From 58f4f24a6e23b68e8c32d44819eceb7dbb2c563d Mon Sep 17 00:00:00 2001 From: Juergen Ributzka Date: Wed, 19 Aug 2015 20:52:55 +0000 Subject: [PATCH] [AArch64][FastISel] Don't fold shifts with UB. We are already falling back to SelectionDAG when encountering an shift with UB. This adds the same checks for shifts with UB that get folded into arithmetic or logical operations. This fixes rdar://problem/22345295. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@245499 91177308-0d34-0410-b5e6-96231b3b80d8 --- lib/Target/AArch64/AArch64FastISel.cpp | 51 +++++-- .../CodeGen/AArch64/fast-isel-folded-shift.ll | 125 ++++++++++++++++++ test/CodeGen/AArch64/fast-isel-logic-op.ll | 2 +- 3 files changed, 164 insertions(+), 14 deletions(-) create mode 100644 test/CodeGen/AArch64/fast-isel-folded-shift.ll diff --git a/lib/Target/AArch64/AArch64FastISel.cpp b/lib/Target/AArch64/AArch64FastISel.cpp index ca4b66a76a5..cc5d415a25c 100644 --- a/lib/Target/AArch64/AArch64FastISel.cpp +++ b/lib/Target/AArch64/AArch64FastISel.cpp @@ -1178,7 +1178,7 @@ unsigned AArch64FastISel::emitAddSub(bool UseAdd, MVT RetVT, const Value *LHS, } // Check if the mul can be folded into the instruction. - if (RHS->hasOneUse() && isValueAvailable(RHS)) + if (RHS->hasOneUse() && isValueAvailable(RHS)) { if (isMulPowOf2(RHS)) { const Value *MulLHS = cast(RHS)->getOperand(0); const Value *MulRHS = cast(RHS)->getOperand(1); @@ -1193,12 +1193,16 @@ unsigned AArch64FastISel::emitAddSub(bool UseAdd, MVT RetVT, const Value *LHS, if (!RHSReg) return 0; bool RHSIsKill = hasTrivialKill(MulLHS); - return emitAddSub_rs(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, RHSIsKill, - AArch64_AM::LSL, ShiftVal, SetFlags, WantResult); + ResultReg = emitAddSub_rs(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, + RHSIsKill, AArch64_AM::LSL, ShiftVal, SetFlags, + WantResult); + if (ResultReg) + return ResultReg; } + } // Check if the shift can be folded into the instruction. - if (RHS->hasOneUse() && isValueAvailable(RHS)) + if (RHS->hasOneUse() && isValueAvailable(RHS)) { if (const auto *SI = dyn_cast(RHS)) { if (const auto *C = dyn_cast(SI->getOperand(1))) { AArch64_AM::ShiftExtendType ShiftType = AArch64_AM::InvalidShiftExtend; @@ -1214,12 +1218,15 @@ unsigned AArch64FastISel::emitAddSub(bool UseAdd, MVT RetVT, const Value *LHS, if (!RHSReg) return 0; bool RHSIsKill = hasTrivialKill(SI->getOperand(0)); - return emitAddSub_rs(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, - RHSIsKill, ShiftType, ShiftVal, SetFlags, - WantResult); + ResultReg = emitAddSub_rs(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, + RHSIsKill, ShiftType, ShiftVal, SetFlags, + WantResult); + if (ResultReg) + return ResultReg; } } } + } unsigned RHSReg = getRegForValue(RHS); if (!RHSReg) @@ -1323,6 +1330,10 @@ unsigned AArch64FastISel::emitAddSub_rs(bool UseAdd, MVT RetVT, unsigned LHSReg, if (RetVT != MVT::i32 && RetVT != MVT::i64) return 0; + // Don't deal with undefined shifts. + if (ShiftImm >= RetVT.getSizeInBits()) + return 0; + static const unsigned OpcTable[2][2][2] = { { { AArch64::SUBWrs, AArch64::SUBXrs }, { AArch64::ADDWrs, AArch64::ADDXrs } }, @@ -1360,6 +1371,9 @@ unsigned AArch64FastISel::emitAddSub_rx(bool UseAdd, MVT RetVT, unsigned LHSReg, if (RetVT != MVT::i32 && RetVT != MVT::i64) return 0; + if (ShiftImm >= 4) + return 0; + static const unsigned OpcTable[2][2][2] = { { { AArch64::SUBWrx, AArch64::SUBXrx }, { AArch64::ADDWrx, AArch64::ADDXrx } }, @@ -1542,7 +1556,7 @@ unsigned AArch64FastISel::emitLogicalOp(unsigned ISDOpc, MVT RetVT, return ResultReg; // Check if the mul can be folded into the instruction. - if (RHS->hasOneUse() && isValueAvailable(RHS)) + if (RHS->hasOneUse() && isValueAvailable(RHS)) { if (isMulPowOf2(RHS)) { const Value *MulLHS = cast(RHS)->getOperand(0); const Value *MulRHS = cast(RHS)->getOperand(1); @@ -1558,12 +1572,15 @@ unsigned AArch64FastISel::emitLogicalOp(unsigned ISDOpc, MVT RetVT, if (!RHSReg) return 0; bool RHSIsKill = hasTrivialKill(MulLHS); - return emitLogicalOp_rs(ISDOpc, RetVT, LHSReg, LHSIsKill, RHSReg, - RHSIsKill, ShiftVal); + ResultReg = emitLogicalOp_rs(ISDOpc, RetVT, LHSReg, LHSIsKill, RHSReg, + RHSIsKill, ShiftVal); + if (ResultReg) + return ResultReg; } + } // Check if the shift can be folded into the instruction. - if (RHS->hasOneUse() && isValueAvailable(RHS)) + if (RHS->hasOneUse() && isValueAvailable(RHS)) { if (const auto *SI = dyn_cast(RHS)) if (const auto *C = dyn_cast(SI->getOperand(1))) { uint64_t ShiftVal = C->getZExtValue(); @@ -1571,9 +1588,12 @@ unsigned AArch64FastISel::emitLogicalOp(unsigned ISDOpc, MVT RetVT, if (!RHSReg) return 0; bool RHSIsKill = hasTrivialKill(SI->getOperand(0)); - return emitLogicalOp_rs(ISDOpc, RetVT, LHSReg, LHSIsKill, RHSReg, - RHSIsKill, ShiftVal); + ResultReg = emitLogicalOp_rs(ISDOpc, RetVT, LHSReg, LHSIsKill, RHSReg, + RHSIsKill, ShiftVal); + if (ResultReg) + return ResultReg; } + } unsigned RHSReg = getRegForValue(RHS); if (!RHSReg) @@ -1646,6 +1666,11 @@ unsigned AArch64FastISel::emitLogicalOp_rs(unsigned ISDOpc, MVT RetVT, { AArch64::ORRWrs, AArch64::ORRXrs }, { AArch64::EORWrs, AArch64::EORXrs } }; + + // Don't deal with undefined shifts. + if (ShiftImm >= RetVT.getSizeInBits()) + return 0; + const TargetRegisterClass *RC; unsigned Opc; switch (RetVT.SimpleTy) { diff --git a/test/CodeGen/AArch64/fast-isel-folded-shift.ll b/test/CodeGen/AArch64/fast-isel-folded-shift.ll new file mode 100644 index 00000000000..b881ef5c6d5 --- /dev/null +++ b/test/CodeGen/AArch64/fast-isel-folded-shift.ll @@ -0,0 +1,125 @@ +; RUN: llc -mtriple=aarch64-apple-darwin -fast-isel=1 -verify-machineinstrs < %s | FileCheck %s + +; Test invalid shift values. This will fall-back to SDAG. +; AND +define zeroext i8 @and_rs_i8(i8 signext %a, i8 signext %b) { +; CHECK-LABEL: and_rs_i8 +; CHECK: and [[REG:w[0-9]+]], w0, w8 +; CHECK-NEXT: and {{w[0-9]+}}, [[REG]], #0xff + %1 = shl i8 %b, 8 + %2 = and i8 %a, %1 + ret i8 %2 +} + +define