Commit 8e069d62 authored by Jakob Kummerow's avatar Jakob Kummerow Committed by V8 LUCI CQ

[wasm][simd] Fix SpillAdjacentFpRegisters...

...to honor the {pinned} list under all circumstances.

Drive-by: DEBUG-mode helpers to print FunctionSig and LiftoffRegList
objects to stdout.

Fixed: chromium:1356718
Change-Id: I487db12294f687790cec1d658d7a7d754f3c2f99
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3859752Reviewed-by: 's avatarClemens Backes <clemensb@chromium.org>
Auto-Submit: Jakob Kummerow <jkummerow@chromium.org>
Commit-Queue: Jakob Kummerow <jkummerow@chromium.org>
Cr-Commit-Position: refs/heads/main@{#82815}
parent f1e800f0
......@@ -1274,6 +1274,14 @@ void LiftoffAssembler::MoveToReturnLocations(
#endif
}
#if DEBUG
void LiftoffRegList::Print() const {
std::ostringstream os;
os << *this << "\n";
PrintF("%s", os.str().c_str());
}
#endif
#ifdef ENABLE_SLOW_DCHECKS
bool LiftoffAssembler::ValidateCacheState() const {
uint32_t register_use_count[kAfterMaxLiftoffRegCode] = {0};
......@@ -1337,6 +1345,10 @@ LiftoffRegister LiftoffAssembler::SpillAdjacentFpRegisters(
if (last_fp.fp().code() % 2 == 0) {
pinned.set(last_fp);
}
// If half of an adjacent pair is pinned, consider the whole pair pinned.
// Otherwise the code below would potentially spill the pinned register
// (after first spilling the unpinned half of the pair).
pinned = pinned.SpreadSetBitsToAdjacentFpRegs();
// We can try to optimize the spilling here:
// 1. Try to get a free fp register, either:
......
......@@ -344,6 +344,8 @@ class LiftoffRegList {
// Sets all even numbered fp registers.
static constexpr uint64_t kEvenFpSetMask = uint64_t{0x5555555555555555}
<< kAfterMaxLiftoffGpRegCode;
static constexpr uint64_t kOddFpSetMask = uint64_t{0xAAAAAAAAAAAAAAAA}
<< kAfterMaxLiftoffGpRegCode;
constexpr LiftoffRegList() = default;
......@@ -426,6 +428,15 @@ class LiftoffRegList {
return !GetAdjacentFpRegsSet().is_empty();
}
// Returns a list where if any part of an adjacent pair of FP regs was set,
// both are set in the result. For example, [1, 4] is turned into [0, 1, 4, 5]
// because (0, 1) and (4, 5) are adjacent pairs.
constexpr LiftoffRegList SpreadSetBitsToAdjacentFpRegs() const {
storage_t odd_regs = regs_ & kOddFpSetMask;
storage_t even_regs = regs_ & kEvenFpSetMask;
return FromBits(regs_ | (odd_regs >> 1) | ((even_regs << 1) & kFpMask));
}
constexpr bool operator==(const LiftoffRegList other) const {
return regs_ == other.regs_;
}
......@@ -461,7 +472,7 @@ class LiftoffRegList {
inline Iterator begin() const;
inline Iterator end() const;
static LiftoffRegList FromBits(storage_t bits) {
static constexpr LiftoffRegList FromBits(storage_t bits) {
DCHECK_EQ(bits, bits & (kGpMask | kFpMask));
return LiftoffRegList(bits);
}
......@@ -472,6 +483,10 @@ class LiftoffRegList {
return LiftoffRegList{bits};
}
#if DEBUG
void Print() const;
#endif
private:
// Unchecked constructor. Only use for valid bits.
explicit constexpr LiftoffRegList(storage_t bits) : regs_(bits) {}
......
......@@ -29,6 +29,21 @@ base::Optional<wasm::ValueKind> WasmReturnTypeFromSignature(
}
}
#if DEBUG
V8_EXPORT_PRIVATE extern void PrintFunctionSig(const wasm::FunctionSig* sig) {
std::ostringstream os;
os << sig->parameter_count() << " parameters:\n";
for (size_t i = 0; i < sig->parameter_count(); i++) {
os << " " << i << ": " << sig->GetParam(i) << "\n";
}
os << sig->return_count() << " returns:\n";
for (size_t i = 0; i < sig->return_count(); i++) {
os << " " << i << ": " << sig->GetReturn() << "\n";
}
PrintF("%s", os.str().c_str());
}
#endif
} // namespace wasm
} // namespace internal
} // namespace v8
// Copyright 2022 the V8 project authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.
d8.file.execute("test/mjsunit/wasm/wasm-module-builder.js");
let builder = new WasmModuleBuilder();
builder.addFunction('crash', makeSig([
kWasmF64, kWasmS128, kWasmS128, kWasmS128, kWasmS128, kWasmS128, kWasmS128,
kWasmF64, kWasmI32
], [])).addBody([kExprNop]);
builder.instantiate();
......@@ -25,7 +25,8 @@
#include "src/execution/riscv/frame-constants-riscv.h"
#endif
#include "src/base/macros.h"
#include "src/wasm/baseline/liftoff-register.h"
#include "testing/gtest/include/gtest/gtest.h"
namespace v8 {
namespace internal {
......@@ -38,6 +39,32 @@ static_assert(kLiftoffAssemblerGpCacheRegs ==
static_assert(kLiftoffAssemblerFpCacheRegs ==
WasmDebugBreakFrameConstants::kPushedFpRegs);
class WasmRegisterTest : public ::testing::Test {};
TEST_F(WasmRegisterTest, SpreadSetBitsToAdjacentFpRegs) {
LiftoffRegList input(
// GP reg selection criteria: an even and an odd register belonging to
// separate adjacent pairs, and contained in kLiftoffAssemblerGpCacheRegs
// for the given platform.
#if V8_TARGET_ARCH_S390X || V8_TARGET_ARCH_PPC64
LiftoffRegister::from_code(kGpReg, 4),
LiftoffRegister::from_code(kGpReg, 7),
#else
LiftoffRegister::from_code(kGpReg, 1),
LiftoffRegister::from_code(kGpReg, 2),
#endif
LiftoffRegister::from_code(kFpReg, 1),
LiftoffRegister::from_code(kFpReg, 4));
// GP regs are left alone, FP regs are spread to adjacent pairs starting
// at an even index: 1 → (0, 1) and 4 → (4, 5).
LiftoffRegList expected =
input | LiftoffRegList(LiftoffRegister::from_code(kFpReg, 0),
LiftoffRegister::from_code(kFpReg, 5));
LiftoffRegList actual = input.SpreadSetBitsToAdjacentFpRegs();
EXPECT_EQ(expected, actual);
}
} // namespace wasm
} // namespace internal
} // namespace v8
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment