[llvm] Add optimisation for SVE intrinsics with no active lanes (PR #73964)

[Paul Walker via llvm-commits]

================
@@ -1891,91 +1944,117 @@ AArch64TTIImpl::instCombineIntrinsic(InstCombiner &IC,
   case Intrinsic::aarch64_sve_ptest_last:
     return instCombineSVEPTest(IC, II);
   case Intrinsic::aarch64_sve_fabd:
-    return instCombineSVEAllActive(II, Intrinsic::aarch64_sve_fabd_u);
+  case Intrinsic::aarch64_sve_fabd_u:
+    return instCombineSVEAllOrNoActive(IC, II, Intrinsic::aarch64_sve_fabd_u);
   case Intrinsic::aarch64_sve_fadd:
     return instCombineSVEVectorFAdd(IC, II);
   case Intrinsic::aarch64_sve_fadd_u:
     return instCombineSVEVectorFAddU(IC, II);
   case Intrinsic::aarch64_sve_fdiv:
-    return instCombineSVEAllActive(II, Intrinsic::aarch64_sve_fdiv_u);
+  case Intrinsic::aarch64_sve_fdiv_u:
+    return instCombineSVEAllOrNoActive(IC, II, Intrinsic::aarch64_sve_fdiv_u);
   case Intrinsic::aarch64_sve_fmax:
-    return instCombineSVEAllActive(II, Intrinsic::aarch64_sve_fmax_u);
+  case Intrinsic::aarch64_sve_fmax_u:
+    return instCombineSVEAllOrNoActive(IC, II, Intrinsic::aarch64_sve_fmax_u);
   case Intrinsic::aarch64_sve_fmaxnm:
-    return instCombineSVEAllActive(II, Intrinsic::aarch64_sve_fmaxnm_u);
+  case Intrinsic::aarch64_sve_fmaxnm_u:
+    return instCombineSVEAllOrNoActive(IC, II, Intrinsic::aarch64_sve_fmaxnm_u);
   case Intrinsic::aarch64_sve_fmin:
-    return instCombineSVEAllActive(II, Intrinsic::aarch64_sve_fmin_u);
+  case Intrinsic::aarch64_sve_fmin_u:
+    return instCombineSVEAllOrNoActive(IC, II, Intrinsic::aarch64_sve_fmin_u);
   case Intrinsic::aarch64_sve_fminnm:
-    return instCombineSVEAllActive(II, Intrinsic::aarch64_sve_fminnm_u);
+  case Intrinsic::aarch64_sve_fminnm_u:
+    return instCombineSVEAllOrNoActive(IC, II, Intrinsic::aarch64_sve_fminnm_u);
   case Intrinsic::aarch64_sve_fmla:
-    return instCombineSVEAllActive(II, Intrinsic::aarch64_sve_fmla_u);
+  case Intrinsic::aarch64_sve_fmla_u:
+    return instCombineSVEAllOrNoActive(IC, II, Intrinsic::aarch64_sve_fmla_u);
   case Intrinsic::aarch64_sve_fmls:
-    return instCombineSVEAllActive(II, Intrinsic::aarch64_sve_fmls_u);
+  case Intrinsic::aarch64_sve_fmls_u:
+    return instCombineSVEAllOrNoActive(IC, II, Intrinsic::aarch64_sve_fmls_u);
   case Intrinsic::aarch64_sve_fmul:
   case Intrinsic::aarch64_sve_fmul_u:
     return instCombineSVEVectorMul(IC, II, Intrinsic::aarch64_sve_fmul_u);
   case Intrinsic::aarch64_sve_fmulx:
-    return instCombineSVEAllActive(II, Intrinsic::aarch64_sve_fmulx_u);
+  case Intrinsic::aarch64_sve_fmulx_u:
+    return instCombineSVEAllOrNoActive(IC, II, Intrinsic::aarch64_sve_fmulx_u);
   case Intrinsic::aarch64_sve_fnmla:
-    return instCombineSVEAllActive(II, Intrinsic::aarch64_sve_fnmla_u);
+  case Intrinsic::aarch64_sve_fnmla_u:
+    return instCombineSVEAllOrNoActive(IC, II, Intrinsic::aarch64_sve_fnmla_u);
   case Intrinsic::aarch64_sve_fnmls:
-    return instCombineSVEAllActive(II, Intrinsic::aarch64_sve_fnmls_u);
+  case Intrinsic::aarch64_sve_fnmls_u:
+    return instCombineSVEAllOrNoActive(IC, II, Intrinsic::aarch64_sve_fnmls_u);
   case Intrinsic::aarch64_sve_fsub:
     return instCombineSVEVectorFSub(IC, II);
   case Intrinsic::aarch64_sve_fsub_u:
