[llvm] [AMDGPU] Refactored code for handling dynamic allocas with growing up stack (PR #119168)

via llvm-commits llvm-commits at lists.llvm.org
Mon Dec 9 00:25:43 PST 2024


https://github.com/easyonaadit updated https://github.com/llvm/llvm-project/pull/119168

>From 19e4f79a094b1a7ce44970a506287d2137205801 Mon Sep 17 00:00:00 2001
From: easyonaadit <aaditya.alokdeshpande at amd.com>
Date: Sat, 7 Dec 2024 16:27:59 +0530
Subject: [PATCH 1/2] changes to old code to refactor them, alignment, zero
 sized allocas, returning correct start location

---
 llvm/lib/Target/AMDGPU/SIISelLowering.cpp    | 35 +++++++----
 llvm/test/CodeGen/AMDGPU/non-entry-alloca.ll | 64 +++++++++-----------
 2 files changed, 52 insertions(+), 47 deletions(-)

diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index fc8bbb154d035d..92f8a8f8099967 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -4023,10 +4023,25 @@ SDValue SITargetLowering::lowerDYNAMIC_STACKALLOCImpl(SDValue Op,
   Chain = DAG.getCALLSEQ_START(Chain, 0, 0, dl);
 
   SDValue Size = Tmp2.getOperand(1);
-  SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
-  Chain = SP.getValue(1);
+  // Start address of the dynamically sized stack object
+  SDValue SPOld = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
+  Chain = SPOld.getValue(1);
   MaybeAlign Alignment = cast<ConstantSDNode>(Tmp3)->getMaybeAlignValue();
   const TargetFrameLowering *TFL = Subtarget->getFrameLowering();
+  // First we need to align the start address of the stack object to the required alignment.
+  Align StackAlign = TFL->getStackAlign();
+  if (Alignment && *Alignment > StackAlign) {
+    // formula for aligning address `SPold` to alignment boundry `align` => alignedSP = (SPold + (align - 1)) & ~(align - 1) 
+    SDValue AlignedValue = DAG.getConstant(Alignment->value(), dl, VT); // the alignment boundry we want to align to
+    SDValue StackAlignMask = DAG.getNode(ISD::SUB, dl, VT, AlignedValue, // StackAlignMask = (align - 1)
+                              DAG.getConstant(1, dl, VT));
+    Tmp1 = DAG.getNode(ISD::ADD, dl, VT, SPOld, StackAlignMask); // Tmp1 = (SPold + (align - 1))
+    Tmp1 = DAG.getNode( // Tmp1 now holds the start address aligned to the required value
+        ISD::AND, dl, VT, Tmp1,
+        DAG.getSignedConstant(-(uint64_t)Alignment->value()
+                                  << Subtarget->getWavefrontSizeLog2(),
+                              dl, VT));
+  }
   unsigned Opc =
       TFL->getStackGrowthDirection() == TargetFrameLowering::StackGrowsUp
           ? ISD::ADD
@@ -4035,19 +4050,17 @@ SDValue SITargetLowering::lowerDYNAMIC_STACKALLOCImpl(SDValue Op,
   SDValue ScaledSize = DAG.getNode(
       ISD::SHL, dl, VT, Size,
       DAG.getConstant(Subtarget->getWavefrontSizeLog2(), dl, MVT::i32));
+  // incase the value in %n at runtime is 0, we need to handle that case. There should not be a 0 sized stack object. 
+  ScaledSize = DAG.getNode( // size = max(size, 0)
+      ISD::UMAX, dl, VT, ScaledSize, 
+      DAG.getConstant(1, dl, VT));
 
-  Align StackAlign = TFL->getStackAlign();
-  Tmp1 = DAG.getNode(Opc, dl, VT, SP, ScaledSize); // Value
-  if (Alignment && *Alignment > StackAlign) {
-    Tmp1 = DAG.getNode(
-        ISD::AND, dl, VT, Tmp1,
-        DAG.getSignedConstant(-(uint64_t)Alignment->value()
-                                  << Subtarget->getWavefrontSizeLog2(),
-                              dl, VT));
-  }
+  Tmp1 = DAG.getNode(Opc, dl, VT, SPOld, ScaledSize); // Value
 
   Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
   Tmp2 = DAG.getCALLSEQ_END(Chain, 0, 0, SDValue(), dl);
+  // Set Tmp1 to point to the start address of this stack object. 
+  Tmp1 = SPOld;
 
   return DAG.getMergeValues({Tmp1, Tmp2}, dl);
 }
diff --git a/llvm/test/CodeGen/AMDGPU/non-entry-alloca.ll b/llvm/test/CodeGen/AMDGPU/non-entry-alloca.ll
index 85096eb63f46e1..0477d55e9baa36 100644
--- a/llvm/test/CodeGen/AMDGPU/non-entry-alloca.ll
+++ b/llvm/test/CodeGen/AMDGPU/non-entry-alloca.ll
@@ -30,15 +30,14 @@ define amdgpu_kernel void @kernel_non_entry_block_static_alloca_uniformly_reache
 ; MUBUF-NEXT:    s_cmp_lg_u32 s9, 0
 ; MUBUF-NEXT:    s_cbranch_scc1 .LBB0_3
 ; MUBUF-NEXT:  ; %bb.2: ; %bb.1
-; MUBUF-NEXT:    s_add_i32 s6, s32, 0x1000
-; MUBUF-NEXT:    s_lshl_b32 s7, s10, 2
-; MUBUF-NEXT:    s_mov_b32 s32, s6
+; MUBUF-NEXT:    s_mov_b32 s6, s32
 ; MUBUF-NEXT:    v_mov_b32_e32 v1, 0
-; MUBUF-NEXT:    v_mov_b32_e32 v2, s6
-; MUBUF-NEXT:    v_mov_b32_e32 v3, 1
+; MUBUF-NEXT:    v_mov_b32_e32 v2, 1
+; MUBUF-NEXT:    s_lshl_b32 s7, s10, 2
+; MUBUF-NEXT:    s_add_i32 s32, s6, 0x1000
+; MUBUF-NEXT:    buffer_store_dword v1, off, s[0:3], s6
+; MUBUF-NEXT:    buffer_store_dword v2, off, s[0:3], s6 offset:4
 ; MUBUF-NEXT:    s_add_i32 s6, s6, s7
-; MUBUF-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
-; MUBUF-NEXT:    buffer_store_dword v3, v2, s[0:3], 0 offen offset:4
 ; MUBUF-NEXT:    v_mov_b32_e32 v2, s6
 ; MUBUF-NEXT:    buffer_load_dword v2, v2, s[0:3], 0 offen
 ; MUBUF-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x0
@@ -66,11 +65,11 @@ define amdgpu_kernel void @kernel_non_entry_block_static_alloca_uniformly_reache
 ; FLATSCR-NEXT:    s_cmp_lg_u32 s5, 0
 ; FLATSCR-NEXT:    s_cbranch_scc1 .LBB0_3
 ; FLATSCR-NEXT:  ; %bb.2: ; %bb.1
-; FLATSCR-NEXT:    s_add_i32 s2, s32, 0x1000
+; FLATSCR-NEXT:    s_mov_b32 s2, s32
 ; FLATSCR-NEXT:    v_mov_b32_e32 v1, 0
 ; FLATSCR-NEXT:    v_mov_b32_e32 v2, 1
 ; FLATSCR-NEXT:    s_lshl_b32 s3, s6, 2
-; FLATSCR-NEXT:    s_mov_b32 s32, s2
+; FLATSCR-NEXT:    s_add_i32 s32, s2, 0x1000
 ; FLATSCR-NEXT:    scratch_store_dwordx2 off, v[1:2], s2
 ; FLATSCR-NEXT:    s_add_i32 s2, s2, s3
 ; FLATSCR-NEXT:    scratch_load_dword v2, off, s2
@@ -131,16 +130,14 @@ define amdgpu_kernel void @kernel_non_entry_block_static_alloca_uniformly_reache
 ; MUBUF-NEXT:    s_cmp_lg_u32 s4, 0
 ; MUBUF-NEXT:    s_cbranch_scc1 .LBB1_2
 ; MUBUF-NEXT:  ; %bb.1: ; %bb.0
-; MUBUF-NEXT:    s_add_i32 s4, s32, 0x1000
-; MUBUF-NEXT:    s_and_b32 s4, s4, 0xfffff000
-; MUBUF-NEXT:    s_lshl_b32 s5, s5, 2
-; MUBUF-NEXT:    s_mov_b32 s32, s4
+; MUBUF-NEXT:    s_mov_b32 s4, s32
