[ARM,MVE] Generate the right instruction for vmaxnmq_m_f16.

[simon_tatham]

Summary

Due to a copy-paste error in the isel patterns, the predicated version
of this intrinsic was expanding to the VMAXNMT.F32 instruction
instead of VMAXNMT.F16. Similarly for vminnm.

[dmgreen]

LGTM

llvm/lib/Target/ARM/ARMInstrMVE.td

def : Pat<(v8f16 (fmaxnum (v8f16 MQPR:$val1), (v8f16 MQPR:$val2))),
             (v8f16 (MVE_VMAXNMf16 (v8f16 MQPR:$val1), (v8f16 MQPR:$val2)))>;
   def : Pat<(v4f32 (int_arm_mve_max_predicated (v4f32 MQPR:$val1), (v4f32 MQPR:$val2),
                           (v4i1 VCCR:$mask), (v4f32 MQPR:$inactive))),
             (v4f32 (MVE_VMAXNMf32 (v4f32 MQPR:$val1), (v4f32 MQPR:$val2),
                           ARMVCCThen, (v4i1 VCCR:$mask),
                           (v4f32 MQPR:$inactive)))>;
   def : Pat<(v8f16 (int_arm_mve_max_predicated (v8f16 MQPR:$val1), (v8f16 MQPR:$val2),
                           (v8i1 VCCR:$mask), (v8f16 MQPR:$inactive))),
-            (v8f16 (MVE_VMAXNMf32 (v8f16 MQPR:$val1), (v8f16 MQPR:$val2),
+            (v8f16 (MVE_VMAXNMf16 (v8f16 MQPR:$val1), (v8f16 MQPR:$val2),
                           ARMVCCThen, (v8i1 VCCR:$mask),
                           (v8f16 MQPR:$inactive)))>;
 }
 
 def MVE_VMINNMf32 : MVE_VMINMAXNM<"vminnm", "f32", 0b0, 0b1>;
 def MVE_VMINNMf16 : MVE_VMINMAXNM<"vminnm", "f16", 0b1, 0b1>;
 
 let Predicates = [HasMVEFloat] in {
   def : Pat<(v4f32 (fminnum (v4f32 MQPR:$val1), (v4f32 MQPR:$val2))),
             (v4f32 (MVE_VMINNMf32 (v4f32 MQPR:$val1), (v4f32 MQPR:$val2)))>;
   def : Pat<(v8f16 (fminnum (v8f16 MQPR:$val1), (v8f16 MQPR:$val2))),
             (v8f16 (MVE_VMINNMf16 (v8f16 MQPR:$val1), (v8f16 MQPR:$val2)))>;
   def : Pat<(v4f32 (int_arm_mve_min_predicated (v4f32 MQPR:$val1), (v4f32 MQPR:$val2),
                           (v4i1 VCCR:$mask), (v4f32 MQPR:$inactive))),
             (v4f32 (MVE_VMINNMf32 (v4f32 MQPR:$val1), (v4f32 MQPR:$val2),
                           ARMVCCThen, (v4i1 VCCR:$mask),
                           (v4f32 MQPR:$inactive)))>;
   def : Pat<(v8f16 (int_arm_mve_min_predicated (v8f16 MQPR:$val1), (v8f16 MQPR:$val2),
                           (v8i1 VCCR:$mask), (v8f16 MQPR:$inactive))),
-            (v8f16 (MVE_VMINNMf32 (v8f16 MQPR:$val1), (v8f16 MQPR:$val2),
+            (v8f16 (MVE_VMINNMf16 (v8f16 MQPR:$val1), (v8f16 MQPR:$val2),
                           ARMVCCThen, (v8i1 VCCR:$mask),
                           (v8f16 MQPR:$inactive)))>;
 }
 
 
 class MVE_VMINMAX<string iname, string suffix, bit U, bits<2> size,
               bit bit_4, list<dag> pattern=[]>
   : MVE_comp<NoItinerary, iname, suffix, "", pattern> {

