[AMDGPU] Do not fold into v_accvpr_mov/write/read (#120475)

In SIFoldOperands, leave copies for moving between agpr and vgpr registers. The register coalescer is able to handle the copies more efficiently than v_accvgpr_mov, v_accvgpr_write, and v_accvgpr_read. Otherwise, the compiler generates unneccesary instructions such as v_accvgpr_mov a0, a0.
llvm · Jan 7, 2025 · 17c8c1c · 17c8c1c
1 parent 8178d3c
commit 17c8c1c
Show file tree

Hide file tree

Showing 4 changed files with 537 additions and 13 deletions.
diff --git a/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp b/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
@@ -1096,21 +1096,8 @@ void SIFoldOperandsImpl::foldOperand(
           B.addImm(Defs[I].second);
         }
         LLVM_DEBUG(dbgs() << "Folded " << *UseMI);
-        return;
       }
 
-      if (Size != 4)
-        return;
-
-      Register Reg0 = UseMI->getOperand(0).getReg();
-      Register Reg1 = UseMI->getOperand(1).getReg();
-      if (TRI->isAGPR(*MRI, Reg0) && TRI->isVGPR(*MRI, Reg1))
-        UseMI->setDesc(TII->get(AMDGPU::V_ACCVGPR_WRITE_B32_e64));
-      else if (TRI->isVGPR(*MRI, Reg0) && TRI->isAGPR(*MRI, Reg1))
-        UseMI->setDesc(TII->get(AMDGPU::V_ACCVGPR_READ_B32_e64));
-      else if (ST->hasGFX90AInsts() && TRI->isAGPR(*MRI, Reg0) &&
-               TRI->isAGPR(*MRI, Reg1))
-        UseMI->setDesc(TII->get(AMDGPU::V_ACCVGPR_MOV_B32));
       return;
     }
 