 ; MUBUF-NEXT:    v_mov_b32_e32 v1, 0
-; MUBUF-NEXT:    v_mov_b32_e32 v2, s4
-; MUBUF-NEXT:    v_mov_b32_e32 v3, 1
+; MUBUF-NEXT:    v_mov_b32_e32 v2, 1
+; MUBUF-NEXT:    s_lshl_b32 s5, s5, 2
+; MUBUF-NEXT:    s_add_i32 s32, s4, 0x1000
+; MUBUF-NEXT:    buffer_store_dword v1, off, s[0:3], s4
+; MUBUF-NEXT:    buffer_store_dword v2, off, s[0:3], s4 offset:4
 ; MUBUF-NEXT:    s_add_i32 s4, s4, s5
-; MUBUF-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
-; MUBUF-NEXT:    buffer_store_dword v3, v2, s[0:3], 0 offen offset:4
 ; MUBUF-NEXT:    v_mov_b32_e32 v2, s4
 ; MUBUF-NEXT:    buffer_load_dword v2, v2, s[0:3], 0 offen
 ; MUBUF-NEXT:    s_load_dwordx2 s[4:5], s[8:9], 0x0
@@ -165,12 +162,11 @@ define amdgpu_kernel void @kernel_non_entry_block_static_alloca_uniformly_reache
 ; FLATSCR-NEXT:    s_cmp_lg_u32 s0, 0
 ; FLATSCR-NEXT:    s_cbranch_scc1 .LBB1_2
 ; FLATSCR-NEXT:  ; %bb.1: ; %bb.0
-; FLATSCR-NEXT:    s_add_i32 s0, s32, 0x1000
 ; FLATSCR-NEXT:    v_mov_b32_e32 v1, 0
-; FLATSCR-NEXT:    s_and_b32 s0, s0, 0xfffff000
+; FLATSCR-NEXT:    s_mov_b32 s0, s32
 ; FLATSCR-NEXT:    v_mov_b32_e32 v2, 1
 ; FLATSCR-NEXT:    s_lshl_b32 s1, s1, 2
-; FLATSCR-NEXT:    s_mov_b32 s32, s0
+; FLATSCR-NEXT:    s_add_i32 s32, s0, 0x1000
 ; FLATSCR-NEXT:    scratch_store_dwordx2 off, v[1:2], s0
 ; FLATSCR-NEXT:    s_add_i32 s0, s0, s1
 ; FLATSCR-NEXT:    scratch_load_dword v2, off, s0
@@ -230,16 +226,15 @@ define void @func_non_entry_block_static_alloca_align4(ptr addrspace(1) %out, i3
 ; MUBUF-NEXT:    s_and_b64 exec, exec, vcc
 ; MUBUF-NEXT:    s_cbranch_execz .LBB2_3
 ; MUBUF-NEXT:  ; %bb.2: ; %bb.1
-; MUBUF-NEXT:    s_add_i32 s6, s32, 0x1000
+; MUBUF-NEXT:    s_mov_b32 s6, s32
 ; MUBUF-NEXT:    v_mov_b32_e32 v2, 0
-; MUBUF-NEXT:    v_mov_b32_e32 v3, s6
-; MUBUF-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
+; MUBUF-NEXT:    buffer_store_dword v2, off, s[0:3], s6
 ; MUBUF-NEXT:    v_mov_b32_e32 v2, 1
-; MUBUF-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen offset:4
+; MUBUF-NEXT:    buffer_store_dword v2, off, s[0:3], s6 offset:4
 ; MUBUF-NEXT:    v_lshl_add_u32 v2, v4, 2, s6
 ; MUBUF-NEXT:    buffer_load_dword v2, v2, s[0:3], 0 offen
 ; MUBUF-NEXT:    v_and_b32_e32 v3, 0x3ff, v31
-; MUBUF-NEXT:    s_mov_b32 s32, s6
+; MUBUF-NEXT:    s_add_i32 s32, s6, 0x1000
 ; MUBUF-NEXT:    s_waitcnt vmcnt(0)
 ; MUBUF-NEXT:    v_add_u32_e32 v2, v2, v3
 ; MUBUF-NEXT:    global_store_dword v[0:1], v2, off
@@ -266,14 +261,14 @@ define void @func_non_entry_block_static_alloca_align4(ptr addrspace(1) %out, i3
 ; FLATSCR-NEXT:    s_and_b64 exec, exec, vcc
 ; FLATSCR-NEXT:    s_cbranch_execz .LBB2_3
