Constant Fold logf128 calls (#90611)

This is a second attempt to land #84501 which failed on several targets.

This patch adds the HAS_IEE754_FLOAT128 define which makes the check for
typedef'ing float128 more precise by checking whether __uint128_t is
available and checking if the host does not use __ibm128 which is
prevalent on power pc targets and replaces IEEE754 float128s.

Author: MDevereau

llvm/CMakeLists.txt

@@ -560,6 +560,8 @@ 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.")

llvm/cmake/config-ix.cmake

@@ -247,6 +247,17 @@ 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)

llvm/include/llvm/ADT/APFloat.h

@@ -19,6 +19,7 @@
 #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)                             \
@@ -354,6 +355,9 @@ 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;
 
   /// @}
@@ -1218,6 +1222,15 @@ 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
+
   /// Converts this APFloat to host float value.
   ///
   /// \pre The APFloat must be built using semantics, that can be represented by

llvm/include/llvm/ADT/APInt.h

@@ -17,6 +17,7 @@
 
 #include "llvm/Support/Compiler.h"
 #include "llvm/Support/MathExtras.h"
+#include "llvm/Support/float128.h"
 #include <cassert>
 #include <climits>
 #include <cstring>
@@ -1677,6 +1678,13 @@ 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

llvm/include/llvm/Config/llvm-config.h.cmake

@@ -198,4 +198,7 @@
 /* Define if plugins enabled */
 #cmakedefine LLVM_ENABLE_PLUGINS
 
+/* Define if logf128 is available */
+#cmakedefine LLVM_HAS_LOGF128
+
 #endif

llvm/include/llvm/Support/float128.h

@@ -0,0 +1,26 @@
+//===-- llvm/Support/float128.h - Compiler abstraction support --*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_FLOAT128
+#define LLVM_FLOAT128
+
+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__))
+#define HAS_IEE754_FLOAT128
+typedef _Float128 float128;
+#endif
+
+} // namespace llvm
+#endif // LLVM_FLOAT128

llvm/lib/Analysis/CMakeLists.txt

@@ -159,3 +159,9 @@ 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()

llvm/lib/Analysis/ConstantFolding.cpp

@@ -2087,6 +2087,17 @@ static Constant *ConstantFoldScalarCall1(StringRef Name,
     if (IntrinsicID == Intrinsic::canonicalize)
       return constantFoldCanonicalize(Ty, Call, U);
 
+#if defined(HAS_IEE754_FLOAT128) && defined(HAS_LOGF128)
+    if (Ty->isFP128Ty()) {
+      switch (IntrinsicID) {
+      default:
+        return nullptr;
+      case Intrinsic::log:
+        return ConstantFP::get(Ty, logf128(Op->getValueAPF().convertToQuad()));
+      }
+    }
+#endif
+
     if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
       return nullptr;
 

llvm/lib/Support/APFloat.cpp

@@ -3665,6 +3665,15 @@ 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")
@@ -5260,6 +5269,21 @@ 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();
+}
+#endif
+
 float APFloat::convertToFloat() const {
   if (&getSemantics() == (const llvm::fltSemantics *)&semIEEEsingle)
     return getIEEE().convertToFloat();

llvm/test/CMakeLists.txt

@@ -26,6 +26,7 @@ llvm_canonicalize_cmake_booleans(
   LLVM_TOOL_LLVM_DRIVER_BUILD
   LLVM_INCLUDE_SPIRV_TOOLS_TESTS
   LLVM_APPEND_VC_REV
+  LLVM_HAS_LOGF128
   )
 
