maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/llvm/unittests/IR/PatternMatch.cpp

565 lines
21 KiB
C++
Raw Normal View History

PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
//===---- llvm/unittest/IR/PatternMatch.cpp - PatternMatch unit tests ----===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
Re-sort #include lines for unittests. This uses a slightly modified clang-format (https://reviews.llvm.org/D33932) to keep primary headers at the top and handle new utility headers like 'gmock' consistently with other utility headers. No other change was made. I did no manual edits, all of this is clang-format. This should allow other changes to have more clear and focused diffs, and is especially motivated by moving some headers into more focused libraries. llvm-svn: 304786
2017-06-06 19:06:56 +08:00
#include "llvm/IR/PatternMatch.h"
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
#include "llvm/ADT/STLExtras.h"
#include "llvm/Analysis/ValueTracking.h"
#include "llvm/IR/BasicBlock.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/DerivedTypes.h"
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
#include "llvm/IR/Function.h"
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
#include "llvm/IR/IRBuilder.h"
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
#include "llvm/IR/Instructions.h"
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
#include "llvm/IR/LLVMContext.h"
#include "llvm/IR/MDBuilder.h"
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
#include "llvm/IR/Module.h"
[Modules] Move the NoFolder into the IR library as it creates instructions. llvm-svn: 202834
2014-03-04 20:05:47 +08:00
#include "llvm/IR/NoFolder.h"
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
#include "llvm/IR/Operator.h"
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
#include "llvm/IR/Type.h"
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
#include "gtest/gtest.h"
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
using namespace llvm;
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
using namespace llvm::PatternMatch;
namespace {
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
struct PatternMatchTest : ::testing::Test {
LLVMContext Ctx;
Replace OwningPtr<T> with std::unique_ptr<T>. This compiles with no changes to clang/lld/lldb with MSVC and includes overloads to various functions which are used by those projects and llvm which have OwningPtr's as parameters. This should allow out of tree projects some time to move. There are also no changes to libs/Target, which should help out of tree targets have time to move, if necessary. llvm-svn: 203083
2014-03-06 13:51:42 +08:00
std::unique_ptr<Module> M;
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
Function *F;
BasicBlock *BB;
Remove PreserveNames template parameter from IRBuilder This reapplies r263258, which was reverted in r263321 because of issues on Clang side. From: Mehdi Amini <mehdi.amini@apple.com> llvm-svn: 263393
2016-03-14 05:05:13 +08:00
IRBuilder<NoFolder> IRB;
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
PatternMatchTest()
: M(new Module("PatternMatchTestModule", Ctx)),
F(Function::Create(
FunctionType::get(Type::getVoidTy(Ctx), /* IsVarArg */ false),
Function::ExternalLinkage, "f", M.get())),
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
BB(BasicBlock::Create(Ctx, "entry", F)), IRB(BB) {}
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
};
Add in a unittest for the one-use pattern matcher. llvm-svn: 198552
2014-01-05 17:14:53 +08:00
TEST_F(PatternMatchTest, OneUse) {
// Build up a little tree of values:
//
// One = (1 + 2) + 42
// Two = One + 42
// Leaf = (Two + 8) + (Two + 13)
Value *One = IRB.CreateAdd(IRB.CreateAdd(IRB.getInt32(1), IRB.getInt32(2)),
IRB.getInt32(42));
Value *Two = IRB.CreateAdd(One, IRB.getInt32(42));
Value *Leaf = IRB.CreateAdd(IRB.CreateAdd(Two, IRB.getInt32(8)),
IRB.CreateAdd(Two, IRB.getInt32(13)));
Value *V;
EXPECT_TRUE(m_OneUse(m_Value(V)).match(One));
EXPECT_EQ(One, V);
EXPECT_FALSE(m_OneUse(m_Value()).match(Two));
EXPECT_FALSE(m_OneUse(m_Value()).match(Leaf));
}
[PatternMatch] Stabilize the matching order of commutative matchers Summary: Currently, we 1. match `LHS` matcher to the `first` operand of binary operator, 2. and then match `RHS` matcher to the `second` operand of binary operator. If that does not match, we swap the `LHS` and `RHS` matchers: 1. match `RHS` matcher to the `first` operand of binary operator, 2. and then match `LHS` matcher to the `second` operand of binary operator. This works ok. But it complicates writing of commutative matchers, where one would like to match (`m_Value()`) the value on one side, and use (`m_Specific()`) it on the other side. This is additionally complicated by the fact that `m_Specific()` stores the `Value *`, not `Value **`, so it won't work at all out of the box. The last problem is trivially solved by adding a new `m_c_Specific()` that stores the `Value **`, not `Value *`. I'm choosing to add a new matcher, not change the existing one because i guess all the current users are ok with existing behavior, and this additional pointer indirection may have performance drawbacks. Also, i'm storing pointer, not reference, because for some mysterious-to-me reason it did not work with the reference. The first one appears