[llvm] Constant Fold logf128 calls (PR #90611)
Matthew Devereau via llvm-commits
llvm-commits at lists.llvm.org
Wed May 1 07:02:09 PDT 2024
https://github.com/MDevereau updated https://github.com/llvm/llvm-project/pull/90611
>From 62df07e6fd930fff060a338bd2f12aa21a10e94c Mon Sep 17 00:00:00 2001
From: Matt Devereau <matthew.devereau at arm.com>
Date: Tue, 30 Apr 2024 13:06:57 +0000
Subject: [PATCH] Constant Fold logf128 calls
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.
---
llvm/CMakeLists.txt | 2 +
llvm/cmake/config-ix.cmake | 11 ++
llvm/include/llvm/ADT/APFloat.h | 13 ++
llvm/include/llvm/ADT/APInt.h | 8 ++
llvm/include/llvm/Config/llvm-config.h.cmake | 3 +
llvm/include/llvm/Support/float128.h | 26 ++++
llvm/lib/Analysis/CMakeLists.txt | 6 +
llvm/lib/Analysis/ConstantFolding.cpp | 11 ++
llvm/lib/Support/APFloat.cpp | 24 ++++
llvm/test/CMakeLists.txt | 1 +
.../InstSimplify/ConstProp/logf128.ll | 126 ++++++++++++++++++
llvm/test/lit.cfg.py | 3 +
llvm/test/lit.site.cfg.py.in | 1 +
13 files changed, 235 insertions(+)
create mode 100644 llvm/include/llvm/Support/float128.h
create mode 100644 llvm/test/Transforms/InstSimplify/ConstProp/logf128.ll
diff --git a/llvm/CMakeLists.txt b/llvm/CMakeLists.txt
index c06e661573ed42..b7b158996c2323 100644
--- a/llvm/CMakeLists.txt
+++ b/llvm/CMakeLists.txt
@@ -562,6 +562,8 @@ 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.")
diff --git a/llvm/cmake/config-ix.cmake b/llvm/cmake/config-ix.cmake
index bf1b110245bb2f..ce07cab7a56c6b 100644
--- a/llvm/cmake/config-ix.cmake
+++ b/llvm/cmake/config-ix.cmake
@@ -257,6 +257,17 @@ else()
set(LLVM_ENABLE_TERMINFO 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)
diff --git a/llvm/include/llvm/ADT/APFloat.h b/llvm/include/llvm/ADT/APFloat.h
index deb74cb2fdeb1e..44a301ecc99280 100644
--- a/llvm/include/llvm/ADT/APFloat.h
+++ b/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
diff --git a/llvm/include/llvm/ADT/APInt.h b/llvm/include/llvm/ADT/APInt.h
index 8d3c029b2e7e91..4e543d6e7618b5 100644
--- a/llvm/include/llvm/ADT/APInt.h
+++ b/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>
@@ -1670,6 +1671,13 @@ 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
diff --git a/llvm/include/llvm/Config/llvm-config.h.cmake b/llvm/include/llvm/Config/llvm-config.h.cmake
index 6605ea60df99e1..629977cc11d683 100644
--- a/llvm/include/llvm/Config/llvm-config.h.cmake
+++ b/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
diff --git a/llvm/include/llvm/Support/float128.h b/llvm/include/llvm/Support/float128.h
new file mode 100644
index 00000000000000..e15a98dc5a6779
--- /dev/null
+++ b/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
diff --git a/llvm/lib/Analysis/CMakeLists.txt b/llvm/lib/Analysis/CMakeLists.txt
index 474b8d20fde16f..74476cb5440c61 100644
--- a/llvm/lib/Analysis/CMakeLists.txt
+++ b/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()
diff --git a/llvm/lib/Analysis/ConstantFolding.cpp b/llvm/lib/Analysis/ConstantFolding.cpp
index 749374a3aa48af..bb4b334b96ee6e 100644
--- a/llvm/lib/Analysis/ConstantFolding.cpp
+++ b/llvm/lib/Analysis/ConstantFolding.cpp
@@ -2089,6 +2089,17 @@ 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;
diff --git a/llvm/lib/Support/APFloat.cpp b/llvm/lib/Support/APFloat.cpp
index 0a4f5ac01553f1..52bf3b91c55af9 100644
--- a/llvm/lib/Support/APFloat.cpp
+++ b/llvm/lib/Support/APFloat.cpp
@@ -3670,6 +3670,15 @@ 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")
@@ -5265,6 +5274,21 @@ 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();
diff --git a/llvm/test/CMakeLists.txt b/llvm/test/CMakeLists.txt
index eb401351141641..b55fea144f2fa5 100644
--- a/llvm/test/CMakeLists.txt
+++ b/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(
diff --git a/llvm/test/Transforms/InstSimplify/ConstProp/logf128.ll b/llvm/test/Transforms/InstSimplify/ConstProp/logf128.ll
new file mode 100644
index 00000000000000..da56997f693822
--- /dev/null
+++ b/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
+}
diff --git a/llvm/test/lit.cfg.py b/llvm/test/lit.cfg.py
index affd87b98c1410..fe1262893212fb 100644
--- a/llvm/test/lit.cfg.py
+++ b/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")
diff --git a/llvm/test/lit.site.cfg.py.in b/llvm/test/lit.site.cfg.py.in
index 60a68b0edaf933..0968f6214772d0 100644
--- a/llvm/test/lit.site.cfg.py.in
+++ b/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)
More information about the llvm-commits
mailing list