[llvm] [CodeGenPrepare] Reverse the canonicalization of isInf/isNanOrInf (PR #81572)

[Matt Arsenault via llvm-commits]

================
@@ -173,8 +177,10 @@ define amdgpu_kernel void @test_isfinite_pattern_0(ptr addrspace(1) nocapture %o
 ; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
 ; GFX11-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    v_cmp_class_f32_e64 s2, s2, 0x1f8
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_cmp_o_f32_e64 s3, s2, s2
+; GFX11-NEXT:    v_cmp_class_f32_e64 s2, s2, 0x1fb
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_b32 s2, s3, s2
----------------
arsenm wrote:

Regression 

https://github.com/llvm/llvm-project/pull/81572