trivial, too. Currently, we 1. match `LHS` matcher to the `first` operand of binary operator, 2. and then match `RHS` matcher to the `second` operand of binary operator. If that does not match, we swap the ~~`LHS` and `RHS` matchers~~ **operands**: 1. match ~~`RHS`~~ **`LHS`** matcher to the ~~`first`~~ **`second`** operand of binary operator, 2. and then match ~~`LHS`~~ **`RHS`** matcher to the ~~`second`~ **`first`** operand of binary operator. Surprisingly, `$ ninja check-llvm` still passes with this. But i expect the bots will disagree.. The motivational unittest is included. I'd like to use this in D45664. Reviewers: spatel, craig.topper, arsenm, RKSimon Reviewed By: craig.topper Subscribers: xbolva00, wdng, llvm-commits Differential Revision: https://reviews.llvm.org/D45828 llvm-svn: 331085
2018-04-28 05:23:20 +08:00
TEST_F(PatternMatchTest, CommutativeDeferredValue) {
Value *X = IRB.getInt32(1);
Value *Y = IRB.getInt32(2);
{
Value *tX = X;
EXPECT_TRUE(match(X, m_Deferred(tX)));
EXPECT_FALSE(match(Y, m_Deferred(tX)));
}
{
const Value *tX = X;
EXPECT_TRUE(match(X, m_Deferred(tX)));
EXPECT_FALSE(match(Y, m_Deferred(tX)));
}
{
Value *const tX = X;
EXPECT_TRUE(match(X, m_Deferred(tX)));
EXPECT_FALSE(match(Y, m_Deferred(tX)));
}
{
const Value *const tX = X;
EXPECT_TRUE(match(X, m_Deferred(tX)));
EXPECT_FALSE(match(Y, m_Deferred(tX)));
}
{
Value *tX = nullptr;
EXPECT_TRUE(match(IRB.CreateAnd(X, X), m_And(m_Value(tX), m_Deferred(tX))));
EXPECT_EQ(tX, X);
}
{
Value *tX = nullptr;
EXPECT_FALSE(
match(IRB.CreateAnd(X, Y), m_c_And(m_Value(tX), m_Deferred(tX))));
}
auto checkMatch = [X, Y](Value *Pattern) {
Value *tX = nullptr, *tY = nullptr;
EXPECT_TRUE(match(
Pattern, m_c_And(m_Value(tX), m_c_And(m_Deferred(tX), m_Value(tY)))));
EXPECT_EQ(tX, X);
EXPECT_EQ(tY, Y);
};
checkMatch(IRB.CreateAnd(X, IRB.CreateAnd(X, Y)));
checkMatch(IRB.CreateAnd(X, IRB.CreateAnd(Y, X)));
checkMatch(IRB.CreateAnd(IRB.CreateAnd(X, Y), X));
checkMatch(IRB.CreateAnd(IRB.CreateAnd(Y, X), X));
}
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
TEST_F(PatternMatchTest, FloatingPointOrderedMin) {
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
Type *FltTy = IRB.getFloatTy();
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
Value *L = ConstantFP::get(FltTy, 1.0);
Value *R = ConstantFP::get(FltTy, 2.0);
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
Value *MatchL, *MatchR;
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
// Test OLT.
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
EXPECT_TRUE(m_OrdFMin(m_Value(MatchL), m_Value(MatchR))
.match(IRB.CreateSelect(IRB.CreateFCmpOLT(L, R), L, R)));
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
EXPECT_EQ(L, MatchL);
EXPECT_EQ(R, MatchR);
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
// Test OLE.
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
EXPECT_TRUE(m_OrdFMin(m_Value(MatchL), m_Value(MatchR))
.match(IRB.CreateSelect(IRB.CreateFCmpOLE(L, R), L, R)));
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
EXPECT_EQ(L, MatchL);
EXPECT_EQ(R, MatchR);
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
// Test no match on OGE.
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
EXPECT_FALSE(m_OrdFMin(m_Value(MatchL), m_Value(MatchR))
.match(IRB.CreateSelect(IRB.CreateFCmpOGE(L, R), L, R)));
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
// Test no match on OGT.
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
EXPECT_FALSE(m_OrdFMin(m_Value(MatchL), m_Value(MatchR))
.match(IRB.CreateSelect(IRB.CreateFCmpOGT(L, R), L, R)));
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
Fix m_[Ord|Unord][FMin|FMax] matchers to correctly match ordering. Previously, the matching was done incorrectly for the case where operands for FCmpInst and SelectInst were in opposite order. Patch by Andrei Elovikov. Differential Revision: https://reviews.llvm.org/D33185 llvm-svn: 305308
2017-06-14 01:18:45 +08:00
// Test inverted selects. Note, that this "inverts" the ordering, e.g.:
// %cmp = fcmp oge L, R
// %min = select %cmp R, L
// Given L == NaN
// the above is expanded to %cmp == false ==> %min = L
// which is true for UnordFMin, not OrdFMin, so test that:
// [OU]GE with inverted select.
EXPECT_FALSE(m_OrdFMin(m_Value(MatchL), m_Value(MatchR))
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
.match(IRB.CreateSelect(IRB.CreateFCmpOGE(L, R), R, L)));
Fix m_[Ord|Unord][FMin|FMax] matchers to correctly match ordering. Previously, the matching was done incorrectly for the case where operands for FCmpInst and SelectInst were in opposite order. Patch by Andrei Elovikov. Differential Revision: https://reviews.llvm.org/D33185 llvm-svn: 305308
2017-06-14 01:18:45 +08:00
EXPECT_TRUE(m_OrdFMin(m_Value(MatchL), m_Value(MatchR))
.match(IRB.CreateSelect(IRB.CreateFCmpUGE(L, R), R, L)));
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
EXPECT_EQ(L, MatchL);
EXPECT_EQ(R, MatchR);
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
Fix m_[Ord|Unord][FMin|FMax] matchers to correctly match ordering. Previously, the matching was done incorrectly for the case where operands for FCmpInst and SelectInst were in opposite order. Patch by Andrei Elovikov. Differential Revision: https://reviews.llvm.org/D33185 llvm-svn: 305308
2017-06-14 01:18:45 +08:00
// [OU]GT with inverted select.