llvm/test/CodeGen/Thumb2/mve-intrinsics/vmaxnmq.ll

 
 declare <4 x float> @llvm.maxnum.v4f32(<4 x float>, <4 x float>) #1
 
 define arm_aapcs_vfpcc <8 x half> @test_vmaxnmq_m_f16(<8 x half> %inactive, <8 x half> %a, <8 x half> %b, i16 zeroext %p) local_unnamed_addr #0 {
 ; CHECK-LABEL: test_vmaxnmq_m_f16:
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    vmsr p0, r0
 ; CHECK-NEXT:    vpst
-; CHECK-NEXT:    vmaxnmt.f32 q0, q1, q2
+; CHECK-NEXT:    vmaxnmt.f16 q0, q1, q2
 ; CHECK-NEXT:    bx lr
 entry:
   %0 = zext i16 %p to i32
   %1 = tail call <8 x i1> @llvm.arm.mve.pred.i2v.v8i1(i32 %0)
   %2 = tail call <8 x half> @llvm.arm.mve.max.predicated.v8f16.v8i1(<8 x half> %a, <8 x half> %b, <8 x i1> %1, <8 x half> %inactive)
   ret <8 x half> %2
 }
 

 
 declare <4 x float> @llvm.arm.mve.max.predicated.v4f32.v4i1(<4 x float>, <4 x float>, <4 x i1>, <4 x float>) #2
 
 define arm_aapcs_vfpcc <8 x half> @test_vmaxnmq_x_f16(<8 x half> %a, <8 x half> %b, i16 zeroext %p) local_unnamed_addr #0 {
 ; CHECK-LABEL: test_vmaxnmq_x_f16:
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    vmsr p0, r0
 ; CHECK-NEXT:    vpst
-; CHECK-NEXT:    vmaxnmt.f32 q0, q0, q1
+; CHECK-NEXT:    vmaxnmt.f16 q0, q0, q1
 ; CHECK-NEXT:    bx lr
 entry:
   %0 = zext i16 %p to i32
   %1 = tail call <8 x i1> @llvm.arm.mve.pred.i2v.v8i1(i32 %0)
   %2 = tail call <8 x half> @llvm.arm.mve.max.predicated.v8f16.v8i1(<8 x half> %a, <8 x half> %b, <8 x i1> %1, <8 x half> undef)
   ret <8 x half> %2
 }
 

llvm/test/CodeGen/Thumb2/mve-intrinsics/vminnmq.ll

 
 declare <4 x float> @llvm.minnum.v4f32(<4 x float>, <4 x float>) #1
 
 define arm_aapcs_vfpcc <8 x half> @test_vminnmq_m_f16(<8 x half> %inactive, <8 x half> %a, <8 x half> %b, i16 zeroext %p) local_unnamed_addr #0 {
 ; CHECK-LABEL: test_vminnmq_m_f16:
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    vmsr p0, r0
 ; CHECK-NEXT:    vpst
-; CHECK-NEXT:    vminnmt.f32 q0, q1, q2
+; CHECK-NEXT:    vminnmt.f16 q0, q1, q2
 ; CHECK-NEXT:    bx lr
 entry:
   %0 = zext i16 %p to i32
   %1 = tail call <8 x i1> @llvm.arm.mve.pred.i2v.v8i1(i32 %0)
   %2 = tail call <8 x half> @llvm.arm.mve.min.predicated.v8f16.v8i1(<8 x half> %a, <8 x half> %b, <8 x i1> %1, <8 x half> %inactive)
   ret <8 x half> %2
 }
 

 
 declare <4 x float> @llvm.arm.mve.min.predicated.v4f32.v4i1(<4 x float>, <4 x float>, <4 x i1>, <4 x float>) #2
 
 define arm_aapcs_vfpcc <8 x half> @test_vminnmq_x_f16(<8 x half> %a, <8 x half> %b, i16 zeroext %p) local_unnamed_addr #0 {
 ; CHECK-LABEL: test_vminnmq_x_f16:
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    vmsr p0, r0
 ; CHECK-NEXT:    vpst
-; CHECK-NEXT:    vminnmt.f32 q0, q0, q1
+; CHECK-NEXT:    vminnmt.f16 q0, q0, q1
 ; CHECK-NEXT:    bx lr
 entry:
   %0 = zext i16 %p to i32
   %1 = tail call <8 x i1> @llvm.arm.mve.pred.i2v.v8i1(i32 %0)
   %2 = tail call <8 x half> @llvm.arm.mve.min.predicated.v8f16.v8i1(<8 x half> %a, <8 x half> %b, <8 x i1> %1, <8 x half> undef)
   ret <8 x half> %2
 }