[llvm] 14c7e4a - Enable logf128 constant folding for hosts with 128bit long double (#104929)
via llvm-commits
llvm-commits at lists.llvm.org
Thu Aug 22 02:13:03 PDT 2024
Author: Matthew Devereau
Date: 2024-08-22T10:12:59+01:00
New Revision: 14c7e4a1844904f3db9b2dc93b722925a8c66b27
URL: https://github.com/llvm/llvm-project/commit/14c7e4a1844904f3db9b2dc93b722925a8c66b27
DIFF: https://github.com/llvm/llvm-project/commit/14c7e4a1844904f3db9b2dc93b722925a8c66b27.diff
LOG: Enable logf128 constant folding for hosts with 128bit long double (#104929)
This is a reland of (#96287). This patch attempts to reduce the reverted
patch's clang compile time by removing #includes of float128.h and
inlining convertToQuad functions instead.
Added:
Modified:
llvm/CMakeLists.txt
llvm/cmake/config-ix.cmake
llvm/include/llvm/ADT/APFloat.h
llvm/include/llvm/ADT/APInt.h
llvm/include/llvm/Support/float128.h
llvm/lib/Analysis/CMakeLists.txt
llvm/lib/Analysis/ConstantFolding.cpp
llvm/lib/Support/APFloat.cpp
Removed:
################################################################################
diff --git a/llvm/CMakeLists.txt b/llvm/CMakeLists.txt
index d681b1ccab6299..b03d89a43c34b0 100644
--- a/llvm/CMakeLists.txt
+++ b/llvm/CMakeLists.txt
@@ -560,8 +560,6 @@ set(LLVM_USE_STATIC_ZSTD FALSE CACHE BOOL "Use static version of zstd. Can be TR
set(LLVM_ENABLE_CURL "OFF" CACHE STRING "Use libcurl for the HTTP client if available. Can be ON, OFF, or FORCE_ON")
-set(LLVM_HAS_LOGF128 "OFF" CACHE STRING "Use logf128 to constant fold fp128 logarithm calls. Can be ON, OFF, or FORCE_ON")
-
set(LLVM_ENABLE_HTTPLIB "OFF" CACHE STRING "Use cpp-httplib HTTP server library if available. Can be ON, OFF, or FORCE_ON")
set(LLVM_Z3_INSTALL_DIR "" CACHE STRING "Install directory of the Z3 solver.")
diff --git a/llvm/cmake/config-ix.cmake b/llvm/cmake/config-ix.cmake
index f76eacb9d51366..471dd1615c2e7b 100644
--- a/llvm/cmake/config-ix.cmake
+++ b/llvm/cmake/config-ix.cmake
@@ -246,17 +246,6 @@ else()
set(HAVE_LIBEDIT 0)
endif()
-if(LLVM_HAS_LOGF128)
- include(CheckCXXSymbolExists)
- check_cxx_symbol_exists(logf128 math.h HAS_LOGF128)
-
- if(LLVM_HAS_LOGF128 STREQUAL FORCE_ON AND NOT HAS_LOGF128)
- message(FATAL_ERROR "Failed to configure logf128")
- endif()
-
- set(LLVM_HAS_LOGF128 "${HAS_LOGF128}")
-endif()
-
# function checks
check_symbol_exists(arc4random "stdlib.h" HAVE_DECL_ARC4RANDOM)
find_package(Backtrace)
@@ -270,6 +259,13 @@ if(C_SUPPORTS_WERROR_UNGUARDED_AVAILABILITY_NEW)
set(CMAKE_REQUIRED_FLAGS "${CMAKE_REQUIRED_FLAGS} -Werror=unguarded-availability-new")
endif()
+check_cxx_symbol_exists(logf128 cmath HAS_LOGF128)
+check_symbol_exists(__powerpc64le__ "" __PPC64LE)
+if(HAS_LOGF128 AND NOT __PPC64LE)
+ set(LLVM_HAS_LOGF128 On)
+ add_compile_definitions(HAS_LOGF128)
+endif()
+
# Determine whether we can register EH tables.
check_symbol_exists(__register_frame "${CMAKE_CURRENT_LIST_DIR}/unwind.h" HAVE_REGISTER_FRAME)
check_symbol_exists(__deregister_frame "${CMAKE_CURRENT_LIST_DIR}/unwind.h" HAVE_DEREGISTER_FRAME)
diff --git a/llvm/include/llvm/ADT/APFloat.h b/llvm/include/llvm/ADT/APFloat.h
index 7039e961bff82d..925d03d4c06670 100644
--- a/llvm/include/llvm/ADT/APFloat.h
+++ b/llvm/include/llvm/ADT/APFloat.h
@@ -19,7 +19,6 @@
#include "llvm/ADT/ArrayRef.h"
#include "llvm/ADT/FloatingPointMode.h"
#include "llvm/Support/ErrorHandling.h"
-#include "llvm/Support/float128.h"
#include <memory>
#define APFLOAT_DISPATCH_ON_SEMANTICS(METHOD_CALL) \
@@ -378,9 +377,6 @@ class IEEEFloat final : public APFloatBase {
Expected<opStatus> convertFromString(StringRef, roundingMode);
APInt bitcastToAPInt() const;
double convertToDouble() const;
-#ifdef HAS_IEE754_FLOAT128
- float128 convertToQuad() const;
-#endif
float convertToFloat() const;
/// @}
@@ -1274,14 +1270,9 @@ class APFloat : public APFloatBase {
/// shorter semantics, like IEEEsingle and others.
double convertToDouble() const;
- /// Converts this APFloat to host float value.