EXPECT_FALSE(m_OrdFMin(m_Value(MatchL), m_Value(MatchR))
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
.match(IRB.CreateSelect(IRB.CreateFCmpOGT(L, R), R, L)));
Fix m_[Ord|Unord][FMin|FMax] matchers to correctly match ordering. Previously, the matching was done incorrectly for the case where operands for FCmpInst and SelectInst were in opposite order. Patch by Andrei Elovikov. Differential Revision: https://reviews.llvm.org/D33185 llvm-svn: 305308
2017-06-14 01:18:45 +08:00
EXPECT_TRUE(m_OrdFMin(m_Value(MatchL), m_Value(MatchR))
.match(IRB.CreateSelect(IRB.CreateFCmpUGT(L, R), R, L)));
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
EXPECT_EQ(L, MatchL);
EXPECT_EQ(R, MatchR);
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
}
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
TEST_F(PatternMatchTest, FloatingPointOrderedMax) {
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
Type *FltTy = IRB.getFloatTy();
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
Value *L = ConstantFP::get(FltTy, 1.0);
Value *R = ConstantFP::get(FltTy, 2.0);
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
Value *MatchL, *MatchR;
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
// Test OGT.
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
EXPECT_TRUE(m_OrdFMax(m_Value(MatchL), m_Value(MatchR))
.match(IRB.CreateSelect(IRB.CreateFCmpOGT(L, R), L, R)));
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
EXPECT_EQ(L, MatchL);
EXPECT_EQ(R, MatchR);
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
// Test OGE.
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
EXPECT_TRUE(m_OrdFMax(m_Value(MatchL), m_Value(MatchR))
.match(IRB.CreateSelect(IRB.CreateFCmpOGE(L, R), L, R)));
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
EXPECT_EQ(L, MatchL);
EXPECT_EQ(R, MatchR);
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
// Test no match on OLE.
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
EXPECT_FALSE(m_OrdFMax(m_Value(MatchL), m_Value(MatchR))
.match(IRB.CreateSelect(IRB.CreateFCmpOLE(L, R), L, R)));
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
// Test no match on OLT.
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
EXPECT_FALSE(m_OrdFMax(m_Value(MatchL), m_Value(MatchR))
.match(IRB.CreateSelect(IRB.CreateFCmpOLT(L, R), L, R)));
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
Fix m_[Ord|Unord][FMin|FMax] matchers to correctly match ordering. Previously, the matching was done incorrectly for the case where operands for FCmpInst and SelectInst were in opposite order. Patch by Andrei Elovikov. Differential Revision: https://reviews.llvm.org/D33185 llvm-svn: 305308
2017-06-14 01:18:45 +08:00
// Test inverted selects. Note, that this "inverts" the ordering, e.g.:
// %cmp = fcmp ole L, R
// %max = select %cmp, R, L
// Given L == NaN,
// the above is expanded to %cmp == false ==> %max == L
// which is true for UnordFMax, not OrdFMax, so test that:
// [OU]LE with inverted select.
EXPECT_FALSE(m_OrdFMax(m_Value(MatchL), m_Value(MatchR))
.match(IRB.CreateSelect(IRB.CreateFCmpOLE(L, R), R, L)));
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
EXPECT_TRUE(m_OrdFMax(m_Value(MatchL), m_Value(MatchR))
Fix m_[Ord|Unord][FMin|FMax] matchers to correctly match ordering. Previously, the matching was done incorrectly for the case where operands for FCmpInst and SelectInst were in opposite order. Patch by Andrei Elovikov. Differential Revision: https://reviews.llvm.org/D33185 llvm-svn: 305308
2017-06-14 01:18:45 +08:00
.match(IRB.CreateSelect(IRB.CreateFCmpULE(L, R), R, L)));
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
EXPECT_EQ(L, MatchL);
EXPECT_EQ(R, MatchR);
Fix m_[Ord|Unord][FMin|FMax] matchers to correctly match ordering. Previously, the matching was done incorrectly for the case where operands for FCmpInst and SelectInst were in opposite order. Patch by Andrei Elovikov. Differential Revision: https://reviews.llvm.org/D33185 llvm-svn: 305308
2017-06-14 01:18:45 +08:00
// [OUT]LT with inverted select.
EXPECT_FALSE(m_OrdFMax(m_Value(MatchL), m_Value(MatchR))
.match(IRB.CreateSelect(IRB.CreateFCmpOLT(L, R), R, L)));
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
EXPECT_TRUE(m_OrdFMax(m_Value(MatchL), m_Value(MatchR))
Fix m_[Ord|Unord][FMin|FMax] matchers to correctly match ordering. Previously, the matching was done incorrectly for the case where operands for FCmpInst and SelectInst were in opposite order. Patch by Andrei Elovikov. Differential Revision: https://reviews.llvm.org/D33185 llvm-svn: 305308
2017-06-14 01:18:45 +08:00
.match(IRB.CreateSelect(IRB.CreateFCmpULT(L, R), R, L)));
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
EXPECT_EQ(L, MatchL);
EXPECT_EQ(R, MatchR);
}
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
TEST_F(PatternMatchTest, FloatingPointUnorderedMin) {
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
Type *FltTy = IRB.getFloatTy();
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
Value *L = ConstantFP::get(FltTy, 1.0);
Value *R = ConstantFP::get(FltTy, 2.0);
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
Value *MatchL, *MatchR;
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
// Test ULT.
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
EXPECT_TRUE(m_UnordFMin(m_Value(MatchL), m_Value(MatchR))
.match(IRB.CreateSelect(IRB.CreateFCmpULT(L, R), L, R)));
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
EXPECT_EQ(L, MatchL);
EXPECT_EQ(R, MatchR);
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
// Test ULE.
