diff --git a/lib/Target/Hexagon/HexagonFrameLowering.cpp b/lib/Target/Hexagon/HexagonFrameLowering.cpp index f0b7f587379..2b0ceaa6625 100644 --- a/lib/Target/Hexagon/HexagonFrameLowering.cpp +++ b/lib/Target/Hexagon/HexagonFrameLowering.cpp @@ -1051,10 +1051,26 @@ int HexagonFrameLowering::getFrameIndexReference(const MachineFunction &MF, bool HasExtraAlign = HRI.needsStackRealignment(MF); bool NoOpt = MF.getTarget().getOptLevel() == CodeGenOpt::None; + unsigned FrameSize = MFI.getStackSize(); unsigned SP = HRI.getStackRegister(), FP = HRI.getFrameRegister(); auto &HMFI = *MF.getInfo(); unsigned AP = HMFI.getStackAlignBasePhysReg(); - unsigned FrameSize = MFI.getStackSize(); + // It may happen that AP will be absent even HasAlloca && HasExtraAlign + // is true. HasExtraAlign may be set because of vector spills, without + // aligned locals or aligned outgoing function arguments. Since vector + // spills will ultimately be "unaligned", it is safe to use FP as the + // base register. + // In fact, in such a scenario the stack is actually not required to be + // aligned, although it may end up being aligned anyway, since this + // particular case is not easily detectable. The alignment will be + // unnecessary, but not incorrect. + // Unfortunately there is no quick way to verify that the above is + // indeed the case (and that it's not a result of an error), so just + // assume that missing AP will be replaced by FP. + // (A better fix would be to rematerialize AP from FP and always align + // vector spills.) + if (AP == 0) + AP = FP; bool UseFP = false, UseAP = false; // Default: use SP (except at -O0). // Use FP at -O0, except when there are objects with extra alignment. diff --git a/lib/Target/Hexagon/HexagonInstrInfo.cpp b/lib/Target/Hexagon/HexagonInstrInfo.cpp index f43101fa456..fec2dc5ce30 100644 --- a/lib/Target/Hexagon/HexagonInstrInfo.cpp +++ b/lib/Target/Hexagon/HexagonInstrInfo.cpp @@ -94,10 +94,6 @@ static cl::opt UseDFAHazardRec("dfa-hazard-rec", /// /// Constants for Hexagon instructions. /// -const int Hexagon_MEMV_OFFSET_MAX_128B = 896; // #s4: -8*128...7*128 -const int Hexagon_MEMV_OFFSET_MIN_128B = -1024; // #s4 -const int Hexagon_MEMV_OFFSET_MAX = 448; // #s4: -8*64...7*64 -const int Hexagon_MEMV_OFFSET_MIN = -512; // #s4 const int Hexagon_MEMW_OFFSET_MAX = 4095; const int Hexagon_MEMW_OFFSET_MIN = -4096; const int Hexagon_MEMD_OFFSET_MAX = 8191; @@ -2443,8 +2439,7 @@ bool HexagonInstrInfo::isValidOffset(unsigned Opcode, int Offset, case Hexagon::V6_vS32b_ai: case Hexagon::V6_vL32Ub_ai: case Hexagon::V6_vS32Ub_ai: - return (Offset >= Hexagon_MEMV_OFFSET_MIN) && - (Offset <= Hexagon_MEMV_OFFSET_MAX); + return isShiftedInt<4,6>(Offset); case Hexagon::PS_vstorerq_ai_128B: case Hexagon::PS_vstorerw_ai_128B: @@ -2454,8 +2449,7 @@ bool HexagonInstrInfo::isValidOffset(unsigned Opcode, int Offset, case Hexagon::V6_vS32b_ai_128B: case Hexagon::V6_vL32Ub_ai_128B: case Hexagon::V6_vS32Ub_ai_128B: - return (Offset >= Hexagon_MEMV_OFFSET_MIN_128B) && - (Offset <= Hexagon_MEMV_OFFSET_MAX_128B); + return isShiftedInt<4,7>(Offset); case Hexagon::J2_loop0i: case Hexagon::J2_loop1i: diff --git a/test/CodeGen/Hexagon/stack-align-reset.ll b/test/CodeGen/Hexagon/stack-align-reset.ll new file mode 100644 index 00000000000..0d028fb95b2 --- /dev/null +++ b/test/CodeGen/Hexagon/stack-align-reset.ll @@ -0,0 +1,51 @@ +; RUN: llc -march=hexagon < %s | FileCheck %s +; +; This used to crash. +; CHECK: call f1 + +target triple = "hexagon-unknown--elf" + +%struct.0 = type { [5 x i32] } +%struct.2 = type { i32, i32, i32, %struct.1* } +%struct.1 = type { i16*, i32, i32, i32 } + +@g0 = external hidden unnamed_addr constant [52 x i8], align 1 +@g1 = external hidden unnamed_addr constant [3 x i8], align 1 + +declare extern_weak void @f0(i32, i8*, i32, i8*, ...) #0 +declare void @f1(%struct.0*, i32) #0 + +define void @fred(i8* %a0) #0 { +b1: + %v2 = alloca %struct.0, align 4 + %v3 = alloca %struct.2, i32 undef, align 8 + br i1 undef, label %b5, label %b4 + +b4: ; preds = %b1 + br label %b7 + +b5: ; preds = %b5, %b1 + %v6 = getelementptr inbounds %struct.2, %struct.2* %v3, i32 undef, i32 3 + store %struct.1* undef, %struct.1** %v6, align 4 + br label %b5 + +b7: ; preds = %b10, %b4 + %v8 = call i32 @llvm.hexagon.V6.extractw(<16 x i32> zeroinitializer, i32 0) + br i1 icmp eq (void (i32, i8*, i32, i8*, ...)* @f0, void (i32, i8*, i32, i8*, ...)* null), label %b11, label %b9 + +b9: ; preds = %b7 + call void (i32, i8*, i32, i8*, ...) @f0(i32 2, i8* getelementptr inbounds ([52 x i8], [52 x i8]* @g0, i32 0, i32 0), i32 2346, i8* getelementptr inbounds ([3 x i8], [3 x i8]* @g1, i32 0, i32 0), i32 %v8) + unreachable + +b10: ; preds = %b11 + call void @f1(%struct.0* nonnull %v2, i32 28) + br label %b7 + +b11: ; preds = %b11, %b7 + br i1 undef, label %b10, label %b11 +} + +declare i32 @llvm.hexagon.V6.extractw(<16 x i32>, i32) #1 + +attributes #0 = { nounwind "target-cpu"="hexagonv60" "target-features"="+hvx,-hvx-double" } +attributes #1 = { nounwind readnone } diff --git a/test/CodeGen/Hexagon/vec-vararg-align.ll b/test/CodeGen/Hexagon/vec-vararg-align.ll new file mode 100644 index 00000000000..d4c6bd3ef61 --- /dev/null +++ b/test/CodeGen/Hexagon/vec-vararg-align.ll @@ -0,0 +1,30 @@ +; RUN: llc -march=hexagon < %s | FileCheck %s +; +; Check that the stack is aligned according to the outgoing function arguments. +; CHECK: r29 = and(r29,#-128) + +target triple = "hexagon-unknown--elf" + +@.str = private unnamed_addr constant [32 x i8] c"\0AMixed Vectors, Pairs, int flt\0A\00", align 1 +@.str.1 = private unnamed_addr constant [11 x i8] c"\0AVar args\0A\00", align 1 +@gVec0 = common global <16 x i32> zeroinitializer, align 64 +@gVec10 = common global <32 x i32> zeroinitializer, align 128 +@gi1 = common global i32 0, align 4 +@gf1 = common global float 0.000000e+00, align 4 + +define i32 @main() #0 { +b0: + %v1 = call i32 (i8*, ...) @printf(i8* getelementptr inbounds ([32 x i8], [32 x i8]* @.str, i32 0, i32 0)) #0 + %v2 = load <16 x i32>, <16 x i32>* @gVec0, align 64 + %v3 = load <32 x i32>, <32 x i32>* @gVec10, align 128 + %v4 = load i32, i32* @gi1, align 4 + %v5 = load float, float* @gf1, align 4 + %v6 = fpext float %v5 to double + call void (i8*, i32, ...) @VarVec1(i8* getelementptr inbounds ([11 x i8], [11 x i8]* @.str.1, i32 0, i32 0), i32 4, <16 x i32> %v2, <32 x i32> %v3, i32 %v4, double %v6) + ret i32 0 +} + +declare i32 @printf(i8*, ...) #0 +declare void @VarVec1(i8*, i32, ...) #0 + +attributes #0 = { nounwind "target-cpu"="hexagonv60" "target-features"="+hvx" }