[AArch64] Disable GlobalISel/FastISel for more SME functions

[dmgreen]

Summary

The patch D136361 disabled GlobalISel and FastISel for some SME functions, as the saving and restoring of SM is not yet handled. There were several tests added for fp128 fadd, which will be expanded to a libcall, that only happened to work by accident and did not handle other cases such as f32/f64 frem libcalls.

This extends the cases where GlobalISel / FastISel is disabled for functions with SME attributes, under the assumption that it is difficult to tell what will become a libcall reliably, and so should fall back for all function until GlobalISel and/or FastISel can handle them.

[sdesmalen]

llvm/lib/Target/AArch64/AArch64FastISel.cpp

5190–5191

This condition is now redundant.

llvm/lib/Target/AArch64/GISel/AArch64CallLowering.cpp

535–536

This condition is now redundant.

[dmgreen]

Oh right - that was meant to be hasStreamingCompatibleInterface. I don't know these interfaces very well, let me know if that sounds wrong.

[sdesmalen]

In D158490#4605854, @dmgreen wrote:

Oh right - that was meant to be hasStreamingCompatibleInterface. I don't know these interfaces very well, let me know if that sounds wrong.

That makes more sense. Can you add a test where the caller has "aarch64_pstate_sm_compatible" as well?

It's a bit unfortunate we can't predict in advance if any operations will be expanded to function calls, but indeed better to be conservative.

[dmgreen]

Added a frem_call_sm_compat test.

[sdesmalen]

Thanks for fixing!

llvm/lib/Target/AArch64/AArch64FastISel.cpp

bool AArch64FastISel::fastSelectInstruction(const Instruction *I) {
   // fall-back to target-independent instruction selection.
   return selectOperator(I, I->getOpcode());
 }
 
 FastISel *AArch64::createFastISel(FunctionLoweringInfo &FuncInfo,
                                         const TargetLibraryInfo *LibInfo) {
 
   SMEAttrs CallerAttrs(*FuncInfo.Fn);
-  if (CallerAttrs.hasZAState() ||
-      (!CallerAttrs.hasStreamingInterface() && CallerAttrs.hasStreamingBody()))
+  if (CallerAttrs.hasZAState() || CallerAttrs.hasStreamingInterfaceOrBody() ||
+      CallerAttrs.hasStreamingInterface())
     return nullptr;
   return new AArch64FastISel(FuncInfo, LibInfo);
 }

llvm/lib/Target/AArch64/GISel/AArch64CallLowering.cpp

if (F.getReturnType()->isScalableTy() ||
     return true;
   const auto &ST = MF.getSubtarget<AArch64Subtarget>();
   if (!ST.hasNEON() || !ST.hasFPARMv8()) {
     LLVM_DEBUG(dbgs() << "Falling back to SDAG because we don't support no-NEON\n");
     return true;
   }
 
   SMEAttrs Attrs(F);
-  if (Attrs.hasNewZAInterface() ||
-      (!Attrs.hasStreamingInterface() && Attrs.hasStreamingBody()))
+  if (Attrs.hasZAState() || Attrs.hasStreamingInterfaceOrBody() ||
+      Attrs.hasStreamingInterface())
     return true;
 
   return false;
 }
 
 void AArch64CallLowering::saveVarArgRegisters(
     MachineIRBuilder &MIRBuilder, CallLowering::IncomingValueHandler &Handler,
     CCState &CCInfo) const {

llvm/test/CodeGen/AArch64/sme-disable-gisel-fisel.ll

 entry:
   %call = call double @streaming_callee(double %x) "aarch64_pstate_sm_enabled"
   %add = fadd double %call, 4.200000e+01
   ret double %add
 }
 
