================
@@ -0,0 +1,136 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
UTC_ARGS: --version 5
+; RUN: llc -global-isel=0 -march=amdgcn -mcpu=gfx1200 < %s | FileCheck
--check-prefixes=GCN,SDAG %s
+; RUN: llc -global-isel=1 -march=amdgcn -mcpu=gfx1200 < %s | FileCheck
--check-prefixes=GCN,GISEL %s
+
+define amdgpu_ps void @prefetch_data_sgpr_base_sgpr_len(ptr addrspace(4) inreg
%ptr, i32 inreg %len) {
+; GCN-LABEL: prefetch_data_sgpr_base_sgpr_len:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_prefetch_data s[0:1], 0x0, s2, 0
+; GCN-NEXT: s_endpgm
+entry:
+ tail call void @llvm.amdgcn.s.prefetch.data.p4(ptr addrspace(4) %ptr, i32
%len)
+ ret void
+}
+
+define amdgpu_ps void @prefetch_data_sgpr_imm_base_sgpr_len(ptr addrspace(4)
inreg %ptr, i32 inreg %len) {
+; GCN-LABEL: prefetch_data_sgpr_imm_base_sgpr_len:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_prefetch_data s[0:1], 0x200, s2, 0
+; GCN-NEXT: s_endpgm
+entry:
+ %gep = getelementptr i32, ptr addrspace(4) %ptr, i32 128
+ tail call void @llvm.amdgcn.s.prefetch.data.p4(ptr addrspace(4) %gep, i32
%len)
+ ret void
+}
+
+define amdgpu_ps void @prefetch_data_sgpr_base_imm_len(ptr addrspace(4) inreg
%ptr) {
+; GCN-LABEL: prefetch_data_sgpr_base_imm_len:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_prefetch_data s[0:1], 0x0, null, 31
+; GCN-NEXT: s_endpgm
+entry:
+ tail call void @llvm.amdgcn.s.prefetch.data.p4(ptr addrspace(4) %ptr, i32 31)
+ ret void
+}
+
+define amdgpu_ps void @prefetch_data_sgpr_imm_base_imm_len(ptr addrspace(4)
inreg %ptr) {
+; GCN-LABEL: prefetch_data_sgpr_imm_base_imm_len:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_prefetch_data s[0:1], 0x200, null, 31
+; GCN-NEXT: s_endpgm
+entry:
+ %gep = getelementptr i32, ptr addrspace(4) %ptr, i32 128
+ tail call void @llvm.amdgcn.s.prefetch.data.p4(ptr addrspace(4) %gep, i32 31)
+ ret void
+}
+
+define amdgpu_ps void @prefetch_data_vgpr_base_sgpr_len(ptr addrspace(4) %ptr,
i32 inreg %len) {
+; GCN-LABEL: prefetch_data_vgpr_base_sgpr_len:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: v_readfirstlane_b32 s2, v0
+; GCN-NEXT: v_readfirstlane_b32 s3, v1
+; GCN-NEXT: s_prefetch_data s[2:3], 0x0, s0, 0
+; GCN-NEXT: s_endpgm
+entry:
+ tail call void @llvm.amdgcn.s.prefetch.data.p4(ptr addrspace(4) %ptr, i32
%len)
+ ret void
+}
+
+define amdgpu_ps void @prefetch_data_vgpr_imm_base_sgpr_len(ptr addrspace(4)
%ptr, i32 inreg %len) {
+; SDAG-LABEL: prefetch_data_vgpr_imm_base_sgpr_len:
+; SDAG: ; %bb.0: ; %entry
+; SDAG-NEXT: v_readfirstlane_b32 s2, v0
+; SDAG-NEXT: v_readfirstlane_b32 s3, v1
+; SDAG-NEXT: s_prefetch_data s[2:3], 0x200, s0, 0
+; SDAG-NEXT: s_endpgm
+;
+; GISEL-LABEL: prefetch_data_vgpr_imm_base_sgpr_len:
+; GISEL: ; %bb.0: ; %entry
+; GISEL-NEXT: v_add_co_u32 v0, vcc_lo, 0x200, v0
+; GISEL-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GISEL-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) |
instid1(VALU_DEP_2)
+; GISEL-NEXT: v_readfirstlane_b32 s2, v0
+; GISEL-NEXT: v_readfirstlane_b32 s3, v1
+; GISEL-NEXT: s_prefetch_data s[2:3], 0x0, s0, 0
+; GISEL-NEXT: s_endpgm
+entry:
+ %gep = getelementptr i32, ptr addrspace(4) %ptr, i32 128
+ tail call void @llvm.amdgcn.s.prefetch.data.p4(ptr addrspace(4) %gep, i32
%len)
+ ret void
+}
+
+define amdgpu_ps void @prefetch_data_sgpr_base_vgpr_len(ptr addrspace(4) inreg
%ptr, i32 %len) {
+; GCN-LABEL: prefetch_data_sgpr_base_vgpr_len:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: v_readfirstlane_b32 s2, v0
+; GCN-NEXT: s_prefetch_data s[0:1], 0x0, s2, 0
+; GCN-NEXT: s_endpgm
+entry:
+ tail call void @llvm.amdgcn.s.prefetch.data.p4(ptr addrspace(4) %ptr, i32
%len)
+ ret void
+}
+
+define amdgpu_ps void @prefetch_data_sgpr_base_imm_len_global(ptr addrspace(1)
inreg %ptr) {
+; GCN-LABEL: prefetch_data_sgpr_base_imm_len_global:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_prefetch_data s[0:1], 0x0, null, 31
----------------
rampitec wrote:
It is scalar prefetch, it takes SGPR address and fills scalar cache. Yes, AS1
is also isFlatGlobal and a scalar load may be issued for AS1 as well should the
address be uniform and cache r/o. It is really up to the intrinsic user to
reason if prefetching into the scalar cache would be useful or not.
Then the pointer could be VGPR, but this VGPR can contain a uniform value. That
is perfectly legal and happens. To legalize this you could either issue a
readfirstlane or do a waterfall loop. I cannot imagine anyone wanting a
prefetching waterfall loop. Using readfirstlane on the other hand is reasonable
assuming the value may be dynamically uniform. In the 'worst' case it will
prefetch less than a waterfall loop would, but prefetch is generally
discardable.
https://github.com/llvm/llvm-project/pull/107133
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