 ; FLATSCR-NEXT:  ; %bb.2: ; %bb.1
-; FLATSCR-NEXT:    s_add_i32 s2, s32, 0x1000
+; FLATSCR-NEXT:    s_mov_b32 s2, s32
 ; FLATSCR-NEXT:    v_mov_b32_e32 v2, 0
 ; FLATSCR-NEXT:    v_mov_b32_e32 v3, 1
 ; FLATSCR-NEXT:    scratch_store_dwordx2 off, v[2:3], s2
 ; FLATSCR-NEXT:    v_lshl_add_u32 v2, v4, 2, s2
 ; FLATSCR-NEXT:    scratch_load_dword v2, v2, off
 ; FLATSCR-NEXT:    v_and_b32_e32 v3, 0x3ff, v31
-; FLATSCR-NEXT:    s_mov_b32 s32, s2
+; FLATSCR-NEXT:    s_add_i32 s32, s2, 0x1000
 ; FLATSCR-NEXT:    s_waitcnt vmcnt(0)
 ; FLATSCR-NEXT:    v_add_u32_e32 v2, v2, v3
 ; FLATSCR-NEXT:    global_store_dword v[0:1], v2, off
@@ -324,17 +319,15 @@ define void @func_non_entry_block_static_alloca_align64(ptr addrspace(1) %out, i
 ; MUBUF-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; MUBUF-NEXT:    s_cbranch_execz .LBB3_2
 ; MUBUF-NEXT:  ; %bb.1: ; %bb.0
-; MUBUF-NEXT:    s_add_i32 s6, s32, 0x1000
-; MUBUF-NEXT:    s_and_b32 s6, s6, 0xfffff000
+; MUBUF-NEXT:    s_mov_b32 s6, s32
 ; MUBUF-NEXT:    v_mov_b32_e32 v2, 0
-; MUBUF-NEXT:    v_mov_b32_e32 v4, s6
-; MUBUF-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
+; MUBUF-NEXT:    buffer_store_dword v2, off, s[0:3], s6
 ; MUBUF-NEXT:    v_mov_b32_e32 v2, 1
-; MUBUF-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen offset:4
+; MUBUF-NEXT:    buffer_store_dword v2, off, s[0:3], s6 offset:4
 ; MUBUF-NEXT:    v_lshl_add_u32 v2, v3, 2, s6
 ; MUBUF-NEXT:    buffer_load_dword v2, v2, s[0:3], 0 offen
 ; MUBUF-NEXT:    v_and_b32_e32 v3, 0x3ff, v31
-; MUBUF-NEXT:    s_mov_b32 s32, s6
+; MUBUF-NEXT:    s_add_i32 s32, s6, 0x1000
 ; MUBUF-NEXT:    s_waitcnt vmcnt(0)
 ; MUBUF-NEXT:    v_add_u32_e32 v2, v2, v3
 ; MUBUF-NEXT:    global_store_dword v[0:1], v2, off
@@ -358,15 +351,14 @@ define void @func_non_entry_block_static_alloca_align64(ptr addrspace(1) %out, i
 ; FLATSCR-NEXT:    s_and_saveexec_b64 s[0:1], vcc
 ; FLATSCR-NEXT:    s_cbranch_execz .LBB3_2
 ; FLATSCR-NEXT:  ; %bb.1: ; %bb.0
-; FLATSCR-NEXT:    s_add_i32 s2, s32, 0x1000
-; FLATSCR-NEXT:    s_and_b32 s2, s2, 0xfffff000
+; FLATSCR-NEXT:    s_mov_b32 s2, s32
 ; FLATSCR-NEXT:    v_mov_b32_e32 v4, 0
 ; FLATSCR-NEXT:    v_mov_b32_e32 v5, 1
 ; FLATSCR-NEXT:    scratch_store_dwordx2 off, v[4:5], s2
 ; FLATSCR-NEXT:    v_lshl_add_u32 v2, v3, 2, s2
 ; FLATSCR-NEXT:    scratch_load_dword v2, v2, off
 ; FLATSCR-NEXT:    v_and_b32_e32 v3, 0x3ff, v31
-; FLATSCR-NEXT:    s_mov_b32 s32, s2
+; FLATSCR-NEXT:    s_add_i32 s32, s2, 0x1000
 ; FLATSCR-NEXT:    s_waitcnt vmcnt(0)
 ; FLATSCR-NEXT:    v_add_u32_e32 v2, v2, v3
 ; FLATSCR-NEXT:    global_store_dword v[0:1], v2, off