 
 define double @streaming_caller_nonstreaming_callee(double %x) nounwind noinline optnone "aarch64_pstate_sm_enabled" {
-; CHECK-FISEL-LABEL: streaming_caller_nonstreaming_callee:
-; CHECK-FISEL:       // %bb.0: // %entry
-; CHECK-FISEL-NEXT:    sub sp, sp, #96
-; CHECK-FISEL-NEXT:    stp d15, d14, [sp, #16] // 16-byte Folded Spill
-; CHECK-FISEL-NEXT:    stp d13, d12, [sp, #32] // 16-byte Folded Spill
-; CHECK-FISEL-NEXT:    stp d11, d10, [sp, #48] // 16-byte Folded Spill
-; CHECK-FISEL-NEXT:    stp d9, d8, [sp, #64] // 16-byte Folded Spill
-; CHECK-FISEL-NEXT:    str x30, [sp, #80] // 8-byte Folded Spill
-; CHECK-FISEL-NEXT:    str d0, [sp, #8] // 8-byte Folded Spill
-; CHECK-FISEL-NEXT:    smstop sm
-; CHECK-FISEL-NEXT:    ldr d0, [sp, #8] // 8-byte Folded Reload
-; CHECK-FISEL-NEXT:    bl normal_callee
-; CHECK-FISEL-NEXT:    str d0, [sp, #88] // 8-byte Folded Spill
-; CHECK-FISEL-NEXT:    smstart sm
-; CHECK-FISEL-NEXT:    adrp x8, .LCPI1_0
-; CHECK-FISEL-NEXT:    ldr d0, [x8, :lo12:.LCPI1_0]
-; CHECK-FISEL-NEXT:    ldr d1, [sp, #88] // 8-byte Folded Reload
-; CHECK-FISEL-NEXT:    fadd d0, d1, d0
-; CHECK-FISEL-NEXT:    ldr x30, [sp, #80] // 8-byte Folded Reload
-; CHECK-FISEL-NEXT:    ldp d9, d8, [sp, #64] // 16-byte Folded Reload
-; CHECK-FISEL-NEXT:    ldp d11, d10, [sp, #48] // 16-byte Folded Reload
-; CHECK-FISEL-NEXT:    ldp d13, d12, [sp, #32] // 16-byte Folded Reload
-; CHECK-FISEL-NEXT:    ldp d15, d14, [sp, #16] // 16-byte Folded Reload
-; CHECK-FISEL-NEXT:    add sp, sp, #96
-; CHECK-FISEL-NEXT:    ret
-;
-; CHECK-GISEL-LABEL: streaming_caller_nonstreaming_callee:
-; CHECK-GISEL:       // %bb.0: // %entry
-; CHECK-GISEL-NEXT:    sub sp, sp, #96
-; CHECK-GISEL-NEXT:    stp d15, d14, [sp, #16] // 16-byte Folded Spill
-; CHECK-GISEL-NEXT:    stp d13, d12, [sp, #32] // 16-byte Folded Spill
-; CHECK-GISEL-NEXT:    stp d11, d10, [sp, #48] // 16-byte Folded Spill
-; CHECK-GISEL-NEXT:    stp d9, d8, [sp, #64] // 16-byte Folded Spill
-; CHECK-GISEL-NEXT:    str x30, [sp, #80] // 8-byte Folded Spill
-; CHECK-GISEL-NEXT:    str d0, [sp, #8] // 8-byte Folded Spill
-; CHECK-GISEL-NEXT:    smstop sm
-; CHECK-GISEL-NEXT:    ldr d0, [sp, #8] // 8-byte Folded Reload
-; CHECK-GISEL-NEXT:    bl normal_callee
-; CHECK-GISEL-NEXT:    str d0, [sp, #88] // 8-byte Folded Spill
-; CHECK-GISEL-NEXT:    smstart sm
-; CHECK-GISEL-NEXT:    mov x8, #4631107791820423168 // =0x4045000000000000
-; CHECK-GISEL-NEXT:    fmov d0, x8
-; CHECK-GISEL-NEXT:    ldr d1, [sp, #88] // 8-byte Folded Reload
-; CHECK-GISEL-NEXT:    fadd d0, d1, d0
-; CHECK-GISEL-NEXT:    ldr x30, [sp, #80] // 8-byte Folded Reload
-; CHECK-GISEL-NEXT:    ldp d9, d8, [sp, #64] // 16-byte Folded Reload
-; CHECK-GISEL-NEXT:    ldp d11, d10, [sp, #48] // 16-byte Folded Reload
-; CHECK-GISEL-NEXT:    ldp d13, d12, [sp, #32] // 16-byte Folded Reload
-; CHECK-GISEL-NEXT:    ldp d15, d14, [sp, #16] // 16-byte Folded Reload
-; CHECK-GISEL-NEXT:    add sp, sp, #96
-; CHECK-GISEL-NEXT:    ret
+; CHECK-COMMON-LABEL: streaming_caller_nonstreaming_callee:
+; CHECK-COMMON:       // %bb.0: // %entry
+; CHECK-COMMON-NEXT:    sub sp, sp, #96
+; CHECK-COMMON-NEXT:    stp d15, d14, [sp, #16] // 16-byte Folded Spill
+; CHECK-COMMON-NEXT:    stp d13, d12, [sp, #32] // 16-byte Folded Spill
+; CHECK-COMMON-NEXT:    stp d11, d10, [sp, #48] // 16-byte Folded Spill
+; CHECK-COMMON-NEXT:    stp d9, d8, [sp, #64] // 16-byte Folded Spill
+; CHECK-COMMON-NEXT:    str x30, [sp, #80] // 8-byte Folded Spill
+; CHECK-COMMON-NEXT:    str d0, [sp, #8] // 8-byte Folded Spill
+; CHECK-COMMON-NEXT:    smstop sm
+; CHECK-COMMON-NEXT:    ldr d0, [sp, #8] // 8-byte Folded Reload
+; CHECK-COMMON-NEXT:    bl normal_callee
+; CHECK-COMMON-NEXT:    str d0, [sp, #88] // 8-byte Folded Spill
+; CHECK-COMMON-NEXT:    smstart sm
+; CHECK-COMMON-NEXT:    mov x8, #4631107791820423168 // =0x4045000000000000
+; CHECK-COMMON-NEXT:    fmov d0, x8
+; CHECK-COMMON-NEXT:    ldr d1, [sp, #88] // 8-byte Folded Reload
+; CHECK-COMMON-NEXT:    fadd d0, d1, d0
+; CHECK-COMMON-NEXT:    ldr x30, [sp, #80] // 8-byte Folded Reload
+; CHECK-COMMON-NEXT:    ldp d9, d8, [sp, #64] // 16-byte Folded Reload
+; CHECK-COMMON-NEXT:    ldp d11, d10, [sp, #48] // 16-byte Folded Reload
+; CHECK-COMMON-NEXT:    ldp d13, d12, [sp, #32] // 16-byte Folded Reload
+; CHECK-COMMON-NEXT:    ldp d15, d14, [sp, #16] // 16-byte Folded Reload
+; CHECK-COMMON-NEXT:    add sp, sp, #96
+; CHECK-COMMON-NEXT:    ret
 entry:
   %call = call double @normal_callee(double %x)
   %add = fadd double %call, 4.200000e+01
   ret double %add
 }
 
