[clang] [clang][SPIR-V] Always add convergence intrinsics (PR #88918)
https://github.com/Keenuts closed https://github.com/llvm/llvm-project/pull/88918 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [clang][SPIR-V] Always add convergence intrinsics (PR #88918)
Keenuts wrote: rebased on main, local tests are passing, waiting on CI to merge. https://github.com/llvm/llvm-project/pull/88918 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [clang][SPIR-V] Always add convergence intrinsics (PR #88918)
https://github.com/Keenuts updated
https://github.com/llvm/llvm-project/pull/88918
From 440cdfa4132a969702348c32f2810924012c5ea6 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Nathan=20Gau=C3=ABr?=
Date: Mon, 15 Apr 2024 17:05:40 +0200
Subject: [PATCH 1/6] [clang][SPIR-V] Always add convervence intrinsics
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
PR #80680 added bits in the codegen to lazily add convergence intrinsics
when required. This logic relied on the LoopStack. The issue is
when parsing the condition, the loopstack doesn't yet reflect the
correct values, as expected since we are not yet in the loop.
However, convergence tokens should sometimes already be available.
The solution which seemed the simplest is to greedily generate the
tokens when we generate SPIR-V.
Fixes #88144
Signed-off-by: Nathan Gauër
---
clang/lib/CodeGen/CGBuiltin.cpp | 88 +
clang/lib/CodeGen/CGCall.cpp | 3 +
clang/lib/CodeGen/CGStmt.cpp | 94 ++
clang/lib/CodeGen/CodeGenFunction.cpp | 9 ++
clang/lib/CodeGen/CodeGenFunction.h | 9 +-
.../builtins/RWBuffer-constructor.hlsl| 1 -
.../CodeGenHLSL/convergence/do.while.hlsl | 90 +
clang/test/CodeGenHLSL/convergence/for.hlsl | 121 ++
clang/test/CodeGenHLSL/convergence/while.hlsl | 119 +
9 files changed, 445 insertions(+), 89 deletions(-)
create mode 100644 clang/test/CodeGenHLSL/convergence/do.while.hlsl
create mode 100644 clang/test/CodeGenHLSL/convergence/for.hlsl
create mode 100644 clang/test/CodeGenHLSL/convergence/while.hlsl
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index f9ee93049b12d..e251091c6ce3e 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -1141,91 +1141,8 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
-// Returns the first convergence entry/loop/anchor instruction found in |BB|.
-// std::nullptr otherwise.
-llvm::IntrinsicInst *getConvergenceToken(llvm::BasicBlock *BB) {
- for (auto : *BB) {
-auto *II = dyn_cast();
-if (II && isConvergenceControlIntrinsic(II->getIntrinsicID()))
- return II;
- }
- return nullptr;
-}
-
} // namespace
-llvm::CallBase *
-CodeGenFunction::addConvergenceControlToken(llvm::CallBase *Input,
-llvm::Value *ParentToken) {
- llvm::Value *bundleArgs[] = {ParentToken};
- llvm::OperandBundleDef OB("convergencectrl", bundleArgs);
- auto Output = llvm::CallBase::addOperandBundle(
- Input, llvm::LLVMContext::OB_convergencectrl, OB, Input);
- Input->replaceAllUsesWith(Output);
- Input->eraseFromParent();
- return Output;
-}
-
-llvm::IntrinsicInst *
-CodeGenFunction::emitConvergenceLoopToken(llvm::BasicBlock *BB,
- llvm::Value *ParentToken) {
- CGBuilderTy::InsertPoint IP = Builder.saveIP();
- Builder.SetInsertPoint(>front());
- auto CB = Builder.CreateIntrinsic(
- llvm::Intrinsic::experimental_convergence_loop, {}, {});
- Builder.restoreIP(IP);
-
- auto I = addConvergenceControlToken(CB, ParentToken);
- return cast(I);
-}
-
-llvm::IntrinsicInst *
-CodeGenFunction::getOrEmitConvergenceEntryToken(llvm::Function *F) {
- auto *BB = >getEntryBlock();
- auto *token = getConvergenceToken(BB);
- if (token)
-return token;
-
- // Adding a convergence token requires the function to be marked as
- // convergent.
- F->setConvergent();
-
- CGBuilderTy::InsertPoint IP = Builder.saveIP();
- Builder.SetInsertPoint(>front());
- auto I = Builder.CreateIntrinsic(
- llvm::Intrinsic::experimental_convergence_entry, {}, {});
- assert(isa(I));
- Builder.restoreIP(IP);
-
- return cast(I);
-}
-
-llvm::IntrinsicInst *
-CodeGenFunction::getOrEmitConvergenceLoopToken(const LoopInfo *LI) {
- assert(LI != nullptr);
-
- auto *token = getConvergenceToken(LI->getHeader());
- if (token)
-return token;
-
- llvm::IntrinsicInst *PII =
- LI->getParent()
- ? emitConvergenceLoopToken(
-LI->getHeader(),
getOrEmitConvergenceLoopToken(LI->getParent()))
- : getOrEmitConvergenceEntryToken(LI->getHeader()->getParent());
-
- return emitConvergenceLoopToken(LI->getHeader(), PII);
-}
-
-llvm::CallBase *
-CodeGenFunction::addControlledConvergenceToken(llvm::CallBase *Input) {
- llvm::Value *ParentToken =
- LoopStack.hasInfo()
- ? getOrEmitConvergenceLoopToken(())
- : getOrEmitConvergenceEntryToken(Input->getFunction());
- return addConvergenceControlToken(Input, ParentToken);
-}
-
BitTest BitTest::decodeBitTestBuiltin(unsigned BuiltinID) {
switch (BuiltinID) {
// Main portable variants.
@@ -18402,12 +18319,9 @@ Value *CodeGenFunction::EmitHLSLBuiltinExpr(unsigned
BuiltinID,
ArrayRef{Op0}, nullptr, "dx.rsqrt");
}
case
[clang] [clang][SPIR-V] Always add convergence intrinsics (PR #88918)
@@ -1586,6 +1586,12 @@ class CodeGenModule : public CodeGenTypeCache {
void AddGlobalDtor(llvm::Function *Dtor, int Priority = 65535,
bool IsDtorAttrFunc = false);
+ // Return whether structured convergence intrinsics should be generated for
+ // this target.
+ bool shouldEmitConvergenceTokens() const {
+return getTriple().isSPIRVLogical();
Keenuts wrote:
That makes sense. Added a TODO line!
https://github.com/llvm/llvm-project/pull/88918
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [clang][SPIR-V] Always add convergence intrinsics (PR #88918)
https://github.com/Keenuts updated
https://github.com/llvm/llvm-project/pull/88918
From a8bf6fe83a1c145ef81ee30471dc51de1b5354ef Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Nathan=20Gau=C3=ABr?=
Date: Mon, 15 Apr 2024 17:05:40 +0200
Subject: [PATCH 1/6] [clang][SPIR-V] Always add convervence intrinsics
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
PR #80680 added bits in the codegen to lazily add convergence intrinsics
when required. This logic relied on the LoopStack. The issue is
when parsing the condition, the loopstack doesn't yet reflect the
correct values, as expected since we are not yet in the loop.
However, convergence tokens should sometimes already be available.
The solution which seemed the simplest is to greedily generate the
tokens when we generate SPIR-V.
Fixes #88144
Signed-off-by: Nathan Gauër
---
clang/lib/CodeGen/CGBuiltin.cpp | 88 +
clang/lib/CodeGen/CGCall.cpp | 3 +
clang/lib/CodeGen/CGStmt.cpp | 94 ++
clang/lib/CodeGen/CodeGenFunction.cpp | 9 ++
clang/lib/CodeGen/CodeGenFunction.h | 9 +-
.../builtins/RWBuffer-constructor.hlsl| 1 -
.../CodeGenHLSL/convergence/do.while.hlsl | 90 +
clang/test/CodeGenHLSL/convergence/for.hlsl | 121 ++
clang/test/CodeGenHLSL/convergence/while.hlsl | 119 +
9 files changed, 445 insertions(+), 89 deletions(-)
create mode 100644 clang/test/CodeGenHLSL/convergence/do.while.hlsl
create mode 100644 clang/test/CodeGenHLSL/convergence/for.hlsl
create mode 100644 clang/test/CodeGenHLSL/convergence/while.hlsl
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index 8e31652f4dabe..fb5904558bbae 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -1141,91 +1141,8 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
-// Returns the first convergence entry/loop/anchor instruction found in |BB|.
-// std::nullptr otherwise.
-llvm::IntrinsicInst *getConvergenceToken(llvm::BasicBlock *BB) {
- for (auto : *BB) {
-auto *II = dyn_cast();
-if (II && isConvergenceControlIntrinsic(II->getIntrinsicID()))
- return II;
- }
- return nullptr;
-}
-
} // namespace
-llvm::CallBase *
-CodeGenFunction::addConvergenceControlToken(llvm::CallBase *Input,
-llvm::Value *ParentToken) {
- llvm::Value *bundleArgs[] = {ParentToken};
- llvm::OperandBundleDef OB("convergencectrl", bundleArgs);
- auto Output = llvm::CallBase::addOperandBundle(
- Input, llvm::LLVMContext::OB_convergencectrl, OB, Input);
- Input->replaceAllUsesWith(Output);
- Input->eraseFromParent();
- return Output;
-}
-
-llvm::IntrinsicInst *
-CodeGenFunction::emitConvergenceLoopToken(llvm::BasicBlock *BB,
- llvm::Value *ParentToken) {
- CGBuilderTy::InsertPoint IP = Builder.saveIP();
- Builder.SetInsertPoint(>front());
- auto CB = Builder.CreateIntrinsic(
- llvm::Intrinsic::experimental_convergence_loop, {}, {});
- Builder.restoreIP(IP);
-
- auto I = addConvergenceControlToken(CB, ParentToken);
- return cast(I);
-}
-
-llvm::IntrinsicInst *
-CodeGenFunction::getOrEmitConvergenceEntryToken(llvm::Function *F) {
- auto *BB = >getEntryBlock();
- auto *token = getConvergenceToken(BB);
- if (token)
-return token;
-
- // Adding a convergence token requires the function to be marked as
- // convergent.
- F->setConvergent();
-
- CGBuilderTy::InsertPoint IP = Builder.saveIP();
- Builder.SetInsertPoint(>front());
- auto I = Builder.CreateIntrinsic(
- llvm::Intrinsic::experimental_convergence_entry, {}, {});
- assert(isa(I));
- Builder.restoreIP(IP);
-
- return cast(I);
-}
-
-llvm::IntrinsicInst *
-CodeGenFunction::getOrEmitConvergenceLoopToken(const LoopInfo *LI) {
- assert(LI != nullptr);
-
- auto *token = getConvergenceToken(LI->getHeader());
- if (token)
-return token;
-
- llvm::IntrinsicInst *PII =
- LI->getParent()
- ? emitConvergenceLoopToken(
-LI->getHeader(),
getOrEmitConvergenceLoopToken(LI->getParent()))
- : getOrEmitConvergenceEntryToken(LI->getHeader()->getParent());
-
- return emitConvergenceLoopToken(LI->getHeader(), PII);
-}
-
-llvm::CallBase *
-CodeGenFunction::addControlledConvergenceToken(llvm::CallBase *Input) {
- llvm::Value *ParentToken =
- LoopStack.hasInfo()
- ? getOrEmitConvergenceLoopToken(())
- : getOrEmitConvergenceEntryToken(Input->getFunction());
- return addConvergenceControlToken(Input, ParentToken);
-}
-
BitTest BitTest::decodeBitTestBuiltin(unsigned BuiltinID) {
switch (BuiltinID) {
// Main portable variants.
@@ -18400,12 +18317,9 @@ Value *CodeGenFunction::EmitHLSLBuiltinExpr(unsigned
BuiltinID,
ArrayRef{Op0}, nullptr, "dx.rsqrt");
}
case
[clang] [clang][SPIR-V] Always add convergence intrinsics (PR #88918)
Keenuts wrote: Thanks for the reviews. Waiting for 1 approval from MS and I'll merge https://github.com/llvm/llvm-project/pull/88918 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [clang][SPIR-V] Always add convergence intrinsics (PR #88918)
@@ -1586,6 +1586,12 @@ class CodeGenModule : public CodeGenTypeCache {
void AddGlobalDtor(llvm::Function *Dtor, int Priority = 65535,
bool IsDtorAttrFunc = false);
+ // Return whether structured convergence intrinsics should be generated for
+ // this target.
+ bool shouldEmitConvergenceTokens() const {
+return getTriple().isSPIRVLogical();
Keenuts wrote:
Not necessarily. The SPIR-V compute (as OpenCL not Vulkan compute) doesn't
require a structured CFG, so this could remain forever if they don't see any
benefit in generating a SCFG.
https://github.com/llvm/llvm-project/pull/88918
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [clang][SPIR-V] Always add convervence intrinsics (PR #88918)
Keenuts wrote: Hi all, rebased on main, and addressed the comments. This commits changes the register order on SPIR-V vs DXIL, which required me to fix the mad+lerp intrinsic tests. Should be NFC, just storing the register name in a CHECK variable. https://github.com/llvm/llvm-project/pull/88918 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [clang][SPIR-V] Always add convervence intrinsics (PR #88918)
https://github.com/Keenuts updated
https://github.com/llvm/llvm-project/pull/88918
From a8bf6fe83a1c145ef81ee30471dc51de1b5354ef Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Nathan=20Gau=C3=ABr?=
Date: Mon, 15 Apr 2024 17:05:40 +0200
Subject: [PATCH 1/5] [clang][SPIR-V] Always add convervence intrinsics
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
PR #80680 added bits in the codegen to lazily add convergence intrinsics
when required. This logic relied on the LoopStack. The issue is
when parsing the condition, the loopstack doesn't yet reflect the
correct values, as expected since we are not yet in the loop.
However, convergence tokens should sometimes already be available.
The solution which seemed the simplest is to greedily generate the
tokens when we generate SPIR-V.
Fixes #88144
Signed-off-by: Nathan Gauër
---
clang/lib/CodeGen/CGBuiltin.cpp | 88 +
clang/lib/CodeGen/CGCall.cpp | 3 +
clang/lib/CodeGen/CGStmt.cpp | 94 ++
clang/lib/CodeGen/CodeGenFunction.cpp | 9 ++
clang/lib/CodeGen/CodeGenFunction.h | 9 +-
.../builtins/RWBuffer-constructor.hlsl| 1 -
.../CodeGenHLSL/convergence/do.while.hlsl | 90 +
clang/test/CodeGenHLSL/convergence/for.hlsl | 121 ++
clang/test/CodeGenHLSL/convergence/while.hlsl | 119 +
9 files changed, 445 insertions(+), 89 deletions(-)
create mode 100644 clang/test/CodeGenHLSL/convergence/do.while.hlsl
create mode 100644 clang/test/CodeGenHLSL/convergence/for.hlsl
create mode 100644 clang/test/CodeGenHLSL/convergence/while.hlsl
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index 8e31652f4dabef..fb5904558bbae6 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -1141,91 +1141,8 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
-// Returns the first convergence entry/loop/anchor instruction found in |BB|.
-// std::nullptr otherwise.
-llvm::IntrinsicInst *getConvergenceToken(llvm::BasicBlock *BB) {
- for (auto : *BB) {
-auto *II = dyn_cast();
-if (II && isConvergenceControlIntrinsic(II->getIntrinsicID()))
- return II;
- }
- return nullptr;
-}
-
} // namespace
-llvm::CallBase *
-CodeGenFunction::addConvergenceControlToken(llvm::CallBase *Input,
-llvm::Value *ParentToken) {
- llvm::Value *bundleArgs[] = {ParentToken};
- llvm::OperandBundleDef OB("convergencectrl", bundleArgs);
- auto Output = llvm::CallBase::addOperandBundle(
- Input, llvm::LLVMContext::OB_convergencectrl, OB, Input);
- Input->replaceAllUsesWith(Output);
- Input->eraseFromParent();
- return Output;
-}
-
-llvm::IntrinsicInst *
-CodeGenFunction::emitConvergenceLoopToken(llvm::BasicBlock *BB,
- llvm::Value *ParentToken) {
- CGBuilderTy::InsertPoint IP = Builder.saveIP();
- Builder.SetInsertPoint(>front());
- auto CB = Builder.CreateIntrinsic(
- llvm::Intrinsic::experimental_convergence_loop, {}, {});
- Builder.restoreIP(IP);
-
- auto I = addConvergenceControlToken(CB, ParentToken);
- return cast(I);
-}
-
-llvm::IntrinsicInst *
-CodeGenFunction::getOrEmitConvergenceEntryToken(llvm::Function *F) {
- auto *BB = >getEntryBlock();
- auto *token = getConvergenceToken(BB);
- if (token)
-return token;
-
- // Adding a convergence token requires the function to be marked as
- // convergent.