diff --git a/llvm/test/CodeGen/AMDGPU/no-fold-accvgpr-mov.ll b/llvm/test/CodeGen/AMDGPU/no-fold-accvgpr-mov.ll
@@ -0,0 +1,120 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=gfx942 %s -o - | FileCheck %s --check-prefixes=GFX942
+; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=gfx908 %s -o - | FileCheck %s --check-prefixes=GFX908
+
+define amdgpu_kernel void @matmul_kernel(i32 %a0, i32 %a1) {
+; GFX942-LABEL: matmul_kernel:
+; GFX942:       ; %bb.0: ; %entry
+; GFX942-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX942-NEXT:    v_mov_b32_e32 v1, 0
+; GFX942-NEXT:    s_mov_b32 s2, 0
+; GFX942-NEXT:    v_accvgpr_write_b32 a0, v1
+; GFX942-NEXT:    s_mov_b32 s3, 0
+; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NEXT:    s_cmp_lg_u32 s0, 0
+; GFX942-NEXT:    s_cselect_b64 s[0:1], -1, 0
+; GFX942-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s[0:1]
+; GFX942-NEXT:    v_cmp_ne_u32_e64 s[0:1], 1, v0
+; GFX942-NEXT:    s_branch .LBB0_2
+; GFX942-NEXT:  .LBB0_1: ; %bb2
+; GFX942-NEXT:    ; in Loop: Header=BB0_2 Depth=1
+; GFX942-NEXT:    s_or_b32 s4, s3, 1
+; GFX942-NEXT:    s_ashr_i32 s5, s3, 31
+; GFX942-NEXT:    s_mov_b32 s3, s2
+; GFX942-NEXT:    v_mov_b64_e32 v[4:5], s[2:3]
+; GFX942-NEXT:    v_accvgpr_read_b32 v0, a0
+; GFX942-NEXT:    v_mov_b32_e32 v2, v1
+; GFX942-NEXT:    v_mov_b32_e32 v3, v1
+; GFX942-NEXT:    v_accvgpr_write_b32 a0, v0
+; GFX942-NEXT:    v_accvgpr_write_b32 a1, v1
+; GFX942-NEXT:    v_accvgpr_write_b32 a2, v2
+; GFX942-NEXT:    v_accvgpr_write_b32 a3, v3
+; GFX942-NEXT:    s_and_b32 s3, s5, s4
+; GFX942-NEXT:    s_nop 0
+; GFX942-NEXT:    v_mfma_f32_16x16x16_f16 a[0:3], v[4:5], v[4:5], a[0:3]
+; GFX942-NEXT:    s_cbranch_execz .LBB0_4
+; GFX942-NEXT:  .LBB0_2: ; %bb
+; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX942-NEXT:    s_and_b64 vcc, exec, s[0:1]
+; GFX942-NEXT:    s_cbranch_vccz .LBB0_1
+; GFX942-NEXT:  ; %bb.3:
+; GFX942-NEXT:    ; implicit-def: $sgpr3
+; GFX942-NEXT:  .LBB0_4: ; %common.ret
+; GFX942-NEXT:    s_endpgm
+;
+; GFX908-LABEL: matmul_kernel:
+; GFX908:       ; %bb.0: ; %entry
+; GFX908-NEXT:    s_load_dwordx2 s[0:1], s[8:9], 0x0
+; GFX908-NEXT:    v_mov_b32_e32 v1, 0
+; GFX908-NEXT:    s_mov_b32 s2, 0
+; GFX908-NEXT:    s_mov_b32 s3, 0
+; GFX908-NEXT:    v_accvgpr_write_b32 a0, v1
+; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX908-NEXT:    s_cmp_lg_u32 s0, 0
+; GFX908-NEXT:    s_cselect_b64 s[0:1], -1, 0
+; GFX908-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s[0:1]
+; GFX908-NEXT:    v_cmp_ne_u32_e64 s[0:1], 1, v0
+; GFX908-NEXT:    s_branch .LBB0_2
+; GFX908-NEXT:  .LBB0_1: ; %bb2
+; GFX908-NEXT:    ; in Loop: Header=BB0_2 Depth=1
+; GFX908-NEXT:    s_or_b32 s4, s3, 1
+; GFX908-NEXT:    s_ashr_i32 s5, s3, 31
+; GFX908-NEXT:    s_mov_b32 s3, s2
+; GFX908-NEXT:    s_nop 3
+; GFX908-NEXT:    v_accvgpr_read_b32 v0, a0
+; GFX908-NEXT:    v_mov_b32_e32 v5, s3
+; GFX908-NEXT:    v_mov_b32_e32 v4, s2
+; GFX908-NEXT:    v_mov_b32_e32 v2, v1
+; GFX908-NEXT:    v_mov_b32_e32 v3, v1
+; GFX908-NEXT:    v_accvgpr_write_b32 a0, v0
+; GFX908-NEXT:    v_accvgpr_write_b32 a1, v1
+; GFX908-NEXT:    v_accvgpr_write_b32 a2, v2
+; GFX908-NEXT:    v_accvgpr_write_b32 a3, v3
+; GFX908-NEXT:    s_and_b32 s3, s5, s4
+; GFX908-NEXT:    v_mfma_f32_16x16x16f16 a[0:3], v[4:5], v[4:5], a[0:3]
+; GFX908-NEXT:    s_cbranch_execz .LBB0_4
+; GFX908-NEXT:  .LBB0_2: ; %bb
+; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT:    s_and_b64 vcc, exec, s[0:1]
+; GFX908-NEXT:    s_cbranch_vccz .LBB0_1
+; GFX908-NEXT:  ; %bb.3:
+; GFX908-NEXT:    ; implicit-def: $sgpr3
+; GFX908-NEXT:  .LBB0_4: ; %common.ret
+; GFX908-NEXT:    s_endpgm
+entry:
+  br label %bb
+
+bb:
+  %i = phi { float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float } [ %i10, %bb2 ], [ zeroinitializer, %entry ]
+  %i1 = phi i32 [ %i5, %bb2 ], [ 0, %entry ]
+  %c0 = icmp ne i32 %a0, 0
+  br i1 %c0, label %bb2, label %bb11
+
+bb2:
+  %i3 = or i32 %i1, 1
+  %i4 = icmp slt i32 %i1, 0
+  %i5 = select i1 %i4, i32 %i3, i32 0
+  %i6 = extractvalue { float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float } %i, 123
+  %i7 = insertelement <4 x float> zeroinitializer, float %i6, i32 0
+  %i8 = call <4 x float> @llvm.amdgcn.mfma.f32.16x16x16f16(<4 x half> zeroinitializer, <4 x half> zeroinitializer, <4 x float> %i7, i32 0, i32 0, i32 0)
+  %i9 = extractelement <4 x float> %i8, i32 0
+  %i10 = insertvalue { float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float } zeroinitializer, float %i9, 123
+  br label %bb
+
+bb11:
+  %c1 = icmp ne i32 %a1, 0
+  br i1 %c1, label %bb12, label %common.ret
+
+common.ret:
+  ret void
+
+bb12:
+  %i13 = extractvalue { float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float, float } %i, 0
+  %i14 = insertelement <4 x float> zeroinitializer, float %i13, i32 0
+  %i15 = insertelement <4 x float> %i14, float 0.000000e+00, i32 0
+  %i16 = insertelement <4 x float> %i15, float 0.000000e+00, i32 0
+  br label %common.ret
+}
+
+; Function Attrs: convergent nocallback nofree nosync nounwind willreturn memory(none)
+declare <4 x float> @llvm.amdgcn.mfma.f32.16x16x16f16(<4 x half>, <4 x half>, <4 x float>, i32 immarg, i32 immarg, i32 immarg)