[llvm] Allow shifting instead of masks if the types are legal. (PR #162732)

[Matt Arsenault via llvm-commits]

================
@@ -121,12 +121,23 @@ define i32 @bzhi32_c0(i32 %val, i32 %numlowbits) nounwind {
 }
 
 define i32 @bzhi32_c0_clamp(i32 %val, i32 %numlowbits) nounwind {
-; GCN-LABEL: bzhi32_c0_clamp:
-; GCN:       ; %bb.0:
-; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT:    v_and_b32_e32 v1, 31, v1
-; GCN-NEXT:    v_bfe_u32 v0, v0, 0, v1
-; GCN-NEXT:    s_setpc_b64 s[30:31]
+; SI-LABEL: bzhi32_c0_clamp:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v1, 31, v1
+; SI-NEXT:    v_sub_i32_e32 v1, vcc, 32, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v0, v1, v0
+; SI-NEXT:    v_lshrrev_b32_e32 v0, v1, v0
----------------
arsenm wrote:

Regression, broke a BFE pattern? 

https://github.com/llvm/llvm-project/pull/162732