- F->setConvergent();
-
- CGBuilderTy::InsertPoint IP = Builder.saveIP();
- Builder.SetInsertPoint(>front());
- auto I = Builder.CreateIntrinsic(
- llvm::Intrinsic::experimental_convergence_entry, {}, {});
- assert(isa(I));
- Builder.restoreIP(IP);
-
- return cast(I);
-}
-
-llvm::IntrinsicInst *
-CodeGenFunction::getOrEmitConvergenceLoopToken(const LoopInfo *LI) {
- assert(LI != nullptr);
-
- auto *token = getConvergenceToken(LI->getHeader());
- if (token)
-return token;
-
- llvm::IntrinsicInst *PII =
- LI->getParent()
- ? emitConvergenceLoopToken(
-LI->getHeader(),
getOrEmitConvergenceLoopToken(LI->getParent()))
- : getOrEmitConvergenceEntryToken(LI->getHeader()->getParent());
-
- return emitConvergenceLoopToken(LI->getHeader(), PII);
-}
-
-llvm::CallBase *
-CodeGenFunction::addControlledConvergenceToken(llvm::CallBase *Input) {
- llvm::Value *ParentToken =
- LoopStack.hasInfo()
- ? getOrEmitConvergenceLoopToken(())
- : getOrEmitConvergenceEntryToken(Input->getFunction());
- return addConvergenceControlToken(Input, ParentToken);
-}
-
BitTest BitTest::decodeBitTestBuiltin(unsigned BuiltinID) {
switch (BuiltinID) {
// Main portable variants.
@@ -18400,12 +18317,9 @@ Value *CodeGenFunction::EmitHLSLBuiltinExpr(unsigned
BuiltinID,
ArrayRef{Op0}, nullptr, "dx.rsqrt");
}
case
[clang] [clang][SPIR-V] Always add convervence intrinsics (PR #88918)
@@ -3101,3 +3130,68 @@ CodeGenFunction::GenerateCapturedStmtFunction(const
CapturedStmt ) {
return F;
}
+
+namespace {
+// Returns the first convergence entry/loop/anchor instruction found in |BB|.
+// std::nullptr otherwise.
+llvm::IntrinsicInst *getConvergenceToken(llvm::BasicBlock *BB) {
+ for (auto : *BB) {
+auto *II = dyn_cast();
+if (II && llvm::isConvergenceControlIntrinsic(II->getIntrinsicID()))
+ return II;
+ }
+ return nullptr;
+}
+
+} // namespace
+
+llvm::CallBase *
+CodeGenFunction::addConvergenceControlToken(llvm::CallBase *Input,
+llvm::Value *ParentToken) {
+ llvm::Value *bundleArgs[] = {ParentToken};
+ llvm::OperandBundleDef OB("convergencectrl", bundleArgs);
+ auto Output = llvm::CallBase::addOperandBundle(
+ Input, llvm::LLVMContext::OB_convergencectrl, OB, Input);
+ Input->replaceAllUsesWith(Output);
+ Input->eraseFromParent();
+ return Output;
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::emitConvergenceLoopToken(llvm::BasicBlock *BB,
+ llvm::Value *ParentToken) {
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+
+ if (BB->empty())
+Builder.SetInsertPoint(BB);
+ else
+Builder.SetInsertPoint(>front());
+
+ auto CB = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_loop, {}, {});
+ Builder.restoreIP(IP);
+
+ auto I = addConvergenceControlToken(CB, ParentToken);
Keenuts wrote:
Right, replaced the auto usage
https://github.com/llvm/llvm-project/pull/88918
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [clang][SPIR-V] Always add convervence intrinsics (PR #88918)
https://github.com/Keenuts updated
https://github.com/llvm/llvm-project/pull/88918
From 94d76dcdfac88d1d50fe705406c0280c33766e15 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Nathan=20Gau=C3=ABr?=
Date: Mon, 15 Apr 2024 17:05:40 +0200
Subject: [PATCH 1/4] [clang][SPIR-V] Always add convervence intrinsics
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
PR #80680 added bits in the codegen to lazily add convergence intrinsics
when required. This logic relied on the LoopStack. The issue is
when parsing the condition, the loopstack doesn't yet reflect the
correct values, as expected since we are not yet in the loop.
However, convergence tokens should sometimes already be available.
The solution which seemed the simplest is to greedily generate the
tokens when we generate SPIR-V.
Fixes #88144
Signed-off-by: Nathan Gauër
---
clang/lib/CodeGen/CGBuiltin.cpp | 88 +
clang/lib/CodeGen/CGCall.cpp | 3 +
clang/lib/CodeGen/CGStmt.cpp | 94 ++
clang/lib/CodeGen/CodeGenFunction.cpp | 9 ++
clang/lib/CodeGen/CodeGenFunction.h | 9 +-
.../builtins/RWBuffer-constructor.hlsl| 1 -
.../CodeGenHLSL/convergence/do.while.hlsl | 90 +
clang/test/CodeGenHLSL/convergence/for.hlsl | 121 ++
clang/test/CodeGenHLSL/convergence/while.hlsl | 119 +
9 files changed, 445 insertions(+), 89 deletions(-)
create mode 100644 clang/test/CodeGenHLSL/convergence/do.while.hlsl
create mode 100644 clang/test/CodeGenHLSL/convergence/for.hlsl
create mode 100644 clang/test/CodeGenHLSL/convergence/while.hlsl
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index df7502b8def5314..f5d40a1555fcb57 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -1133,91 +1133,8 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
-// Returns the first convergence entry/loop/anchor instruction found in |BB|.
-// std::nullptr otherwise.
-llvm::IntrinsicInst *getConvergenceToken(llvm::BasicBlock *BB) {
- for (auto : *BB) {
-auto *II = dyn_cast();
-if (II && isConvergenceControlIntrinsic(II->getIntrinsicID()))
- return II;
- }
- return nullptr;
-}
-
} // namespace
-llvm::CallBase *
-CodeGenFunction::addConvergenceControlToken(llvm::CallBase *Input,
-llvm::Value *ParentToken) {
- llvm::Value *bundleArgs[] = {ParentToken};
- llvm::OperandBundleDef OB("convergencectrl", bundleArgs);
- auto Output = llvm::CallBase::addOperandBundle(
- Input, llvm::LLVMContext::OB_convergencectrl, OB, Input);
- Input->replaceAllUsesWith(Output);
- Input->eraseFromParent();
- return Output;
-}
-
-llvm::IntrinsicInst *
-CodeGenFunction::emitConvergenceLoopToken(llvm::BasicBlock *BB,
- llvm::Value *ParentToken) {
- CGBuilderTy::InsertPoint IP = Builder.saveIP();
- Builder.SetInsertPoint(>front());
- auto CB = Builder.CreateIntrinsic(
- llvm::Intrinsic::experimental_convergence_loop, {}, {});
- Builder.restoreIP(IP);
-
- auto I = addConvergenceControlToken(CB, ParentToken);
- return cast(I);
-}
-
-llvm::IntrinsicInst *
-CodeGenFunction::getOrEmitConvergenceEntryToken(llvm::Function *F) {
- auto *BB = >getEntryBlock();
- auto *token = getConvergenceToken(BB);
- if (token)
-return token;
-
- // Adding a convergence token requires the function to be marked as
- // convergent.
- F->setConvergent();
-
- CGBuilderTy::InsertPoint IP = Builder.saveIP();
- Builder.SetInsertPoint(>front());
- auto I = Builder.CreateIntrinsic(
- llvm::Intrinsic::experimental_convergence_entry, {}, {});
- assert(isa(I));
- Builder.restoreIP(IP);
-
- return cast(I);
-}
-
-llvm::IntrinsicInst *
-CodeGenFunction::getOrEmitConvergenceLoopToken(const LoopInfo *LI) {
- assert(LI != nullptr);
-
- auto *token = getConvergenceToken(LI->getHeader());
- if (token)
-return token;
-
- llvm::IntrinsicInst *PII =
- LI->getParent()
- ? emitConvergenceLoopToken(
-LI->getHeader(),
getOrEmitConvergenceLoopToken(LI->getParent()))
- : getOrEmitConvergenceEntryToken(LI->getHeader()->getParent());
-
- return emitConvergenceLoopToken(LI->getHeader(), PII);
-}
-
-llvm::CallBase *
-CodeGenFunction::addControlledConvergenceToken(llvm::CallBase *Input) {
- llvm::Value *ParentToken =
- LoopStack.hasInfo()
- ? getOrEmitConvergenceLoopToken(())
- : getOrEmitConvergenceEntryToken(Input->getFunction());
- return addConvergenceControlToken(Input, ParentToken);
-}
-
BitTest BitTest::decodeBitTestBuiltin(unsigned BuiltinID) {
switch (BuiltinID) {
// Main portable variants.
@@ -18306,12 +18223,9 @@ Value *CodeGenFunction::EmitHLSLBuiltinExpr(unsigned
BuiltinID,
ArrayRef{Op0}, nullptr, "dx.rsqrt");
}
case
[clang] [llvm] [SPIRV] Add tan intrinsic part 3 (PR #90278)
https://github.com/Keenuts approved this pull request. https://github.com/llvm/llvm-project/pull/90278 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [clang][SPIR-V] Always add convervence intrinsics (PR #88918)
@@ -1109,6 +1124,10 @@ void CodeGenFunction::EmitForStmt(const ForStmt , llvm::BasicBlock *CondBlock = CondDest.getBlock(); EmitBlock(CondBlock); + if (getTarget().getTriple().isSPIRVLogical()) Keenuts wrote: Changed those for a module-level function which hides the target specific bit behind a more generic check. Let me know if that's not what you had in mind! https://github.com/llvm/llvm-project/pull/88918 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [clang][SPIR-V] Always add convervence intrinsics (PR #88918)
https://github.com/Keenuts updated
https://github.com/llvm/llvm-project/pull/88918
From 94d76dcdfac88d1d50fe705406c0280c33766e15 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Nathan=20Gau=C3=ABr?=
Date: Mon, 15 Apr 2024 17:05:40 +0200
Subject: [PATCH 1/3] [clang][SPIR-V] Always add convervence intrinsics
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
PR #80680 added bits in the codegen to lazily add convergence intrinsics
when required. This logic relied on the LoopStack. The issue is
when parsing the condition, the loopstack doesn't yet reflect the
correct values, as expected since we are not yet in the loop.
However, convergence tokens should sometimes already be available.
The solution which seemed the simplest is to greedily generate the
tokens when we generate SPIR-V.
Fixes #88144
Signed-off-by: Nathan Gauër
---
clang/lib/CodeGen/CGBuiltin.cpp | 88 +
clang/lib/CodeGen/CGCall.cpp | 3 +
clang/lib/CodeGen/CGStmt.cpp | 94 ++
clang/lib/CodeGen/CodeGenFunction.cpp | 9 ++
clang/lib/CodeGen/CodeGenFunction.h | 9 +-
.../builtins/RWBuffer-constructor.hlsl| 1 -
.../CodeGenHLSL/convergence/do.while.hlsl | 90 +
clang/test/CodeGenHLSL/convergence/for.hlsl | 121 ++
clang/test/CodeGenHLSL/convergence/while.hlsl | 119 +
9 files changed, 445 insertions(+), 89 deletions(-)
create mode 100644 clang/test/CodeGenHLSL/convergence/do.while.hlsl
create mode 100644 clang/test/CodeGenHLSL/convergence/for.hlsl
create mode 100644 clang/test/CodeGenHLSL/convergence/while.hlsl
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index df7502b8def531..f5d40a1555fcb5 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -1133,91 +1133,8 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
-// Returns the first convergence entry/loop/anchor instruction found in |BB|.
-// std::nullptr otherwise.
-llvm::IntrinsicInst *getConvergenceToken(llvm::BasicBlock *BB) {
- for (auto : *BB) {
-auto *II = dyn_cast();
-if (II && isConvergenceControlIntrinsic(II->getIntrinsicID()))
- return II;
- }
- return nullptr;
-}
-
} // namespace
-llvm::CallBase *
-CodeGenFunction::addConvergenceControlToken(llvm::CallBase *Input,
-llvm::Value *ParentToken) {
- llvm::Value *bundleArgs[] = {ParentToken};
- llvm::OperandBundleDef OB("convergencectrl", bundleArgs);
- auto Output = llvm::CallBase::addOperandBundle(
- Input, llvm::LLVMContext::OB_convergencectrl, OB, Input);
- Input->replaceAllUsesWith(Output);
- Input->eraseFromParent();
- return Output;
-}
-
-llvm::IntrinsicInst *
-CodeGenFunction::emitConvergenceLoopToken(llvm::BasicBlock *BB,
- llvm::Value *ParentToken) {
- CGBuilderTy::InsertPoint IP = Builder.saveIP();
- Builder.SetInsertPoint(>front());
- auto CB = Builder.CreateIntrinsic(
- llvm::Intrinsic::experimental_convergence_loop, {}, {});
- Builder.restoreIP(IP);
-
- auto I = addConvergenceControlToken(CB, ParentToken);
- return cast(I);
-}
-
-llvm::IntrinsicInst *
-CodeGenFunction::getOrEmitConvergenceEntryToken(llvm::Function *F) {
- auto *BB = >getEntryBlock();
- auto *token = getConvergenceToken(BB);
- if (token)
-return token;
-
- // Adding a convergence token requires the function to be marked as
- // convergent.
- F->setConvergent();
-
- CGBuilderTy::InsertPoint IP = Builder.saveIP();
- Builder.SetInsertPoint(>front());
- auto I = Builder.CreateIntrinsic(
- llvm::Intrinsic::experimental_convergence_entry, {}, {});
- assert(isa(I));
- Builder.restoreIP(IP);
-
- return cast(I);
-}
-
-llvm::IntrinsicInst *
-CodeGenFunction::getOrEmitConvergenceLoopToken(const LoopInfo *LI) {
- assert(LI != nullptr);
-
- auto *token = getConvergenceToken(LI->getHeader());
- if (token)
-return token;
-
- llvm::IntrinsicInst *PII =
- LI->getParent()
- ? emitConvergenceLoopToken(
-LI->getHeader(),
getOrEmitConvergenceLoopToken(LI->getParent()))
- : getOrEmitConvergenceEntryToken(LI->getHeader()->getParent());
-
- return emitConvergenceLoopToken(LI->getHeader(), PII);
-}
-
-llvm::CallBase *
-CodeGenFunction::addControlledConvergenceToken(llvm::CallBase *Input) {
- llvm::Value *ParentToken =
- LoopStack.hasInfo()
- ? getOrEmitConvergenceLoopToken(())
- : getOrEmitConvergenceEntryToken(Input->getFunction());
- return addConvergenceControlToken(Input, ParentToken);
-}
-
BitTest BitTest::decodeBitTestBuiltin(unsigned BuiltinID) {
switch (BuiltinID) {
// Main portable variants.
@@ -18306,12 +18223,9 @@ Value *CodeGenFunction::EmitHLSLBuiltinExpr(unsigned
BuiltinID,
ArrayRef{Op0}, nullptr, "dx.rsqrt");
}
case
[clang] [clang][SPIR-V] Always add convervence intrinsics (PR #88918)
@@ -1109,6 +1124,10 @@ void CodeGenFunction::EmitForStmt(const ForStmt , llvm::BasicBlock *CondBlock = CondDest.getBlock(); EmitBlock(CondBlock); + if (getTarget().getTriple().isSPIRVLogical()) Keenuts wrote: Right, something at the module level, which hides this into a more genering function, like the `shouldEmitRTTI`. Will fix. https://github.com/llvm/llvm-project/pull/88918 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [clang][SPIR-V] Set AS for the SPIR-V logical triple (PR #88939)
Keenuts wrote: Thanks all! Agree with Bogner, let's unblock the tests first. As for the address space for globals, this isn't something we have looked into yet, so I'd be in favor of keeping the same behavior as the SPIRN flavor until we have a reason to diverge (as in "thought about this issue" ) https://github.com/llvm/llvm-project/pull/88939 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [clang][SPIR-V] Always add convervence intrinsics (PR #88918)
https://github.com/Keenuts updated
https://github.com/llvm/llvm-project/pull/88918
From 94d76dcdfac88d1d50fe705406c0280c33766e15 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Nathan=20Gau=C3=ABr?=
Date: Mon, 15 Apr 2024 17:05:40 +0200
Subject: [PATCH 1/2] [clang][SPIR-V] Always add convervence intrinsics
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
PR #80680 added bits in the codegen to lazily add convergence intrinsics
when required. This logic relied on the LoopStack. The issue is
when parsing the condition, the loopstack doesn't yet reflect the
correct values, as expected since we are not yet in the loop.
However, convergence tokens should sometimes already be available.
The solution which seemed the simplest is to greedily generate the
tokens when we generate SPIR-V.
Fixes #88144
Signed-off-by: Nathan Gauër
---
clang/lib/CodeGen/CGBuiltin.cpp | 88 +
clang/lib/CodeGen/CGCall.cpp | 3 +
clang/lib/CodeGen/CGStmt.cpp | 94 ++
clang/lib/CodeGen/CodeGenFunction.cpp | 9 ++
clang/lib/CodeGen/CodeGenFunction.h | 9 +-
.../builtins/RWBuffer-constructor.hlsl| 1 -
.../CodeGenHLSL/convergence/do.while.hlsl | 90 +
clang/test/CodeGenHLSL/convergence/for.hlsl | 121 ++
clang/test/CodeGenHLSL/convergence/while.hlsl | 119 +
9 files changed, 445 insertions(+), 89 deletions(-)
create mode 100644 clang/test/CodeGenHLSL/convergence/do.while.hlsl
create mode 100644 clang/test/CodeGenHLSL/convergence/for.hlsl
create mode 100644 clang/test/CodeGenHLSL/convergence/while.hlsl
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index df7502b8def531..f5d40a1555fcb5 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -1133,91 +1133,8 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
-// Returns the first convergence entry/loop/anchor instruction found in |BB|.
