ExpandLargeFpConvert.cpp | Explore in Territory

//===--- ExpandLargeFpConvert.cpp - Expand large fp convert----------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//

// This pass expands ‘fptoui .. to’, ‘fptosi .. to’, ‘uitofp .. to’,
// ‘sitofp .. to’ instructions with a bitwidth above a threshold into
// auto-generated functions. This is useful for targets like x86_64 that cannot
// lower fp convertions with more than 128 bits.
//
//===----------------------------------------------------------------------===//

#include "llvm/CodeGen/ExpandLargeFpConvert.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/StringExtras.h"
#include "llvm/Analysis/GlobalsModRef.h"
#include "llvm/CodeGen/Passes.h"
#include "llvm/CodeGen/TargetLowering.h"
#include "llvm/CodeGen/TargetPassConfig.h"
#include "llvm/CodeGen/TargetSubtargetInfo.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/InstIterator.h"
#include "llvm/IR/PassManager.h"
#include "llvm/InitializePasses.h"
#include "llvm/Pass.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Target/TargetMachine.h"

usingnamespacellvm;

static cl::opt<unsigned>
    ExpandFpConvertBits("expand-fp-convert-bits", cl::Hidden,
                     cl::init(llvm::IntegerType::MAX_INT_BITS),
                     cl::desc("fp convert instructions on integers with "
                              "more than <N> bits are expanded."));

/// Generate code to convert a fp number to integer, replacing FPToS(U)I with
/// the generated code. This currently generates code similarly to compiler-rt's
/// implementations.
///
/// An example IR generated from compiler-rt/fixsfdi.c looks like below:
/// define dso_local i64 @foo(float noundef %a) local_unnamed_addr #0 {
/// entry:
///   %0 = bitcast float %a to i32
///   %conv.i = zext i32 %0 to i64
///   %tobool.not = icmp sgt i32 %0, -1
///   %conv = select i1 %tobool.not, i64 1, i64 -1
///   %and = lshr i64 %conv.i, 23
///   %shr = and i64 %and, 255
///   %and2 = and i64 %conv.i, 8388607
///   %or = or i64 %and2, 8388608
///   %cmp = icmp ult i64 %shr, 127
///   br i1 %cmp, label %cleanup, label %if.end
///
/// if.end:                                           ; preds = %entry
///   %sub = add nuw nsw i64 %shr, 4294967169
///   %conv5 = and i64 %sub, 4294967232
///   %cmp6.not = icmp eq i64 %conv5, 0
///   br i1 %cmp6.not, label %if.end12, label %if.then8
///
/// if.then8:                                         ; preds = %if.end
///   %cond11 = select i1 %tobool.not, i64 9223372036854775807, i64 -9223372036854775808
///   br label %cleanup
///
/// if.end12:                                         ; preds = %if.end
///   %cmp13 = icmp ult i64 %shr, 150
///   br i1 %cmp13, label %if.then15, label %if.else
///
/// if.then15:                                        ; preds = %if.end12
///   %sub16 = sub nuw nsw i64 150, %shr
///   %shr17 = lshr i64 %or, %sub16
///   %mul = mul nsw i64 %shr17, %conv
///   br label %cleanup
///
/// if.else:                                          ; preds = %if.end12
///   %sub18 = add nsw i64 %shr, -150
///   %shl = shl i64 %or, %sub18
///   %mul19 = mul nsw i64 %shl, %conv
///   br label %cleanup
///
/// cleanup:                                          ; preds = %entry, %if.else, %if.then15, %if.then8
///   %retval.0 = phi i64 [ %cond11, %if.then8 ], [ %mul, %if.then15 ], [ %mul19, %if.else ], [ 0, %entry ]
///   ret i64 %retval.0
/// }
///
/// Replace fp to integer with generated code.
static void expandFPToI(Instruction *FPToI) { … }