- ///
- /// \pre The APFloat must be built using semantics, that can be represented by
- /// the host float type without loss of precision. It can be IEEEquad and
- /// shorter semantics, like IEEEdouble and others.
-#ifdef HAS_IEE754_FLOAT128
- float128 convertToQuad() const;
-#endif
+ /// Return true if this APFloat has quadruple precision floating point
+ /// semantics
+ bool isValidIEEEQuad() const;
/// Converts this APFloat to host float value.
///
diff --git a/llvm/include/llvm/ADT/APInt.h b/llvm/include/llvm/ADT/APInt.h
index 65ba3f15305c78..13837413ae49fe 100644
--- a/llvm/include/llvm/ADT/APInt.h
+++ b/llvm/include/llvm/ADT/APInt.h
@@ -17,7 +17,6 @@
#include "llvm/Support/Compiler.h"
#include "llvm/Support/MathExtras.h"
-#include "llvm/Support/float128.h"
#include <cassert>
#include <climits>
#include <cstring>
@@ -1677,13 +1676,6 @@ class [[nodiscard]] APInt {
/// any bit width. Exactly 64 bits will be translated.
double bitsToDouble() const { return llvm::bit_cast<double>(getWord(0)); }
-#ifdef HAS_IEE754_FLOAT128
- float128 bitsToQuad() const {
- __uint128_t ul = ((__uint128_t)U.pVal[1] << 64) + U.pVal[0];
- return llvm::bit_cast<float128>(ul);
- }
-#endif
-
/// Converts APInt bits to a float
///
/// The conversion does not do a translation from integer to float, it just
diff --git a/llvm/include/llvm/Support/float128.h b/llvm/include/llvm/Support/float128.h
index e15a98dc5a6779..618b320086ba59 100644
--- a/llvm/include/llvm/Support/float128.h
+++ b/llvm/include/llvm/Support/float128.h
@@ -9,18 +9,16 @@
#ifndef LLVM_FLOAT128
#define LLVM_FLOAT128
+#include <cmath>
+
namespace llvm {
-#if defined(__clang__) && defined(__FLOAT128__) && \
- defined(__SIZEOF_INT128__) && !defined(__LONG_DOUBLE_IBM128__)
-#define HAS_IEE754_FLOAT128
-typedef __float128 float128;
-#elif defined(__FLOAT128__) && defined(__SIZEOF_INT128__) && \
- !defined(__LONG_DOUBLE_IBM128__) && \
- (defined(__GNUC__) || defined(__GNUG__))
+#ifdef HAS_LOGF128
+#if !defined(__LONG_DOUBLE_IBM128__) && (__SIZEOF_INT128__ == 16)
+typedef decltype(logf128(0.)) float128;
#define HAS_IEE754_FLOAT128
-typedef _Float128 float128;
#endif
+#endif // HAS_LOGF128
} // namespace llvm
#endif // LLVM_FLOAT128
diff --git a/llvm/lib/Analysis/CMakeLists.txt b/llvm/lib/Analysis/CMakeLists.txt
index 393803fad89383..3127f45cc54cb1 100644
--- a/llvm/lib/Analysis/CMakeLists.txt
+++ b/llvm/lib/Analysis/CMakeLists.txt
@@ -162,9 +162,3 @@ add_llvm_component_library(LLVMAnalysis
Support
TargetParser
)
-
-include(CheckCXXSymbolExists)
-check_cxx_symbol_exists(logf128 math.h HAS_LOGF128)
-if(HAS_LOGF128)
- target_compile_definitions(LLVMAnalysis PRIVATE HAS_LOGF128)
-endif()
diff --git a/llvm/lib/Analysis/ConstantFolding.cpp b/llvm/lib/Analysis/ConstantFolding.cpp
index defcacdfa8b105..81c4d4ec5be412 100644
--- a/llvm/lib/Analysis/ConstantFolding.cpp
+++ b/llvm/lib/Analysis/ConstantFolding.cpp
@@ -54,6 +54,7 @@
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/KnownBits.h"
#include "llvm/Support/MathExtras.h"
+#include "llvm/Support/float128.h"
#include <cassert>
#include <cerrno>
#include <cfenv>
@@ -1741,7 +1742,7 @@ Constant *GetConstantFoldFPValue(double V, Type *Ty) {
llvm_unreachable("Can only constant fold half/float/double");
}
-#if defined(HAS_IEE754_FLOAT128) && defined(HAS_LOGF128)
+#if defined(HAS_IEE754_FLOAT128)