-// std::nullptr otherwise.
-llvm::IntrinsicInst *getConvergenceToken(llvm::BasicBlock *BB) {
- for (auto : *BB) {
-auto *II = dyn_cast();
-if (II && isConvergenceControlIntrinsic(II->getIntrinsicID()))
- return II;
- }
- return nullptr;
-}
-
} // namespace
-llvm::CallBase *
-CodeGenFunction::addConvergenceControlToken(llvm::CallBase *Input,
-llvm::Value *ParentToken) {
- llvm::Value *bundleArgs[] = {ParentToken};
- llvm::OperandBundleDef OB("convergencectrl", bundleArgs);
- auto Output = llvm::CallBase::addOperandBundle(
- Input, llvm::LLVMContext::OB_convergencectrl, OB, Input);
- Input->replaceAllUsesWith(Output);
- Input->eraseFromParent();
- return Output;
-}
-
-llvm::IntrinsicInst *
-CodeGenFunction::emitConvergenceLoopToken(llvm::BasicBlock *BB,
- llvm::Value *ParentToken) {
- CGBuilderTy::InsertPoint IP = Builder.saveIP();
- Builder.SetInsertPoint(>front());
- auto CB = Builder.CreateIntrinsic(
- llvm::Intrinsic::experimental_convergence_loop, {}, {});
- Builder.restoreIP(IP);
-
- auto I = addConvergenceControlToken(CB, ParentToken);
- return cast(I);
-}
-
-llvm::IntrinsicInst *
-CodeGenFunction::getOrEmitConvergenceEntryToken(llvm::Function *F) {
- auto *BB = >getEntryBlock();
- auto *token = getConvergenceToken(BB);
- if (token)
-return token;
-
- // Adding a convergence token requires the function to be marked as
- // convergent.
- F->setConvergent();
-
- CGBuilderTy::InsertPoint IP = Builder.saveIP();
- Builder.SetInsertPoint(>front());
- auto I = Builder.CreateIntrinsic(
- llvm::Intrinsic::experimental_convergence_entry, {}, {});
- assert(isa(I));
- Builder.restoreIP(IP);
-
- return cast(I);
-}
-
-llvm::IntrinsicInst *
-CodeGenFunction::getOrEmitConvergenceLoopToken(const LoopInfo *LI) {
- assert(LI != nullptr);
-
- auto *token = getConvergenceToken(LI->getHeader());
- if (token)
-return token;
-
- llvm::IntrinsicInst *PII =
- LI->getParent()
- ? emitConvergenceLoopToken(
-LI->getHeader(),
getOrEmitConvergenceLoopToken(LI->getParent()))
- : getOrEmitConvergenceEntryToken(LI->getHeader()->getParent());
-
- return emitConvergenceLoopToken(LI->getHeader(), PII);
-}
-
-llvm::CallBase *
-CodeGenFunction::addControlledConvergenceToken(llvm::CallBase *Input) {
- llvm::Value *ParentToken =
- LoopStack.hasInfo()
- ? getOrEmitConvergenceLoopToken(())
- : getOrEmitConvergenceEntryToken(Input->getFunction());
- return addConvergenceControlToken(Input, ParentToken);
-}
-
BitTest BitTest::decodeBitTestBuiltin(unsigned BuiltinID) {
switch (BuiltinID) {
// Main portable variants.
@@ -18306,12 +18223,9 @@ Value *CodeGenFunction::EmitHLSLBuiltinExpr(unsigned
BuiltinID,
ArrayRef{Op0}, nullptr, "dx.rsqrt");
}
case
[clang] [clang][SPIR-V] Always add convervence intrinsics (PR #88918)
@@ -4987,7 +4990,11 @@ class CodeGenFunction : public CodeGenTypeCache {
const llvm::Twine = "");
// Adds a convergence_ctrl token to |Input| and emits the required parent
// convergence instructions.
- llvm::CallBase *addControlledConvergenceToken(llvm::CallBase *Input);
+ template
+ CallType *addControlledConvergenceToken(CallType *Input) {
+return dyn_cast(
Keenuts wrote:
Forgot to fix the use, and yes, should be a cast. Bad habits...
https://github.com/llvm/llvm-project/pull/88918
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [clang][SPIR-V] Always add convervence intrinsics (PR #88918)
https://github.com/Keenuts created
https://github.com/llvm/llvm-project/pull/88918
PR #80680 added bits in the codegen to lazily add convergence intrinsics when
required. This logic relied on the LoopStack. The issue is when parsing the
condition, the loopstack doesn't yet reflect the correct values, as expected
since we are not yet in the loop.
However, convergence tokens should sometimes already be available. The solution
which seemed the simplest is to greedily generate the tokens when we generate
SPIR-V.
Fixes #88144
From 94d76dcdfac88d1d50fe705406c0280c33766e15 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Nathan=20Gau=C3=ABr?=
Date: Mon, 15 Apr 2024 17:05:40 +0200
Subject: [PATCH] [clang][SPIR-V] Always add convervence intrinsics
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
PR #80680 added bits in the codegen to lazily add convergence intrinsics
when required. This logic relied on the LoopStack. The issue is
when parsing the condition, the loopstack doesn't yet reflect the
correct values, as expected since we are not yet in the loop.
However, convergence tokens should sometimes already be available.
The solution which seemed the simplest is to greedily generate the
tokens when we generate SPIR-V.
Fixes #88144
Signed-off-by: Nathan Gauër
---
clang/lib/CodeGen/CGBuiltin.cpp | 88 +
clang/lib/CodeGen/CGCall.cpp | 3 +
clang/lib/CodeGen/CGStmt.cpp | 94 ++
clang/lib/CodeGen/CodeGenFunction.cpp | 9 ++
clang/lib/CodeGen/CodeGenFunction.h | 9 +-
.../builtins/RWBuffer-constructor.hlsl| 1 -
.../CodeGenHLSL/convergence/do.while.hlsl | 90 +
clang/test/CodeGenHLSL/convergence/for.hlsl | 121 ++
clang/test/CodeGenHLSL/convergence/while.hlsl | 119 +
9 files changed, 445 insertions(+), 89 deletions(-)
create mode 100644 clang/test/CodeGenHLSL/convergence/do.while.hlsl
create mode 100644 clang/test/CodeGenHLSL/convergence/for.hlsl
create mode 100644 clang/test/CodeGenHLSL/convergence/while.hlsl
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index df7502b8def531..f5d40a1555fcb5 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -1133,91 +1133,8 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
-// Returns the first convergence entry/loop/anchor instruction found in |BB|.
-// std::nullptr otherwise.
-llvm::IntrinsicInst *getConvergenceToken(llvm::BasicBlock *BB) {
- for (auto : *BB) {
-auto *II = dyn_cast();
-if (II && isConvergenceControlIntrinsic(II->getIntrinsicID()))
- return II;
- }
- return nullptr;
-}
-
} // namespace
-llvm::CallBase *
-CodeGenFunction::addConvergenceControlToken(llvm::CallBase *Input,
-llvm::Value *ParentToken) {
- llvm::Value *bundleArgs[] = {ParentToken};
- llvm::OperandBundleDef OB("convergencectrl", bundleArgs);
- auto Output = llvm::CallBase::addOperandBundle(
- Input, llvm::LLVMContext::OB_convergencectrl, OB, Input);
- Input->replaceAllUsesWith(Output);
- Input->eraseFromParent();
- return Output;
-}
-
-llvm::IntrinsicInst *
-CodeGenFunction::emitConvergenceLoopToken(llvm::BasicBlock *BB,
- llvm::Value *ParentToken) {
- CGBuilderTy::InsertPoint IP = Builder.saveIP();
- Builder.SetInsertPoint(>front());
- auto CB = Builder.CreateIntrinsic(
- llvm::Intrinsic::experimental_convergence_loop, {}, {});
- Builder.restoreIP(IP);
-
- auto I = addConvergenceControlToken(CB, ParentToken);
- return cast(I);
-}
-
-llvm::IntrinsicInst *
-CodeGenFunction::getOrEmitConvergenceEntryToken(llvm::Function *F) {
- auto *BB = >getEntryBlock();
- auto *token = getConvergenceToken(BB);
- if (token)
-return token;
-
- // Adding a convergence token requires the function to be marked as
- // convergent.
- F->setConvergent();
-
- CGBuilderTy::InsertPoint IP = Builder.saveIP();
- Builder.SetInsertPoint(>front());
- auto I = Builder.CreateIntrinsic(
- llvm::Intrinsic::experimental_convergence_entry, {}, {});
- assert(isa(I));
- Builder.restoreIP(IP);
-
- return cast(I);
-}
-
-llvm::IntrinsicInst *
-CodeGenFunction::getOrEmitConvergenceLoopToken(const LoopInfo *LI) {
- assert(LI != nullptr);
-
- auto *token = getConvergenceToken(LI->getHeader());
- if (token)
-return token;
-
- llvm::IntrinsicInst *PII =
- LI->getParent()
- ? emitConvergenceLoopToken(
-LI->getHeader(),
getOrEmitConvergenceLoopToken(LI->getParent()))
- : getOrEmitConvergenceEntryToken(LI->getHeader()->getParent());
-
- return emitConvergenceLoopToken(LI->getHeader(), PII);
-}
-
-llvm::CallBase *
-CodeGenFunction::addControlledConvergenceToken(llvm::CallBase *Input) {
- llvm::Value *ParentToken =
- LoopStack.hasInfo()
-
[clang] [llvm] [HLSL][SPIRV] Add any intrinsic lowering (PR #88325)
https://github.com/Keenuts approved this pull request. https://github.com/llvm/llvm-project/pull/88325 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [clang][CodeGen] Add AS for Globals to SPIR & SPIRV datalayouts (PR #88455)
https://github.com/Keenuts approved this pull request. https://github.com/llvm/llvm-project/pull/88455 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [HLSL][DXIL][SPIRV] Implementation of an abstraction for intrinsic selection of HLSL backends (PR #87171)
https://github.com/Keenuts edited https://github.com/llvm/llvm-project/pull/87171 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [HLSL][DXIL][SPIRV] Implementation of an abstraction for intrinsic selection of HLSL backends (PR #87171)
https://github.com/Keenuts edited https://github.com/llvm/llvm-project/pull/87171 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [HLSL][DXIL][SPIRV] Implementation of an abstraction for intrinsic selection of HLSL backends (PR #87171)
@@ -0,0 +1,44 @@
+
+//===- CGHLSLUtils.h - Utility functions for HLSL CodeGen ---*- C++
-*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM
Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===--===//
+//
+// This File Provides utility function for HLSL code generation.
+// It is used to abstract away implementation details of backends.
+//
+//===--===//
+
+#ifndef LLVM_CLANG_LIB_CODEGEN_CGHLSLUTILS_H
+#define LLVM_CLANG_LIB_CODEGEN_CGHLSLUTILS_H
+
+#include "llvm/IR/Intrinsics.h"
+#include "llvm/IR/IntrinsicsDirectX.h"
+#include "llvm/IR/IntrinsicsSPIRV.h"
+
+// Define the function generator macro
+#define GENERATE_HLSL_INTRINSIC_FUNCTION(name)
\
+ static llvm::Intrinsic::ID get_hlsl_##name##_intrinsic(
\
+ const llvm::Triple::ArchType Arch) {
\
+switch (Arch) {
\
+case llvm::Triple::dxil:
\
+ return llvm::Intrinsic::dx_##name;
\
+case llvm::Triple::spirv:
\
+ return llvm::Intrinsic::spv_##name;
\
+default:
\
+ llvm_unreachable("Input semantic not supported by target");
\
+}
\
+ }
+
+class CGHLSLUtils {
+public:
+ GENERATE_HLSL_INTRINSIC_FUNCTION(all)
Keenuts wrote:
Why are those functions in a class and not a namespace? Does the class provides
something a namespace wouldn't?
https://github.com/llvm/llvm-project/pull/87171
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [HLSL][DXIL][SPIRV] Implementation of an abstraction for intrinsic selection of HLSL backends (PR #87171)
@@ -100,6 +100,118 @@ double3 abs(double3); _HLSL_BUILTIN_ALIAS(__builtin_elementwise_abs) double4 abs(double4); +//===--===// +// all builtins +//===--===// + +/// \fn bool all(T x) +/// \brief Returns True if all components of the \a x parameter are non-zero; +/// otherwise, false. \param x The input value. + +#ifdef __HLSL_ENABLE_16_BIT +_HLSL_AVAILABILITY(shadermodel, 6.2) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(int16_t); +_HLSL_AVAILABILITY(shadermodel, 6.2) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(int16_t2); +_HLSL_AVAILABILITY(shadermodel, 6.2) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(int16_t3); +_HLSL_AVAILABILITY(shadermodel, 6.2) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(int16_t4); +_HLSL_AVAILABILITY(shadermodel, 6.2) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(uint16_t); +_HLSL_AVAILABILITY(shadermodel, 6.2) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(uint16_t2); +_HLSL_AVAILABILITY(shadermodel, 6.2) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(uint16_t3); +_HLSL_AVAILABILITY(shadermodel, 6.2) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(uint16_t4); +#endif + +_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(half); +_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(half2); +_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(half3); +_HLSL_16BIT_AVAILABILITY(shadermodel, 6.2) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(half4); + +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(bool); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(bool2); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(bool3); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(bool4); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) + +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(int); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(int2); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(int3); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(int4); + +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(uint); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(uint2); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(uint3); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(uint4); + +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(float); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(float2); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(float3); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(float4); + +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(int64_t); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(int64_t2); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(int64_t3); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) +bool all(int64_t4); Keenuts wrote: Not sure if that's considered cleaner, but could those repetition be avoid with something like: ```cpp #define _HLSL_ALIAS_BUILTIN_FOR_VECTOR(BuiltIn, BaseType) \ BuiltIn(BaseType); \ BuiltIn(BaseType ## 2); \ BuiltIn(BaseType ## 3); \ BuiltIn(BaseType ## 4) ``` Which would be used like so: ```cpp #define _DEFINE_BUILTIN_HLSL_ELEMENTWISE_ALL(Type) \ _HLSL_AVAILABILITY(shadermodel, 6.2)\ _HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_all) \ bool all(Type) _HLSL_ALIAS_BUILTIN_FOR_VECTOR(_DEFINE_BUILTIN_HLSL_ELEMENTWISE_ALL, int); _HLSL_ALIAS_BUILTIN_FOR_VECTOR(_DEFINE_BUILTIN_HLSL_ELEMENTWISE_ALL, uint); _HLSL_ALIAS_BUILTIN_FOR_VECTOR(_DEFINE_BUILTIN_HLSL_ELEMENTWISE_ALL, float); _HLSL_ALIAS_BUILTIN_FOR_VECTOR(_DEFINE_BUILTIN_HLSL_ELEMENTWISE_ALL, double); #define _DEFINE_BUILTIN_HLSL_ELEMENTWISE_ANY(Type) \ _HLSL_AVAILABILITY(shadermodel, 6.2)\ _HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_any) \ bool all(Type) _HLSL_ALIAS_BUILTIN_FOR_VECTOR(_DEFINE_BUILTIN_HLSL_ELEMENTWISE_ANY, int); _HLSL_ALIAS_BUILTIN_FOR_VECTOR(_DEFINE_BUILTIN_HLSL_ELEMENTWISE_ANY, uint); _HLSL_ALIAS_BUILTIN_FOR_VECTOR(_DEFINE_BUILTIN_HLSL_ELEMENTWISE_ANY, float); _HLSL_ALIAS_BUILTIN_FOR_VECTOR(_DEFINE_BUILTIN_HLSL_ELEMENTWISE_ANY, double); ``` https://github.com/llvm/llvm-project/pull/87171 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [HLSL][DXIL][SPIRV] Implementation of an abstraction for intrinsic selection of HLSL backends (PR #87171)
@@ -0,0 +1,44 @@ + +//===- CGHLSLUtils.h - Utility functions for HLSL CodeGen ---*- C++ -*-===// +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// +//===--===// +// +// This File Provides utility function for HLSL code generation. Keenuts wrote: nit: ```suggestion // This file provides utility functions for HLSL code generation. ``` https://github.com/llvm/llvm-project/pull/87171 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [HLSL][DXIL][SPIRV] Implementation of an abstraction for intrinsic selection of HLSL backends (PR #87171)
@@ -0,0 +1,95 @@ +; RUN: llc -O0 -mtriple=spirv-unknown-unknown %s -o - | FileCheck %s +; Note: The validator is wrong it wants the return to be a bool vector when it is bool scalar return Keenuts wrote: AFAIK the validator is not wrong, it complains for 2 reasons: `OpAll` cannot be used with a single boolean as input: It **must** be a vector of boolean. And SPIR-V doesn't allow vector of 1 value. Hence, `all(my_bool)` shall be translated to `return my_bool`. Then, for int/long intrinsics: HLSL only checks if the value (any type) is non-zero for all vector items. for SPIR-V, `OpAll` only works for boolean vectors. This means the translation is not that straightforward. What is being done in DXC is the following: ``` %const = OpConstantComposite %v2int %int_0 %int_0 %tmp = OpINotEqual %v2bool %variable %const %res = OpAll %bool %tmp ``` https://github.com/llvm/llvm-project/pull/87171 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [HLSL][DXIL][SPIRV] Implementation of an abstraction for intrinsic selection of HLSL backends (PR #87171)
@@ -0,0 +1,44 @@
+
+//===- CGHLSLUtils.h - Utility functions for HLSL CodeGen ---*- C++
-*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM
Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===--===//
+//
+// This File Provides utility function for HLSL code generation.