 define double @locally_streaming_caller_normal_callee(double %x) nounwind noinline optnone "aarch64_pstate_sm_body" {
 ; CHECK-COMMON-LABEL: locally_streaming_caller_normal_callee:

 ; CHECK-COMMON-NEXT:    ldr q0, [sp, #16] // 16-byte Folded Reload
 ; CHECK-COMMON-NEXT:    ldr x30, [sp, #96] // 8-byte Folded Reload
 ; CHECK-COMMON-NEXT:    add sp, sp, #112
 ; CHECK-COMMON-NEXT:    ret
   %res = fadd fp128 %a, %b
   ret fp128 %res
 }
 
-; FIXME: As above this should use Selection DAG to make sure the libcall call is lowered correctly.
+; As above this should use Selection DAG to make sure the libcall call is lowered correctly.
 define double @frem_call_za(double %a, double %b) "aarch64_pstate_za_shared" nounwind {
-; CHECK-FISEL-LABEL: frem_call_za:
-; CHECK-FISEL:       // %bb.0:
-; CHECK-FISEL-NEXT:    stp x29, x30, [sp, #-16]! // 16-byte Folded Spill
-; CHECK-FISEL-NEXT:    mov x29, sp
-; CHECK-FISEL-NEXT:    sub sp, sp, #16
-; CHECK-FISEL-NEXT:    rdsvl x8, #1
-; CHECK-FISEL-NEXT:    mov x9, sp
-; CHECK-FISEL-NEXT:    mul x8, x8, x8
-; CHECK-FISEL-NEXT:    sub x9, x9, x8
-; CHECK-FISEL-NEXT:    mov sp, x9
-; CHECK-FISEL-NEXT:    sub x10, x29, #16
-; CHECK-FISEL-NEXT:    stur x9, [x29, #-16]
-; CHECK-FISEL-NEXT:    sturh w8, [x29, #-8]
-; CHECK-FISEL-NEXT:    msr TPIDR2_EL0, x10
-; CHECK-FISEL-NEXT:    bl fmod
-; CHECK-FISEL-NEXT:    smstart za
-; CHECK-FISEL-NEXT:    sub x0, x29, #16
-; CHECK-FISEL-NEXT:    mrs x8, TPIDR2_EL0
-; CHECK-FISEL-NEXT:    cbnz x8, .LBB10_2
-; CHECK-FISEL-NEXT:  // %bb.1:
-; CHECK-FISEL-NEXT:    bl __arm_tpidr2_restore
-; CHECK-FISEL-NEXT:  .LBB10_2:
-; CHECK-FISEL-NEXT:    msr TPIDR2_EL0, xzr
-; CHECK-FISEL-NEXT:    mov sp, x29
-; CHECK-FISEL-NEXT:    ldp x29, x30, [sp], #16 // 16-byte Folded Reload
-; CHECK-FISEL-NEXT:    ret
-;
-; CHECK-GISEL-LABEL: frem_call_za:
-; CHECK-GISEL:       // %bb.0:
-; CHECK-GISEL-NEXT:    str x30, [sp, #-16]! // 8-byte Folded Spill
-; CHECK-GISEL-NEXT:    bl fmod
-; CHECK-GISEL-NEXT:    ldr x30, [sp], #16 // 8-byte Folded Reload
-; CHECK-GISEL-NEXT:    ret
+; CHECK-COMMON-LABEL: frem_call_za:
+; CHECK-COMMON:       // %bb.0:
+; CHECK-COMMON-NEXT:    stp x29, x30, [sp, #-16]! // 16-byte Folded Spill
+; CHECK-COMMON-NEXT:    mov x29, sp
+; CHECK-COMMON-NEXT:    sub sp, sp, #16
+; CHECK-COMMON-NEXT:    rdsvl x8, #1
+; CHECK-COMMON-NEXT:    mov x9, sp
+; CHECK-COMMON-NEXT:    mul x8, x8, x8
+; CHECK-COMMON-NEXT:    sub x9, x9, x8
+; CHECK-COMMON-NEXT:    mov sp, x9
+; CHECK-COMMON-NEXT:    sub x10, x29, #16
+; CHECK-COMMON-NEXT:    stur x9, [x29, #-16]
+; CHECK-COMMON-NEXT:    sturh w8, [x29, #-8]
+; CHECK-COMMON-NEXT:    msr TPIDR2_EL0, x10
+; CHECK-COMMON-NEXT:    bl fmod
+; CHECK-COMMON-NEXT:    smstart za
+; CHECK-COMMON-NEXT:    sub x0, x29, #16
+; CHECK-COMMON-NEXT:    mrs x8, TPIDR2_EL0
+; CHECK-COMMON-NEXT:    cbnz x8, .LBB10_2
+; CHECK-COMMON-NEXT:  // %bb.1:
+; CHECK-COMMON-NEXT:    bl __arm_tpidr2_restore
+; CHECK-COMMON-NEXT:  .LBB10_2:
+; CHECK-COMMON-NEXT:    msr TPIDR2_EL0, xzr
+; CHECK-COMMON-NEXT:    mov sp, x29
+; CHECK-COMMON-NEXT:    ldp x29, x30, [sp], #16 // 16-byte Folded Reload
+; CHECK-COMMON-NEXT:    ret
   %res = frem double %a, %b
   ret double %res
 }
 
