From ad3d81d320408a5baf50d61c6f666e0d854502d2 Mon Sep 17 00:00:00 2001 From: Tim Northover Date: Fri, 7 Mar 2014 11:04:30 +0000 Subject: [PATCH] CodeGenPrep: sink extends of illegal types into use block. This helps the instruction selector to lower an i64 * i64 -> i128 multiplication into a single instruction on targets which support it. Patch by Manuel Jacob. llvm-svn: 203230 --- llvm/lib/CodeGen/CodeGenPrepare.cpp | 92 ++++++++++++++-------- llvm/lib/Target/X86/README.txt | 48 ----------- llvm/test/CodeGen/X86/mul128_sext_loop.ll | 32 ++++++++ .../CodeGenPrepare/X86/extend-sink-hoist.ll | 46 +++++++++++ 4 files changed, 136 insertions(+), 82 deletions(-) create mode 100644 llvm/test/CodeGen/X86/mul128_sext_loop.ll create mode 100644 llvm/test/Transforms/CodeGenPrepare/X86/extend-sink-hoist.ll diff --git a/llvm/lib/CodeGen/CodeGenPrepare.cpp b/llvm/lib/CodeGen/CodeGenPrepare.cpp index cb99049..e02d478 100644 --- a/llvm/lib/CodeGen/CodeGenPrepare.cpp +++ b/llvm/lib/CodeGen/CodeGenPrepare.cpp @@ -129,6 +129,7 @@ typedef DenseMap InstrToOrigTy; bool OptimizeMemoryInst(Instruction *I, Value *Addr, Type *AccessTy); bool OptimizeInlineAsmInst(CallInst *CS); bool OptimizeCallInst(CallInst *CI); + bool SinkExtExpand(CastInst *I); bool MoveExtToFormExtLoad(Instruction *I); bool OptimizeExtUses(Instruction *I); bool OptimizeSelectInst(SelectInst *SI); @@ -465,40 +466,8 @@ void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) { DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n"); } -/// OptimizeNoopCopyExpression - If the specified cast instruction is a noop -/// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC), -/// sink it into user blocks to reduce the number of virtual -/// registers that must be created and coalesced. -/// -/// Return true if any changes are made. -/// -static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){ - // If this is a noop copy, - EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType()); - EVT DstVT = TLI.getValueType(CI->getType()); - - // This is an fp<->int conversion? - if (SrcVT.isInteger() != DstVT.isInteger()) - return false; - - // If this is an extension, it will be a zero or sign extension, which - // isn't a noop. - if (SrcVT.bitsLT(DstVT)) return false; - - // If these values will be promoted, find out what they will be promoted - // to. This helps us consider truncates on PPC as noop copies when they - // are. - if (TLI.getTypeAction(CI->getContext(), SrcVT) == - TargetLowering::TypePromoteInteger) - SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT); - if (TLI.getTypeAction(CI->getContext(), DstVT) == - TargetLowering::TypePromoteInteger) - DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT); - - // If, after promotion, these are the same types, this is a noop copy. - if (SrcVT != DstVT) - return false; - +/// SinkCast - Sink the specified cast instruction into its user blocks +static bool SinkCast(CastInst *CI) { BasicBlock *DefBB = CI->getParent(); /// InsertedCasts - Only insert a cast in each block once. @@ -548,6 +517,43 @@ static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){ return MadeChange; } +/// OptimizeNoopCopyExpression - If the specified cast instruction is a noop +/// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC), +/// sink it into user blocks to reduce the number of virtual +/// registers that must be created and coalesced. +/// +/// Return true if any changes are made. +/// +static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){ + // If this is a noop copy, + EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType()); + EVT DstVT = TLI.getValueType(CI->getType()); + + // This is an fp<->int conversion? + if (SrcVT.isInteger() != DstVT.isInteger()) + return false; + + // If this is an extension, it will be a zero or sign extension, which + // isn't a noop. + if (SrcVT.bitsLT(DstVT)) return false; + + // If these values will be promoted, find out what they will be promoted + // to. This helps us consider truncates on PPC as noop copies when they + // are. + if (TLI.getTypeAction(CI->getContext(), SrcVT) == + TargetLowering::TypePromoteInteger) + SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT); + if (TLI.getTypeAction(CI->getContext(), DstVT) == + TargetLowering::TypePromoteInteger) + DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT); + + // If, after promotion, these are the same types, this is a noop copy. + if (SrcVT != DstVT) + return false; + + return SinkCast(CI); +} + /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce /// the number of virtual registers that must be created and coalesced. This is /// a clear win except on targets with multiple condition code registers @@ -2522,6 +2528,16 @@ bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) { return MadeChange; } +/// SinkExtExpand - Sink a zext or sext into its user blocks if the target type +/// doesn't fit in one register +bool CodeGenPrepare::SinkExtExpand(CastInst *CI) { + if (TLI && + TLI->getTypeAction(CI->getContext(), TLI->getValueType(CI->getType())) == + TargetLowering::TypeExpandInteger) + return SinkCast(CI); + return false; +} + /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same /// basic block as the load, unless conditions are unfavorable. This allows /// SelectionDAG to fold the extend into the load. @@ -2535,6 +2551,12 @@ bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) { if (LI->getParent() == I->getParent()) return false; + // Do not undo the optimization in SinkExtExpand + if (TLI && + TLI->getTypeAction(I->getContext(), TLI->getValueType(I->getType())) == + TargetLowering::TypeExpandInteger) + return false; + // If the load has other users and the truncate is not free, this probably // isn't worthwhile. if (!LI->hasOneUse() && @@ -2821,6 +2843,8 @@ bool CodeGenPrepare::OptimizeInst(Instruction *I) { return true; if (isa(I) || isa(I)) { + if (SinkExtExpand(CI)) + return true; bool MadeChange = MoveExtToFormExtLoad(I); return MadeChange | OptimizeExtUses(I); } diff --git a/llvm/lib/Target/X86/README.txt b/llvm/lib/Target/X86/README.txt index a30f8c7..52d3c01 100644 --- a/llvm/lib/Target/X86/README.txt +++ b/llvm/lib/Target/X86/README.txt @@ -1444,54 +1444,6 @@ it would be nice to produce "into" someday. //===---------------------------------------------------------------------===// -This code: - -void vec_mpys1(int y[], const int x[], int scaler) { -int i; -for (i = 0; i < 150; i++) - y[i] += (((long long)scaler * (long long)x[i]) >> 31); -} - -Compiles to this loop with GCC 3.x: - -.L5: - movl %ebx, %eax - imull (%edi,%ecx,4) - shrdl $31, %edx, %eax - addl %eax, (%esi,%ecx,4) - incl %ecx - cmpl $149, %ecx - jle .L5 - -llvm-gcc compiles it to the much uglier: - -LBB1_1: ## bb1 - movl 24(%esp), %eax - movl (%eax,%edi,4), %ebx - movl %ebx, %ebp - imull %esi, %ebp - movl %ebx, %eax - mull %ecx - addl %ebp, %edx - sarl $31, %ebx - imull %ecx, %ebx - addl %edx, %ebx - shldl $1, %eax, %ebx - movl 20(%esp), %eax - addl %ebx, (%eax,%edi,4) - incl %edi - cmpl $150, %edi - jne LBB1_1 ## bb1 - -The issue is that we hoist the cast of "scaler" to long long outside of the -loop, the value comes into the loop as two values, and -RegsForValue::getCopyFromRegs doesn't know how to put an AssertSext on the -constructed BUILD_PAIR which represents the cast value. - -This can be handled by making CodeGenPrepare sink the cast. - -//===---------------------------------------------------------------------===// - Test instructions can be eliminated by using EFLAGS values from arithmetic instructions. This is currently not done for mul, and, or, xor, neg, shl, sra, srl, shld, shrd, atomic ops, and others. It is also currently not done diff --git a/llvm/test/CodeGen/X86/mul128_sext_loop.ll b/llvm/test/CodeGen/X86/mul128_sext_loop.ll new file mode 100644 index 0000000..a516f03 --- /dev/null +++ b/llvm/test/CodeGen/X86/mul128_sext_loop.ll @@ -0,0 +1,32 @@ +; RUN: llc < %s -march=x86-64 | FileCheck %s + +define void @test(i64* nocapture %arr, i64 %arrsize, i64 %factor) nounwind uwtable { + %1 = icmp sgt i64 %arrsize, 0 + br i1 %1, label %.lr.ph, label %._crit_edge + +.lr.ph: ; preds = %0 + %2 = sext i64 %factor to i128 + br label %3 + +;