[llvm] 8860daf - AMDGPU: Handle a few missing cases in getAddrModeArguments
Matt Arsenault via llvm-commits
llvm-commits at lists.llvm.org
Tue Jul 28 17:22:57 PDT 2020
Author: Matt Arsenault
Date: 2020-07-28T20:22:38-04:00
New Revision: 8860daf0ed248960c3bc26208b7a38b3e1aadb1d
URL: https://github.com/llvm/llvm-project/commit/8860daf0ed248960c3bc26208b7a38b3e1aadb1d
DIFF: https://github.com/llvm/llvm-project/commit/8860daf0ed248960c3bc26208b7a38b3e1aadb1d.diff
LOG: AMDGPU: Handle a few missing cases in getAddrModeArguments
Added:
llvm/test/CodeGen/AMDGPU/cgp-addressing-modes-gfx1030.ll
llvm/test/CodeGen/AMDGPU/cgp-addressing-modes-gfx908.ll
Modified:
llvm/lib/Target/AMDGPU/SIISelLowering.cpp
llvm/test/CodeGen/AMDGPU/cgp-addressing-modes.ll
Removed:
################################################################################
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index c92065f7495a..4c9fce92f9a2 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -1175,9 +1175,13 @@ bool SITargetLowering::getAddrModeArguments(IntrinsicInst *II,
case Intrinsic::amdgcn_atomic_dec:
case Intrinsic::amdgcn_ds_ordered_add:
case Intrinsic::amdgcn_ds_ordered_swap:
+ case Intrinsic::amdgcn_ds_append:
+ case Intrinsic::amdgcn_ds_consume:
case Intrinsic::amdgcn_ds_fadd:
case Intrinsic::amdgcn_ds_fmin:
- case Intrinsic::amdgcn_ds_fmax: {
+ case Intrinsic::amdgcn_ds_fmax:
+ case Intrinsic::amdgcn_global_atomic_fadd:
+ case Intrinsic::amdgcn_global_atomic_csub: {
Value *Ptr = II->getArgOperand(0);
AccessTy = II->getType();
Ops.push_back(Ptr);
diff --git a/llvm/test/CodeGen/AMDGPU/cgp-addressing-modes-gfx1030.ll b/llvm/test/CodeGen/AMDGPU/cgp-addressing-modes-gfx1030.ll
new file mode 100644
index 000000000000..0f655dadfa11
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/cgp-addressing-modes-gfx1030.ll
@@ -0,0 +1,75 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: opt -S -codegenprepare -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1030 < %s | FileCheck -check-prefix=OPT %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1030 < %s | FileCheck -check-prefix=GCN %s
+
+; Make sure we match the addressing mode offset of csub intrinsics across blocks.
+
+define amdgpu_kernel void @test_sink_small_offset_global_atomic_csub_i32(i32 addrspace(1)* %out, i32 addrspace(1)* %in) {
+; OPT-LABEL: @test_sink_small_offset_global_atomic_csub_i32(
+; OPT-NEXT: entry:
+; OPT-NEXT: [[OUT_GEP:%.*]] = getelementptr i32, i32 addrspace(1)* [[OUT:%.*]], i32 999999
+; OPT-NEXT: [[TID:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0) #3
+; OPT-NEXT: [[CMP:%.*]] = icmp eq i32 [[TID]], 0
+; OPT-NEXT: br i1 [[CMP]], label [[ENDIF:%.*]], label [[IF:%.*]]
+; OPT: if:
+; OPT-NEXT: [[TMP0:%.*]] = bitcast i32 addrspace(1)* [[IN:%.*]] to i8 addrspace(1)*
+; OPT-NEXT: [[SUNKADDR:%.*]] = getelementptr i8, i8 addrspace(1)* [[TMP0]], i64 28
