From: Jakob Stoklund Olesen Date: Mon, 14 May 2012 21:10:25 +0000 (+0000) Subject: Fix PR12821. X-Git-Url: http://plrg.eecs.uci.edu/git/?a=commitdiff_plain;h=4d10829e1296b7aea0455e000fc318b147182c1c;p=oota-llvm.git Fix PR12821. RAFast must add an operand when it is rewriting a sub-register def that isn't a read-modify-write. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@156777 91177308-0d34-0410-b5e6-96231b3b80d8 --- diff --git a/lib/CodeGen/RegAllocFast.cpp b/lib/CodeGen/RegAllocFast.cpp index 70fbfd4c5ca..15485b13dad 100644 --- a/lib/CodeGen/RegAllocFast.cpp +++ b/lib/CodeGen/RegAllocFast.cpp @@ -674,6 +674,12 @@ bool RAFast::setPhysReg(MachineInstr *MI, unsigned OpNum, unsigned PhysReg) { MI->addRegisterKilled(PhysReg, TRI, true); return true; } + + // A of a sub-register requires an implicit def of the full + // register. + if (MO.isDef() && MO.isUndef()) + MI->addRegisterDefined(PhysReg, TRI); + return MO.isDead(); } diff --git a/test/CodeGen/X86/liveness-local-regalloc.ll b/test/CodeGen/X86/liveness-local-regalloc.ll index b469d0837dc..52f6e5e2848 100644 --- a/test/CodeGen/X86/liveness-local-regalloc.ll +++ b/test/CodeGen/X86/liveness-local-regalloc.ll @@ -1,5 +1,6 @@ -; RUN: llc < %s -O3 -regalloc=fast -mtriple=x86_64-apple-darwin10 +; RUN: llc < %s -regalloc=fast -verify-machineinstrs -mtriple=x86_64-apple-darwin10 ; +; PR12821 %0 = type { i32, i8*, i8*, %1*, i8*, i64, i64, i32, i32, i32, i32, [1024 x i8] } %1 = type { i8*, i32, i32, i16, i16, %2, i32, i8*, i32 (i8*)*, i32 (i8*, i8*, i32)*, i64 (i8*, i64, i32)*, i32 (i8*, i8*, i32)*, %2, %3*, i32, [3 x i8], [1 x i8], %2, i32, i64 } @@ -58,3 +59,34 @@ infloop: ; preds = %infloop, %bb3 infloop1: ; preds = %infloop1, %bb5 br label %infloop1 } + + +; RAFast would forget to add a super-register when rewriting: +; %vreg10:sub_32bit = COPY %R9D +; This trips up the machine code verifier. +define void @autogen_SD24657(i8*, i32*, i64*, i32, i64, i8) { +BB: + %A4 = alloca <16 x i16> + %A3 = alloca double + %A2 = alloca <2 x i8> + %A1 = alloca i1 + %A = alloca i32 + %L = load i8* %0 + store i8 -37, i8* %0 + %E = extractelement <4 x i64> zeroinitializer, i32 2 + %Shuff = shufflevector <4 x i64> zeroinitializer, <4 x i64> zeroinitializer, <4 x i32> + %I = insertelement <2 x i8> , i8 %5, i32 1 + %B = fadd float 0x45CDF5B1C0000000, 0x45CDF5B1C0000000 + %FC = uitofp i32 275048 to double + %Sl = select i1 true, <2 x i8> %I, <2 x i8> + %Cmp = icmp slt i64 0, %E + br label %CF + +CF: ; preds = %BB + store i8 %5, i8* %0 + store <2 x i8> %I, <2 x i8>* %A2 + store i8 %5, i8* %0 + store i8 %5, i8* %0 + store i8 %5, i8* %0 + ret void +}