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
EXPECT_TRUE(m_UnordFMin(m_Value(MatchL), m_Value(MatchR))
.match(IRB.CreateSelect(IRB.CreateFCmpULE(L, R), L, R)));
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
EXPECT_EQ(L, MatchL);
EXPECT_EQ(R, MatchR);
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
// Test no match on UGE.
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
EXPECT_FALSE(m_UnordFMin(m_Value(MatchL), m_Value(MatchR))
.match(IRB.CreateSelect(IRB.CreateFCmpUGE(L, R), L, R)));
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
// Test no match on UGT.
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
EXPECT_FALSE(m_UnordFMin(m_Value(MatchL), m_Value(MatchR))
.match(IRB.CreateSelect(IRB.CreateFCmpUGT(L, R), L, R)));
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
Fix m_[Ord|Unord][FMin|FMax] matchers to correctly match ordering. Previously, the matching was done incorrectly for the case where operands for FCmpInst and SelectInst were in opposite order. Patch by Andrei Elovikov. Differential Revision: https://reviews.llvm.org/D33185 llvm-svn: 305308
2017-06-14 01:18:45 +08:00
// Test inverted selects. Note, that this "inverts" the ordering, e.g.:
// %cmp = fcmp uge L, R
// %min = select %cmp R, L
// Given L == NaN
// the above is expanded to %cmp == true ==> %min = R
// which is true for OrdFMin, not UnordFMin, so test that:
// [UO]GE with inverted select.
EXPECT_FALSE(m_UnordFMin(m_Value(MatchL), m_Value(MatchR))
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
.match(IRB.CreateSelect(IRB.CreateFCmpUGE(L, R), R, L)));
Fix m_[Ord|Unord][FMin|FMax] matchers to correctly match ordering. Previously, the matching was done incorrectly for the case where operands for FCmpInst and SelectInst were in opposite order. Patch by Andrei Elovikov. Differential Revision: https://reviews.llvm.org/D33185 llvm-svn: 305308
2017-06-14 01:18:45 +08:00
EXPECT_TRUE(m_UnordFMin(m_Value(MatchL), m_Value(MatchR))
.match(IRB.CreateSelect(IRB.CreateFCmpOGE(L, R), R, L)));
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
EXPECT_EQ(L, MatchL);
EXPECT_EQ(R, MatchR);
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
Fix m_[Ord|Unord][FMin|FMax] matchers to correctly match ordering. Previously, the matching was done incorrectly for the case where operands for FCmpInst and SelectInst were in opposite order. Patch by Andrei Elovikov. Differential Revision: https://reviews.llvm.org/D33185 llvm-svn: 305308
2017-06-14 01:18:45 +08:00
// [UO]GT with inverted select.
EXPECT_FALSE(m_UnordFMin(m_Value(MatchL), m_Value(MatchR))
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
.match(IRB.CreateSelect(IRB.CreateFCmpUGT(L, R), R, L)));
Fix m_[Ord|Unord][FMin|FMax] matchers to correctly match ordering. Previously, the matching was done incorrectly for the case where operands for FCmpInst and SelectInst were in opposite order. Patch by Andrei Elovikov. Differential Revision: https://reviews.llvm.org/D33185 llvm-svn: 305308
2017-06-14 01:18:45 +08:00
EXPECT_TRUE(m_UnordFMin(m_Value(MatchL), m_Value(MatchR))
.match(IRB.CreateSelect(IRB.CreateFCmpOGT(L, R), R, L)));
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
EXPECT_EQ(L, MatchL);
EXPECT_EQ(R, MatchR);
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
}
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
TEST_F(PatternMatchTest, FloatingPointUnorderedMax) {
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
Type *FltTy = IRB.getFloatTy();
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
Value *L = ConstantFP::get(FltTy, 1.0);
Value *R = ConstantFP::get(FltTy, 2.0);
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
Value *MatchL, *MatchR;
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
// Test UGT.
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
EXPECT_TRUE(m_UnordFMax(m_Value(MatchL), m_Value(MatchR))
.match(IRB.CreateSelect(IRB.CreateFCmpUGT(L, R), L, R)));
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
EXPECT_EQ(L, MatchL);
EXPECT_EQ(R, MatchR);
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
// Test UGE.
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
EXPECT_TRUE(m_UnordFMax(m_Value(MatchL), m_Value(MatchR))
.match(IRB.CreateSelect(IRB.CreateFCmpUGE(L, R), L, R)));
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
EXPECT_EQ(L, MatchL);
EXPECT_EQ(R, MatchR);
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
// Test no match on ULE.
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
EXPECT_FALSE(m_UnordFMax(m_Value(MatchL), m_Value(MatchR))
.match(IRB.CreateSelect(IRB.CreateFCmpULE(L, R), L, R)));
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
// Test no match on ULT.
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
EXPECT_FALSE(m_UnordFMax(m_Value(MatchL), m_Value(MatchR))
.match(IRB.CreateSelect(IRB.CreateFCmpULT(L, R), L, R)));
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
Fix m_[Ord|Unord][FMin|FMax] matchers to correctly match ordering. Previously, the matching was done incorrectly for the case where operands for FCmpInst and SelectInst were in opposite order. Patch by Andrei Elovikov. Differential Revision: https://reviews.llvm.org/D33185 llvm-svn: 305308
2017-06-14 01:18:45 +08:00
// Test inverted selects. Note, that this "inverts" the ordering, e.g.:
// %cmp = fcmp ule L, R
// %max = select %cmp R, L
// Given L == NaN
// the above is expanded to %cmp == true ==> %max = R
// which is true for OrdFMax, not UnordFMax, so test that:
// [UO]LE with inverted select.