>From 969399068837617e642ff954f7b7c7900ba5e5d9 Mon Sep 17 00:00:00 2001
From: easyonaadit <aaditya.alokdeshpande at amd.com>
Date: Mon, 9 Dec 2024 13:54:33 +0530
Subject: [PATCH 2/2] updated code

---
 llvm/lib/Target/AMDGPU/SIISelLowering.cpp | 23 +++++++++++------------
 1 file changed, 11 insertions(+), 12 deletions(-)

diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 92f8a8f8099967..8e490c40e5a175 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -4002,8 +4002,9 @@ SDValue SITargetLowering::LowerCall(CallLoweringInfo &CLI,
                          InVals, /*IsThisReturn=*/false, SDValue());
 }
 
-// This is identical to the default implementation in ExpandDYNAMIC_STACKALLOC,
-// except for applying the wave size scale to the increment amount.
+// This is similar to the default implementation in ExpandDYNAMIC_STACKALLOC,
+// except for considering a growing up stack and applying the wave size scale
+// to the increment amount.
 SDValue SITargetLowering::lowerDYNAMIC_STACKALLOCImpl(SDValue Op,
                                                       SelectionDAG &DAG) const {
   const MachineFunction &MF = DAG.getMachineFunction();
@@ -4032,15 +4033,15 @@ SDValue SITargetLowering::lowerDYNAMIC_STACKALLOCImpl(SDValue Op,
   Align StackAlign = TFL->getStackAlign();
   if (Alignment && *Alignment > StackAlign) {
     // formula for aligning address `SPold` to alignment boundry `align` => alignedSP = (SPold + (align - 1)) & ~(align - 1) 
-    SDValue AlignedValue = DAG.getConstant(Alignment->value(), dl, VT); // the alignment boundry we want to align to
-    SDValue StackAlignMask = DAG.getNode(ISD::SUB, dl, VT, AlignedValue, // StackAlignMask = (align - 1)
+    SDValue ScaledAlignment = DAG.getSignedConstant((uint64_t)Alignment->value()
+                                  << Subtarget->getWavefrontSizeLog2(),
+                                  dl, VT);
+    // SDValue AlignedValue = DAG.getConstant(Alignment->value(), dl, VT); // the alignment boundry we want to align to
+    SDValue StackAlignMask = DAG.getNode(ISD::SUB, dl, VT, ScaledAlignment, // StackAlignMask = (align - 1)
                               DAG.getConstant(1, dl, VT));
     Tmp1 = DAG.getNode(ISD::ADD, dl, VT, SPOld, StackAlignMask); // Tmp1 = (SPold + (align - 1))
     Tmp1 = DAG.getNode( // Tmp1 now holds the start address aligned to the required value
-        ISD::AND, dl, VT, Tmp1,
-        DAG.getSignedConstant(-(uint64_t)Alignment->value()
-                                  << Subtarget->getWavefrontSizeLog2(),
-                              dl, VT));
+        ISD::AND, dl, VT, Tmp1, ScaledAlignment);
   }
   unsigned Opc =
       TFL->getStackGrowthDirection() == TargetFrameLowering::StackGrowsUp
@@ -4051,7 +4052,7 @@ SDValue SITargetLowering::lowerDYNAMIC_STACKALLOCImpl(SDValue Op,
       ISD::SHL, dl, VT, Size,
       DAG.getConstant(Subtarget->getWavefrontSizeLog2(), dl, MVT::i32));
   // incase the value in %n at runtime is 0, we need to handle that case. There should not be a 0 sized stack object. 
-  ScaledSize = DAG.getNode( // size = max(size, 0)
+  ScaledSize = DAG.getNode( // size = max(size, 1)
       ISD::UMAX, dl, VT, ScaledSize, 
       DAG.getConstant(1, dl, VT));
 
@@ -4059,10 +4060,8 @@ SDValue SITargetLowering::lowerDYNAMIC_STACKALLOCImpl(SDValue Op,
 
   Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
   Tmp2 = DAG.getCALLSEQ_END(Chain, 0, 0, SDValue(), dl);
-  // Set Tmp1 to point to the start address of this stack object. 
-  Tmp1 = SPOld;
 
-  return DAG.getMergeValues({Tmp1, Tmp2}, dl);
+  return DAG.getMergeValues({SPOld, Tmp2}, dl); // return start address of the stack object
 }
 
 SDValue SITargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,



More information about the llvm-commits mailing list