+// It is used to abstract away implementation details of backends.
+//
+//===--===//
+
+#ifndef LLVM_CLANG_LIB_CODEGEN_CGHLSLUTILS_H
+#define LLVM_CLANG_LIB_CODEGEN_CGHLSLUTILS_H
+
+#include "llvm/IR/Intrinsics.h"
+#include "llvm/IR/IntrinsicsDirectX.h"
+#include "llvm/IR/IntrinsicsSPIRV.h"
+
+// Define the function generator macro
+#define GENERATE_HLSL_INTRINSIC_FUNCTION(name)
\
+ static llvm::Intrinsic::ID get_hlsl_##name##_intrinsic(
\
+ const llvm::Triple::ArchType Arch) {
\
+switch (Arch) {
\
+case llvm::Triple::dxil:
\
+ return llvm::Intrinsic::dx_##name;
\
+case llvm::Triple::spirv:
\
+ return llvm::Intrinsic::spv_##name;
\
+default:
\
+ llvm_unreachable("Input semantic not supported by target");
\
Keenuts wrote:
Shall the error be `"Intrinsic " #name " not supported by target architecture"`?
https://github.com/llvm/llvm-project/pull/87171
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [HLSL][clang] Move hlsl_wave_get_lane_index to EmitHLSLBuiltinExpr (PR #87131)
Keenuts wrote: Hi! Thanks @farzonl and @marcauberer for making sure SPIR-V backend didn't break, really appreciate it! https://github.com/llvm/llvm-project/pull/87131 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
https://github.com/Keenuts closed https://github.com/llvm/llvm-project/pull/80680 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
Keenuts wrote: Local tests for SPIR-V & DXIL pass https://github.com/llvm/llvm-project/pull/80680 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
https://github.com/Keenuts updated
https://github.com/llvm/llvm-project/pull/80680
From dc008167980ca0a479d2cdceeeb1ab6cd4983ec3 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Nathan=20Gau=C3=ABr?=
Date: Fri, 2 Feb 2024 16:38:46 +0100
Subject: [PATCH 1/9] [clang][HLSL][SPRI-V] Add convergence intrinsics
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
HLSL has wave operations and other kind of function which required the
control flow to either be converged, or respect certain constraints as
where and how to re-converge.
At the HLSL level, the convergence are mostly obvious: the control flow
is expected to re-converge at the end of a scope.
Once translated to IR, HLSL scopes disapear. This means we need a way to
communicate convergence restrictions down to the backend.
For this, the SPIR-V backend uses convergence intrinsics. So this commit
adds some code to generate convergence intrinsics when required.
This commit is not to be submitted as-is (lacks testing), but
should serve as a basis for an upcoming RFC.
Signed-off-by: Nathan Gauër
---
clang/lib/CodeGen/CGBuiltin.cpp | 102 +++
clang/lib/CodeGen/CGCall.cpp | 4 ++
clang/lib/CodeGen/CGLoopInfo.h | 8 ++-
clang/lib/CodeGen/CodeGenFunction.h | 19 +
llvm/include/llvm/IR/IntrinsicInst.h | 13
5 files changed, 145 insertions(+), 1 deletion(-)
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index fdb517eb254d3b..76fe8c1d40a419 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -1131,8 +1131,97 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
+
+// Returns the first convergence entry/loop/anchor instruction found in |BB|.
+// std::nullopt otherwise.
+std::optional getConvergenceToken(llvm::BasicBlock *BB)
{
+ for (auto : *BB) {
+auto *II = dyn_cast();
+if (II && isConvergenceControlIntrinsic(II->getIntrinsicID()))
+ return II;
+ }
+ return std::nullopt;
+}
+
} // namespace
+llvm::CallBase *
+CodeGenFunction::AddConvergenceControlAttr(llvm::CallBase *Input,
+ llvm::Value *ParentToken) {
+ llvm::Value *bundleArgs[] = {ParentToken};
+ llvm::OperandBundleDef OB("convergencectrl", bundleArgs);
+ auto Output = llvm::CallBase::addOperandBundle(
+ Input, llvm::LLVMContext::OB_convergencectrl, OB, Input);
+ Input->replaceAllUsesWith(Output);
+ Input->eraseFromParent();
+ return Output;
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::EmitConvergenceLoop(llvm::BasicBlock *BB,
+ llvm::Value *ParentToken) {
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+ Builder.SetInsertPoint(>front());
+ auto CB = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_loop, {}, {});
+ Builder.restoreIP(IP);
+
+ auto I = AddConvergenceControlAttr(CB, ParentToken);
+ // Controlled convergence is incompatible with uncontrolled convergence.
+ // Removing any old attributes.
+ I->setNotConvergent();
+
+ assert(isa(I));
+ return dyn_cast(I);
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::getOrEmitConvergenceEntryToken(llvm::Function *F) {
+ auto *BB = >getEntryBlock();
+ auto token = getConvergenceToken(BB);
+ if (token.has_value())
+return token.value();
+
+ // Adding a convergence token requires the function to be marked as
+ // convergent.
+ F->setConvergent();
+
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+ Builder.SetInsertPoint(>front());
+ auto I = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_entry, {}, {});
+ assert(isa(I));
+ Builder.restoreIP(IP);
+
+ return dyn_cast(I);
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::getOrEmitConvergenceLoopToken(const LoopInfo *LI) {
+ assert(LI != nullptr);
+
+ auto token = getConvergenceToken(LI->getHeader());
+ if (token.has_value())
+return *token;
+
+ llvm::IntrinsicInst *PII =
+ LI->getParent()
+ ? EmitConvergenceLoop(LI->getHeader(),
+getOrEmitConvergenceLoopToken(LI->getParent()))
+ : getOrEmitConvergenceEntryToken(LI->getHeader()->getParent());
+
+ return EmitConvergenceLoop(LI->getHeader(), PII);
+}
+
+llvm::CallBase *
+CodeGenFunction::AddControlledConvergenceAttr(llvm::CallBase *Input) {
+ llvm::Value *ParentToken =
+ LoopStack.hasInfo()
+ ? getOrEmitConvergenceLoopToken(())
+ : getOrEmitConvergenceEntryToken(Input->getFunction());
+ return AddConvergenceControlAttr(Input, ParentToken);
+}
+
BitTest BitTest::decodeBitTestBuiltin(unsigned BuiltinID) {
switch (BuiltinID) {
// Main portable variants.
@@ -5803,6 +5892,19 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl
GD, unsigned BuiltinID,
{NDRange, Kernel, Block}));
}
+ case Builtin::BI__builtin_hlsl_wave_active_count_bits: {
+llvm::Type *BoolTy =
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
Keenuts wrote: Thanks all the the reviews! We have 3 LGTMs and an ack from Arsenm, so I'm going to rebase on main, wait for the bots & tests, and if all is green, merge this. https://github.com/llvm/llvm-project/pull/80680 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
@@ -1130,8 +1130,92 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
+
+// Returns the first convergence entry/loop/anchor instruction found in |BB|.
+// std::nullptr otherwise.
+llvm::IntrinsicInst *getConvergenceToken(llvm::BasicBlock *BB) {
+ for (auto : *BB) {
+auto *II = dyn_cast();
+if (II && isConvergenceControlIntrinsic(II->getIntrinsicID()))
+ return II;
+ }
+ return nullptr;
+}
+
} // namespace
+llvm::CallBase *
+CodeGenFunction::AddConvergenceControlAttr(llvm::CallBase *Input,
Keenuts wrote:
Thanks, done, Added `Token` to function name end, changed `Attr` to `Token`,
and fixed case the first letters.
https://github.com/llvm/llvm-project/pull/80680
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
https://github.com/Keenuts updated
https://github.com/llvm/llvm-project/pull/80680
From afbe709931942b3970f92884022e250c1e7eb84f Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Nathan=20Gau=C3=ABr?=
Date: Fri, 2 Feb 2024 16:38:46 +0100
Subject: [PATCH 1/9] [clang][HLSL][SPRI-V] Add convergence intrinsics
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
HLSL has wave operations and other kind of function which required the
control flow to either be converged, or respect certain constraints as
where and how to re-converge.
At the HLSL level, the convergence are mostly obvious: the control flow
is expected to re-converge at the end of a scope.
Once translated to IR, HLSL scopes disapear. This means we need a way to
communicate convergence restrictions down to the backend.
For this, the SPIR-V backend uses convergence intrinsics. So this commit
adds some code to generate convergence intrinsics when required.
This commit is not to be submitted as-is (lacks testing), but
should serve as a basis for an upcoming RFC.
Signed-off-by: Nathan Gauër
---
clang/lib/CodeGen/CGBuiltin.cpp | 102 +++
clang/lib/CodeGen/CGCall.cpp | 4 ++
clang/lib/CodeGen/CGLoopInfo.h | 8 ++-
clang/lib/CodeGen/CodeGenFunction.h | 19 +
llvm/include/llvm/IR/IntrinsicInst.h | 13
5 files changed, 145 insertions(+), 1 deletion(-)
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index 2eaceeba617700..9cc630cd05785a 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -1130,8 +1130,97 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
+
+// Returns the first convergence entry/loop/anchor instruction found in |BB|.
+// std::nullopt otherwise.
+std::optional getConvergenceToken(llvm::BasicBlock *BB)
{
+ for (auto : *BB) {
+auto *II = dyn_cast();
+if (II && isConvergenceControlIntrinsic(II->getIntrinsicID()))
+ return II;
+ }
+ return std::nullopt;
+}
+
} // namespace
+llvm::CallBase *
+CodeGenFunction::AddConvergenceControlAttr(llvm::CallBase *Input,
+ llvm::Value *ParentToken) {
+ llvm::Value *bundleArgs[] = {ParentToken};
+ llvm::OperandBundleDef OB("convergencectrl", bundleArgs);
+ auto Output = llvm::CallBase::addOperandBundle(
+ Input, llvm::LLVMContext::OB_convergencectrl, OB, Input);
+ Input->replaceAllUsesWith(Output);
+ Input->eraseFromParent();
+ return Output;
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::EmitConvergenceLoop(llvm::BasicBlock *BB,
+ llvm::Value *ParentToken) {
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+ Builder.SetInsertPoint(>front());
+ auto CB = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_loop, {}, {});
+ Builder.restoreIP(IP);
+
+ auto I = AddConvergenceControlAttr(CB, ParentToken);
+ // Controlled convergence is incompatible with uncontrolled convergence.
+ // Removing any old attributes.
+ I->setNotConvergent();
+
+ assert(isa(I));
+ return dyn_cast(I);
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::getOrEmitConvergenceEntryToken(llvm::Function *F) {
+ auto *BB = >getEntryBlock();
+ auto token = getConvergenceToken(BB);
+ if (token.has_value())
+return token.value();
+
+ // Adding a convergence token requires the function to be marked as
+ // convergent.
+ F->setConvergent();
+
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+ Builder.SetInsertPoint(>front());
+ auto I = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_entry, {}, {});
+ assert(isa(I));
+ Builder.restoreIP(IP);
+
+ return dyn_cast(I);
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::getOrEmitConvergenceLoopToken(const LoopInfo *LI) {
+ assert(LI != nullptr);
+
+ auto token = getConvergenceToken(LI->getHeader());
+ if (token.has_value())
+return *token;
+
+ llvm::IntrinsicInst *PII =
+ LI->getParent()
+ ? EmitConvergenceLoop(LI->getHeader(),
+getOrEmitConvergenceLoopToken(LI->getParent()))
+ : getOrEmitConvergenceEntryToken(LI->getHeader()->getParent());
+
+ return EmitConvergenceLoop(LI->getHeader(), PII);
+}
+
+llvm::CallBase *
+CodeGenFunction::AddControlledConvergenceAttr(llvm::CallBase *Input) {
+ llvm::Value *ParentToken =
+ LoopStack.hasInfo()
+ ? getOrEmitConvergenceLoopToken(())
+ : getOrEmitConvergenceEntryToken(Input->getFunction());
+ return AddConvergenceControlAttr(Input, ParentToken);
+}
+
BitTest BitTest::decodeBitTestBuiltin(unsigned BuiltinID) {
switch (BuiltinID) {
// Main portable variants.
@@ -5801,6 +5890,19 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl
GD, unsigned BuiltinID,
{NDRange, Kernel, Block}));
}
+ case Builtin::BI__builtin_hlsl_wave_active_count_bits: {
+llvm::Type *BoolTy =
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
https://github.com/Keenuts edited https://github.com/llvm/llvm-project/pull/80680 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
Keenuts wrote: @ssahasra it is up to you then https://github.com/llvm/llvm-project/pull/80680 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
Keenuts wrote: @arsenm would you be fine with those codegen changes as-is? Given that the convergent/no-convergent switch will be done later, depending on when the required IR change is merged? https://github.com/llvm/llvm-project/pull/80680 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
Keenuts wrote: Rebases on main (almost, HEAD is slightly broken), and added back the convergence attribute. The backend changes are ready for this intrinsic. https://github.com/llvm/llvm-project/pull/80680 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
https://github.com/Keenuts updated
https://github.com/llvm/llvm-project/pull/80680
From afbe709931942b3970f92884022e250c1e7eb84f Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Nathan=20Gau=C3=ABr?=
Date: Fri, 2 Feb 2024 16:38:46 +0100
Subject: [PATCH 1/8] [clang][HLSL][SPRI-V] Add convergence intrinsics
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
HLSL has wave operations and other kind of function which required the
control flow to either be converged, or respect certain constraints as
where and how to re-converge.
At the HLSL level, the convergence are mostly obvious: the control flow
is expected to re-converge at the end of a scope.
Once translated to IR, HLSL scopes disapear. This means we need a way to
communicate convergence restrictions down to the backend.
For this, the SPIR-V backend uses convergence intrinsics. So this commit
adds some code to generate convergence intrinsics when required.
This commit is not to be submitted as-is (lacks testing), but
should serve as a basis for an upcoming RFC.
Signed-off-by: Nathan Gauër
---
clang/lib/CodeGen/CGBuiltin.cpp | 102 +++
clang/lib/CodeGen/CGCall.cpp | 4 ++
clang/lib/CodeGen/CGLoopInfo.h | 8 ++-
clang/lib/CodeGen/CodeGenFunction.h | 19 +
llvm/include/llvm/IR/IntrinsicInst.h | 13
5 files changed, 145 insertions(+), 1 deletion(-)
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index 2eaceeba617700..9cc630cd05785a 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -1130,8 +1130,97 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
+
+// Returns the first convergence entry/loop/anchor instruction found in |BB|.
+// std::nullopt otherwise.
+std::optional getConvergenceToken(llvm::BasicBlock *BB)
{
+ for (auto : *BB) {
+auto *II = dyn_cast();
+if (II && isConvergenceControlIntrinsic(II->getIntrinsicID()))
+ return II;
+ }
+ return std::nullopt;
+}
+
} // namespace
+llvm::CallBase *
+CodeGenFunction::AddConvergenceControlAttr(llvm::CallBase *Input,
+ llvm::Value *ParentToken) {
+ llvm::Value *bundleArgs[] = {ParentToken};
+ llvm::OperandBundleDef OB("convergencectrl", bundleArgs);
+ auto Output = llvm::CallBase::addOperandBundle(
+ Input, llvm::LLVMContext::OB_convergencectrl, OB, Input);
+ Input->replaceAllUsesWith(Output);
+ Input->eraseFromParent();
+ return Output;
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::EmitConvergenceLoop(llvm::BasicBlock *BB,
+ llvm::Value *ParentToken) {
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+ Builder.SetInsertPoint(>front());
+ auto CB = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_loop, {}, {});
+ Builder.restoreIP(IP);
+
+ auto I = AddConvergenceControlAttr(CB, ParentToken);
+ // Controlled convergence is incompatible with uncontrolled convergence.
+ // Removing any old attributes.
+ I->setNotConvergent();
+
+ assert(isa(I));
+ return dyn_cast(I);
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::getOrEmitConvergenceEntryToken(llvm::Function *F) {
+ auto *BB = >getEntryBlock();
+ auto token = getConvergenceToken(BB);
+ if (token.has_value())
+return token.value();
+
+ // Adding a convergence token requires the function to be marked as
+ // convergent.
