[llvm] Add `llvm.vector.partial.reduce.fadd` intrinsic (PR #159776)

[Paul Walker via llvm-commits]

================
@@ -0,0 +1,93 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
+; RUN: llc -mattr=+sve2 < %s | FileCheck %s --check-prefixes=CHECK,SVE2
+; RUN: llc -global-isel -global-isel-abort=2 -mattr=+sve2 < %s | FileCheck %s --check-prefixes=CHECK,SVE2
+; RUN: llc -mattr=+sve2p1 < %s | FileCheck %s --check-prefixes=CHECK,SVE2P1
+; RUN: llc -global-isel -global-isel-abort=2 -mattr=+sve2p1 < %s | FileCheck %s --check-prefixes=CHECK,SVE2P1
+
+target triple = "aarch64-linux-gnu"
+
+define <vscale x 4 x float> @fdot_wide_nxv4f32(<vscale x 4 x float> %acc, <vscale x 8 x half> %a, <vscale x 8 x half> %b) {
+; SVE2-LABEL: fdot_wide_nxv4f32:
+; SVE2:       // %bb.0: // %entry
+; SVE2-NEXT:    uunpklo z3.s, z1.h
+; SVE2-NEXT:    uunpklo z4.s, z2.h
+; SVE2-NEXT:    ptrue p0.s
+; SVE2-NEXT:    uunpkhi z1.s, z1.h
+; SVE2-NEXT:    uunpkhi z2.s, z2.h
+; SVE2-NEXT:    fcvt z3.s, p0/m, z3.h
+; SVE2-NEXT:    fcvt z4.s, p0/m, z4.h
+; SVE2-NEXT:    fcvt z1.s, p0/m, z1.h
+; SVE2-NEXT:    fcvt z2.s, p0/m, z2.h
+; SVE2-NEXT:    fmul z3.s, z3.s, z4.s
+; SVE2-NEXT:    fmul z1.s, z1.s, z2.s
+; SVE2-NEXT:    fadd z0.s, z0.s, z3.s
+; SVE2-NEXT:    fadd z0.s, z0.s, z1.s
+; SVE2-NEXT:    ret
+;
+; SVE2P1-LABEL: fdot_wide_nxv4f32:
+; SVE2P1:       // %bb.0: // %entry
+; SVE2P1-NEXT:    fdot z0.s, z1.h, z2.h
+; SVE2P1-NEXT:    ret
+entry:
+  %a.wide = fpext <vscale x 8 x half> %a to <vscale x 8 x float>
+  %b.wide = fpext <vscale x 8 x half> %b to <vscale x 8 x float>
+  %mult = fmul <vscale x 8 x float> %a.wide, %b.wide
+  %partial.reduce = call <vscale x 4 x float> @llvm.vector.partial.reduce.fadd(<vscale x 4 x float> %acc, <vscale x 8 x float> %mult)
+  ret <vscale x 4 x float> %partial.reduce
----------------
paulwalker-arm wrote:

Yes, along with other fast math flags like reassociation.

https://github.com/llvm/llvm-project/pull/159776