EXPECT_FALSE(m_UnordFMax(m_Value(MatchL), m_Value(MatchR))
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
.match(IRB.CreateSelect(IRB.CreateFCmpULE(L, R), R, L)));
Fix m_[Ord|Unord][FMin|FMax] matchers to correctly match ordering. Previously, the matching was done incorrectly for the case where operands for FCmpInst and SelectInst were in opposite order. Patch by Andrei Elovikov. Differential Revision: https://reviews.llvm.org/D33185 llvm-svn: 305308
2017-06-14 01:18:45 +08:00
EXPECT_TRUE(m_UnordFMax(m_Value(MatchL), m_Value(MatchR))
.match(IRB.CreateSelect(IRB.CreateFCmpOLE(L, R), R, L)));
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
EXPECT_EQ(L, MatchL);
EXPECT_EQ(R, MatchR);
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
Fix m_[Ord|Unord][FMin|FMax] matchers to correctly match ordering. Previously, the matching was done incorrectly for the case where operands for FCmpInst and SelectInst were in opposite order. Patch by Andrei Elovikov. Differential Revision: https://reviews.llvm.org/D33185 llvm-svn: 305308
2017-06-14 01:18:45 +08:00
// [UO]LT with inverted select.
EXPECT_FALSE(m_UnordFMax(m_Value(MatchL), m_Value(MatchR))
Use a shorter name for the IRBuilder member. This will help the tests I'm adding next be a lot more readable. llvm-svn: 198534
2014-01-05 10:23:11 +08:00
.match(IRB.CreateSelect(IRB.CreateFCmpULT(L, R), R, L)));
Fix m_[Ord|Unord][FMin|FMax] matchers to correctly match ordering. Previously, the matching was done incorrectly for the case where operands for FCmpInst and SelectInst were in opposite order. Patch by Andrei Elovikov. Differential Revision: https://reviews.llvm.org/D33185 llvm-svn: 305308
2017-06-14 01:18:45 +08:00
EXPECT_TRUE(m_UnordFMax(m_Value(MatchL), m_Value(MatchR))
.match(IRB.CreateSelect(IRB.CreateFCmpOLT(L, R), R, L)));
Simplify the PatternMatch unittest by giving it a module, function, and basic block to hold instructions, and managing all of their lifetimes in a fixture. This makes it easy to sink the expectations into the test cases themselves which also makes things a bit more explicit and clearer IMO. llvm-svn: 198532
2014-01-05 10:07:20 +08:00
EXPECT_EQ(L, MatchL);
EXPECT_EQ(R, MatchR);
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
}
Add support to the pattern match library for matching NSW and NUW instructions. I needed this for a quick experiment I was making, and while I've no idea if that will ever get committed, I didn't want to throw away the pattern match code and for anyone else to have to write it again. I've added unittests to make sure this works correctly. In fun news, this also uncovered the IRBuilder bug. Doh! llvm-svn: 198541
2014-01-05 11:28:29 +08:00
TEST_F(PatternMatchTest, OverflowingBinOps) {
Value *L = IRB.getInt32(1);
Value *R = IRB.getInt32(2);
Value *MatchL, *MatchR;
EXPECT_TRUE(
m_NSWAdd(m_Value(MatchL), m_Value(MatchR)).match(IRB.CreateNSWAdd(L, R)));
EXPECT_EQ(L, MatchL);
EXPECT_EQ(R, MatchR);
[C++11] Use 'nullptr'. llvm-svn: 210442
2014-06-09 06:29:17 +08:00
MatchL = MatchR = nullptr;
Add support to the pattern match library for matching NSW and NUW instructions. I needed this for a quick experiment I was making, and while I've no idea if that will ever get committed, I didn't want to throw away the pattern match code and for anyone else to have to write it again. I've added unittests to make sure this works correctly. In fun news, this also uncovered the IRBuilder bug. Doh! llvm-svn: 198541
2014-01-05 11:28:29 +08:00
EXPECT_TRUE(
m_NSWSub(m_Value(MatchL), m_Value(MatchR)).match(IRB.CreateNSWSub(L, R)));
EXPECT_EQ(L, MatchL);
EXPECT_EQ(R, MatchR);
[C++11] Use 'nullptr'. llvm-svn: 210442
2014-06-09 06:29:17 +08:00
MatchL = MatchR = nullptr;
Add support to the pattern match library for matching NSW and NUW instructions. I needed this for a quick experiment I was making, and while I've no idea if that will ever get committed, I didn't want to throw away the pattern match code and for anyone else to have to write it again. I've added unittests to make sure this works correctly. In fun news, this also uncovered the IRBuilder bug. Doh! llvm-svn: 198541
2014-01-05 11:28:29 +08:00
EXPECT_TRUE(
m_NSWMul(m_Value(MatchL), m_Value(MatchR)).match(IRB.CreateNSWMul(L, R)));
EXPECT_EQ(L, MatchL);
EXPECT_EQ(R, MatchR);
[C++11] Use 'nullptr'. llvm-svn: 210442
