[llvm] [AArch64] Generate zeroing forms of certain SVE2.2 instructions (7/11) (PR #116833)

[Momchil Velikov via llvm-commits]

================
@@ -0,0 +1,258 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mattr=+sve2   < %s | FileCheck %s
+; RUN: llc -mattr=+sve2p2 < %s | FileCheck %s -check-prefix CHECK-2p2
+
+; RUN: llc -mattr=+sme2   -force-streaming < %s | FileCheck %s
+; RUN: llc -mattr=+sme2p2 -force-streaming < %s | FileCheck %s -check-prefix CHECK-2p2
+
+target triple = "aarch64-linux"
+
+define <vscale x 8 x i16> @test_svlogb_f16_x_1(<vscale x 8 x i1> %pg, <vscale x 8 x half> %x) {
+; CHECK-LABEL: test_svlogb_f16_x_1:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    flogb z0.h, p0/m, z0.h
+; CHECK-NEXT:    ret
+;
+; CHECK-2p2-LABEL: test_svlogb_f16_x_1:
+; CHECK-2p2:       // %bb.0: // %entry
+; CHECK-2p2-NEXT:    flogb z0.h, p0/z, z0.h
+; CHECK-2p2-NEXT:    ret
+entry:
+  %0 = tail call <vscale x 8 x i16> @llvm.aarch64.sve.flogb.nxv8f16(<vscale x 8 x i16> poison, <vscale x 8 x i1> %pg, <vscale x 8 x half> %x)
+  ret <vscale x 8 x i16> %0
+}
+
+define <vscale x 8 x i16> @test_svlogb_f16_x_2(<vscale x 8 x i1> %pg, double %z0, <vscale x 8 x half> %x) {
+; CHECK-LABEL: test_svlogb_f16_x_2:
+; CHECK:       // %bb.0: // %entry
----------------
momchil-velikov wrote:

Change to the zeroing version to avoid the false dependency on the destination register.
In `flogb z0.h, p0/m, z1.h` `z0` is input/output, but in `flogb z0.h, p0/z, z1.h` it's only output.
For some of the instructions in this series of patches the `movprx` optimisation is not implemented, thuszeroing forms, if available, are even more efficient.

https://github.com/llvm/llvm-project/pull/116833