+ F->setConvergent();
+
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+ Builder.SetInsertPoint(>front());
+ auto I = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_entry, {}, {});
+ assert(isa(I));
+ Builder.restoreIP(IP);
+
+ return dyn_cast(I);
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::getOrEmitConvergenceLoopToken(const LoopInfo *LI) {
+ assert(LI != nullptr);
+
+ auto token = getConvergenceToken(LI->getHeader());
+ if (token.has_value())
+return *token;
+
+ llvm::IntrinsicInst *PII =
+ LI->getParent()
+ ? EmitConvergenceLoop(LI->getHeader(),
+getOrEmitConvergenceLoopToken(LI->getParent()))
+ : getOrEmitConvergenceEntryToken(LI->getHeader()->getParent());
+
+ return EmitConvergenceLoop(LI->getHeader(), PII);
+}
+
+llvm::CallBase *
+CodeGenFunction::AddControlledConvergenceAttr(llvm::CallBase *Input) {
+ llvm::Value *ParentToken =
+ LoopStack.hasInfo()
+ ? getOrEmitConvergenceLoopToken(())
+ : getOrEmitConvergenceEntryToken(Input->getFunction());
+ return AddConvergenceControlAttr(Input, ParentToken);
+}
+
BitTest BitTest::decodeBitTestBuiltin(unsigned BuiltinID) {
switch (BuiltinID) {
// Main portable variants.
@@ -5801,6 +5890,19 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl
GD, unsigned BuiltinID,
{NDRange, Kernel, Block}));
}
+ case Builtin::BI__builtin_hlsl_wave_active_count_bits: {
+llvm::Type *BoolTy =
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
@@ -1295,11 +1295,13 @@ double4 trunc(double4); /// true, across all active lanes in the current wave. _HLSL_AVAILABILITY(shadermodel, 6.0) _HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_count_bits) +__attribute__((convergent)) Keenuts wrote: Right, so in that case, I'll add the convergent attribute again, and later down the road, once we have the noconvergent-default in place, we'll be able to flip this back. Thanks all for the context and explanations! https://github.com/llvm/llvm-project/pull/80680 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
https://github.com/Keenuts updated
https://github.com/llvm/llvm-project/pull/80680
From 818ccfd0258602fdd0630823bb2b8af0507749d5 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Nathan=20Gau=C3=ABr?=
Date: Fri, 2 Feb 2024 16:38:46 +0100
Subject: [PATCH 1/7] [clang][HLSL][SPRI-V] Add convergence intrinsics
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
HLSL has wave operations and other kind of function which required the
control flow to either be converged, or respect certain constraints as
where and how to re-converge.
At the HLSL level, the convergence are mostly obvious: the control flow
is expected to re-converge at the end of a scope.
Once translated to IR, HLSL scopes disapear. This means we need a way to
communicate convergence restrictions down to the backend.
For this, the SPIR-V backend uses convergence intrinsics. So this commit
adds some code to generate convergence intrinsics when required.
This commit is not to be submitted as-is (lacks testing), but
should serve as a basis for an upcoming RFC.
Signed-off-by: Nathan Gauër
---
clang/lib/CodeGen/CGBuiltin.cpp | 102 +++
clang/lib/CodeGen/CGCall.cpp | 4 ++
clang/lib/CodeGen/CGLoopInfo.h | 8 ++-
clang/lib/CodeGen/CodeGenFunction.h | 19 +
llvm/include/llvm/IR/IntrinsicInst.h | 13
5 files changed, 145 insertions(+), 1 deletion(-)
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index 20c35757939152..ba5e27a5d4668c 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -1130,8 +1130,97 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
+
+// Returns the first convergence entry/loop/anchor instruction found in |BB|.
+// std::nullopt otherwise.
+std::optional getConvergenceToken(llvm::BasicBlock *BB)
{
+ for (auto : *BB) {
+auto *II = dyn_cast();
+if (II && isConvergenceControlIntrinsic(II->getIntrinsicID()))
+ return II;
+ }
+ return std::nullopt;
+}
+
} // namespace
+llvm::CallBase *
+CodeGenFunction::AddConvergenceControlAttr(llvm::CallBase *Input,
+ llvm::Value *ParentToken) {
+ llvm::Value *bundleArgs[] = {ParentToken};
+ llvm::OperandBundleDef OB("convergencectrl", bundleArgs);
+ auto Output = llvm::CallBase::addOperandBundle(
+ Input, llvm::LLVMContext::OB_convergencectrl, OB, Input);
+ Input->replaceAllUsesWith(Output);
+ Input->eraseFromParent();
+ return Output;
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::EmitConvergenceLoop(llvm::BasicBlock *BB,
+ llvm::Value *ParentToken) {
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+ Builder.SetInsertPoint(>front());
+ auto CB = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_loop, {}, {});
+ Builder.restoreIP(IP);
+
+ auto I = AddConvergenceControlAttr(CB, ParentToken);
+ // Controlled convergence is incompatible with uncontrolled convergence.
+ // Removing any old attributes.
+ I->setNotConvergent();
+
+ assert(isa(I));
+ return dyn_cast(I);
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::getOrEmitConvergenceEntryToken(llvm::Function *F) {
+ auto *BB = >getEntryBlock();
+ auto token = getConvergenceToken(BB);
+ if (token.has_value())
+return token.value();
+
+ // Adding a convergence token requires the function to be marked as
+ // convergent.
+ F->setConvergent();
+
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+ Builder.SetInsertPoint(>front());
+ auto I = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_entry, {}, {});
+ assert(isa(I));
+ Builder.restoreIP(IP);
+
+ return dyn_cast(I);
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::getOrEmitConvergenceLoopToken(const LoopInfo *LI) {
+ assert(LI != nullptr);
+
+ auto token = getConvergenceToken(LI->getHeader());
+ if (token.has_value())
+return *token;
+
+ llvm::IntrinsicInst *PII =
+ LI->getParent()
+ ? EmitConvergenceLoop(LI->getHeader(),
+getOrEmitConvergenceLoopToken(LI->getParent()))
+ : getOrEmitConvergenceEntryToken(LI->getHeader()->getParent());
+
+ return EmitConvergenceLoop(LI->getHeader(), PII);
+}
+
+llvm::CallBase *
+CodeGenFunction::AddControlledConvergenceAttr(llvm::CallBase *Input) {
+ llvm::Value *ParentToken =
+ LoopStack.hasInfo()
+ ? getOrEmitConvergenceLoopToken(())
+ : getOrEmitConvergenceEntryToken(Input->getFunction());
+ return AddConvergenceControlAttr(Input, ParentToken);
+}
+
BitTest BitTest::decodeBitTestBuiltin(unsigned BuiltinID) {
switch (BuiltinID) {
// Main portable variants.
@@ -5698,6 +5787,19 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl
GD, unsigned BuiltinID,
{NDRange, Kernel, Block}));
}
+ case Builtin::BI__builtin_hlsl_wave_active_count_bits: {
+llvm::Type *BoolTy =
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
@@ -1295,11 +1295,13 @@ double4 trunc(double4); /// true, across all active lanes in the current wave. _HLSL_AVAILABILITY(shadermodel, 6.0) _HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_count_bits) +__attribute__((convergent)) Keenuts wrote: @llvm-beanz FYI. Seems like I'll remove it then. https://github.com/llvm/llvm-project/pull/80680 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
https://github.com/Keenuts updated
https://github.com/llvm/llvm-project/pull/80680
From 818ccfd0258602fdd0630823bb2b8af0507749d5 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Nathan=20Gau=C3=ABr?=
Date: Fri, 2 Feb 2024 16:38:46 +0100
Subject: [PATCH 1/6] [clang][HLSL][SPRI-V] Add convergence intrinsics
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
HLSL has wave operations and other kind of function which required the
control flow to either be converged, or respect certain constraints as
where and how to re-converge.
At the HLSL level, the convergence are mostly obvious: the control flow
is expected to re-converge at the end of a scope.
Once translated to IR, HLSL scopes disapear. This means we need a way to
communicate convergence restrictions down to the backend.
For this, the SPIR-V backend uses convergence intrinsics. So this commit
adds some code to generate convergence intrinsics when required.
This commit is not to be submitted as-is (lacks testing), but
should serve as a basis for an upcoming RFC.
Signed-off-by: Nathan Gauër
---
clang/lib/CodeGen/CGBuiltin.cpp | 102 +++
clang/lib/CodeGen/CGCall.cpp | 4 ++
clang/lib/CodeGen/CGLoopInfo.h | 8 ++-
clang/lib/CodeGen/CodeGenFunction.h | 19 +
llvm/include/llvm/IR/IntrinsicInst.h | 13
5 files changed, 145 insertions(+), 1 deletion(-)
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index 20c35757939152..ba5e27a5d4668c 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -1130,8 +1130,97 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
+
+// Returns the first convergence entry/loop/anchor instruction found in |BB|.
+// std::nullopt otherwise.
+std::optional getConvergenceToken(llvm::BasicBlock *BB)
{
+ for (auto : *BB) {
+auto *II = dyn_cast();
+if (II && isConvergenceControlIntrinsic(II->getIntrinsicID()))
+ return II;
+ }
+ return std::nullopt;
+}
+
} // namespace
+llvm::CallBase *
+CodeGenFunction::AddConvergenceControlAttr(llvm::CallBase *Input,
+ llvm::Value *ParentToken) {
+ llvm::Value *bundleArgs[] = {ParentToken};
+ llvm::OperandBundleDef OB("convergencectrl", bundleArgs);
+ auto Output = llvm::CallBase::addOperandBundle(
+ Input, llvm::LLVMContext::OB_convergencectrl, OB, Input);
+ Input->replaceAllUsesWith(Output);
+ Input->eraseFromParent();
+ return Output;
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::EmitConvergenceLoop(llvm::BasicBlock *BB,
+ llvm::Value *ParentToken) {
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+ Builder.SetInsertPoint(>front());
+ auto CB = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_loop, {}, {});
+ Builder.restoreIP(IP);
+
+ auto I = AddConvergenceControlAttr(CB, ParentToken);
+ // Controlled convergence is incompatible with uncontrolled convergence.
+ // Removing any old attributes.
+ I->setNotConvergent();
+
+ assert(isa(I));
+ return dyn_cast(I);
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::getOrEmitConvergenceEntryToken(llvm::Function *F) {
+ auto *BB = >getEntryBlock();
+ auto token = getConvergenceToken(BB);
+ if (token.has_value())
+return token.value();
+
+ // Adding a convergence token requires the function to be marked as
+ // convergent.
+ F->setConvergent();
+
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+ Builder.SetInsertPoint(>front());
+ auto I = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_entry, {}, {});
+ assert(isa(I));
+ Builder.restoreIP(IP);
+
+ return dyn_cast(I);
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::getOrEmitConvergenceLoopToken(const LoopInfo *LI) {
+ assert(LI != nullptr);
+
+ auto token = getConvergenceToken(LI->getHeader());
+ if (token.has_value())
+return *token;
+
+ llvm::IntrinsicInst *PII =
+ LI->getParent()
+ ? EmitConvergenceLoop(LI->getHeader(),
+getOrEmitConvergenceLoopToken(LI->getParent()))
+ : getOrEmitConvergenceEntryToken(LI->getHeader()->getParent());
+
+ return EmitConvergenceLoop(LI->getHeader(), PII);
+}
+
+llvm::CallBase *
+CodeGenFunction::AddControlledConvergenceAttr(llvm::CallBase *Input) {
+ llvm::Value *ParentToken =
+ LoopStack.hasInfo()
+ ? getOrEmitConvergenceLoopToken(())
+ : getOrEmitConvergenceEntryToken(Input->getFunction());
+ return AddConvergenceControlAttr(Input, ParentToken);
+}
+
BitTest BitTest::decodeBitTestBuiltin(unsigned BuiltinID) {
switch (BuiltinID) {
// Main portable variants.
@@ -5698,6 +5787,19 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl
GD, unsigned BuiltinID,
{NDRange, Kernel, Block}));
}
+ case Builtin::BI__builtin_hlsl_wave_active_count_bits: {
+llvm::Type *BoolTy =
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
@@ -1130,8 +1130,96 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
+
+// Returns the first convergence entry/loop/anchor instruction found in |BB|.
+// std::nullopt otherwise.
+std::optional getConvergenceToken(llvm::BasicBlock *BB)
{
+ for (auto : *BB) {
Keenuts wrote:
Good to know! I'd be in favor to implement this map once we have a merged use
case.
-> This way, we can have an independent NFC PR which focuses on this
optimization/refactoring by adding either an analysis or the bits in
`llvm::BasicBlock` to quickly fetch those.
https://github.com/llvm/llvm-project/pull/80680
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
@@ -1297,5 +1297,10 @@ _HLSL_AVAILABILITY(shadermodel, 6.0) _HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_count_bits) uint WaveActiveCountBits(bool Val); +/// \brief Returns the index of the current lane within the current wave. +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_get_lane_index) +uint WaveGetLaneIndex(); Keenuts wrote: Yes, good catch, done https://github.com/llvm/llvm-project/pull/80680 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
https://github.com/Keenuts updated
https://github.com/llvm/llvm-project/pull/80680
From 818ccfd0258602fdd0630823bb2b8af0507749d5 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Nathan=20Gau=C3=ABr?=
Date: Fri, 2 Feb 2024 16:38:46 +0100
Subject: [PATCH 1/5] [clang][HLSL][SPRI-V] Add convergence intrinsics
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
HLSL has wave operations and other kind of function which required the
control flow to either be converged, or respect certain constraints as
where and how to re-converge.
At the HLSL level, the convergence are mostly obvious: the control flow
is expected to re-converge at the end of a scope.
Once translated to IR, HLSL scopes disapear. This means we need a way to
communicate convergence restrictions down to the backend.
For this, the SPIR-V backend uses convergence intrinsics. So this commit
adds some code to generate convergence intrinsics when required.
This commit is not to be submitted as-is (lacks testing), but
should serve as a basis for an upcoming RFC.
Signed-off-by: Nathan Gauër
---
clang/lib/CodeGen/CGBuiltin.cpp | 102 +++
clang/lib/CodeGen/CGCall.cpp | 4 ++
clang/lib/CodeGen/CGLoopInfo.h | 8 ++-
clang/lib/CodeGen/CodeGenFunction.h | 19 +
llvm/include/llvm/IR/IntrinsicInst.h | 13
5 files changed, 145 insertions(+), 1 deletion(-)
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index 20c35757939152..ba5e27a5d4668c 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -1130,8 +1130,97 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
+
+// Returns the first convergence entry/loop/anchor instruction found in |BB|.
+// std::nullopt otherwise.
+std::optional getConvergenceToken(llvm::BasicBlock *BB)
{
+ for (auto : *BB) {
+auto *II = dyn_cast();
+if (II && isConvergenceControlIntrinsic(II->getIntrinsicID()))
+ return II;
+ }
+ return std::nullopt;
+}
+
} // namespace
+llvm::CallBase *
+CodeGenFunction::AddConvergenceControlAttr(llvm::CallBase *Input,
+ llvm::Value *ParentToken) {
+ llvm::Value *bundleArgs[] = {ParentToken};
+ llvm::OperandBundleDef OB("convergencectrl", bundleArgs);
+ auto Output = llvm::CallBase::addOperandBundle(
+ Input, llvm::LLVMContext::OB_convergencectrl, OB, Input);
+ Input->replaceAllUsesWith(Output);
+ Input->eraseFromParent();
+ return Output;
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::EmitConvergenceLoop(llvm::BasicBlock *BB,
+ llvm::Value *ParentToken) {
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+ Builder.SetInsertPoint(>front());
+ auto CB = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_loop, {}, {});
+ Builder.restoreIP(IP);
+
+ auto I = AddConvergenceControlAttr(CB, ParentToken);
+ // Controlled convergence is incompatible with uncontrolled convergence.
+ // Removing any old attributes.
+ I->setNotConvergent();
+
+ assert(isa(I));
+ return dyn_cast(I);
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::getOrEmitConvergenceEntryToken(llvm::Function *F) {
+ auto *BB = >getEntryBlock();
+ auto token = getConvergenceToken(BB);
+ if (token.has_value())
+return token.value();
+
+ // Adding a convergence token requires the function to be marked as
+ // convergent.
+ F->setConvergent();
+
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+ Builder.SetInsertPoint(>front());
+ auto I = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_entry, {}, {});
+ assert(isa(I));
+ Builder.restoreIP(IP);
+
+ return dyn_cast(I);
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::getOrEmitConvergenceLoopToken(const LoopInfo *LI) {
+ assert(LI != nullptr);
+
+ auto token = getConvergenceToken(LI->getHeader());
+ if (token.has_value())
+return *token;
+
+ llvm::IntrinsicInst *PII =
+ LI->getParent()
+ ? EmitConvergenceLoop(LI->getHeader(),
+getOrEmitConvergenceLoopToken(LI->getParent()))
+ : getOrEmitConvergenceEntryToken(LI->getHeader()->getParent());
+
+ return EmitConvergenceLoop(LI->getHeader(), PII);
+}
+
+llvm::CallBase *
+CodeGenFunction::AddControlledConvergenceAttr(llvm::CallBase *Input) {
+ llvm::Value *ParentToken =
+ LoopStack.hasInfo()
+ ? getOrEmitConvergenceLoopToken(())
+ : getOrEmitConvergenceEntryToken(Input->getFunction());
+ return AddConvergenceControlAttr(Input, ParentToken);
+}
+
BitTest BitTest::decodeBitTestBuiltin(unsigned BuiltinID) {
switch (BuiltinID) {
// Main portable variants.
