to be mapped in order to get the deep copy right and the dereferencing code work properly. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: https://reviews.llvm.org/D22787 llvm-svn: 276934
2016-07-28 06:52:16 +08:00
// CK29-DAG: [[VAR00]] = getelementptr inbounds [[SSB]], [[SSB]]* [[VAR0]], i32 0, i32 1
// CK29-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
// CK29-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK29-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[SSA]]****
// CK29-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [[SSA]]***
// CK29-DAG: store [[SSA]]*** [[VAR00]], [[SSA]]**** [[CBP1]]
// CK29-DAG: store [[SSA]]** [[VAR1:%.+]], [[SSA]]*** [[CP1]]
[OpenMP] Add support to map member expressions with references to pointers. Summary: This patch add support to map pointers through references in class members. Although a reference does not have storage that a user can access, it still has to be mapped in order to get the deep copy right and the dereferencing code work properly. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: https://reviews.llvm.org/D22787 llvm-svn: 276934
2016-07-28 06:52:16 +08:00
// CK29-DAG: [[VAR1]] = load [[SSA]]**, [[SSA]]*** [[VAR00]],
// CK29-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 2
// CK29-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 2
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK29-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to [[SSA]]***
// CK29-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to double****
// CK29-DAG: store [[SSA]]** [[VAR1]], [[SSA]]*** [[CBP2]]
// CK29-DAG: store double*** [[VAR2:%.+]], double**** [[CP2]]
[OpenMP] Add support to map member expressions with references to pointers. Summary: This patch add support to map pointers through references in class members. Although a reference does not have storage that a user can access, it still has to be mapped in order to get the deep copy right and the dereferencing code work properly. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: https://reviews.llvm.org/D22787 llvm-svn: 276934
2016-07-28 06:52:16 +08:00
// CK29-DAG: [[VAR2]] = getelementptr inbounds [[SSA]], [[SSA]]* %{{.+}}, i32 0, i32 1
// CK29-DAG: [[BP3:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 3
// CK29-DAG: [[P3:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 3
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK29-DAG: [[CBP3:%.+]] = bitcast i8** [[BP3]] to double****
// CK29-DAG: [[CP3:%.+]] = bitcast i8** [[P3]] to double***
// CK29-DAG: store double*** [[VAR2]], double**** [[CBP3]]
// CK29-DAG: store double** [[VAR3:%.+]], double*** [[CP3]]
[OpenMP] Add support to map member expressions with references to pointers. Summary: This patch add support to map pointers through references in class members. Although a reference does not have storage that a user can access, it still has to be mapped in order to get the deep copy right and the dereferencing code work properly. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: https://reviews.llvm.org/D22787 llvm-svn: 276934
2016-07-28 06:52:16 +08:00
// CK29-DAG: [[VAR3]] = load double**, double*** [[VAR2]],
// CK29-DAG: [[BP4:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 4
// CK29-DAG: [[P4:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 4
[OPENMP][DEBUG] Generate second function with correct arg types. Currently, if the some of the parameters are captured by value, this argument is converted to uintptr_t type and thus we loosing the debug info about real type of the argument (captured variable): ``` void @.outlined_function.(uintptr %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.outlined_function.(uintptr %a.casted.val) ... ``` To resolve this problem, in debug mode a speciall external wrapper function is generated, that calls the outlined function with the correct parameters types: ``` void @.wrapper.(uintptr %par) { %a = alloca i32 %cast = bitcast i32* %a to uintptr* store uintptr %par, uintptr *%cast %a.val = load i32, i32* %a call void @.outlined_function.(i32 %a) ret void } void @.outlined_function.(i32 %par); ... %a = alloca i32 %a.casted = alloca uintptr %cast = bitcast uintptr* %a.casted to i32* %a.val = load i32, i32 *%a store i32 %a.val, i32 *%cast %a.casted.val = load uintptr, uintptr* %a.casted call void @.wrapper.(uintptr %a.casted.val) ... ``` llvm-svn: 306697
2017-06-30 00:43:05 +08:00
// CK29-DAG: [[CBP4:%.+]] = bitcast i8** [[BP4]] to double***
// CK29-DAG: [[CP4:%.+]] = bitcast i8** [[P4]] to double**
// CK29-DAG: store double** [[VAR3]], double*** [[CBP4]]
// CK29-DAG: store double* [[VAR4:%.+]], double** [[CP4]]
[OpenMP] Add support to map member expressions with references to pointers. Summary: This patch add support to map pointers through references in class members. Although a reference does not have storage that a user can access, it still has to be mapped in order to get the deep copy right and the dereferencing code work properly. Reviewers: hfinkel, carlo.bertolli, arpith-jacob, kkwli0, ABataev Subscribers: caomhin, cfe-commits Differential Revision: https://reviews.llvm.org/D22787 llvm-svn: 276934
2016-07-28 06:52:16 +08:00
// CK29-DAG: [[VAR4]] = getelementptr inbounds double, double* [[VAR44:%.+]], i{{.+}} 0
// CK29-DAG: [[VAR44]] = load double*, double**
// CK29: call void [[CALL00:@.+]]([[SSB]]* {{[^,]+}})
#pragma omp target map(pr->pr[:10])
{
pr->pr++;
}
}
};
void explicit_maps_member_pointer_references(SSA *sap) {
double *d;
SSA sa(d);
SSB sb(sap);
sb.foo();
}
#endif
[OpenMP] Update target directive codegen to use 4.5 implicit data mappings. Summary: This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed. Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available. Reviewers: hfinkel, rjmccall, ABataev Subscribers: cfe-commits, carlo.bertolli Differential Revision: http://reviews.llvm.org/D14940 llvm-svn: 254521
2015-12-03 01:44:43 +08:00
#endif