zeroext i16 @and_rs_i16(i16 signext %a, i16 signext %b) { +; CHECK-LABEL: and_rs_i16 +; CHECK: and [[REG:w[0-9]+]], w0, w8 +; CHECK-NEXT: and {{w[0-9]+}}, [[REG]], #0xffff + %1 = shl i16 %b, 16 + %2 = and i16 %a, %1 + ret i16 %2 +} + +define i32 @and_rs_i32(i32 %a, i32 %b) { +; CHECK-LABEL: and_rs_i32 +; CHECK: and w0, w0, w8 + %1 = shl i32 %b, 32 + %2 = and i32 %a, %1 + ret i32 %2 +} + +define i64 @and_rs_i64(i64 %a, i64 %b) { +; CHECK-LABEL: and_rs_i64 +; CHECK: and x0, x0, x8 + %1 = shl i64 %b, 64 + %2 = and i64 %a, %1 + ret i64 %2 +} + +; OR +define zeroext i8 @or_rs_i8(i8 signext %a, i8 signext %b) { +; CHECK-LABEL: or_rs_i8 +; CHECK: orr [[REG:w[0-9]+]], w0, w8 +; CHECK-NEXT: and {{w[0-9]+}}, [[REG]], #0xff + %1 = shl i8 %b, 8 + %2 = or i8 %a, %1 + ret i8 %2 +} + +define zeroext i16 @or_rs_i16(i16 signext %a, i16 signext %b) { +; CHECK-LABEL: or_rs_i16 +; CHECK: orr [[REG:w[0-9]+]], w0, w8 +; CHECK-NEXT: and {{w[0-9]+}}, [[REG]], #0xffff + %1 = shl i16 %b, 16 + %2 = or i16 %a, %1 + ret i16 %2 +} + +define i32 @or_rs_i32(i32 %a, i32 %b) { +; CHECK-LABEL: or_rs_i32 +; CHECK: orr w0, w0, w8 + %1 = shl i32 %b, 32 + %2 = or i32 %a, %1 + ret i32 %2 +} + +define i64 @or_rs_i64(i64 %a, i64 %b) { +; CHECK-LABEL: or_rs_i64 +; CHECK: orr x0, x0, x8 + %1 = shl i64 %b, 64 + %2 = or i64 %a, %1 + ret i64 %2 +} + +; XOR +define zeroext i8 @xor_rs_i8(i8 %a, i8 %b) { +; CHECK-LABEL: xor_rs_i8 +; CHECK: eor [[REG:w[0-9]+]], w0, w8 +; CHECK-NEXT: and {{w[0-9]+}}, [[REG]], #0xff + %1 = shl i8 %b, 8 + %2 = xor i8 %a, %1 + ret i8 %2 +} + +define zeroext i16 @xor_rs_i16(i16 %a, i16 %b) { +; CHECK-LABEL: xor_rs_i16 +; CHECK: eor [[REG:w[0-9]+]], w0, w8 +; CHECK-NEXT: and {{w[0-9]+}}, [[REG]], #0xffff + %1 = shl i16 %b, 16 + %2 = xor i16 %a, %1 + ret i16 %2 +} + +define i32 @xor_rs_i32(i32 %a, i32 %b) { +; CHECK-LABEL: xor_rs_i32 +; CHECK: eor w0, w0, w8 + %1 = shl i32 %b, 32 + %2 = xor i32 %a, %1 + ret i32 %2 +} + +define i64 @xor_rs_i64(i64 %a, i64 %b) { +; CHECK-LABEL: xor_rs_i64 +; CHECK: eor x0, x0, x8 + %1 = shl i64 %b, 64 + %2 = xor i64 %a, %1 + ret i64 %2 +} + +;ADD +define i32 @add_rs_i32(i32 %a, i32 %b) { +; CHECK-LABEL: add_rs_i32 +; CHECK: add w0, w0, w8 + %1 = shl i32 %b, 32 + %2 = add i32 %a, %1 + ret i32 %2 +} + +define i64 @add_rs_i64(i64 %a, i64 %b) { +; CHECK-LABEL: add_rs_i64 +; CHECK: add x0, x0, x8 + %1 = shl i64 %b, 64 + %2 = add i64 %a, %1 + ret i64 %2 +} + diff --git a/test/CodeGen/AArch64/fast-isel-logic-op.ll b/test/CodeGen/AArch64/fast-isel-logic-op.ll index 89c5f2c4802..16d0429fe98 100644 --- a/test/CodeGen/AArch64/fast-isel-logic-op.ll +++ b/test/CodeGen/AArch64/fast-isel-logic-op.ll @@ -1,4 +1,4 @@ -; RUN: llc -mtriple=aarch64-apple-darwin -fast-isel=0 -verify-machineinstrs < %s | FileCheck %s +; RUN: llc -mtriple=aarch64-apple-darwin -fast-isel=0 -verify-machineinstrs < %s | FileCheck %s ; RUN: llc -mtriple=aarch64-apple-darwin -fast-isel=1 -fast-isel-abort=1 -verify-machineinstrs < %s | FileCheck %s ; AND -- 2.34.1