[llvm] [AMDGPU] Add patterns for V_CMP_O/U (PR #69157)

[Jay Foad via llvm-commits]

================
@@ -494,6 +494,121 @@ define amdgpu_kernel void @v_fcmp_f32_ole(ptr addrspace(1) %out, float %src) {
   ret void
 }
 
+define amdgpu_kernel void @v_fcmp_f32_o(ptr addrspace(1) %out, float %src) {
+; SDAG-GFX11-LABEL: v_fcmp_f32_o:
+; SDAG-GFX11:       ; %bb.0:
+; SDAG-GFX11-NEXT:    s_clause 0x1
+; SDAG-GFX11-NEXT:    s_load_b32 s2, s[0:1], 0x2c
+; SDAG-GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
+; SDAG-GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; SDAG-GFX11-NEXT:    v_cmp_o_f32_e64 s2, 0x42c80000, s2
+; SDAG-GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; SDAG-GFX11-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
+; SDAG-GFX11-NEXT:    global_store_b32 v0, v1, s[0:1]
+; SDAG-GFX11-NEXT:    s_nop 0
+; SDAG-GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; SDAG-GFX11-NEXT:    s_endpgm
+;
+; SDAG-GFX10-LABEL: v_fcmp_f32_o:
+; SDAG-GFX10:       ; %bb.0:
+; SDAG-GFX10-NEXT:    s_clause 0x1
+; SDAG-GFX10-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; SDAG-GFX10-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; SDAG-GFX10-NEXT:    v_mov_b32_e32 v0, 0
+; SDAG-GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; SDAG-GFX10-NEXT:    v_cmp_o_f32_e64 s0, 0x42c80000, s4
+; SDAG-GFX10-NEXT:    v_mov_b32_e32 v1, s0
+; SDAG-GFX10-NEXT:    global_store_dword v0, v1, s[2:3]
+; SDAG-GFX10-NEXT:    s_endpgm
+;
+; GISEL-GFX11-LABEL: v_fcmp_f32_o:
+; GISEL-GFX11:       ; %bb.0:
+; GISEL-GFX11-NEXT:    s_clause 0x1
+; GISEL-GFX11-NEXT:    s_load_b32 s2, s[0:1], 0x2c
+; GISEL-GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
+; GISEL-GFX11-NEXT:    v_mov_b32_e32 v1, 0
+; GISEL-GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GISEL-GFX11-NEXT:    v_cmp_o_f32_e64 s2, 0x42c80000, s2
+; GISEL-GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GISEL-GFX11-NEXT:    v_mov_b32_e32 v0, s2
+; GISEL-GFX11-NEXT:    global_store_b32 v1, v0, s[0:1]
+; GISEL-GFX11-NEXT:    s_nop 0
+; GISEL-GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GISEL-GFX11-NEXT:    s_endpgm
+;
+; GISEL-GFX10-LABEL: v_fcmp_f32_o:
+; GISEL-GFX10:       ; %bb.0:
+; GISEL-GFX10-NEXT:    s_clause 0x1
+; GISEL-GFX10-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GISEL-GFX10-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GISEL-GFX10-NEXT:    v_mov_b32_e32 v1, 0
+; GISEL-GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GISEL-GFX10-NEXT:    v_cmp_o_f32_e64 s0, 0x42c80000, s4
+; GISEL-GFX10-NEXT:    v_mov_b32_e32 v0, s0
+; GISEL-GFX10-NEXT:    global_store_dword v1, v0, s[2:3]
+; GISEL-GFX10-NEXT:    s_endpgm
+  %result = call i32 @llvm.amdgcn.fcmp.f32(float %src, float 100.00, i32 7)
+  store i32 %result, ptr addrspace(1) %out
+  ret void
+}
+
+define amdgpu_kernel void @v_fcmp_f32_uo(ptr addrspace(1) %out, float %src) {
+; SDAG-GFX11-LABEL: v_fcmp_f32_uo:
+; SDAG-GFX11:       ; %bb.0:
+; SDAG-GFX11-NEXT:    s_clause 0x1
+; SDAG-GFX11-NEXT:    s_load_b32 s2, s[0:1], 0x2c
+; SDAG-GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
+; SDAG-GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; SDAG-GFX11-NEXT:    v_cmp_u_f32_e64 s2, 0x42c80000, s2
+; SDAG-GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; SDAG-GFX11-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
+; SDAG-GFX11-NEXT:    global_store_b32 v0, v1, s[0:1]
+; SDAG-GFX11-NEXT:    s_nop 0
+; SDAG-GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; SDAG-GFX11-NEXT:    s_endpgm
+;
+; SDAG-GFX10-LABEL: v_fcmp_f32_uo:
+; SDAG-GFX10:       ; %bb.0:
+; SDAG-GFX10-NEXT:    s_clause 0x1
+; SDAG-GFX10-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; SDAG-GFX10-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; SDAG-GFX10-NEXT:    v_mov_b32_e32 v0, 0
+; SDAG-GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; SDAG-GFX10-NEXT:    v_cmp_u_f32_e64 s0, 0x42c80000, s4
+; SDAG-GFX10-NEXT:    v_mov_b32_e32 v1, s0
+; SDAG-GFX10-NEXT:    global_store_dword v0, v1, s[2:3]
+; SDAG-GFX10-NEXT:    s_endpgm
+;
+; GISEL-GFX11-LABEL: v_fcmp_f32_uo:
+; GISEL-GFX11:       ; %bb.0:
+; GISEL-GFX11-NEXT:    s_clause 0x1
+; GISEL-GFX11-NEXT:    s_load_b32 s2, s[0:1], 0x2c
+; GISEL-GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
+; GISEL-GFX11-NEXT:    v_mov_b32_e32 v1, 0
+; GISEL-GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GISEL-GFX11-NEXT:    v_cmp_u_f32_e64 s2, 0x42c80000, s2
+; GISEL-GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GISEL-GFX11-NEXT:    v_mov_b32_e32 v0, s2
+; GISEL-GFX11-NEXT:    global_store_b32 v1, v0, s[0:1]
+; GISEL-GFX11-NEXT:    s_nop 0
+; GISEL-GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GISEL-GFX11-NEXT:    s_endpgm
+;
+; GISEL-GFX10-LABEL: v_fcmp_f32_uo:
+; GISEL-GFX10:       ; %bb.0:
+; GISEL-GFX10-NEXT:    s_clause 0x1
+; GISEL-GFX10-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GISEL-GFX10-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GISEL-GFX10-NEXT:    v_mov_b32_e32 v1, 0
+; GISEL-GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GISEL-GFX10-NEXT:    v_cmp_u_f32_e64 s0, 0x42c80000, s4
+; GISEL-GFX10-NEXT:    v_mov_b32_e32 v0, s0
+; GISEL-GFX10-NEXT:    global_store_dword v1, v0, s[2:3]
+; GISEL-GFX10-NEXT:    s_endpgm
+  %result = call i32 @llvm.amdgcn.fcmp.f32(float %src, float 100.00, i32 8)
----------------
jayfoad wrote:

Can the problem also occur when you use regular fcmp instructions (possibly combined with a ballot intrinsic)? Or does it only affect the llvm.amdgcn.fcmp intrinsic?

The llvm.amdgcn.fcmp intrinsic should be deprecated.

https://github.com/llvm/llvm-project/pull/69157