+; OPT-NEXT: [[TMP1:%.*]] = bitcast i8 addrspace(1)* [[SUNKADDR]] to i32 addrspace(1)*
+; OPT-NEXT: [[VAL:%.*]] = call i32 @llvm.amdgcn.global.atomic.csub.p1i32(i32 addrspace(1)* [[TMP1]], i32 2)
+; OPT-NEXT: br label [[ENDIF]]
+; OPT: endif:
+; OPT-NEXT: [[X:%.*]] = phi i32 [ [[VAL]], [[IF]] ], [ 0, [[ENTRY:%.*]] ]
+; OPT-NEXT: store i32 [[X]], i32 addrspace(1)* [[OUT_GEP]], align 4
+; OPT-NEXT: ret void
+;
+; GCN-LABEL: test_sink_small_offset_global_atomic_csub_i32:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GCN-NEXT: v_mbcnt_lo_u32_b32_e64 v0, -1, 0
+; GCN-NEXT: ; implicit-def: $vcc_hi
+; GCN-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GCN-NEXT: v_mov_b32_e32 v0, 0
+; GCN-NEXT: s_and_saveexec_b32 s4, vcc_lo
+; GCN-NEXT: s_cbranch_execz BB0_2
+; GCN-NEXT: ; %bb.1: ; %if
+; GCN-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-NEXT: v_mov_b32_e32 v0, s2
+; GCN-NEXT: v_mov_b32_e32 v1, s3
+; GCN-NEXT: v_mov_b32_e32 v2, 2
+; GCN-NEXT: global_atomic_csub v0, v[0:1], v2, off offset:28 glc
+; GCN-NEXT: BB0_2: ; %endif
+; GCN-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GCN-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-NEXT: v_add_co_u32_e64 v1, s0, 0x3d0800, s0
+; GCN-NEXT: v_add_co_ci_u32_e64 v2, s0, 0, s1, s0
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: global_store_dword v[1:2], v0, off offset:252
+; GCN-NEXT: s_endpgm
+entry:
+ %out.gep = getelementptr i32, i32 addrspace(1)* %out, i32 999999
+ %in.gep = getelementptr i32, i32 addrspace(1)* %in, i32 7
+ %tid = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0) #0
+ %cmp = icmp eq i32 %tid, 0
+ br i1 %cmp, label %endif, label %if
+
+if:
+ %val = call i32 @llvm.amdgcn.global.atomic.csub.p1i32(i32 addrspace(1)* %in.gep, i32 2)
+ br label %endif
+
+endif:
+ %x = phi i32 [ %val, %if ], [ 0, %entry ]
+ store i32 %x, i32 addrspace(1)* %out.gep
+ br label %done
+
+done:
+ ret void
+}
+
+declare i32 @llvm.amdgcn.global.atomic.csub.p1i32(i32 addrspace(1)* nocapture, i32) #0
+declare i32 @llvm.amdgcn.mbcnt.lo(i32, i32) #1
+declare void @llvm.amdgcn.global.atomic.fadd.p1f32.f32(float addrspace(1)* nocapture, float) #2
+
+attributes #0 = { argmemonly nounwind }
+attributes #1 = { nounwind readnone willreturn }
+attributes #2 = { argmemonly nounwind willreturn }
diff --git a/llvm/test/CodeGen/AMDGPU/cgp-addressing-modes-gfx908.ll b/llvm/test/CodeGen/AMDGPU/cgp-addressing-modes-gfx908.ll
new file mode 100644
index 000000000000..840a4ec3dac8
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/cgp-addressing-modes-gfx908.ll
@@ -0,0 +1,78 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: opt -S -codegenprepare -mtriple=amdgcn-amd-amdhsa -mcpu=gfx908 < %s | FileCheck -check-prefix=OPT %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx908 < %s | FileCheck -check-prefix=GCN %s
+
+; Make sure we match the addressing mode offset of globla.atomic.fadd intrinsics across blocks.