     return instCombineSVEVectorFSubU(IC, II);
   case Intrinsic::aarch64_sve_add:
     return instCombineSVEVectorAdd(IC, II);
   case Intrinsic::aarch64_sve_add_u:
-    return instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_mul_u,
-                                             Intrinsic::aarch64_sve_mla_u>(
-        IC, II, true);
+    return instCombineSVEVectorAddU(IC, II);
   case Intrinsic::aarch64_sve_mla:
-    return instCombineSVEAllActive(II, Intrinsic::aarch64_sve_mla_u);
+  case Intrinsic::aarch64_sve_mla_u:
+    return instCombineSVEAllOrNoActive(IC, II, Intrinsic::aarch64_sve_mla_u);
   case Intrinsic::aarch64_sve_mls:
-    return instCombineSVEAllActive(II, Intrinsic::aarch64_sve_mls_u);
+  case Intrinsic::aarch64_sve_mls_u:
+    return instCombineSVEAllOrNoActive(IC, II, Intrinsic::aarch64_sve_mls_u);
   case Intrinsic::aarch64_sve_mul:
   case Intrinsic::aarch64_sve_mul_u:
     return instCombineSVEVectorMul(IC, II, Intrinsic::aarch64_sve_mul_u);
   case Intrinsic::aarch64_sve_sabd:
-    return instCombineSVEAllActive(II, Intrinsic::aarch64_sve_sabd_u);
+  case Intrinsic::aarch64_sve_sabd_u:
+    return instCombineSVEAllOrNoActive(IC, II, Intrinsic::aarch64_sve_sabd_u);
   case Intrinsic::aarch64_sve_smax:
-    return instCombineSVEAllActive(II, Intrinsic::aarch64_sve_smax_u);
+  case Intrinsic::aarch64_sve_smax_u:
+    return instCombineSVEAllOrNoActive(IC, II, Intrinsic::aarch64_sve_smax_u);
   case Intrinsic::aarch64_sve_smin:
-    return instCombineSVEAllActive(II, Intrinsic::aarch64_sve_smin_u);
+  case Intrinsic::aarch64_sve_smin_u:
+    return instCombineSVEAllOrNoActive(IC, II, Intrinsic::aarch64_sve_smin_u);
   case Intrinsic::aarch64_sve_smulh:
-    return instCombineSVEAllActive(II, Intrinsic::aarch64_sve_smulh_u);
+  case Intrinsic::aarch64_sve_smulh_u:
+    return instCombineSVEAllOrNoActive(IC, II, Intrinsic::aarch64_sve_smulh_u);
   case Intrinsic::aarch64_sve_sub:
     return instCombineSVEVectorSub(IC, II);
   case Intrinsic::aarch64_sve_sub_u:
-    return instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_mul_u,
-                                             Intrinsic::aarch64_sve_mls_u>(
-        IC, II, true);
+    return instCombineSVEVectorSubU(IC, II);
   case Intrinsic::aarch64_sve_uabd:
-    return instCombineSVEAllActive(II, Intrinsic::aarch64_sve_uabd_u);
+  case Intrinsic::aarch64_sve_uabd_u:
+    return instCombineSVEAllOrNoActive(IC, II, Intrinsic::aarch64_sve_uabd_u);
----------------
paulwalker-arm wrote:

I don't like the idea of handling the two intrinsic types together because I recall we only recently split similar instances apart because it was awkward to implement `_u` and non-`_u` specific optimisation. You can already see this for this patch where only half of `instCombineSVEAllOrNoActive` is applicable.  I'd rather have dedicated functions for the specific combine so by all means have `instCombineSVEAllOrNoActive` in place of `instCombineSVEAllActive` but I think the `_u` forms should have a dedicated function.

That said, I do wonder how useful the combines that produce `undef` are from a practical sense.  I would think that a sign the source material is likely bogus rather than a legitimate optimisation opportunity.

https://github.com/llvm/llvm-project/pull/73964