@@ -5698,6 +5787,19 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl
GD, unsigned BuiltinID,
{NDRange, Kernel, Block}));
}
+ case Builtin::BI__builtin_hlsl_wave_active_count_bits: {
+llvm::Type *BoolTy =
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
@@ -4554,6 +4554,13 @@ def HLSLWaveActiveCountBits : LangBuiltin<"HLSL_LANG"> {
let Prototype = "unsigned int(bool)";
}
+// HLSL
+def HLSLWaveGetLaneIndex : LangBuiltin<"HLSL_LANG"> {
+ let Spellings = ["__builtin_hlsl_wave_get_lane_index"];
+ let Attributes = [NoThrow, Const];
Keenuts wrote:
Yes, but this implies touching to the mangling pattern..
https://github.com/llvm/llvm-project/pull/80680
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
https://github.com/Keenuts updated
https://github.com/llvm/llvm-project/pull/80680
From 818ccfd0258602fdd0630823bb2b8af0507749d5 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Nathan=20Gau=C3=ABr?=
Date: Fri, 2 Feb 2024 16:38:46 +0100
Subject: [PATCH 1/4] [clang][HLSL][SPRI-V] Add convergence intrinsics
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
HLSL has wave operations and other kind of function which required the
control flow to either be converged, or respect certain constraints as
where and how to re-converge.
At the HLSL level, the convergence are mostly obvious: the control flow
is expected to re-converge at the end of a scope.
Once translated to IR, HLSL scopes disapear. This means we need a way to
communicate convergence restrictions down to the backend.
For this, the SPIR-V backend uses convergence intrinsics. So this commit
adds some code to generate convergence intrinsics when required.
This commit is not to be submitted as-is (lacks testing), but
should serve as a basis for an upcoming RFC.
Signed-off-by: Nathan Gauër
---
clang/lib/CodeGen/CGBuiltin.cpp | 102 +++
clang/lib/CodeGen/CGCall.cpp | 4 ++
clang/lib/CodeGen/CGLoopInfo.h | 8 ++-
clang/lib/CodeGen/CodeGenFunction.h | 19 +
llvm/include/llvm/IR/IntrinsicInst.h | 13
5 files changed, 145 insertions(+), 1 deletion(-)
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index 20c35757939152..ba5e27a5d4668c 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -1130,8 +1130,97 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
+
+// Returns the first convergence entry/loop/anchor instruction found in |BB|.
+// std::nullopt otherwise.
+std::optional getConvergenceToken(llvm::BasicBlock *BB)
{
+ for (auto : *BB) {
+auto *II = dyn_cast();
+if (II && isConvergenceControlIntrinsic(II->getIntrinsicID()))
+ return II;
+ }
+ return std::nullopt;
+}
+
} // namespace
+llvm::CallBase *
+CodeGenFunction::AddConvergenceControlAttr(llvm::CallBase *Input,
+ llvm::Value *ParentToken) {
+ llvm::Value *bundleArgs[] = {ParentToken};
+ llvm::OperandBundleDef OB("convergencectrl", bundleArgs);
+ auto Output = llvm::CallBase::addOperandBundle(
+ Input, llvm::LLVMContext::OB_convergencectrl, OB, Input);
+ Input->replaceAllUsesWith(Output);
+ Input->eraseFromParent();
+ return Output;
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::EmitConvergenceLoop(llvm::BasicBlock *BB,
+ llvm::Value *ParentToken) {
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+ Builder.SetInsertPoint(>front());
+ auto CB = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_loop, {}, {});
+ Builder.restoreIP(IP);
+
+ auto I = AddConvergenceControlAttr(CB, ParentToken);
+ // Controlled convergence is incompatible with uncontrolled convergence.
+ // Removing any old attributes.
+ I->setNotConvergent();
+
+ assert(isa(I));
+ return dyn_cast(I);
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::getOrEmitConvergenceEntryToken(llvm::Function *F) {
+ auto *BB = >getEntryBlock();
+ auto token = getConvergenceToken(BB);
+ if (token.has_value())
+return token.value();
+
+ // Adding a convergence token requires the function to be marked as
+ // convergent.
+ F->setConvergent();
+
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+ Builder.SetInsertPoint(>front());
+ auto I = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_entry, {}, {});
+ assert(isa(I));
+ Builder.restoreIP(IP);
+
+ return dyn_cast(I);
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::getOrEmitConvergenceLoopToken(const LoopInfo *LI) {
+ assert(LI != nullptr);
+
+ auto token = getConvergenceToken(LI->getHeader());
+ if (token.has_value())
+return *token;
+
+ llvm::IntrinsicInst *PII =
+ LI->getParent()
+ ? EmitConvergenceLoop(LI->getHeader(),
+getOrEmitConvergenceLoopToken(LI->getParent()))
+ : getOrEmitConvergenceEntryToken(LI->getHeader()->getParent());
+
+ return EmitConvergenceLoop(LI->getHeader(), PII);
+}
+
+llvm::CallBase *
+CodeGenFunction::AddControlledConvergenceAttr(llvm::CallBase *Input) {
+ llvm::Value *ParentToken =
+ LoopStack.hasInfo()
+ ? getOrEmitConvergenceLoopToken(())
+ : getOrEmitConvergenceEntryToken(Input->getFunction());
+ return AddConvergenceControlAttr(Input, ParentToken);
+}
+
BitTest BitTest::decodeBitTestBuiltin(unsigned BuiltinID) {
switch (BuiltinID) {
// Main portable variants.
@@ -5698,6 +5787,19 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl
GD, unsigned BuiltinID,
{NDRange, Kernel, Block}));
}
+ case Builtin::BI__builtin_hlsl_wave_active_count_bits: {
+llvm::Type *BoolTy =
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
@@ -5686,6 +5686,10 @@ RValue CodeGenFunction::EmitCall(const CGFunctionInfo
,
if (!CI->getType()->isVoidTy())
CI->setName("call");
+ if (getTarget().getTriple().isSPIRVLogical() &&
+ CI->getCalledFunction()->isConvergent())
Keenuts wrote:
Yes, thanks
https://github.com/llvm/llvm-project/pull/80680
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
@@ -1130,8 +1130,96 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
+
+// Returns the first convergence entry/loop/anchor instruction found in |BB|.
+// std::nullopt otherwise.
+std::optional getConvergenceToken(llvm::BasicBlock *BB)
{
+ for (auto : *BB) {
+auto *II = dyn_cast();
+if (II && isConvergenceControlIntrinsic(II->getIntrinsicID()))
+ return II;
+ }
+ return std::nullopt;
+}
+
} // namespace
+llvm::CallBase *
+CodeGenFunction::AddConvergenceControlAttr(llvm::CallBase *Input,
+ llvm::Value *ParentToken) {
+ llvm::Value *bundleArgs[] = {ParentToken};
+ llvm::OperandBundleDef OB("convergencectrl", bundleArgs);
+ auto Output = llvm::CallBase::addOperandBundle(
+ Input, llvm::LLVMContext::OB_convergencectrl, OB, Input);
+ Input->replaceAllUsesWith(Output);
+ Input->eraseFromParent();
+ return Output;
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::EmitConvergenceLoop(llvm::BasicBlock *BB,
+ llvm::Value *ParentToken) {
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+ Builder.SetInsertPoint(>front());
+ auto CB = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_loop, {}, {});
+ Builder.restoreIP(IP);
+
+ auto I = AddConvergenceControlAttr(CB, ParentToken);
+ // Controlled convergence is incompatible with uncontrolled convergence.
+ // Removing any old attributes.
+ I->setNotConvergent();
Keenuts wrote:
good catch, I have no idea why this is here. I recall something around
validation, but today reading it, it does not make sense. Thanks!
https://github.com/llvm/llvm-project/pull/80680
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
@@ -1130,8 +1130,96 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
+
+// Returns the first convergence entry/loop/anchor instruction found in |BB|.
+// std::nullopt otherwise.
+std::optional getConvergenceToken(llvm::BasicBlock *BB)
{
+ for (auto : *BB) {
+auto *II = dyn_cast();
+if (II && isConvergenceControlIntrinsic(II->getIntrinsicID()))
+ return II;
+ }
+ return std::nullopt;
+}
+
} // namespace
+llvm::CallBase *
+CodeGenFunction::AddConvergenceControlAttr(llvm::CallBase *Input,
+ llvm::Value *ParentToken) {
+ llvm::Value *bundleArgs[] = {ParentToken};
Keenuts wrote:
well, no... tried, and there is an overload which takes either an ArrayRef of a
vector, so the inline array will cause an ambiguity :/
https://github.com/llvm/llvm-project/pull/80680
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
@@ -1130,8 +1130,96 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
+
+// Returns the first convergence entry/loop/anchor instruction found in |BB|.
+// std::nullopt otherwise.
+std::optional getConvergenceToken(llvm::BasicBlock *BB)
{
+ for (auto : *BB) {
+auto *II = dyn_cast();
+if (II && isConvergenceControlIntrinsic(II->getIntrinsicID()))
+ return II;
+ }
+ return std::nullopt;
+}
+
} // namespace
+llvm::CallBase *
+CodeGenFunction::AddConvergenceControlAttr(llvm::CallBase *Input,
+ llvm::Value *ParentToken) {
+ llvm::Value *bundleArgs[] = {ParentToken};
Keenuts wrote:
yep, thanks
https://github.com/llvm/llvm-project/pull/80680
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
@@ -1130,8 +1130,96 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
+
+// Returns the first convergence entry/loop/anchor instruction found in |BB|.
+// std::nullopt otherwise.
+std::optional getConvergenceToken(llvm::BasicBlock *BB)
{
+ for (auto : *BB) {
Keenuts wrote:
Yep, but I'm not sure I have any other option outside from some kind of
caching(either attribute of a BB, or BB->I map). If it's this kind of
alternative you are thinking about, maybe this should be left in a subsequent
PR as an optimization?
https://github.com/llvm/llvm-project/pull/80680
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
@@ -1130,8 +1130,96 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
+
+// Returns the first convergence entry/loop/anchor instruction found in |BB|.
+// std::nullopt otherwise.
+std::optional getConvergenceToken(llvm::BasicBlock *BB)
{
Keenuts wrote:
No strong opinion. My rational is if you read "optional", you know it's not
always there. But I'm fine with the pointer, changed to pointer.
https://github.com/llvm/llvm-project/pull/80680
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
Keenuts wrote: Hi, thanks for the reviews so far! I believe we are ready to move forward on our side. Adding @llvm-beanz for the HLSL part. The builtin I added is mostly to get something we can generate those intrinsics for. I am fine changing the name, or the implementation around it, it's just to have a wave an easy intrinsic to use/test with (no input, simplest one) https://github.com/llvm/llvm-project/pull/80680 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
https://github.com/Keenuts ready_for_review https://github.com/llvm/llvm-project/pull/80680 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
https://github.com/Keenuts updated
https://github.com/llvm/llvm-project/pull/80680
From 818ccfd0258602fdd0630823bb2b8af0507749d5 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Nathan=20Gau=C3=ABr?=
Date: Fri, 2 Feb 2024 16:38:46 +0100
Subject: [PATCH 1/3] [clang][HLSL][SPRI-V] Add convergence intrinsics
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
HLSL has wave operations and other kind of function which required the
control flow to either be converged, or respect certain constraints as
where and how to re-converge.
At the HLSL level, the convergence are mostly obvious: the control flow
is expected to re-converge at the end of a scope.
Once translated to IR, HLSL scopes disapear. This means we need a way to
communicate convergence restrictions down to the backend.
For this, the SPIR-V backend uses convergence intrinsics. So this commit
adds some code to generate convergence intrinsics when required.
This commit is not to be submitted as-is (lacks testing), but
should serve as a basis for an upcoming RFC.
Signed-off-by: Nathan Gauër
---
clang/lib/CodeGen/CGBuiltin.cpp | 102 +++
clang/lib/CodeGen/CGCall.cpp | 4 ++
clang/lib/CodeGen/CGLoopInfo.h | 8 ++-
clang/lib/CodeGen/CodeGenFunction.h | 19 +
llvm/include/llvm/IR/IntrinsicInst.h | 13
5 files changed, 145 insertions(+), 1 deletion(-)
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index 20c35757939152..ba5e27a5d4668c 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -1130,8 +1130,97 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
+
+// Returns the first convergence entry/loop/anchor instruction found in |BB|.
+// std::nullopt otherwise.
+std::optional getConvergenceToken(llvm::BasicBlock *BB)
{
+ for (auto : *BB) {
+auto *II = dyn_cast();
+if (II && isConvergenceControlIntrinsic(II->getIntrinsicID()))
+ return II;
+ }
+ return std::nullopt;
+}
+
} // namespace
+llvm::CallBase *
+CodeGenFunction::AddConvergenceControlAttr(llvm::CallBase *Input,
+ llvm::Value *ParentToken) {
+ llvm::Value *bundleArgs[] = {ParentToken};
+ llvm::OperandBundleDef OB("convergencectrl", bundleArgs);
+ auto Output = llvm::CallBase::addOperandBundle(
+ Input, llvm::LLVMContext::OB_convergencectrl, OB, Input);
+ Input->replaceAllUsesWith(Output);
+ Input->eraseFromParent();
+ return Output;
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::EmitConvergenceLoop(llvm::BasicBlock *BB,
+ llvm::Value *ParentToken) {
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+ Builder.SetInsertPoint(>front());
+ auto CB = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_loop, {}, {});
+ Builder.restoreIP(IP);
+
+ auto I = AddConvergenceControlAttr(CB, ParentToken);
+ // Controlled convergence is incompatible with uncontrolled convergence.
+ // Removing any old attributes.
+ I->setNotConvergent();
+
+ assert(isa(I));
+ return dyn_cast(I);
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::getOrEmitConvergenceEntryToken(llvm::Function *F) {
+ auto *BB = >getEntryBlock();
+ auto token = getConvergenceToken(BB);
+ if (token.has_value())
+return token.value();
+
+ // Adding a convergence token requires the function to be marked as
+ // convergent.
+ F->setConvergent();
+
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+ Builder.SetInsertPoint(>front());
+ auto I = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_entry, {}, {});
+ assert(isa(I));
+ Builder.restoreIP(IP);
+
+ return dyn_cast(I);
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::getOrEmitConvergenceLoopToken(const LoopInfo *LI) {
+ assert(LI != nullptr);
+
+ auto token = getConvergenceToken(LI->getHeader());
+ if (token.has_value())
+return *token;
+
+ llvm::IntrinsicInst *PII =
+ LI->getParent()
+ ? EmitConvergenceLoop(LI->getHeader(),
+getOrEmitConvergenceLoopToken(LI->getParent()))
+ : getOrEmitConvergenceEntryToken(LI->getHeader()->getParent());
+
+ return EmitConvergenceLoop(LI->getHeader(), PII);
+}
+
+llvm::CallBase *
+CodeGenFunction::AddControlledConvergenceAttr(llvm::CallBase *Input) {
+ llvm::Value *ParentToken =
+ LoopStack.hasInfo()
+ ? getOrEmitConvergenceLoopToken(())
+ : getOrEmitConvergenceEntryToken(Input->getFunction());
+ return AddConvergenceControlAttr(Input, ParentToken);
+}
+
BitTest BitTest::decodeBitTestBuiltin(unsigned BuiltinID) {
switch (BuiltinID) {
// Main portable variants.
@@ -5698,6 +5787,19 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl
GD, unsigned BuiltinID,
{NDRange, Kernel, Block}));
}
+ case Builtin::BI__builtin_hlsl_wave_active_count_bits: {
+llvm::Type *BoolTy =
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
https://github.com/Keenuts updated
https://github.com/llvm/llvm-project/pull/80680
From f43134787dc5b1120e3722e03028d0071de22331 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Nathan=20Gau=C3=ABr?=
Date: Fri, 2 Feb 2024 16:38:46 +0100
Subject: [PATCH 1/2] [clang][HLSL][SPRI-V] Add convergence intrinsics
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
HLSL has wave operations and other kind of function which required the
control flow to either be converged, or respect certain constraints as
where and how to re-converge.
At the HLSL level, the convergence are mostly obvious: the control flow
is expected to re-converge at the end of a scope.
Once translated to IR, HLSL scopes disapear. This means we need a way to
communicate convergence restrictions down to the backend.
For this, the SPIR-V backend uses convergence intrinsics. So this commit
adds some code to generate convergence intrinsics when required.
This commit is not to be submitted as-is (lacks testing), but
should serve as a basis for an upcoming RFC.