+
+define amdgpu_kernel void @test_sink_small_offset_global_atomic_fadd_f32(float addrspace(1)* %out, float addrspace(1)* %in) {
+; OPT-LABEL: @test_sink_small_offset_global_atomic_fadd_f32(
+; OPT-NEXT: entry:
+; OPT-NEXT: [[OUT_GEP:%.*]] = getelementptr float, float addrspace(1)* [[OUT:%.*]], i32 999999
+; OPT-NEXT: [[TID:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0) #3
+; OPT-NEXT: [[CMP:%.*]] = icmp eq i32 [[TID]], 0
+; OPT-NEXT: br i1 [[CMP]], label [[ENDIF:%.*]], label [[IF:%.*]]
+; OPT: if:
+; OPT-NEXT: [[TMP0:%.*]] = bitcast float addrspace(1)* [[IN:%.*]] to i8 addrspace(1)*
+; OPT-NEXT: [[SUNKADDR:%.*]] = getelementptr i8, i8 addrspace(1)* [[TMP0]], i64 28
+; OPT-NEXT: [[TMP1:%.*]] = bitcast i8 addrspace(1)* [[SUNKADDR]] to float addrspace(1)*
+; OPT-NEXT: call void @llvm.amdgcn.global.atomic.fadd.p1f32.f32(float addrspace(1)* [[TMP1]], float 2.000000e+00)
+; OPT-NEXT: [[VAL:%.*]] = load volatile float, float addrspace(1)* undef, align 4
+; OPT-NEXT: br label [[ENDIF]]
+; OPT: endif:
+; OPT-NEXT: [[X:%.*]] = phi float [ [[VAL]], [[IF]] ], [ 0.000000e+00, [[ENTRY:%.*]] ]
+; OPT-NEXT: store float [[X]], float addrspace(1)* [[OUT_GEP]], align 4
+; OPT-NEXT: ret void
+;
+; GCN-LABEL: test_sink_small_offset_global_atomic_fadd_f32:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GCN-NEXT: v_mbcnt_lo_u32_b32 v0, -1, 0
+; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
+; GCN-NEXT: v_mov_b32_e32 v0, 0
+; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GCN-NEXT: s_cbranch_execz BB0_2
+; GCN-NEXT: ; %bb.1: ; %if
+; GCN-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-NEXT: v_mov_b32_e32 v0, s2
+; GCN-NEXT: v_mov_b32_e32 v1, s3
+; GCN-NEXT: v_mov_b32_e32 v2, 2.0
+; GCN-NEXT: global_atomic_add_f32 v[0:1], v2, off offset:28
+; GCN-NEXT: global_load_dword v0, v[0:1], off
+; GCN-NEXT: BB0_2: ; %endif
+; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
+; GCN-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-NEXT: v_mov_b32_e32 v1, s0
+; GCN-NEXT: v_add_co_u32_e32 v1, vcc, 0x3d0000, v1
+; GCN-NEXT: v_mov_b32_e32 v2, s1
+; GCN-NEXT: v_addc_co_u32_e32 v2, vcc, 0, v2, vcc
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: global_store_dword v[1:2], v0, off offset:2300
+; GCN-NEXT: s_endpgm
+entry:
+ %out.gep = getelementptr float, float addrspace(1)* %out, i32 999999
+ %in.gep = getelementptr float, float addrspace(1)* %in, i32 7
+ %tid = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0) #0
+ %cmp = icmp eq i32 %tid, 0
+ br i1 %cmp, label %endif, label %if
+
+if:
+ call void @llvm.amdgcn.global.atomic.fadd.p1f32.f32(float addrspace(1)* %in.gep, float 2.0)
+ %val = load volatile float, float addrspace(1)* undef
+ br label %endif
+
+endif:
+ %x = phi float [ %val, %if ], [ 0.0, %entry ]
+ store float %x, float addrspace(1)* %out.gep
+ br label %done
+
+done:
+ ret void
+}
+
+declare i32 @llvm.amdgcn.mbcnt.lo(i32, i32) #1
+declare void @llvm.amdgcn.global.atomic.fadd.p1f32.f32(float addrspace(1)* nocapture, float) #2
+