/// Generate code to convert a fp number to integer, replacing S(U)IToFP with
/// the generated code. This currently generates code similarly to compiler-rt's
/// implementations. This implementation has an implicit assumption that integer
/// width is larger than fp.
///
/// An example IR generated from compiler-rt/floatdisf.c looks like below:
/// define dso_local float @__floatdisf(i64 noundef %a) local_unnamed_addr #0 {
/// entry:
///   %cmp = icmp eq i64 %a, 0
///   br i1 %cmp, label %return, label %if.end
///
/// if.end:                                           ; preds = %entry
///   %shr = ashr i64 %a, 63
///   %xor = xor i64 %shr, %a
///   %sub = sub nsw i64 %xor, %shr
///   %0 = tail call i64 @llvm.ctlz.i64(i64 %sub, i1 true), !range !5
///   %cast = trunc i64 %0 to i32
///   %sub1 = sub nuw nsw i32 64, %cast
///   %sub2 = xor i32 %cast, 63
///   %cmp3 = icmp ult i32 %cast, 40
///   br i1 %cmp3, label %if.then4, label %if.else
///
/// if.then4:                                         ; preds = %if.end
///   switch i32 %sub1, label %sw.default [
///     i32 25, label %sw.bb
///     i32 26, label %sw.epilog
///   ]
///
/// sw.bb:                                            ; preds = %if.then4
///   %shl = shl i64 %sub, 1
///   br label %sw.epilog
///
/// sw.default:                                       ; preds = %if.then4
///   %sub5 = sub nsw i64 38, %0
///   %sh_prom = and i64 %sub5, 4294967295
///   %shr6 = lshr i64 %sub, %sh_prom
///   %shr9 = lshr i64 274877906943, %0
///   %and = and i64 %shr9, %sub
///   %cmp10 = icmp ne i64 %and, 0
///   %conv11 = zext i1 %cmp10 to i64
///   %or = or i64 %shr6, %conv11
///   br label %sw.epilog
///
/// sw.epilog:                                        ; preds = %sw.default, %if.then4, %sw.bb
///   %a.addr.0 = phi i64 [ %or, %sw.default ], [ %sub, %if.then4 ], [ %shl, %sw.bb ]
///   %1 = lshr i64 %a.addr.0, 2
///   %2 = and i64 %1, 1
///   %or16 = or i64 %2, %a.addr.0
///   %inc = add nsw i64 %or16, 1
///   %3 = and i64 %inc, 67108864
///   %tobool.not = icmp eq i64 %3, 0
///   %spec.select.v = select i1 %tobool.not, i64 2, i64 3
///   %spec.select = ashr i64 %inc, %spec.select.v
///   %spec.select56 = select i1 %tobool.not, i32 %sub2, i32 %sub1
///   br label %if.end26
///
/// if.else:                                          ; preds = %if.end
///   %sub23 = add nuw nsw i64 %0, 4294967256
///   %sh_prom24 = and i64 %sub23, 4294967295
///   %shl25 = shl i64 %sub, %sh_prom24
///   br label %if.end26
///
/// if.end26:                                         ; preds = %sw.epilog, %if.else
///   %a.addr.1 = phi i64 [ %shl25, %if.else ], [ %spec.select, %sw.epilog ]
///   %e.0 = phi i32 [ %sub2, %if.else ], [ %spec.select56, %sw.epilog ]
///   %conv27 = trunc i64 %shr to i32
///   %and28 = and i32 %conv27, -2147483648
///   %add = shl nuw nsw i32 %e.0, 23
///   %shl29 = add nuw nsw i32 %add, 1065353216
///   %conv31 = trunc i64 %a.addr.1 to i32
///   %and32 = and i32 %conv31, 8388607
///   %or30 = or i32 %and32, %and28
///   %or33 = or i32 %or30, %shl29
///   %4 = bitcast i32 %or33 to float
///   br label %return
///
/// return:                                           ; preds = %entry, %if.end26
///   %retval.0 = phi float [ %4, %if.end26 ], [ 0.000000e+00, %entry ]
///   ret float %retval.0
/// }
///
/// Replace integer to fp with generated code.
static void expandIToFP(Instruction *IToFP) { … }

static void scalarize(Instruction *I, SmallVectorImpl<Instruction *> &Replace) { … }

static bool runImpl(Function &F, const TargetLowering &TLI) { … }

namespace {
class ExpandLargeFpConvertLegacyPass : public FunctionPass { … };
} // namespace

PreservedAnalyses ExpandLargeFpConvertPass::run(Function &F,
                                                FunctionAnalysisManager &FAM) { … }

char ExpandLargeFpConvertLegacyPass::ID = …;
INITIALIZE_PASS_BEGIN(ExpandLargeFpConvertLegacyPass, "expand-large-fp-convert",
                      "Expand large fp convert", false, false)
INITIALIZE_PASS_END(ExpandLargeFpConvertLegacyPass, "expand-large-fp-convert",
                    "Expand large fp convert", false, false)

FunctionPass *llvm::createExpandLargeFpConvertPass() { … }
llvm/llvm/lib/CodeGen/ExpandLargeFpConvert.cpp