2014-06-09 06:29:17 +08:00
MatchL = MatchR = nullptr;
Add support to the pattern match library for matching NSW and NUW instructions. I needed this for a quick experiment I was making, and while I've no idea if that will ever get committed, I didn't want to throw away the pattern match code and for anyone else to have to write it again. I've added unittests to make sure this works correctly. In fun news, this also uncovered the IRBuilder bug. Doh! llvm-svn: 198541
2014-01-05 11:28:29 +08:00
EXPECT_TRUE(m_NSWShl(m_Value(MatchL), m_Value(MatchR)).match(
IRB.CreateShl(L, R, "", /* NUW */ false, /* NSW */ true)));
EXPECT_EQ(L, MatchL);
EXPECT_EQ(R, MatchR);
EXPECT_TRUE(
m_NUWAdd(m_Value(MatchL), m_Value(MatchR)).match(IRB.CreateNUWAdd(L, R)));
EXPECT_EQ(L, MatchL);
EXPECT_EQ(R, MatchR);
[C++11] Use 'nullptr'. llvm-svn: 210442
2014-06-09 06:29:17 +08:00
MatchL = MatchR = nullptr;
Add support to the pattern match library for matching NSW and NUW instructions. I needed this for a quick experiment I was making, and while I've no idea if that will ever get committed, I didn't want to throw away the pattern match code and for anyone else to have to write it again. I've added unittests to make sure this works correctly. In fun news, this also uncovered the IRBuilder bug. Doh! llvm-svn: 198541
2014-01-05 11:28:29 +08:00
EXPECT_TRUE(
m_NUWSub(m_Value(MatchL), m_Value(MatchR)).match(IRB.CreateNUWSub(L, R)));
EXPECT_EQ(L, MatchL);
EXPECT_EQ(R, MatchR);
[C++11] Use 'nullptr'. llvm-svn: 210442
2014-06-09 06:29:17 +08:00
MatchL = MatchR = nullptr;
Add support to the pattern match library for matching NSW and NUW instructions. I needed this for a quick experiment I was making, and while I've no idea if that will ever get committed, I didn't want to throw away the pattern match code and for anyone else to have to write it again. I've added unittests to make sure this works correctly. In fun news, this also uncovered the IRBuilder bug. Doh! llvm-svn: 198541
2014-01-05 11:28:29 +08:00
EXPECT_TRUE(
m_NUWMul(m_Value(MatchL), m_Value(MatchR)).match(IRB.CreateNUWMul(L, R)));
EXPECT_EQ(L, MatchL);
EXPECT_EQ(R, MatchR);
[C++11] Use 'nullptr'. llvm-svn: 210442
2014-06-09 06:29:17 +08:00
MatchL = MatchR = nullptr;
Add support to the pattern match library for matching NSW and NUW instructions. I needed this for a quick experiment I was making, and while I've no idea if that will ever get committed, I didn't want to throw away the pattern match code and for anyone else to have to write it again. I've added unittests to make sure this works correctly. In fun news, this also uncovered the IRBuilder bug. Doh! llvm-svn: 198541
2014-01-05 11:28:29 +08:00
EXPECT_TRUE(m_NUWShl(m_Value(MatchL), m_Value(MatchR)).match(
IRB.CreateShl(L, R, "", /* NUW */ true, /* NSW */ false)));
EXPECT_EQ(L, MatchL);
EXPECT_EQ(R, MatchR);
EXPECT_FALSE(m_NSWAdd(m_Value(), m_Value()).match(IRB.CreateAdd(L, R)));
EXPECT_FALSE(m_NSWAdd(m_Value(), m_Value()).match(IRB.CreateNUWAdd(L, R)));
EXPECT_FALSE(m_NSWAdd(m_Value(), m_Value()).match(IRB.CreateNSWSub(L, R)));
EXPECT_FALSE(m_NSWSub(m_Value(), m_Value()).match(IRB.CreateSub(L, R)));
EXPECT_FALSE(m_NSWSub(m_Value(), m_Value()).match(IRB.CreateNUWSub(L, R)));
EXPECT_FALSE(m_NSWSub(m_Value(), m_Value()).match(IRB.CreateNSWAdd(L, R)));
EXPECT_FALSE(m_NSWMul(m_Value(), m_Value()).match(IRB.CreateMul(L, R)));
EXPECT_FALSE(m_NSWMul(m_Value(), m_Value()).match(IRB.CreateNUWMul(L, R)));
EXPECT_FALSE(m_NSWMul(m_Value(), m_Value()).match(IRB.CreateNSWAdd(L, R)));
EXPECT_FALSE(m_NSWShl(m_Value(), m_Value()).match(IRB.CreateShl(L, R)));
EXPECT_FALSE(m_NSWShl(m_Value(), m_Value()).match(
IRB.CreateShl(L, R, "", /* NUW */ true, /* NSW */ false)));
EXPECT_FALSE(m_NSWShl(m_Value(), m_Value()).match(IRB.CreateNSWAdd(L, R)));
EXPECT_FALSE(m_NUWAdd(m_Value(), m_Value()).match(IRB.CreateAdd(L, R)));
EXPECT_FALSE(m_NUWAdd(m_Value(), m_Value()).match(IRB.CreateNSWAdd(L, R)));
EXPECT_FALSE(m_NUWAdd(m_Value(), m_Value()).match(IRB.CreateNUWSub(L, R)));
EXPECT_FALSE(m_NUWSub(m_Value(), m_Value()).match(IRB.CreateSub(L, R)));
EXPECT_FALSE(m_NUWSub(m_Value(), m_Value()).match(IRB.CreateNSWSub(L, R)));
EXPECT_FALSE(m_NUWSub(m_Value(), m_Value()).match(IRB.CreateNUWAdd(L, R)));
EXPECT_FALSE(m_NUWMul(m_Value(), m_Value()).match(IRB.CreateMul(L, R)));