Constant *GetConstantFoldFPValue128(float128 V, Type *Ty) {
if (Ty->isFP128Ty())
return ConstantFP::get(Ty, V);
@@ -1781,11 +1782,25 @@ Constant *ConstantFoldFP(double (*NativeFP)(double), const APFloat &V,
return GetConstantFoldFPValue(Result, Ty);
}
-#if defined(HAS_IEE754_FLOAT128) && defined(HAS_LOGF128)
+#if defined(HAS_IEE754_FLOAT128)
+float128 ConvertToQuad(const APFloat &Apf) {
+ APInt Api = Apf.bitcastToAPInt();
+ __uint128_t Uint128 =
+ ((__uint128_t)Api.extractBitsAsZExtValue(64, 64) << 64) +
+ Api.extractBitsAsZExtValue(64, 0);
+ return llvm::bit_cast<float128>(Uint128);
+}
+#endif
+
+#if defined(HAS_IEE754_FLOAT128)
Constant *ConstantFoldFP128(float128 (*NativeFP)(float128), const APFloat &V,
Type *Ty) {
llvm_fenv_clearexcept();
- float128 Result = NativeFP(V.convertToQuad());
+ if (!V.isValidIEEEQuad())
+ return nullptr;
+
+ float128 Result = NativeFP(ConvertToQuad(V));
+
if (llvm_fenv_testexcept()) {
llvm_fenv_clearexcept();
return nullptr;
@@ -2114,13 +2129,16 @@ static Constant *ConstantFoldScalarCall1(StringRef Name,
if (IntrinsicID == Intrinsic::canonicalize)
return constantFoldCanonicalize(Ty, Call, U);
-#if defined(HAS_IEE754_FLOAT128) && defined(HAS_LOGF128)
+#if defined(HAS_IEE754_FLOAT128)
if (Ty->isFP128Ty()) {
if (IntrinsicID == Intrinsic::log) {
- float128 Result = logf128(Op->getValueAPF().convertToQuad());
+ APFloat Value = Op->getValueAPF();
+ if (!Value.isValidIEEEQuad())
+ return nullptr;
+
+ float128 Result = logf128(ConvertToQuad(Value));
return GetConstantFoldFPValue128(Result, Ty);
}
-
LibFunc Fp128Func = NotLibFunc;
if (TLI->getLibFunc(Name, Fp128Func) && TLI->has(Fp128Func) &&
Fp128Func == LibFunc_logl)
diff --git a/llvm/lib/Support/APFloat.cpp b/llvm/lib/Support/APFloat.cpp
index 7f68c5ab9b7cf7..2ddf99f56f88d5 100644
--- a/llvm/lib/Support/APFloat.cpp
+++ b/llvm/lib/Support/APFloat.cpp
@@ -3749,15 +3749,6 @@ double IEEEFloat::convertToDouble() const {
return api.bitsToDouble();
}
-#ifdef HAS_IEE754_FLOAT128
-float128 IEEEFloat::convertToQuad() const {
- assert(semantics == (const llvm::fltSemantics *)&semIEEEquad &&
- "Float semantics are not IEEEquads");
- APInt api = bitcastToAPInt();
- return api.bitsToQuad();
-}
-#endif
-
/// Integer bit is explicit in this format. Intel hardware (387 and later)
/// does not support these bit patterns:
/// exponent = all 1's, integer bit 0, significand 0 ("pseudoinfinity")
@@ -5406,20 +5397,9 @@ double APFloat::convertToDouble() const {
return Temp.getIEEE().convertToDouble();
}
-#ifdef HAS_IEE754_FLOAT128
-float128 APFloat::convertToQuad() const {
- if (&getSemantics() == (const llvm::fltSemantics *)&semIEEEquad)
- return getIEEE().convertToQuad();
- assert(getSemantics().isRepresentableBy(semIEEEquad) &&
- "Float semantics is not representable by IEEEquad");
- APFloat Temp = *this;
- bool LosesInfo;
- opStatus St = Temp.convert(semIEEEquad, rmNearestTiesToEven, &LosesInfo);
- assert(!(St & opInexact) && !LosesInfo && "Unexpected imprecision");
- (void)St;
- return Temp.getIEEE().convertToQuad();
+bool APFloat::isValidIEEEQuad() const {
+ return (&getSemantics() == (const llvm::fltSemantics *)&semIEEEquad);
}
-#endif
float APFloat::convertToFloat() const {
if (&getSemantics() == (const llvm::fltSemantics *)&semIEEEsingle)
More information about the llvm-commits
mailing list