1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 4
2 ; RUN: llc < %s -mtriple=riscv64 -mattr=+v | FileCheck %s
4 ; This showcases a miscompile that was fixed in #90573:
5 ; - The memset will be type-legalized to a 512 bit store + 2 x 128 bit stores.
6 ; - the load and store of q aliases the upper 128 bits store of p.
7 ; - The aliasing 128 bit store will be between the chain of the scalar
10 ; t54: ch = store<(store (s512) into %ir.p, align 1)> t0, ...
11 ; t51: ch = store<(store (s128) into %ir.p + 64, align 1)> t0, ...
13 ; t44: i64,ch = load<(load (s32) from %ir.q), sext from i32> t0, ...
14 ; t50: ch = store<(store (s128) into %ir.p + 80, align 1)> t44:1, ...
15 ; t46: ch = store<(store (s32) into %ir.q), trunc to i32> t50, ...
17 ; Previously, the scalar load/store was incorrectly combined away:
19 ; t54: ch = store<(store (s512) into %ir.p, align 1)> t0, ...
20 ; t51: ch = store<(store (s128) into %ir.p + 64, align 1)> t0, ...
23 ; t50: ch = store<(store (s128) into %ir.p + 80, align 1)> t44:1, ...
25 ; See also pr83017.ll: This is the same code, but relies on vscale_range instead
26 ; of -riscv-v-vector-bits-max=128.
27 define void @f(ptr %p) vscale_range(2,2) {
30 ; CHECK-NEXT: lw a1, 84(a0)
31 ; CHECK-NEXT: addi a2, a0, 80
32 ; CHECK-NEXT: vsetivli zero, 16, e8, m1, ta, ma
33 ; CHECK-NEXT: vmv.v.i v8, 0
34 ; CHECK-NEXT: vs1r.v v8, (a2)
35 ; CHECK-NEXT: vsetvli a2, zero, e8, m4, ta, ma
36 ; CHECK-NEXT: vmv.v.i v12, 0
37 ; CHECK-NEXT: vs4r.v v12, (a0)
38 ; CHECK-NEXT: addi a2, a0, 64
39 ; CHECK-NEXT: vs1r.v v8, (a2)
40 ; CHECK-NEXT: sw a1, 84(a0)
42 %q = getelementptr inbounds i8, ptr %p, i64 84
44 call void @llvm.memset.p0.i64(ptr %p, i8 0, i64 96, i1 false)