Signed-off-by: Nathan Gauër
---
clang/lib/CodeGen/CGBuiltin.cpp | 102 +++
clang/lib/CodeGen/CGCall.cpp | 4 ++
clang/lib/CodeGen/CGLoopInfo.h | 8 ++-
clang/lib/CodeGen/CodeGenFunction.h | 19 +
llvm/include/llvm/IR/IntrinsicInst.h | 13
5 files changed, 145 insertions(+), 1 deletion(-)
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index d8b2115f1e5e3c..82e652959238b4 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -1129,8 +1129,97 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
+
+// Returns the first convergence entry/loop/anchor instruction found in |BB|.
+// std::nullopt otherwise.
+std::optional getConvergenceToken(llvm::BasicBlock *BB)
{
+ for (auto : *BB) {
+auto *II = dyn_cast();
+if (II && isConvergenceControlIntrinsic(II->getIntrinsicID()))
+ return II;
+ }
+ return std::nullopt;
+}
+
} // namespace
+llvm::CallBase *
+CodeGenFunction::AddConvergenceControlAttr(llvm::CallBase *Input,
+ llvm::Value *ParentToken) {
+ llvm::Value *bundleArgs[] = {ParentToken};
+ llvm::OperandBundleDef OB("convergencectrl", bundleArgs);
+ auto Output = llvm::CallBase::addOperandBundle(
+ Input, llvm::LLVMContext::OB_convergencectrl, OB, Input);
+ Input->replaceAllUsesWith(Output);
+ Input->eraseFromParent();
+ return Output;
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::EmitConvergenceLoop(llvm::BasicBlock *BB,
+ llvm::Value *ParentToken) {
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+ Builder.SetInsertPoint(>front());
+ auto CB = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_loop, {}, {});
+ Builder.restoreIP(IP);
+
+ auto I = AddConvergenceControlAttr(CB, ParentToken);
+ // Controlled convergence is incompatible with uncontrolled convergence.
+ // Removing any old attributes.
+ I->setNotConvergent();
+
+ assert(isa(I));
+ return dyn_cast(I);
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::getOrEmitConvergenceEntryToken(llvm::Function *F) {
+ auto *BB = >getEntryBlock();
+ auto token = getConvergenceToken(BB);
+ if (token.has_value())
+return token.value();
+
+ // Adding a convergence token requires the function to be marked as
+ // convergent.
+ F->setConvergent();
+
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+ Builder.SetInsertPoint(>front());
+ auto I = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_entry, {}, {});
+ assert(isa(I));
+ Builder.restoreIP(IP);
+
+ return dyn_cast(I);
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::getOrEmitConvergenceLoopToken(const LoopInfo *LI) {
+ assert(LI != nullptr);
+
+ auto token = getConvergenceToken(LI->getHeader());
+ if (token.has_value())
+return *token;
+
+ llvm::IntrinsicInst *PII =
+ LI->getParent()
+ ? EmitConvergenceLoop(LI->getHeader(),
+getOrEmitConvergenceLoopToken(LI->getParent()))
+ : getOrEmitConvergenceEntryToken(LI->getHeader()->getParent());
+
+ return EmitConvergenceLoop(LI->getHeader(), PII);
+}
+
+llvm::CallBase *
+CodeGenFunction::AddControlledConvergenceAttr(llvm::CallBase *Input) {
+ llvm::Value *ParentToken =
+ LoopStack.hasInfo()
+ ? getOrEmitConvergenceLoopToken(())
+ : getOrEmitConvergenceEntryToken(Input->getFunction());
+ return AddConvergenceControlAttr(Input, ParentToken);
+}
+
BitTest BitTest::decodeBitTestBuiltin(unsigned BuiltinID) {
switch (BuiltinID) {
// Main portable variants.
@@ -5696,6 +5785,19 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl
GD, unsigned BuiltinID,
{NDRange, Kernel, Block}));
}
+ case Builtin::BI__builtin_hlsl_wave_active_count_bits: {
+llvm::Type *BoolTy =
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
https://github.com/Keenuts updated
https://github.com/llvm/llvm-project/pull/80680
From f43134787dc5b1120e3722e03028d0071de22331 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Nathan=20Gau=C3=ABr?=
Date: Fri, 2 Feb 2024 16:38:46 +0100
Subject: [PATCH 1/2] [clang][HLSL][SPRI-V] Add convergence intrinsics
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
HLSL has wave operations and other kind of function which required the
control flow to either be converged, or respect certain constraints as
where and how to re-converge.
At the HLSL level, the convergence are mostly obvious: the control flow
is expected to re-converge at the end of a scope.
Once translated to IR, HLSL scopes disapear. This means we need a way to
communicate convergence restrictions down to the backend.
For this, the SPIR-V backend uses convergence intrinsics. So this commit
adds some code to generate convergence intrinsics when required.
This commit is not to be submitted as-is (lacks testing), but
should serve as a basis for an upcoming RFC.
Signed-off-by: Nathan Gauër
---
clang/lib/CodeGen/CGBuiltin.cpp | 102 +++
clang/lib/CodeGen/CGCall.cpp | 4 ++
clang/lib/CodeGen/CGLoopInfo.h | 8 ++-
clang/lib/CodeGen/CodeGenFunction.h | 19 +
llvm/include/llvm/IR/IntrinsicInst.h | 13
5 files changed, 145 insertions(+), 1 deletion(-)
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index d8b2115f1e5e3c..82e652959238b4 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -1129,8 +1129,97 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
+
+// Returns the first convergence entry/loop/anchor instruction found in |BB|.
+// std::nullopt otherwise.
+std::optional getConvergenceToken(llvm::BasicBlock *BB)
{
+ for (auto : *BB) {
+auto *II = dyn_cast();
+if (II && isConvergenceControlIntrinsic(II->getIntrinsicID()))
+ return II;
+ }
+ return std::nullopt;
+}
+
} // namespace
+llvm::CallBase *
+CodeGenFunction::AddConvergenceControlAttr(llvm::CallBase *Input,
+ llvm::Value *ParentToken) {
+ llvm::Value *bundleArgs[] = {ParentToken};
+ llvm::OperandBundleDef OB("convergencectrl", bundleArgs);
+ auto Output = llvm::CallBase::addOperandBundle(
+ Input, llvm::LLVMContext::OB_convergencectrl, OB, Input);
+ Input->replaceAllUsesWith(Output);
+ Input->eraseFromParent();
+ return Output;
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::EmitConvergenceLoop(llvm::BasicBlock *BB,
+ llvm::Value *ParentToken) {
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+ Builder.SetInsertPoint(>front());
+ auto CB = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_loop, {}, {});
+ Builder.restoreIP(IP);
+
+ auto I = AddConvergenceControlAttr(CB, ParentToken);
+ // Controlled convergence is incompatible with uncontrolled convergence.
+ // Removing any old attributes.
+ I->setNotConvergent();
+
+ assert(isa(I));
+ return dyn_cast(I);
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::getOrEmitConvergenceEntryToken(llvm::Function *F) {
+ auto *BB = >getEntryBlock();
+ auto token = getConvergenceToken(BB);
+ if (token.has_value())
+return token.value();
+
+ // Adding a convergence token requires the function to be marked as
+ // convergent.
+ F->setConvergent();
+
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+ Builder.SetInsertPoint(>front());
+ auto I = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_entry, {}, {});
+ assert(isa(I));
+ Builder.restoreIP(IP);
+
+ return dyn_cast(I);
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::getOrEmitConvergenceLoopToken(const LoopInfo *LI) {
+ assert(LI != nullptr);
+
+ auto token = getConvergenceToken(LI->getHeader());
+ if (token.has_value())
+return *token;
+
+ llvm::IntrinsicInst *PII =
+ LI->getParent()
+ ? EmitConvergenceLoop(LI->getHeader(),
+getOrEmitConvergenceLoopToken(LI->getParent()))
+ : getOrEmitConvergenceEntryToken(LI->getHeader()->getParent());
+
+ return EmitConvergenceLoop(LI->getHeader(), PII);
+}
+
+llvm::CallBase *
+CodeGenFunction::AddControlledConvergenceAttr(llvm::CallBase *Input) {
+ llvm::Value *ParentToken =
+ LoopStack.hasInfo()
+ ? getOrEmitConvergenceLoopToken(())
+ : getOrEmitConvergenceEntryToken(Input->getFunction());
+ return AddConvergenceControlAttr(Input, ParentToken);
+}
+
BitTest BitTest::decodeBitTestBuiltin(unsigned BuiltinID) {
switch (BuiltinID) {
// Main portable variants.
@@ -5696,6 +5785,19 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl
GD, unsigned BuiltinID,
{NDRange, Kernel, Block}));
}
+ case Builtin::BI__builtin_hlsl_wave_active_count_bits: {
+llvm::Type *BoolTy =
[clang] [llvm] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
@@ -1129,8 +1129,97 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
+
+// Returns the first convergence entry/loop/anchor instruction found in |BB|.
+// std::nullopt otherwise.
+std::optional getConvergenceToken(llvm::BasicBlock *BB)
{
+ for (auto : *BB) {
+auto *II = dyn_cast();
+if (II && isConvergenceControlIntrinsic(II->getIntrinsicID()))
+ return II;
+ }
+ return std::nullopt;
+}
+
} // namespace
+llvm::CallBase *
+CodeGenFunction::AddConvergenceControlAttr(llvm::CallBase *Input,
+ llvm::Value *ParentToken) {
+ llvm::Value *bundleArgs[] = {ParentToken};
+ llvm::OperandBundleDef OB("convergencectrl", bundleArgs);
+ auto Output = llvm::CallBase::addOperandBundle(
+ Input, llvm::LLVMContext::OB_convergencectrl, OB, Input);
+ Input->replaceAllUsesWith(Output);
+ Input->eraseFromParent();
+ return Output;
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::EmitConvergenceLoop(llvm::BasicBlock *BB,
+ llvm::Value *ParentToken) {
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+ Builder.SetInsertPoint(>front());
+ auto CB = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_loop, {}, {});
+ Builder.restoreIP(IP);
+
+ auto I = AddConvergenceControlAttr(CB, ParentToken);
+ // Controlled convergence is incompatible with uncontrolled convergence.
+ // Removing any old attributes.
+ I->setNotConvergent();
+
+ assert(isa(I));
+ return dyn_cast(I);
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::getOrEmitConvergenceEntryToken(llvm::Function *F) {
+ auto *BB = >getEntryBlock();
+ auto token = getConvergenceToken(BB);
+ if (token.has_value())
+return token.value();
+
+ // Adding a convergence token requires the function to be marked as
+ // convergent.
+ F->setConvergent();
+
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+ Builder.SetInsertPoint(>front());
+ auto I = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_entry, {}, {});
+ assert(isa(I));
Keenuts wrote:
Oh right, thanks!
https://github.com/llvm/llvm-project/pull/80680
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [HLSL][SPIR-V] Add create.handle intrinsic (PR #81038)
https://github.com/Keenuts approved this pull request. https://github.com/llvm/llvm-project/pull/81038 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[llvm] [clang] [clang][HLSL][SPRI-V] Add convergence intrinsics (PR #80680)
https://github.com/Keenuts created
https://github.com/llvm/llvm-project/pull/80680
HLSL has wave operations and other kind of function which required the control
flow to either be converged, or respect certain constraints as where and how to
re-converge.
At the HLSL level, the convergence are mostly obvious: the control flow is
expected to re-converge at the end of a scope.
Once translated to IR, HLSL scopes disapear. This means we need a way to
communicate convergence restrictions down to the backend.
For this, the SPIR-V backend uses convergence intrinsics. So this commit adds
some code to generate convergence intrinsics when required.
This commit is not to be submitted as-is (lacks testing), but should serve as a
basis for an upcoming RFC.
From 8d653d1af6f624f341e88997682fc271195d8a45 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Nathan=20Gau=C3=ABr?=
Date: Fri, 2 Feb 2024 16:38:46 +0100
Subject: [PATCH] [clang][HLSL][SPRI-V] Add convergence intrinsics
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
HLSL has wave operations and other kind of function which required the
control flow to either be converged, or respect certain constraints as
where and how to re-converge.
At the HLSL level, the convergence are mostly obvious: the control flow
is expected to re-converge at the end of a scope.
Once translated to IR, HLSL scopes disapear. This means we need a way to
communicate convergence restrictions down to the backend.
For this, the SPIR-V backend uses convergence intrinsics. So this commit
adds some code to generate convergence intrinsics when required.
This commit is not to be submitted as-is (lacks testing), but
should serve as a basis for an upcoming RFC.
Signed-off-by: Nathan Gauër
---
clang/lib/CodeGen/CGBuiltin.cpp | 102 +++
clang/lib/CodeGen/CGCall.cpp | 4 ++
clang/lib/CodeGen/CGLoopInfo.h | 8 ++-
clang/lib/CodeGen/CodeGenFunction.h | 19 +
llvm/include/llvm/IR/IntrinsicInst.h | 13
5 files changed, 145 insertions(+), 1 deletion(-)
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index f17e4a83305bf..0de350dc65485 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -1129,8 +1129,97 @@ struct BitTest {
static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
};
+
+// Returns the first convergence entry/loop/anchor instruction found in |BB|.
+// std::nullopt otherwise.
+std::optional getConvergenceToken(llvm::BasicBlock *BB)
{
+ for (auto : *BB) {
+auto *II = dyn_cast();
+if (II && isConvergenceControlIntrinsic(II->getIntrinsicID()))
+ return II;
+ }
+ return std::nullopt;
+}
+
} // namespace
+llvm::CallBase *
+CodeGenFunction::AddConvergenceControlAttr(llvm::CallBase *Input,
+ llvm::Value *ParentToken) {
+ llvm::Value *bundleArgs[] = {ParentToken};
+ llvm::OperandBundleDef OB("convergencectrl", bundleArgs);
+ auto Output = llvm::CallBase::addOperandBundle(
+ Input, llvm::LLVMContext::OB_convergencectrl, OB, Input);
+ Input->replaceAllUsesWith(Output);
+ Input->eraseFromParent();
+ return Output;
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::EmitConvergenceLoop(llvm::BasicBlock *BB,
+ llvm::Value *ParentToken) {
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+ Builder.SetInsertPoint(>front());
+ auto CB = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_loop, {}, {});
+ Builder.restoreIP(IP);
+
+ auto I = AddConvergenceControlAttr(CB, ParentToken);
+ // Controlled convergence is incompatible with uncontrolled convergence.
+ // Removing any old attributes.
+ I->setNotConvergent();
+
+ assert(isa(I));
+ return dyn_cast(I);
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::getOrEmitConvergenceEntryToken(llvm::Function *F) {
+ auto *BB = >getEntryBlock();
+ auto token = getConvergenceToken(BB);
+ if (token.has_value())
+return token.value();
+
+ // Adding a convergence token requires the function to be marked as
+ // convergent.
