diff options
Diffstat (limited to 'tools/profiler/lul/LulDwarfSummariser.cpp')
| -rw-r--r-- | tools/profiler/lul/LulDwarfSummariser.cpp | 359 |
1 files changed, 359 insertions, 0 deletions
diff --git a/tools/profiler/lul/LulDwarfSummariser.cpp b/tools/profiler/lul/LulDwarfSummariser.cpp new file mode 100644 index 0000000000..74c2565dff --- /dev/null +++ b/tools/profiler/lul/LulDwarfSummariser.cpp @@ -0,0 +1,359 @@ +/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */ +/* vim: set ts=8 sts=2 et sw=2 tw=80: */ +/* This Source Code Form is subject to the terms of the Mozilla Public + * License, v. 2.0. If a copy of the MPL was not distributed with this + * file, You can obtain one at http://mozilla.org/MPL/2.0/. */ + +#include "LulDwarfSummariser.h" + +#include "mozilla/Assertions.h" + +// Set this to 1 for verbose logging +#define DEBUG_SUMMARISER 0 + +namespace lul { + +// Do |s64|'s lowest 32 bits sign extend back to |s64| itself? +static inline bool fitsIn32Bits(int64 s64) { + return s64 == ((s64 & 0xffffffff) ^ 0x80000000) - 0x80000000; +} + +// Check a LExpr prefix expression, starting at pfxInstrs[start] up to +// the next PX_End instruction, to ensure that: +// * It only mentions registers that are tracked on this target +// * The start point is sane +// If the expression is ok, return NULL. Else return a pointer +// a const char* holding a bit of text describing the problem. +static const char* +checkPfxExpr(const vector<PfxInstr>* pfxInstrs, int64_t start) +{ + size_t nInstrs = pfxInstrs->size(); + if (start < 0 || start >= (ssize_t)nInstrs) { + return "bogus start point"; + } + size_t i; + for (i = start; i < nInstrs; i++) { + PfxInstr pxi = (*pfxInstrs)[i]; + if (pxi.mOpcode == PX_End) + break; + if (pxi.mOpcode == PX_DwReg && + !registerIsTracked((DW_REG_NUMBER)pxi.mOperand)) { + return "uses untracked reg"; + } + } + return nullptr; // success +} + + +Summariser::Summariser(SecMap* aSecMap, uintptr_t aTextBias, + void(*aLog)(const char*)) + : mSecMap(aSecMap) + , mTextBias(aTextBias) + , mLog(aLog) +{ + mCurrAddr = 0; + mMax1Addr = 0; // Gives an empty range. + + // Initialise the running RuleSet to "haven't got a clue" status. + new (&mCurrRules) RuleSet(); +} + +void +Summariser::Entry(uintptr_t aAddress, uintptr_t aLength) +{ + aAddress += mTextBias; + if (DEBUG_SUMMARISER) { + char buf[100]; + SprintfLiteral(buf, + "LUL Entry(%llx, %llu)\n", + (unsigned long long int)aAddress, + (unsigned long long int)aLength); + mLog(buf); + } + // This throws away any previous summary, that is, assumes + // that the previous summary, if any, has been properly finished + // by a call to End(). + mCurrAddr = aAddress; + mMax1Addr = aAddress + aLength; + new (&mCurrRules) RuleSet(); +} + +void +Summariser::Rule(uintptr_t aAddress, int aNewReg, + LExprHow how, int16_t oldReg, int64_t offset) +{ + aAddress += mTextBias; + if (DEBUG_SUMMARISER) { + char buf[100]; + if (how == NODEREF || how == DEREF) { + bool deref = how == DEREF; + SprintfLiteral(buf, + "LUL 0x%llx old-r%d = %sr%d + %lld%s\n", + (unsigned long long int)aAddress, aNewReg, + deref ? "*(" : "", (int)oldReg, (long long int)offset, + deref ? ")" : ""); + } else if (how == PFXEXPR) { + SprintfLiteral(buf, + "LUL 0x%llx old-r%d = pfx-expr-at %lld\n", + (unsigned long long int)aAddress, aNewReg, + (long long int)offset); + } else { + SprintfLiteral(buf, + "LUL 0x%llx old-r%d = (invalid LExpr!)