+attributes #0 = { argmemonly nounwind }
+attributes #1 = { nounwind readnone willreturn }
+attributes #2 = { argmemonly nounwind willreturn }
diff --git a/llvm/test/CodeGen/AMDGPU/cgp-addressing-modes.ll b/llvm/test/CodeGen/AMDGPU/cgp-addressing-modes.ll
index a2fb893af540..bb621b294bbc 100644
--- a/llvm/test/CodeGen/AMDGPU/cgp-addressing-modes.ll
+++ b/llvm/test/CodeGen/AMDGPU/cgp-addressing-modes.ll
@@ -748,10 +748,65 @@ done:
ret void
}
+; OPT-LABEL: @test_sink_small_offset_ds_append(
+; OPT: %0 = bitcast i32 addrspace(3)* %in to i8 addrspace(3)*
+; OPT: %sunkaddr = getelementptr i8, i8 addrspace(3)* %0, i32 28
+; OPT: %1 = bitcast i8 addrspace(3)* %sunkaddr to i32 addrspace(3)*
+; OPT: %tmp1 = call i32 @llvm.amdgcn.ds.append.p3i32(i32 addrspace(3)* %1, i1 false)
+define amdgpu_kernel void @test_sink_small_offset_ds_append(i32 addrspace(3)* %out, i32 addrspace(3)* %in) {
+entry:
+ %out.gep = getelementptr i32, i32 addrspace(3)* %out, i32 999999
+ %in.gep = getelementptr i32, i32 addrspace(3)* %in, i32 7
+ %tid = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0) #0
+ %tmp0 = icmp eq i32 %tid, 0
+ br i1 %tmp0, label %endif, label %if
+
+if:
+ %tmp1 = call i32 @llvm.amdgcn.ds.append.p3i32(i32 addrspace(3)* %in.gep, i1 false)
+ br label %endif
+
+endif:
+ %x = phi i32 [ %tmp1, %if ], [ 0, %entry ]
+ store i32 %x, i32 addrspace(3)* %out.gep
+ br label %done
+
+done:
+ ret void
+}
+
+; OPT-LABEL: @test_sink_small_offset_ds_consume(
+; OPT: %0 = bitcast i32 addrspace(3)* %in to i8 addrspace(3)*
+; OPT: %sunkaddr = getelementptr i8, i8 addrspace(3)* %0, i32 28
+; OPT: %1 = bitcast i8 addrspace(3)* %sunkaddr to i32 addrspace(3)*
+; OPT: %tmp1 = call i32 @llvm.amdgcn.ds.consume.p3i32(i32 addrspace(3)* %1, i1 false)
+define amdgpu_kernel void @test_sink_small_offset_ds_consume(i32 addrspace(3)* %out, i32 addrspace(3)* %in) {
+entry:
+ %out.gep = getelementptr i32, i32 addrspace(3)* %out, i32 999999
+ %in.gep = getelementptr i32, i32 addrspace(3)* %in, i32 7
+ %tid = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0) #0
+ %tmp0 = icmp eq i32 %tid, 0
+ br i1 %tmp0, label %endif, label %if
+
+if:
+ %tmp1 = call i32 @llvm.amdgcn.ds.consume.p3i32(i32 addrspace(3)* %in.gep, i1 false)
+ br label %endif
+
+endif:
+ %x = phi i32 [ %tmp1, %if ], [ 0, %entry ]
+ store i32 %x, i32 addrspace(3)* %out.gep
+ br label %done
+
+done:
+ ret void
+}
+
declare i32 @llvm.amdgcn.mbcnt.lo(i32, i32) #0
declare i32 @llvm.amdgcn.atomic.inc.i32.p3i32(i32 addrspace(3)* nocapture, i32, i32, i32, i1) #2
declare i32 @llvm.amdgcn.atomic.dec.i32.p3i32(i32 addrspace(3)* nocapture, i32, i32, i32, i1) #2
+declare i32 @llvm.amdgcn.ds.append.p3i32(i32 addrspace(3)* nocapture, i1 immarg) #3
+declare i32 @llvm.amdgcn.ds.consume.p3i32(i32 addrspace(3)* nocapture, i1 immarg) #3
attributes #0 = { nounwind readnone }
attributes #1 = { nounwind }
attributes #2 = { nounwind argmemonly }
+attributes #3 = { argmemonly convergent nounwind willreturn }
More information about the llvm-commits
mailing list