Redesign of loop's body in reverse pass #835
Conversation
…ble out of switch
… node creation when there's a continue stmt
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
| @@ -58,6 +58,8 @@ namespace clad { | |||
| std::set<clang::SourceLocation> m_ToBeRecorded; | |||
| /// A flag indicating if the Stmt we are currently visiting is inside loop. | |||
| bool isInsideLoop = false; | |||
| /// A flag indicating if the Stmt we are currently visiting is inside loop. | |||
| bool hasContStmt = false; | |||
There was a problem hiding this comment.
warning: member variable 'hasContStmt' has protected visibility [cppcoreguidelines-non-private-member-variables-in-classes]
bool hasContStmt = false;
^| @@ -58,6 +58,8 @@ | |||
| std::set<clang::SourceLocation> m_ToBeRecorded; | |||
| /// A flag indicating if the Stmt we are currently visiting is inside loop. | |||
| bool isInsideLoop = false; | |||
| /// A flag indicating if the Stmt we are currently visiting is inside loop. | |||
| bool hasContStmt = false; | |||
There was a problem hiding this comment.
warning: invalid case style for protected member 'hasContStmt' [readability-identifier-naming]
| bool hasContStmt = false; | |
| bool m_hasContStmt = false; |
| /// because we need to register all the switch cases later with the | ||
| /// switch statement that will be used to manage the control flow in | ||
| /// the reverse block. | ||
| llvm::SmallVector<clang::SwitchCase*, 4> m_SwitchCases; |
There was a problem hiding this comment.
warning: member variable 'm_SwitchCases' has public visibility [cppcoreguidelines-non-private-member-variables-in-classes]
llvm::SmallVector<clang::SwitchCase*, 4> m_SwitchCases;
^| /// the reverse block. | ||
| llvm::SmallVector<clang::SwitchCase*, 4> m_SwitchCases; | ||
|
|
||
| ReverseModeVisitor& m_RMV; |
There was a problem hiding this comment.
warning: member variable 'm_RMV' has public visibility [cppcoreguidelines-non-private-member-variables-in-classes]
ReverseModeVisitor& m_RMV;
^| void ReverseModeVisitor::AppendCaseStmts(llvm::SmallVectorImpl<Stmt*>& curBlock, | ||
| llvm::SmallVectorImpl<Stmt*>& cases, | ||
| Stmt* S, bool& afterCase) { | ||
| if (auto CS = dyn_cast_or_null<CompoundStmt>(S)) { |
There was a problem hiding this comment.
warning: 'auto CS' can be declared as 'auto *CS' [llvm-qualified-auto]
| if (auto CS = dyn_cast_or_null<CompoundStmt>(S)) { | |
| if (auto *CS = dyn_cast_or_null<CompoundStmt>(S)) { |
| // hence, we store the original flag's value | ||
| SaveAndRestore<bool> SaveAfterCase(afterCase); | ||
| afterCase = false; | ||
| for (auto stmt : CS->body()) |
There was a problem hiding this comment.
warning: 'auto stmt' can be declared as 'auto *stmt' [llvm-qualified-auto]
| for (auto stmt : CS->body()) | |
| for (auto *stmt : CS->body()) |
| AppendCaseStmts(newBlock, cases, stmt, afterCase); | ||
| if (!newBlock.empty()){ | ||
| auto Stmts_ref = clad_compat::makeArrayRef(newBlock.data(), newBlock.size()); | ||
| auto newCS = clad_compat::CompoundStmt_Create( |
There was a problem hiding this comment.
warning: 'auto newCS' can be declared as 'auto *newCS' [llvm-qualified-auto]
| auto newCS = clad_compat::CompoundStmt_Create( | |
| auto *newCS = clad_compat::CompoundStmt_Create( |
| } else if (isa<CaseStmt>(S)) { | ||
| afterCase = true; | ||
| cases.push_back(S); | ||
| } else if (auto If = dyn_cast_or_null<IfStmt>(S)) { |
There was a problem hiding this comment.
warning: 'auto If' can be declared as 'auto *If' [llvm-qualified-auto]
| } else if (auto If = dyn_cast_or_null<IfStmt>(S)) { | |
| } else if (auto *If = dyn_cast_or_null<IfStmt>(S)) { |
| afterCase = true; | ||
| cases.push_back(S); | ||
| } else if (auto If = dyn_cast_or_null<IfStmt>(S)) { | ||
| if (auto IfThenCS = dyn_cast_or_null<CompoundStmt>(If->getThen())) { |
There was a problem hiding this comment.
