[llvm] [AMDGPU] Tweak sched model latencies for VALU fast-forward cases (PR #198475)

[Jay Foad via llvm-commits]

================
@@ -751,28 +751,28 @@ define <16 x ptr> @addrspacecast_v16p5_to_v16p0(<16 x ptr addrspace(5)> %ptr) {
 ; CI-NEXT:    v_cndmask_b32_e32 v35, 0, v2, vcc
 ; CI-NEXT:    v_cndmask_b32_e32 v32, 0, v31, vcc
 ; CI-NEXT:    v_cmp_ne_u32_e32 vcc, -1, v3
-; CI-NEXT:    v_cmp_ne_u32_e64 s[4:5], -1, v4
-; CI-NEXT:    v_cmp_ne_u32_e64 s[10:11], -1, v7
 ; CI-NEXT:    v_cndmask_b32_e32 v36, 0, v3, vcc
+; CI-NEXT:    v_cmp_ne_u32_e64 s[4:5], -1, v4
 ; CI-NEXT:    v_cndmask_b32_e64 v48, 0, v4, s[4:5]
-; CI-NEXT:    v_cndmask_b32_e64 v37, 0, v5, s[6:7]
+; CI-NEXT:    v_cmp_ne_u32_e64 s[8:9], -1, v6
 ; CI-NEXT:    v_cndmask_b32_e64 v33, 0, v6, s[8:9]
+; CI-NEXT:    v_cmp_ne_u32_e64 s[10:11], -1, v7
 ; CI-NEXT:    v_cndmask_b32_e64 v38, 0, v7, s[10:11]
 ; CI-NEXT:    v_cmp_ne_u32_e64 s[12:13], -1, v8
-; CI-NEXT:    v_cmp_ne_u32_e64 s[14:15], -1, v9
----------------
jayfoad wrote:

No, for all of these v_cmp results there is a second use ~10 lines further down in another v_cndmask. So the lifetimes still overlap.

https://github.com/llvm/llvm-project/pull/198475