-; FIXME: As above this should use Selection DAG to make sure the libcall is lowered correctly.
+; As above this should use Selection DAG to make sure the libcall is lowered correctly.
 define float @frem_call_sm(float %a, float %b) "aarch64_pstate_sm_enabled" nounwind {
 ; CHECK-COMMON-LABEL: frem_call_sm:
 ; CHECK-COMMON:       // %bb.0:
-; CHECK-COMMON-NEXT:    str x30, [sp, #-16]! // 8-byte Folded Spill
+; CHECK-COMMON-NEXT:    stp d15, d14, [sp, #-80]! // 16-byte Folded Spill
+; CHECK-COMMON-NEXT:    stp d13, d12, [sp, #16] // 16-byte Folded Spill
+; CHECK-COMMON-NEXT:    stp d11, d10, [sp, #32] // 16-byte Folded Spill
+; CHECK-COMMON-NEXT:    stp d9, d8, [sp, #48] // 16-byte Folded Spill
+; CHECK-COMMON-NEXT:    str x30, [sp, #64] // 8-byte Folded Spill
+; CHECK-COMMON-NEXT:    stp s0, s1, [sp, #72] // 8-byte Folded Spill
+; CHECK-COMMON-NEXT:    smstop sm
+; CHECK-COMMON-NEXT:    ldp s0, s1, [sp, #72] // 8-byte Folded Reload
 ; CHECK-COMMON-NEXT:    bl fmodf
-; CHECK-COMMON-NEXT:    ldr x30, [sp], #16 // 8-byte Folded Reload
+; CHECK-COMMON-NEXT:    str s0, [sp, #76] // 4-byte Folded Spill
+; CHECK-COMMON-NEXT:    smstart sm
+; CHECK-COMMON-NEXT:    ldp d9, d8, [sp, #48] // 16-byte Folded Reload
+; CHECK-COMMON-NEXT:    ldp d11, d10, [sp, #32] // 16-byte Folded Reload
+; CHECK-COMMON-NEXT:    ldp d13, d12, [sp, #16] // 16-byte Folded Reload
+; CHECK-COMMON-NEXT:    ldr s0, [sp, #76] // 4-byte Folded Reload
+; CHECK-COMMON-NEXT:    ldr x30, [sp, #64] // 8-byte Folded Reload
+; CHECK-COMMON-NEXT:    ldp d15, d14, [sp], #80 // 16-byte Folded Reload
 ; CHECK-COMMON-NEXT:    ret
   %res = frem float %a, %b
   ret float %res
 }