[AMDGPU] 32-bit abs is not legal on R600

[LU-JOHN]

Fix bug introduced in #163907. 32-bit abs is not legal on R600.

[llvmbot]

@llvm/pr-subscribers-backend-amdgpu

Author: None (LU-JOHN)

Changes

Fix bug introduced in #163907. 32-bit abs is not legal on R600.

---

Full diff: https://github.com/llvm/llvm-project/pull/164103.diff

2 Files Affected:

• (modified) llvm/lib/Target/AMDGPU/R600ISelLowering.cpp (+3)
• (added) llvm/test/CodeGen/AMDGPU/abs32.ll (+92)

diff --git a/llvm/lib/Target/AMDGPU/R600ISelLowering.cpp b/llvm/lib/Target/AMDGPU/R600ISelLowering.cpp
index 2aa54c920a046..09ef6ac7bcdf2 100644
--- a/llvm/lib/Target/AMDGPU/R600ISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/R600ISelLowering.cpp
@@ -45,6 +45,9 @@ R600TargetLowering::R600TargetLowering(const TargetMachine &TM,
   // Legalize loads and stores to the private address space.
   setOperationAction(ISD::LOAD, {MVT::i32, MVT::v2i32, MVT::v4i32}, Custom);
 
+  // 32-bit ABS is legal for AMDGPU except for R600
+  setOperationAction(ISD::ABS, MVT::i32, Expand);
+
   // EXTLOAD should be the same as ZEXTLOAD. It is legal for some address
   // spaces, so it is custom lowered to handle those where it isn't.
   for (auto Op : {ISD::SEXTLOAD, ISD::ZEXTLOAD, ISD::EXTLOAD})
diff --git a/llvm/test/CodeGen/AMDGPU/abs32.ll b/llvm/test/CodeGen/AMDGPU/abs32.ll
new file mode 100644
index 0000000000000..b53047fadf382
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/abs32.ll
@@ -0,0 +1,92 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9 %s
+; RUN: llc -mtriple=r600 -mcpu=cypress < %s | FileCheck -check-prefixes=R600 %s
+
+define amdgpu_kernel void @abs_v1(ptr addrspace(1) %out, i32 %arg) {
+; GFX9-LABEL: abs_v1:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_abs_i32 s2, s2
+; GFX9-NEXT:    v_mov_b32_e32 v1, s2
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+;
+; R600-LABEL: abs_v1:
+; R600:       ; %bb.0:
+; R600-NEXT:    ALU 4, @4, KC0[CB0:0-32], KC1[]
+; R600-NEXT:    MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
+; R600-NEXT:    CF_END
+; R600-NEXT:    PAD
+; R600-NEXT:    ALU clause starting at 4:
+; R600-NEXT:     MOV * T0.W, KC0[2].Z,
+; R600-NEXT:     SUB_INT * T1.W, 0.0, PV.W,
+; R600-NEXT:     MAX_INT T0.X, T0.W, PV.W,
+; R600-NEXT:     LSHR * T1.X, KC0[2].Y, literal.x,
+; R600-NEXT:    2(2.802597e-45), 0(0.000000e+00)
+  %res = call i32 @llvm.abs.i32(i32 %arg, i1 false)
+  store i32 %res, ptr addrspace(1) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @abs_v2(ptr addrspace(1) %out, i32 %arg) {
+; GFX9-LABEL: abs_v2:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_abs_i32 s2, s2
+; GFX9-NEXT:    v_mov_b32_e32 v1, s2
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+;
+; R600-LABEL: abs_v2:
+; R600:       ; %bb.0:
+; R600-NEXT:    ALU 3, @4, KC0[CB0:0-32], KC1[]
+; R600-NEXT:    MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
+; R600-NEXT:    CF_END
+; R600-NEXT:    PAD
+; R600-NEXT:    ALU clause starting at 4:
+; R600-NEXT:     SUB_INT * T0.W, 0.0, KC0[2].Z,
+; R600-NEXT:     MAX_INT T0.X, KC0[2].Z, PV.W,
+; R600-NEXT:     LSHR * T1.X, KC0[2].Y, literal.x,
+; R600-NEXT:    2(2.802597e-45), 0(0.000000e+00)
+  %neg = sub i32 0, %arg
+  %cond = icmp sgt i32 %arg, %neg
+  %res = select i1 %cond, i32 %arg, i32 %neg
+  store i32 %res, ptr addrspace(1) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @abs_v3(ptr addrspace(1) %out, i32 %arg) {
+; GFX9-LABEL: abs_v3:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_abs_i32 s2, s2
+; GFX9-NEXT:    v_mov_b32_e32 v1, s2
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+;
+; R600-LABEL: abs_v3:
+; R600:       ; %bb.0:
+; R600-NEXT:    ALU 3, @4, KC0[CB0:0-32], KC1[]
+; R600-NEXT:    MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
+; R600-NEXT:    CF_END
+; R600-NEXT:    PAD
+; R600-NEXT:    ALU clause starting at 4:
+; R600-NEXT:     SUB_INT * T0.W, 0.0, KC0[2].Z,
+; R600-NEXT:     MAX_INT T0.X, PV.W, KC0[2].Z,
+; R600-NEXT:     LSHR * T1.X, KC0[2].Y, literal.x,
+; R600-NEXT:    2(2.802597e-45), 0(0.000000e+00)
+  %neg = sub i32 0, %arg
+  %cond = icmp sgt i32 %neg, %arg
+  %res = select i1 %cond, i32 %neg, i32 %arg
+  store i32 %res, ptr addrspace(1) %out, align 4
+  ret void
+}

[jayfoad]

Er... it would be much simpler to mark it as legal in SIISelLowering instead of in the common base class AMDGPUISelLowering.