warning: 'auto IfThenCS' can be declared as 'auto *IfThenCS' [llvm-qualified-auto]
| if (auto IfThenCS = dyn_cast_or_null<CompoundStmt>(If->getThen())) { | |
| if (auto *IfThenCS = dyn_cast_or_null<CompoundStmt>(If->getThen())) { |
| Stmts thenBlock; | ||
| SaveAndRestore<bool> SaveAfterCase(afterCase); | ||
| afterCase = false; | ||
| for (auto stmt : IfThenCS->body()) |
There was a problem hiding this comment.
warning: 'auto stmt' can be declared as 'auto *stmt' [llvm-qualified-auto]
| for (auto stmt : IfThenCS->body()) | |
| for (auto *stmt : IfThenCS->body()) |
| Stmt* Reverse = clad_compat::IfStmt_Create( | ||
| m_Context, noLoc, If->isConstexpr(), initResult.getStmt_dx(), | ||
| condVarClone, reverseCond, noLoc, noLoc, | ||
| m_Sema.ActOnNullStmt(noLoc).get(), noLoc, elseDiff.getStmt_dx()); |
There was a problem hiding this comment.
Can we invert the condition instead of using a null stmt?
e.g.
if(!cond)
*else stmt*
instead of
if(cond)
;
else
*else stmt*
There was a problem hiding this comment.
I thought about this approach before but the way I believe this would turn out is checking for each binary condition operator (<=, >=, <, >, ==, !=) and replacing it accordingly. I will look into it more though, thanks for the comment.
There was a problem hiding this comment.
Thank you for the pull request. We need to verify it on more cases and discuss it further to make this change ready to merge, as I am currently not confident about the correctness of this design change. The below example fails when Clad is built using the changes in this pull request:
#include "clad/Differentiator/Differentiator.h"
#include <iostream>
#define show(x) std::cout << #x << ": " << x << "\n";
double fn(double u, double v) {
double res = 0;
for (int i = 0; i < 3; i++) {
res += u;
if (i == 1)
continue;
}
return res;
}
int main() {
auto d_fn = clad::gradient(fn);
double u = 3, v = 5;
double du, dv;
du = dv = 0;
d_fn.execute(u, v, &du, &dv);
show(du);
show(dv);
}
Output:
du: 1
dv: 0
Expected output:
du: 3
dv: 0
Clad on master gives correct result.
where the cases are independent and indexed correctly, AST speaking.
In the current design, is the indexing of case statements invalid as per C++ standard?
Yes, I completely agree that more tests need to be executed to verify that all cases are correctly derived. May I suggest creating a table to gather all the possible cases that may occur in order to make sure that we have not missed any? I can start working on it and append whatever else you and @PetroZarytskyi would like. Thank you for noticing it. I indeed did not cover the case where multiple cases share some stmts. Right now the approach I have in mind is keeping track of the stmts added in each level (this is the curBlock list probably at each level) till we hit a case and append them to the case hit. So it would look sth like this: However I understand that this is not very efficient, and if I tried to emulate a fall-through approach it would look similar to the original approach. I will have to investigate further. |
I would suggest to wait because this feature is not a priority right now. This feature would be important to resolve if the current indexing of case statements is invalid as per the C++ standard. By current, I mean the implementation in the master branch. I know the indexing of case statements in the current implementation is quirky and complex but we should be okay as long as it does not violate C++ standard.
Implementing this seems a little complicated and can lead to a significant code-size increase. |
|
Yes I agree, we can revisit this if I find another way to do this using fall-through. Feel free to add this as a Draft PR or even close it if it helps in scheduling the reviews of PRs. |
This PR closes #832. What these changes establish is a cleaner switch body, where the cases are independent and indexed correctly, AST speaking.
This comment contains a comparison of the current version and the proposed one. Cases of nested continue statements have also been tested and the derived body and the results were correct.
The workflow in a nutshell is the following:
thenorelsestmt ofifwith thecaseone when there's acontinuestmt in its bodyifthat doesn't include acontinuestmt is added firstly as it belongs to the previous switch case (outer one)casestmts are brought forward, on the first level of the loop's body, by performing a Depth-First search on its compoundscasestmt and that follow it (belong to this case) are also brought foward, while the rest of the stmts remain to their level and later appended in the beginning of the loop's body (as the main loop's body compound)Keep in mind that since the compilation ends up in a completely different result compared to the current implementation in cases were there's a continue stmt, the
testshave to be updated (so they're expected to currently fail).If the changes are well received, support for the
breakstmt in a similar manner only requires an additional variable.