EXPECT_FALSE(m_NUWMul(m_Value(), m_Value()).match(IRB.CreateNSWMul(L, R)));
EXPECT_FALSE(m_NUWMul(m_Value(), m_Value()).match(IRB.CreateNUWAdd(L, R)));
EXPECT_FALSE(m_NUWShl(m_Value(), m_Value()).match(IRB.CreateShl(L, R)));
EXPECT_FALSE(m_NUWShl(m_Value(), m_Value()).match(
IRB.CreateShl(L, R, "", /* NUW */ false, /* NSW */ true)));
EXPECT_FALSE(m_NUWShl(m_Value(), m_Value()).match(IRB.CreateNUWAdd(L, R)));
}
[PatternMatch] Add m_Store pattern match helper Differential Revision: https://reviews.llvm.org/D48279 llvm-svn: 335100
2018-06-20 15:27:45 +08:00
TEST_F(PatternMatchTest, LoadStoreOps) {
// Create this load/store sequence:
//
// %p = alloca i32*
// %0 = load i32*, i32** %p
// store i32 42, i32* %0
Value *Alloca = IRB.CreateAlloca(IRB.getInt32Ty());
Value *LoadInst = IRB.CreateLoad(Alloca);
Value *FourtyTwo = IRB.getInt32(42);
Value *StoreInst = IRB.CreateStore(FourtyTwo, Alloca);
Value *MatchLoad, *MatchStoreVal, *MatchStorePointer;
EXPECT_TRUE(m_Load(m_Value(MatchLoad)).match(LoadInst));
EXPECT_EQ(Alloca, MatchLoad);
EXPECT_TRUE(m_Load(m_Specific(Alloca)).match(LoadInst));
EXPECT_FALSE(m_Load(m_Value(MatchLoad)).match(Alloca));
EXPECT_TRUE(m_Store(m_Value(MatchStoreVal), m_Value(MatchStorePointer))
.match(StoreInst));
EXPECT_EQ(FourtyTwo, MatchStoreVal);
EXPECT_EQ(Alloca, MatchStorePointer);
EXPECT_FALSE(m_Store(m_Value(MatchStoreVal), m_Value(MatchStorePointer))
.match(Alloca));
EXPECT_TRUE(m_Store(m_SpecificInt(42), m_Specific(Alloca))
.match(StoreInst));
EXPECT_FALSE(m_Store(m_SpecificInt(42), m_Specific(FourtyTwo))
.match(StoreInst));
EXPECT_FALSE(m_Store(m_SpecificInt(43), m_Specific(Alloca))
.match(StoreInst));
}
[PatternMatch] Add matchers for vector operations Summary: There aren't any matchers for the three vector operations: insertelement, extractelement, and shufflevector. This patch adds them as well as corresponding unit tests. llvm-svn: 328709
2018-03-28 23:39:00 +08:00
TEST_F(PatternMatchTest, VectorOps) {
// Build up small tree of vector operations
//
// Val = 0 + 1
// Val2 = Val + 3
// VI1 = insertelement <2 x i8> undef, i8 1, i32 0 = <1, undef>
// VI2 = insertelement <2 x i8> %VI1, i8 %Val2, i8 %Val = <1, 4>
// VI3 = insertelement <2 x i8> %VI1, i8 %Val2, i32 1 = <1, 4>
// VI4 = insertelement <2 x i8> %VI1, i8 2, i8 %Val = <1, 2>
//
// SI1 = shufflevector <2 x i8> %VI1, <2 x i8> undef, zeroinitializer
// SI2 = shufflevector <2 x i8> %VI3, <2 x i8> %VI4, <2 x i8> <i8 0, i8 2>
// SI3 = shufflevector <2 x i8> %VI3, <2 x i8> undef, zeroinitializer
// SI4 = shufflevector <2 x i8> %VI4, <2 x i8> undef, zeroinitializer
//
// SP1 = VectorSplat(2, i8 2)
// SP2 = VectorSplat(2, i8 %Val)
Type *VecTy = VectorType::get(IRB.getInt8Ty(), 2);
Type *i32 = IRB.getInt32Ty();
Type *i32VecTy = VectorType::get(i32, 2);
Value *Val = IRB.CreateAdd(IRB.getInt8(0), IRB.getInt8(1));
Value *Val2 = IRB.CreateAdd(Val, IRB.getInt8(3));
SmallVector<Constant *, 2> VecElemIdxs;
VecElemIdxs.push_back(ConstantInt::get(i32, 0));
VecElemIdxs.push_back(ConstantInt::get(i32, 2));
auto *IdxVec = ConstantVector::get(VecElemIdxs);
Value *UndefVec = UndefValue::get(VecTy);
Value *VI1 = IRB.CreateInsertElement(UndefVec, IRB.getInt8(1), (uint64_t)0);
Value *VI2 = IRB.CreateInsertElement(VI1, Val2, Val);
Value *VI3 = IRB.CreateInsertElement(VI1, Val2, (uint64_t)1);
Value *VI4 = IRB.CreateInsertElement(VI1, IRB.getInt8(2), Val);
Value *EX1 = IRB.CreateExtractElement(VI4, Val);
Value *EX2 = IRB.CreateExtractElement(VI4, (uint64_t)0);
Value *EX3 = IRB.CreateExtractElement(IdxVec, (uint64_t)1);
Value *Zero = ConstantAggregateZero::get(i32VecTy);
Value *SI1 = IRB.CreateShuffleVector(VI1, UndefVec, Zero);
Value *SI2 = IRB.CreateShuffleVector(VI3, VI4, IdxVec);
Value *SI3 = IRB.CreateShuffleVector(VI3, UndefVec, Zero);
Value *SI4 = IRB.CreateShuffleVector(VI4, UndefVec, Zero);
Value *SP1 = IRB.CreateVectorSplat(2, IRB.getInt8(2));
Value *SP2 = IRB.CreateVectorSplat(2, Val);
Value *A = nullptr, *B = nullptr, *C = nullptr;
// Test matching insertelement
EXPECT_TRUE(match(VI1, m_InsertElement(m_Value(), m_Value(), m_Value())));
EXPECT_TRUE(
match(VI1, m_InsertElement(m_Undef(), m_ConstantInt(), m_ConstantInt())));
EXPECT_TRUE(
match(VI1, m_InsertElement(m_Undef(), m_ConstantInt(), m_Zero())));
EXPECT_TRUE(
match(VI1, m_InsertElement(m_Undef(), m_SpecificInt(1), m_Zero())));
