[llvm] [RISCV] Fold vp.reverse(vp.load(ADDR, MASK)) -> vp.strided.load(ADDR, -1, MASK). (PR #123115)

[Philip Reames via llvm-commits]

================
@@ -0,0 +1,114 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=riscv64 -mattr=+f,+v -verify-machineinstrs < %s | FileCheck %s
+
+define <vscale x 2 x float> @test_reverse_load_combiner(<vscale x 2 x float>* %ptr, i32 zeroext %evl) {
+; CHECK-LABEL: test_reverse_load_combiner:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    slli a2, a1, 2
+; CHECK-NEXT:    add a0, a2, a0
+; CHECK-NEXT:    addi a0, a0, -4
+; CHECK-NEXT:    li a2, -4
+; CHECK-NEXT:    vsetvli zero, a1, e32, m1, ta, ma
+; CHECK-NEXT:    vlse32.v v8, (a0), a2
+; CHECK-NEXT:    ret
+  %head = insertelement <vscale x 2 x i1> undef, i1 1, i32 0
+  %allones = shufflevector <vscale x 2 x i1> %head, <vscale x 2 x i1> undef, <vscale x 2 x i32> zeroinitializer
+
+  %load = call <vscale x 2 x float> @llvm.vp.load.nxv2f32.p0nxv2f32(<vscale x 2 x float>* %ptr, <vscale x 2 x i1> %allones, i32 %evl)
+  %rev = call <vscale x 2 x float> @llvm.experimental.vp.reverse.nxv2f32(<vscale x 2 x float> %load, <vscale x 2 x i1> %allones, i32 %evl)
+  ret <vscale x 2 x float> %rev
+}
+
+define <vscale x 2 x float> @test_load_mask_is_vp_reverse(<vscale x 2 x float>* %ptr, i32 zeroext %evl) {
+; CHECK-LABEL: test_load_mask_is_vp_reverse:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a2, zero, e8, mf4, ta, ma
+; CHECK-NEXT:    vmv.v.i v8, 0
+; CHECK-NEXT:    li a2, 1
+; CHECK-NEXT:    vsetvli zero, zero, e8, mf4, tu, ma
+; CHECK-NEXT:    vmv.s.x v8, a2
+; CHECK-NEXT:    slli a2, a1, 2
+; CHECK-NEXT:    add a0, a2, a0
+; CHECK-NEXT:    li a2, -4
+; CHECK-NEXT:    vsetvli zero, zero, e8, mf4, ta, ma
+; CHECK-NEXT:    vand.vi v8, v8, 1
+; CHECK-NEXT:    vmsne.vi v0, v8, 0
+; CHECK-NEXT:    addi a0, a0, -4
+; CHECK-NEXT:    vsetvli zero, a1, e32, m1, ta, ma
+; CHECK-NEXT:    vlse32.v v8, (a0), a2, v0.t
+; CHECK-NEXT:    ret
+  %head = insertelement <vscale x 2 x i1> undef, i1 1, i32 0
+  %allones = shufflevector <vscale x 2 x i1> %head, <vscale x 2 x i1> undef, <vscale x 2 x i32> zeroinitializer
+
+  %loadmask = call <vscale x 2 x i1> @llvm.experimental.vp.reverse.nxv2i1(<vscale x 2 x i1> %head, <vscale x 2 x i1> %allones, i32 %evl)
+  %load = call <vscale x 2 x float> @llvm.vp.load.nxv2f32.p0nxv2f32(<vscale x 2 x float>* %ptr, <vscale x 2 x i1> %loadmask, i32 %evl)
+  %rev = call <vscale x 2 x float> @llvm.experimental.vp.reverse.nxv2f32(<vscale x 2 x float> %load, <vscale x 2 x i1> %allones, i32 %evl)
+  ret <vscale x 2 x float> %rev
+}
+
+define <vscale x 2 x float> @test_load_mask_not_all_one(<vscale x 2 x float>* %ptr, i32 zeroext %evl) {
+; CHECK-LABEL: test_load_mask_not_all_one:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a2, zero, e8, mf4, ta, ma
+; CHECK-NEXT:    vmclr.m v0
+; CHECK-NEXT:    vsetvli zero, a1, e32, m1, ta, ma
+; CHECK-NEXT:    vle32.v v9, (a0), v0.t
+; CHECK-NEXT:    vid.v v8, v0.t
+; CHECK-NEXT:    addi a1, a1, -1
+; CHECK-NEXT:    vrsub.vx v10, v8, a1, v0.t
+; CHECK-NEXT:    vrgather.vv v8, v9, v10, v0.t
+; CHECK-NEXT:    ret
+  %head = insertelement <vscale x 2 x i1> undef, i1 1, i32 1
+  %notallones = shufflevector <vscale x 2 x i1> %head, <vscale x 2 x i1> undef, <vscale x 2 x i32> zeroinitializer
+
+  %load = call <vscale x 2 x float> @llvm.vp.load.nxv2f32.p0nxv2f32(<vscale x 2 x float>* %ptr, <vscale x 2 x i1> %notallones, i32 %evl)
+  %rev = call <vscale x 2 x float> @llvm.experimental.vp.reverse.nxv2f32(<vscale x 2 x float> %load, <vscale x 2 x i1> %notallones, i32 %evl)
+  ret <vscale x 2 x float> %rev
+}
+
+define <vscale x 2 x float> @test_differnet_evl(<vscale x 2 x float>* %ptr, i32 zeroext %evl1, i32 zeroext %evl2) {
----------------
preames wrote:

We could handle this case by using the min of the two evls if useful.  

https://github.com/llvm/llvm-project/pull/123115