\n", + (unsigned long long int)aAddress, aNewReg); + } + mLog(buf); + } + + if (mCurrAddr < aAddress) { + // Flush the existing summary first. + mCurrRules.mAddr = mCurrAddr; + mCurrRules.mLen = aAddress - mCurrAddr; + mSecMap->AddRuleSet(&mCurrRules); + if (DEBUG_SUMMARISER) { + mLog("LUL "); mCurrRules.Print(mLog); + mLog("\n"); + } + mCurrAddr = aAddress; + } + + // If for some reason summarisation fails, either or both of these + // become non-null and point at constant text describing the + // problem. Using two rather than just one avoids complications of + // having to concatenate two strings to produce a complete error message. + const char* reason1 = nullptr; + const char* reason2 = nullptr; + + // |offset| needs to be a 32 bit value that sign extends to 64 bits + // on a 64 bit target. We will need to incorporate |offset| into + // any LExpr made here. So we may as well check it right now. + if (!fitsIn32Bits(offset)) { + reason1 = "offset not in signed 32-bit range"; + goto cant_summarise; + } + + // FIXME: factor out common parts of the arch-dependent summarisers. + +#if defined(LUL_ARCH_arm) + + // ----------------- arm ----------------- // + + // Now, can we add the rule to our summary? This depends on whether + // the registers and the overall expression are representable. This + // is the heart of the summarisation process. + switch (aNewReg) { + + case DW_REG_CFA: + // This is a rule that defines the CFA. The only forms we + // choose to represent are: r7/11/12/13 + offset. The offset + // must fit into 32 bits since 'uintptr_t' is 32 bit on ARM, + // hence there is no need to check it for overflow. + if (how != NODEREF) { + reason1 = "rule for DW_REG_CFA: invalid |how|"; + goto cant_summarise; + } + switch (oldReg) { + case DW_REG_ARM_R7: case DW_REG_ARM_R11: + case DW_REG_ARM_R12: case DW_REG_ARM_R13: + break; + default: + reason1 = "rule for DW_REG_CFA: invalid |oldReg|"; + goto cant_summarise; + } + mCurrRules.mCfaExpr = LExpr(how, oldReg, offset); + break; + + case DW_REG_ARM_R7: case DW_REG_ARM_R11: case DW_REG_ARM_R12: + case DW_REG_ARM_R13: case DW_REG_ARM_R14: case DW_REG_ARM_R15: { + // This is a new rule for R7, R11, R12, R13 (SP), R14 (LR) or + // R15 (the return address). + switch (how) { + case NODEREF: case DEREF: + // Check the old register is one we're tracking. + if (!registerIsTracked((DW_REG_NUMBER)oldReg) && + oldReg != DW_REG_CFA) { + reason1 = "rule for R7/11/12/13/14/15: uses untracked reg"; + goto cant_summarise; + } + break; + case PFXEXPR: { + // Check that the prefix expression only mentions tracked registers. + const vector<PfxInstr>* pfxInstrs = mSecMap->GetPfxInstrs(); + reason2 = checkPfxExpr(pfxInstrs, offset); + if (reason2) { + reason1 = "rule for R7/11/12/13/14/15: "; + goto cant_summarise; + } + break; + } + default: + goto cant_summarise; + } + LExpr expr = LExpr(how, oldReg, offset); + switch (aNewReg) { + case DW_REG_ARM_R7: mCurrRules.mR7expr = expr; break; + case DW_REG_ARM_R11: mCurrRules.mR11expr = expr; break; + case DW_REG_ARM_R12: mCurrRules.mR12expr = expr; break; + case DW_REG_ARM_R13: mCurrRules.mR13expr = expr; break; + case DW_REG_ARM_R14: mCurrRules.mR14expr = expr; break; + case DW_REG_ARM_R15: mCurrRules.mR15expr = expr; break; + default: MOZ_ASSERT(0); + } + break; + } + + default: + // Leave |reason1| and |reason2| unset here. This program point + // is reached so often that it causes a flood of "Can't + // summarise" messages. In any case, we don't really care about + // the fact that this summary would produce a new value for a + // register that we're not tracking. We do on the other hand + // care if the summary's expression *uses* a register that we're + // not tracking. But in that case one of the above failures + // should tell us which. + goto cant_summarise; + } + + // Mark callee-saved registers (r4 .. r11) as unchanged, if there is + // no other information about them. FIXME: do this just once, at + // the point where the ruleset is committed. + if (mCurrRules.mR7expr.mHow == UNKNOWN) { + mCurrRules.mR7expr = LExpr(NODEREF, DW_REG_ARM_R7, 0); + } + if (mCurrRules.mR11expr.mHow == UNKNOWN) { + mCurrRules.mR11expr = LExpr(NODEREF, DW_REG_ARM_R11, 0); + } + if (mCurrRules.mR12expr.mHow == UNKNOWN) { + mCurrRules.mR12expr = LExpr(NODEREF, DW_REG_ARM_R12, 0); + } + + // The old r13 (SP) value before the call is always the same as the + // CFA. + mCurrRules.mR13expr = LExpr(NODEREF, DW_REG_CFA, 0); + + // If there's no information about R15 (the return address), say + // it's a copy of R14 (the link register). + if (mCurrRules.mR15expr.mHow == UNKNOWN) { + mCurrRules.mR15expr = LExpr(NODEREF, DW_REG_ARM_R14, 0); + } + +#elif defined(LUL_ARCH_x64) || defined(LUL_ARCH_x86) + + // ---------------- x64/x86 ---------------- // + + // Now, can we add the rule to our summary? This depends on whether + // the registers and the overall expression are representable. This + // is the heart of the summarisation process. + switch (aNewReg) { + + case DW_REG_CFA: + // This is a rule that defines the CFA. The only forms we can + // represent are: = SP+offset or = FP+offset. + if (how != NODEREF) { + reason1 = "rule for DW_REG_CFA: invalid |how|"; + goto cant_summarise; + } + if (oldReg != DW_REG_INTEL_XSP && oldReg != DW_REG_INTEL_XBP) { + reason1 = "rule for DW_REG_CFA: invalid |oldReg|"; + goto cant_summarise; + } + mCurrRules.mCfaExpr = LExpr(how, oldReg, offset); + break; + + case DW_REG_INTEL_XSP: case DW_REG_INTEL_XBP: case DW_REG_INTEL_XIP: { + // This is a new rule for XSP, XBP or XIP (the return address). + switch (how) { + case NODEREF: case DEREF: + // Check the old register is one we're tracking. + if (!registerIsTracked((DW_REG_NUMBER)oldReg) && + oldReg != DW_REG_CFA) { + reason1 = "rule for XSP/XBP/XIP: uses untracked reg"; + goto cant_summarise; + } + break; + case PFXEXPR: { + // Check that the prefix expression only mentions tracked registers. + const vector<PfxInstr>* pfxInstrs = mSecMap->GetPfxInstrs(); + reason2 = checkPfxExpr(pfxInstrs, offset); + if (reason2) { + reason1 = "rule for XSP/XBP/XIP: "; + goto cant_summarise; + } + break; + } + default: + goto cant_summarise; + } + LExpr expr = LExpr(how, oldReg, offset); + switch (aNewReg) { + case DW_REG_INTEL_XBP: mCurrRules.mXbpExpr = expr; break; + case DW_REG_INTEL_XSP: mCurrRules.mXspExpr = expr; break; + case DW_REG_INTEL_XIP: mCurrRules.mXipExpr = expr; break; + default: MOZ_CRASH("impossible value for aNewReg"); + } + break; + } + + default: + // Leave |reason1| and |reason2| unset here, for the reasons + // explained in the analogous point in the ARM case just above. + goto cant_summarise; + + } + + // On Intel, it seems the old SP value before the call is always the + // same as the CFA. Therefore, in the absence of any other way to + // recover the SP, specify that the CFA should be copied. + if (mCurrRules.mXspExpr.mHow == UNKNOWN) { + mCurrRules.mXspExpr = LExpr(NODEREF, DW_REG_CFA, 0); + } + + // Also, gcc says "Undef" for BP when it is unchanged. + if (mCurrRules.mXbpExpr.mHow == UNKNOWN) { + mCurrRules.mXbpExpr = LExpr(NODEREF, DW_REG_INTEL_XBP, 0); + } + +#else + +# error "Unsupported arch" +#endif + + return; + + cant_summarise: + if (reason1 || reason2) { + char buf[200]; + SprintfLiteral(buf, "LUL can't summarise: " + "SVMA=0x%llx: %s%s, expr=LExpr(%s,%u,%lld)\n", + (unsigned long long int)(aAddress - mTextBias), + reason1 ? reason1 : "", reason2 ? reason2 : "", + NameOf_LExprHow(how), + (unsigned int)oldReg, (long long int)offset); + mLog(buf); + } +} + +uint32_t +Summariser::AddPfxInstr(PfxInstr pfxi) +{ + return mSecMap->AddPfxInstr(pfxi); +} + +void +Summariser::End() +{ + if (DEBUG_SUMMARISER) { + mLog("LUL End\n"); + } + if (mCurrAddr < mMax1Addr) { + mCurrRules.mAddr = mCurrAddr; + mCurrRules.mLen = mMax1Addr - mCurrAddr; + mSecMap->AddRuleSet(&mCurrRules); + if (DEBUG_SUMMARISER) { + mLog("LUL "); mCurrRules.Print(mLog); + mLog("\n"); + } + } +} + +} // namespace lul |