+ F->setConvergent();
+
+ CGBuilderTy::InsertPoint IP = Builder.saveIP();
+ Builder.SetInsertPoint(>front());
+ auto I = Builder.CreateIntrinsic(
+ llvm::Intrinsic::experimental_convergence_entry, {}, {});
+ assert(isa(I));
+ Builder.restoreIP(IP);
+
+ return dyn_cast(I);
+}
+
+llvm::IntrinsicInst *
+CodeGenFunction::getOrEmitConvergenceLoopToken(const LoopInfo *LI) {
+ assert(LI != nullptr);
+
+ auto token = getConvergenceToken(LI->getHeader());
+ if (token.has_value())
+return *token;
+
+ llvm::IntrinsicInst *PII =
+ LI->getParent()
+ ? EmitConvergenceLoop(LI->getHeader(),
+getOrEmitConvergenceLoopToken(LI->getParent()))
+ : getOrEmitConvergenceEntryToken(LI->getHeader()->getParent());
+
+ return EmitConvergenceLoop(LI->getHeader(), PII);
+}
+
+llvm::CallBase *
[clang] [HLSL][SPIR-V] Add support -fspv-target-env opt (PR #78611)
https://github.com/Keenuts approved this pull request. https://github.com/llvm/llvm-project/pull/78611 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [HLSL][SPIR-V] Add Vulkan to target triple (PR #76749)
https://github.com/Keenuts approved this pull request. https://github.com/llvm/llvm-project/pull/76749 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[llvm] [clang] [HLSL][SPIR-V] Add Vulkan to target triple (PR #76749)
https://github.com/Keenuts commented: LGTM for the logic, and choices. Just some small nits https://github.com/llvm/llvm-project/pull/76749 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [HLSL][SPIR-V] Add Vulkan to target triple (PR #76749)
@@ -3,29 +3,39 @@ // Supported targets // // RUN: %clang -target dxil-unknown-shadermodel6.2-compute %s -S -o /dev/null 2>&1 | FileCheck --allow-empty --check-prefix=CHECK-VALID %s -// RUN: %clang -target spirv-unknown-shadermodel6.2-compute %s -S -o /dev/null 2>&1 | FileCheck --allow-empty --check-prefix=CHECK-VALID %s +// RUN: %clang -target spirv-unknown-vulkan-compute %s -S -o /dev/null 2>&1 | FileCheck --allow-empty --check-prefix=CHECK-VALID %s +// RUN: %clang -target spirv-unknown-vulkan1.2-compute %s -S -o /dev/null 2>&1 | FileCheck --allow-empty --check-prefix=CHECK-VALID %s +// RUN: %clang -target spirv-unknown-vulkan1.3-compute %s -S -o /dev/null 2>&1 | FileCheck --allow-empty --check-prefix=CHECK-VALID %s +// RUN: %clang -target spirv1.5-unknown-vulkan1.2-compute %s -S -o /dev/null 2>&1 | FileCheck --allow-empty --check-prefix=CHECK-VALID %s +// RUN: %clang -target spirv1.6-unknown-vulkan1.3-compute %s -S -o /dev/null 2>&1 | FileCheck --allow-empty --check-prefix=CHECK-VALID %s -// Empty shader model +// Empty Vulkan environment // // RUN: not %clang -target spirv %s -S -o /dev/null 2>&1 | FileCheck --check-prefix=CHECK-NO-OS %s -// Invalid shader models +// Invalid Vulkan environment // -// RUN: not %clang -target spirv--unknown %s -S -o /dev/null 2>&1 | FileCheck --check-prefix=CHECK-BAD-OS %s +// RUN: not %clang -target spirv--shadermodel %s -S -o /dev/null 2>&1 | FileCheck --check-prefix=CHECK-BAD-OS %s +// RUN: not %clang -target spirv-unknown-vulkan1.0-compute %s -S -o /dev/null 2>&1 | FileCheck --check-prefix=CHECK-BAD-OS %s Keenuts wrote: I wonder where we should document that we explicitly don't want to support vulkan < 1.2 for now, hence this limit. Maybe here a comment to say "vulkan1.0 is valid, but we chose not to support it?) Same for the tests which checks invalid spirv/vk mix? https://github.com/llvm/llvm-project/pull/76749 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [llvm] [HLSL][SPIR-V] Add Vulkan to target triple (PR #76749)
@@ -4236,20 +4236,35 @@ bool CompilerInvocation::ParseLangArgs(LangOptions
, ArgList ,
// TODO: Revisit restricting SPIR-V to logical once we've figured out how
to
// handle PhysicalStorageBuffer64 memory model
if (T.isDXIL() || T.isSPIRVLogical()) {
- enum { ShaderModel, ShaderStage };
+ enum { ShaderModel, VulkanEnv, ShaderStage };
+ enum { OS, Environment };
+
+ int ExpectedOS = T.isSPIRVLogical() ? VulkanEnv : ShaderModel;
+
if (T.getOSName().empty()) {
Diags.Report(diag::err_drv_hlsl_bad_shader_required_in_target)
-<< ShaderModel << T.str();
- } else if (!T.isShaderModelOS() || T.getOSVersion() == VersionTuple(0)) {
-Diags.Report(diag::err_drv_hlsl_bad_shader_unsupported)
-<< ShaderModel << T.getOSName() << T.str();
+<< ExpectedOS << OS << T.str();
} else if (T.getEnvironmentName().empty()) {
Diags.Report(diag::err_drv_hlsl_bad_shader_required_in_target)
-<< ShaderStage << T.str();
+<< ShaderStage << Environment << T.str();
} else if (!T.isShaderStageEnvironment()) {
Diags.Report(diag::err_drv_hlsl_bad_shader_unsupported)
<< ShaderStage << T.getEnvironmentName() << T.str();
}
+
+ if (T.isDXIL()) {
Keenuts wrote:
Shall this be an:
```cpp
if (T.isDXIL()) {
else if (T.isSPIRVLogical()) {
} else {
llvm_unreachable();
}
```
(In case the condition line 4239 gets changed but not this)
https://github.com/llvm/llvm-project/pull/76749
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[llvm] [clang] [HLSL][SPIR-V] Add Vulkan to target triple (PR #76749)
@@ -1328,6 +1331,31 @@ VersionTuple Triple::getDriverKitVersion() const {
}
}
+VersionTuple Triple::getVulkanVersion() const {
+ if (getArch() != spirv || getOS() != Vulkan)
+llvm_unreachable("invalid Vulkan SPIR-V triple");
+
+ VersionTuple VulkanVersion = getOSVersion();
+ SubArchType SpirvVersion = getSubArch();
+
+ llvm::DenseMap ValidVersionMap = {
Keenuts wrote:
Shall we have a way to differentiate unsupported vulkan version, and
unsupported spirv/vulkan mix?
Or have a longer message saying we only support v1.2+spv1.5 and v1.3+spv1.6?)
https://github.com/llvm/llvm-project/pull/76749
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[llvm] [clang] [HLSL][SPIR-V] Add Vulkan to target triple (PR #76749)
https://github.com/Keenuts edited https://github.com/llvm/llvm-project/pull/76749 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [NFC][Clang] Fix potential deref of end iterator (PR #70193)
https://github.com/Keenuts edited https://github.com/llvm/llvm-project/pull/70193 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [NFC][Clang] Fix potential deref of end iterator (PR #70193)
https://github.com/Keenuts closed https://github.com/llvm/llvm-project/pull/70193 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [NFC][Clang] Fix potential deref of end iterator (PR #70193)
@@ -65,7 +65,7 @@ class CXXFieldCollector {
/// getCurFields - Pointer to array of fields added to the currently parsed
/// class.
- FieldDecl **getCurFields() { return &*(Fields.end() - getCurNumFields()); }
Keenuts wrote:
Thanks for the review and linked pointers!
This seems to be the case. and generated code does only the pointer arithmetic,
no actual load. Closing this as this is not an issue.
btw, I like your C++ trivia on twitter
https://github.com/llvm/llvm-project/pull/70193
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [SPIRV] Add -spirv option to DXC driver (PR #65989)
@@ -2,8 +2,8 @@ // Supported targets // -// RUN: %clang -target dxil-unknown-shadermodel6.2-pixel %s -S -o /dev/null 2>&1 | FileCheck --check-prefix=CHECK-VALID %s -// RUN: %clang -target spirv-unknown-shadermodel6.2-library %s -S -o /dev/null 2>&1 | FileCheck --check-prefix=CHECK-VALID %s +// RUN: %clang -target dxil-unknown-shadermodel6.2-compute %s -S -o /dev/null 2>&1 | FileCheck --allow-empty --check-prefix=CHECK-VALID %s +// RUN: %clang -target spirv-unknown-shadermodel6.2-compute %s -S -o /dev/null 2>&1 | FileCheck --allow-empty --check-prefix=CHECK-VALID %s Keenuts wrote: Shall shadermodel version be replaced with unknown on the spirv line? (and triple checked to refuse it when explicitly used with spirv for now?) https://github.com/llvm/llvm-project/pull/65989 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [HLSL][SPIR-V] Fix clang driver lang target test (PR #70330)
https://github.com/Keenuts approved this pull request. https://github.com/llvm/llvm-project/pull/70330 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [HLSL][SPIR-V] Fix clang driver lang target test (PR #70330)
@@ -2,8 +2,8 @@ // Supported targets // -// RUN: %clang -target dxil-unknown-shadermodel6.2-pixel %s -S -o /dev/null 2>&1 | FileCheck --check-prefix=CHECK-VALID %s -// RUN: %clang -target spirv-unknown-shadermodel6.2-library %s -S -o /dev/null 2>&1 | FileCheck --check-prefix=CHECK-VALID %s +// RUN: %clang -target dxil-unknown-shadermodel6.2-compute %s -S -o /dev/null 2>&1 | FileCheck --allow-empty --check-prefix=CHECK-VALID %s Keenuts wrote: Ahh, thanks for the explanation. Yes, sounds better not to rely on an unchecked warning! Thanks! https://github.com/llvm/llvm-project/pull/70330 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [HLSL][SPIR-V] Fix clang driver lang target test (PR #70330)
https://github.com/Keenuts commented: One question, otherwise LGTM, thanks for this! https://github.com/llvm/llvm-project/pull/70330 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [HLSL][SPIR-V] Fix clang driver lang target test (PR #70330)
@@ -2,8 +2,8 @@ // Supported targets // -// RUN: %clang -target dxil-unknown-shadermodel6.2-pixel %s -S -o /dev/null 2>&1 | FileCheck --check-prefix=CHECK-VALID %s -// RUN: %clang -target spirv-unknown-shadermodel6.2-library %s -S -o /dev/null 2>&1 | FileCheck --check-prefix=CHECK-VALID %s +// RUN: %clang -target dxil-unknown-shadermodel6.2-compute %s -S -o /dev/null 2>&1 | FileCheck --allow-empty --check-prefix=CHECK-VALID %s Keenuts wrote: why is the `--allow-empty` required? Isn't the `CHECK-VALID-NOT` enough to not require this option? https://github.com/llvm/llvm-project/pull/70330 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [HLSL][SPIR-V] Fix clang driver lang target test (PR #70330)
https://github.com/Keenuts edited https://github.com/llvm/llvm-project/pull/70330 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [NFC][Clang] Fix potential deref of end iterator (PR #70193)
https://github.com/Keenuts edited https://github.com/llvm/llvm-project/pull/70193 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [NFC][Clang] Fix potential deref of end iterator (PR #70193)
https://github.com/Keenuts ready_for_review https://github.com/llvm/llvm-project/pull/70193 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [NFC][Clang] Fix potential deref of end iterator (PR #70193)
Keenuts wrote: CI seems OK, except clang format which complains about an unrelated line. Marking as ready. https://github.com/llvm/llvm-project/pull/70193 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [NFC][Clang] Fix potential deref of end iterator (PR #70193)
https://github.com/Keenuts created
https://github.com/llvm/llvm-project/pull/70193
This was found by doing bound-checking on SmallVector iterator usage. When the
count is 0, the end iterator is dereferenced to get its address. This doesn't
seem to be an issue in practice as most of the time, and we are allowed to
deref this address, but I don't think this is correct.
From cde1bc9613fa384e4355d39ea29b705b1140dc83 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Nathan=20Gau=C3=ABr?=
Date: Wed, 25 Oct 2023 12:40:22 +0200
Subject: [PATCH] [NFC][Clang] Fix potential deref of end iterator
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
This was found by doing bound-checking on SmallVector iterator usage.
When the count is 0, the end iterator is dereferenced to get its
address. This doesn't seem to be an issue in practice as most of the
time, and we are allowed to deref this address, but I don't think
this is correct.
Signed-off-by: Nathan Gauër
---
clang/include/clang/Sema/CXXFieldCollector.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/clang/include/clang/Sema/CXXFieldCollector.h
b/clang/include/clang/Sema/CXXFieldCollector.h
index f6ecd9f46e5ebdb..ce066581c93fda7 100644
--- a/clang/include/clang/Sema/CXXFieldCollector.h
+++ b/clang/include/clang/Sema/CXXFieldCollector.h
@@ -65,7 +65,7 @@ class CXXFieldCollector {
/// getCurFields - Pointer to array of fields added to the currently parsed
/// class.
- FieldDecl **getCurFields() { return &*(Fields.end() - getCurNumFields()); }
+ FieldDecl **getCurFields() { return Fields.end() - getCurNumFields(); }
/// FinishClass - Called by Sema::ActOnFinishCXXClassDef.
void FinishClass() {
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [SPIRV] Add -spirv option to DXC driver (PR #65989)
https://github.com/Keenuts review_requested https://github.com/llvm/llvm-project/pull/65989 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] 53b6a16 - [SPIR-V] Add SPIR-V logical triple.
Author: Nathan Gauër
Date: 2023-09-11T10:15:24+02:00
New Revision: 53b6a169e453a2a91d3713ca16fa089853c670a8
URL:
https://github.com/llvm/llvm-project/commit/53b6a169e453a2a91d3713ca16fa089853c670a8
DIFF:
https://github.com/llvm/llvm-project/commit/53b6a169e453a2a91d3713ca16fa089853c670a8.diff
LOG: [SPIR-V] Add SPIR-V logical triple.
Clang implements SPIR-V with both Physical32 and Physical64 addressing
models. This commit adds a new triple value for the Logical
addressing model.
Differential Revision: https://reviews.llvm.org/D155978
Added:
Modified:
clang/lib/Basic/Targets.cpp
clang/lib/Basic/Targets/SPIR.cpp
clang/lib/Basic/Targets/SPIR.h
clang/lib/Frontend/CompilerInvocation.cpp
llvm/include/llvm/TargetParser/Triple.h
llvm/lib/TargetParser/Triple.cpp
llvm/unittests/TargetParser/TripleTest.cpp
Removed:
diff --git a/clang/lib/Basic/Targets.cpp b/clang/lib/Basic/Targets.cpp
index 2afffc463d18be1..69576dbc458d9a1 100644
--- a/clang/lib/Basic/Targets.cpp
+++ b/clang/lib/Basic/Targets.cpp
@@ -665,6 +665,9 @@ std::unique_ptr AllocateTarget(const
llvm::Triple ,
return nullptr;
return std::make_unique(Triple, Opts);
}
+ case llvm::Triple::spirv: {
+return std::make_unique(Triple, Opts);
+ }
case llvm::Triple::spirv32: {
if (os != llvm::Triple::UnknownOS ||
Triple.getEnvironment() != llvm::Triple::UnknownEnvironment)
diff --git a/clang/lib/Basic/Targets/SPIR.cpp
b/clang/lib/Basic/Targets/SPIR.cpp
index 09d482a8b9ef594..dc920177d3a9107 100644
--- a/clang/lib/Basic/Targets/SPIR.cpp
+++ b/clang/lib/Basic/Targets/SPIR.cpp
@@ -33,19 +33,24 @@ void SPIR64TargetInfo::getTargetDefines(const LangOptions
,
DefineStd(Builder, "SPIR64", Opts);
}
+void BaseSPIRVTargetInfo::getTargetDefines(const LangOptions ,
+ MacroBuilder ) const {
+ DefineStd(Builder, "SPIRV", Opts);
+}
+
void SPIRVTargetInfo::getTargetDefines(const LangOptions ,
MacroBuilder ) const {
- DefineStd(Builder, "SPIRV", Opts);
+ BaseSPIRVTargetInfo::getTargetDefines(Opts, Builder);
}
void SPIRV32TargetInfo::getTargetDefines(const LangOptions ,
MacroBuilder ) const {
- SPIRVTargetInfo::getTargetDefines(Opts, Builder);
+ BaseSPIRVTargetInfo::getTargetDefines(Opts, Builder);
DefineStd(Builder, "SPIRV32", Opts);
}
void SPIRV64TargetInfo::getTargetDefines(const LangOptions ,
MacroBuilder ) const {
- SPIRVTargetInfo::getTargetDefines(Opts, Builder);
+ BaseSPIRVTargetInfo::getTargetDefines(Opts, Builder);
DefineStd(Builder, "SPIRV64", Opts);
}
diff --git a/clang/lib/Basic/Targets/SPIR.h b/clang/lib/Basic/Targets/SPIR.h
index a7ea03e7a5dd32b..9ab2b7c60936392 100644
--- a/clang/lib/Basic/Targets/SPIR.h
+++ b/clang/lib/Basic/Targets/SPIR.h
@@ -93,10 +93,6 @@ class LLVM_LIBRARY_VISIBILITY BaseSPIRTargetInfo : public
TargetInfo {
: TargetInfo(Triple) {
assert((Triple.isSPIR() || Triple.isSPIRV()) &&
"Invalid architecture for SPIR or SPIR-V.");
-assert(getTriple().getOS() == llvm::Triple::UnknownOS &&
- "SPIR(-V) target must use unknown OS");
-assert(getTriple().getEnvironment() == llvm::Triple::UnknownEnvironment &&
- "SPIR(-V) target must use unknown environment type");
TLSSupported = false;
VLASupported = false;
LongWidth = LongAlign = 64;
@@ -284,31 +280,53 @@ class LLVM_LIBRARY_VISIBILITY SPIR64TargetInfo : public
SPIRTargetInfo {
MacroBuilder ) const override;
};
-class LLVM_LIBRARY_VISIBILITY SPIRVTargetInfo : public BaseSPIRTargetInfo {
+class LLVM_LIBRARY_VISIBILITY BaseSPIRVTargetInfo : public BaseSPIRTargetInfo {
public:
- SPIRVTargetInfo(const llvm::Triple , const TargetOptions )
+ BaseSPIRVTargetInfo(const llvm::Triple , const TargetOptions )
: BaseSPIRTargetInfo(Triple, Opts) {
assert(Triple.isSPIRV() && "Invalid architecture for SPIR-V.");
-assert(getTriple().getOS() == llvm::Triple::UnknownOS &&
- "SPIR-V target must use unknown OS");
-assert(getTriple().getEnvironment() == llvm::Triple::UnknownEnvironment &&
- "SPIR-V target must use unknown environment type");
+ }
+
+ bool hasFeature(StringRef Feature) const override {
+return Feature == "spirv";
}
void getTargetDefines(const LangOptions ,
MacroBuilder ) const override;
+};
- bool hasFeature(StringRef Feature) const override {
-return Feature == "spirv";
+class LLVM_LIBRARY_VISIBILITY SPIRVTargetInfo : public BaseSPIRVTargetInfo {
+public:
+ SPIRVTargetInfo(const llvm::Triple , const TargetOptions )
+ : BaseSPIRVTargetInfo(Triple, Opts) {
+assert(Triple.getArch() == llvm::Triple::spirv &&
+ "Invalid architecture for Logical SPIR-V.");
