[clang] [AArch64][SME2] Add _x2/_x4 svqrshr builtins. (PR #74100)

Kerry McLaughlin via cfe-commits cfe-commits at lists.llvm.org
Fri Dec 1 08:42:06 PST 2023


================
@@ -2119,6 +2119,21 @@ let TargetGuard = "sme2" in {
   // 2-way and 4-way selects
   def SVSEL_X2  : SInst<"svsel[_{d}_x2]", "2}22", "cUcsUsiUilUlbhfd", MergeNone, "aarch64_sve_sel_x2", [IsStreaming], []>;
   def SVSEL_X4  : SInst<"svsel[_{d}_x4]", "4}44", "cUcsUsiUilUlbhfd", MergeNone, "aarch64_sve_sel_x4", [IsStreaming], []>;
+
+  def SVQRSHRN_X4   : SInst<"svqrshrn[_{0}_{d}_x4]", "q4i", "il",   MergeNone, "aarch64_sve_sqrshrn_x4", [IsStreaming], [ImmCheck<1, ImmCheckShiftRight, 0>]>;
+  def SVUQRSHRN_X4  : SInst<"svqrshrn[_{0}_{d}_x4]", "b4i", "UiUl", MergeNone, "aarch64_sve_uqrshrn_x4", [IsStreaming], [ImmCheck<1, ImmCheckShiftRight, 0>]>;
+
+  // SQRSHR / UQRSHR
+  def SVQRSHR_X2  : SInst<"svqrshr[_{0}_{d}_x2]", "h2i", "i",    MergeNone, "aarch64_sve_sqrshr_x2", [IsStreaming], [ImmCheck<1, ImmCheck1_16>]>;
+  def SVUQRSHR_X2 : SInst<"svqrshr[_{0}_{d}_x2]", "e2i", "Ui",   MergeNone, "aarch64_sve_uqrshr_x2", [IsStreaming], [ImmCheck<1, ImmCheck1_16>]>;
+  def SVQRSHR_X4  : SInst<"svqrshr[_{0}_{d}_x4]", "q4i", "il",   MergeNone, "aarch64_sve_sqrshr_x4", [IsStreaming], [ImmCheck<1, ImmCheckShiftRight, 0>]>;
+  def SVUQRSHR_X4 : SInst<"svqrshr[_{0}_{d}_x4]", "b4i", "UiUl", MergeNone, "aarch64_sve_uqrshr_x4", [IsStreaming], [ImmCheck<1, ImmCheckShiftRight, 0>]>;
+
+  // SQRSHRU
+  def SVSQRSHRU_X2 : SInst<"svqrshru[_{0}_{d}_x2]", "e2i", "i",  MergeNone, "aarch64_sve_sqrshru_x2", [IsStreaming], [ImmCheck<1, ImmCheck1_16>]>;
+  def SVSQRSHRU_X4 : SInst<"svqrshru[_{0}_{d}_x4]", "b4i", "il", MergeNone, "aarch64_sve_sqrshru_x4", [IsStreaming], [ImmCheck<1, ImmCheckShiftRight, 0>]>;
+
+  def SVSQRSHRUN_X4 : SInst<"svqrshrun[_{0}_{d}_x4]", "b4i", "il", MergeNone, "aarch64_sve_sqrshrun_x4", [IsStreaming], [ImmCheck<1, ImmCheckShiftRight, 0>]>;
 }
----------------
kmclaughlin-arm wrote:

The qrshr* builtins have changed again in the latest ACLE update (https://github.com/ARM-software/acle/pull/278):
```
arm_neon.h shift-by-immediate instructions use an _n suffix
to indicate that the shift amount is scalar. arm_sve.h
carried this across to the full/non-overloaded forms of SVE
immediate shifts. It seems worth doing the same here for
consistency, and to protect against vector-vector forms
being added in future.
```
Please can you apply this change here?

https://github.com/llvm/llvm-project/pull/74100


More information about the cfe-commits mailing list