EXPECT_TRUE(match(VI2, m_InsertElement(m_Value(), m_Value(), m_Value())));
EXPECT_FALSE(
match(VI2, m_InsertElement(m_Value(), m_Value(), m_ConstantInt())));
EXPECT_FALSE(
match(VI2, m_InsertElement(m_Value(), m_ConstantInt(), m_Value())));
EXPECT_FALSE(match(VI2, m_InsertElement(m_Constant(), m_Value(), m_Value())));
EXPECT_TRUE(match(VI3, m_InsertElement(m_Value(A), m_Value(B), m_Value(C))));
EXPECT_TRUE(A == VI1);
EXPECT_TRUE(B == Val2);
EXPECT_TRUE(isa<ConstantInt>(C));
A = B = C = nullptr; // reset
// Test matching extractelement
EXPECT_TRUE(match(EX1, m_ExtractElement(m_Value(A), m_Value(B))));
EXPECT_TRUE(A == VI4);
EXPECT_TRUE(B == Val);
A = B = C = nullptr; // reset
EXPECT_FALSE(match(EX1, m_ExtractElement(m_Value(), m_ConstantInt())));
EXPECT_TRUE(match(EX2, m_ExtractElement(m_Value(), m_ConstantInt())));
EXPECT_TRUE(match(EX3, m_ExtractElement(m_Constant(), m_ConstantInt())));
// Test matching shufflevector
EXPECT_TRUE(match(SI1, m_ShuffleVector(m_Value(), m_Undef(), m_Zero())));
EXPECT_TRUE(match(SI2, m_ShuffleVector(m_Value(A), m_Value(B), m_Value(C))));
EXPECT_TRUE(A == VI3);
EXPECT_TRUE(B == VI4);
EXPECT_TRUE(C == IdxVec);
A = B = C = nullptr; // reset
// Test matching the vector splat pattern
EXPECT_TRUE(match(
SI1,
m_ShuffleVector(m_InsertElement(m_Undef(), m_SpecificInt(1), m_Zero()),
m_Undef(), m_Zero())));
EXPECT_FALSE(match(
SI3, m_ShuffleVector(m_InsertElement(m_Undef(), m_Value(), m_Zero()),
m_Undef(), m_Zero())));
EXPECT_FALSE(match(
SI4, m_ShuffleVector(m_InsertElement(m_Undef(), m_Value(), m_Zero()),
m_Undef(), m_Zero())));
EXPECT_TRUE(match(
SP1,
m_ShuffleVector(m_InsertElement(m_Undef(), m_SpecificInt(2), m_Zero()),
m_Undef(), m_Zero())));
EXPECT_TRUE(match(
SP2, m_ShuffleVector(m_InsertElement(m_Undef(), m_Value(A), m_Zero()),
m_Undef(), m_Zero())));
EXPECT_TRUE(A == Val);
}
Add support to paternmatch for simple const Value cases. Pattern match has some paths which can operate on constant instructions, but not all. This adds a version of m_value() to return const Value* and changes ICmp matching to use auto so that it can match both constant and mutable instructions. Tests also included for both mutable and constant ICmpInst matching. This will be used in a future commit to constify ValueTracking.cpp. llvm-svn: 278570
2016-08-13 06:16:05 +08:00
template <typename T> struct MutableConstTest : PatternMatchTest { };
typedef ::testing::Types<std::tuple<Value*, Instruction*>,
std::tuple<const Value*, const Instruction *>>
MutableConstTestTypes;
TYPED_TEST_CASE(MutableConstTest, MutableConstTestTypes);
TYPED_TEST(MutableConstTest, ICmp) {
auto &IRB = PatternMatchTest::IRB;
typedef typename std::tuple_element<0, TypeParam>::type ValueType;
typedef typename std::tuple_element<1, TypeParam>::type InstructionType;
Value *L = IRB.getInt32(1);
Value *R = IRB.getInt32(2);
ICmpInst::Predicate Pred = ICmpInst::ICMP_UGT;
ValueType MatchL;
ValueType MatchR;
ICmpInst::Predicate MatchPred;
EXPECT_TRUE(m_ICmp(MatchPred, m_Value(MatchL), m_Value(MatchR))
.match((InstructionType)IRB.CreateICmp(Pred, L, R)));
EXPECT_EQ(L, MatchL);
EXPECT_EQ(R, MatchR);
}
PatternMatch: Matcher for (un)ordered floating point min/max Add support for matching 'ordered' and 'unordered' floating point min/max constructs. In LLVM we can express min/max functions as a combination of compare and select. We have support for matching such constructs for integers but not for floating point. In floating point math there is no total order because of the presence of 'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics when interpreting floating point compare select combinations as a minimum or maximum function. The resulting 'ordered/unordered' floating point maximum function has to select the same value as the select/fcmp combination it is based on. ordered_max(x,y) = max(x,y) iff x and y are not NaN, y otherwise unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise ordered_min(x,y) = min(x,y) iff x and y are not NaN, y otherwise unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R) construct. Any code using this predicate has to preserve this semantics. A follow-up patch will use this to implement floating point min/max reductions in the vectorizer. radar://13723044 llvm-svn: 181143
2013-05-05 09:54:46 +08:00
} // anonymous namespace.