 configure_lit_site_cfg(

llvm/test/Transforms/InstSimplify/ConstProp/logf128.ll

@@ -0,0 +1,126 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
+; RUN: opt < %s -passes=instsimplify -S | FileCheck %s
+
+; REQUIRES: has_logf128
+declare fp128 @llvm.log.f128(fp128)
+
+define fp128 @log_e_64(){
+; CHECK-LABEL: define fp128 @log_e_64() {
+; CHECK-NEXT:    ret fp128 0xL300000000000000040010A2B23F3BAB7
+;
+  %A = call fp128 @llvm.log.f128(fp128 noundef 0xL00000000000000004005000000000000)
+  ret fp128 %A
+}
+
+define fp128 @log_e_smallest_positive_subnormal_number(){
+; CHECK-LABEL: define fp128 @log_e_smallest_positive_subnormal_number() {
+; CHECK-NEXT:    ret fp128 0xL3000000000000000C00C654628220780
+;
+  %A = call fp128 @llvm.log.f128(fp128 noundef 0xL00000000000000010000000000000000)
+  ret fp128 %A
+}
+
+define fp128 @log_e_largest_subnormal_number(){
+; CHECK-LABEL: define fp128 @log_e_largest_subnormal_number() {
+; CHECK-NEXT:    ret fp128 0xLD000000000000000C00C62D918CE2421
+;
+  %A = call fp128 @llvm.log.f128(fp128 noundef 0xLFFFFFFFFFFFFFFFF0000FFFFFFFFFFFF)
+  ret fp128 %A
+}
+
+define fp128 @log_e_smallest_positive_normal_number(){
+;
+; CHECK-LABEL: define fp128 @log_e_smallest_positive_normal_number() {
+; CHECK-NEXT:    ret fp128 0xLD000000000000000C00C62D918CE2421
+;
+  %A = call fp128 @llvm.log.f128(fp128 noundef 0xL00000000000000000001000000000000)
+  ret fp128 %A
+}
+
+define fp128 @log_e_largest_normal_number(){
+; CHECK-LABEL: define fp128 @log_e_largest_normal_number() {
+; CHECK-NEXT:    ret fp128 0xLF000000000000000400C62E42FEFA39E
+;
+  %A = call fp128 @llvm.log.f128(fp128 noundef 0xLFFFFFFFFFFFFFFFF7FFEFFFFFFFFFFFF)
+  ret fp128 %A
+}
+
+define fp128 @log_e_largest_number_less_than_one(){
+; CHECK-LABEL: define fp128 @log_e_largest_number_less_than_one() {
+; CHECK-NEXT:    ret fp128 0xL0000000000000000BF8E000000000000
+;
+  %A = call fp128 @llvm.log.f128(fp128 noundef 0xLFFFFFFFFFFFFFFFF3FFEFFFFFFFFFFFF)
+  ret fp128 %A
+}
+
+define fp128 @log_e_1(){
+; CHECK-LABEL: define fp128 @log_e_1() {
+; CHECK-NEXT:    ret fp128 0xL00000000000000000000000000000000
+;
+  %A = call fp128 @llvm.log.f128(fp128 noundef 0xL00000000000000003FFF000000000000)
+  ret fp128 %A
+}
+
+define fp128 @log_e_smallest_number_larger_than_one(){
+; CHECK-LABEL: define fp128 @log_e_smallest_number_larger_than_one() {
+; CHECK-NEXT:    ret fp128 0xL00000000000000003F8F000000000000
+;
+  %A = call fp128 @llvm.log.f128(fp128 noundef 0xL00000000000000013FFF000000000000)
+  ret fp128 %A
+}
+
+define fp128 @log_e_negative_2(){
+; CHECK-LABEL: define fp128 @log_e_negative_2() {
+; CHECK-NEXT:    ret fp128 0xL00000000000000007FFF800000000000
+;
+  %A = call fp128 @llvm.log.f128(fp128 noundef 0xL0000000000000000C000000000000000)
+  ret fp128 %A
+}
+
+define fp128 @log_e_0(){
+; CHECK-LABEL: define fp128 @log_e_0() {
+; CHECK-NEXT:    ret fp128 0xL0000000000000000FFFF000000000000
+;
+  %A = call fp128 @llvm.log.f128(fp128 noundef 0xL00000000000000000000000000000000)
+  ret fp128 %A
+}
+
+define fp128 @log_e_negative_0(){
+; CHECK-LABEL: define fp128 @log_e_negative_0() {
+; CHECK-NEXT:    ret fp128 0xL0000000000000000FFFF000000000000
+;
+  %A = call fp128 @llvm.log.f128(fp128 noundef 0xL00000000000000008000000000000000)
+  ret fp128 %A
+}
+
+define fp128 @log_e_infinity(){
+; CHECK-LABEL: define fp128 @log_e_infinity() {
+; CHECK-NEXT:    ret fp128 0xL00000000000000007FFF000000000000
+;
+  %A = call fp128 @llvm.log.f128(fp128 noundef 0xL00000000000000007FFF000000000000)
+  ret fp128 %A
+}
+
+define fp128 @log_e_negative_infinity(){
+; CHECK-LABEL: define fp128 @log_e_negative_infinity() {
+; CHECK-NEXT:    ret fp128 0xL00000000000000007FFF800000000000
+;
+  %A = call fp128 @llvm.log.f128(fp128 noundef 0xL0000000000000000FFFF000000000000)
+  ret fp128 %A
+}
+
+define fp128 @log_e_nan(){
+; CHECK-LABEL: define fp128 @log_e_nan() {
+; CHECK-NEXT:    ret fp128 0xL00000000000000007FFF800000000001
+;
+  %A = call fp128 @llvm.log.f128(fp128 noundef 0xL00000000000000007FFF000000000001)
+  ret fp128 %A
+}
+
+define <2 x fp128> @log_e_negative_2_vector(){
+; CHECK-LABEL: define <2 x fp128> @log_e_negative_2_vector() {
+; CHECK-NEXT:    ret <2 x fp128> <fp128 0xL00000000000000007FFF800000000000, fp128 0xL00000000000000007FFF800000000000>
+;
+  %A = call <2 x fp128> @llvm.log.v2f128(<2 x fp128> <fp128 0xL0000000000000000C000000000000000, fp128 0xL0000000000000000C000000000000001>)
+  ret <2 x fp128> %A
+}

llvm/test/lit.cfg.py

@@ -617,3 +617,6 @@ def have_ld64_plugin_support():
 # "OBJECT_MODE" to 'any' by default on AIX OS.
 if "system-aix" in config.available_features:
     config.environment["OBJECT_MODE"] = "any"
+
+if config.has_logf128:
+    config.available_features.add("has_logf128")

llvm/test/lit.site.cfg.py.in

@@ -63,6 +63,7 @@ config.have_llvm_driver = @LLVM_TOOL_LLVM_DRIVER_BUILD@
 config.spirv_tools_tests = @LLVM_INCLUDE_SPIRV_TOOLS_TESTS@
 config.have_vc_rev = @LLVM_APPEND_VC_REV@
 config.force_vc_rev = "@LLVM_FORCE_VC_REVISION@"
+config.has_logf128 = @LLVM_HAS_LOGF128@
 
 import lit.llvm
 lit.llvm.initialize(lit_config, config)