[AMDGPU][SDAG] Add target-specific ISD::PTRADD combines

[ritter-x2a]

This patch adds several (AMDGPU-)target-specific DAG combines for
ISD::PTRADD nodes that reproduce existing similar transforms for
ISD::ADD nodes. There is no functional change intended for the existing
target-specific PTRADD combine.

For SWDEV-516125.

[ritter-x2a]

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

• [AMDGPU][SDAG] Handle ISD::PTRADD in various special cases #145330
• [AMDGPU][SDAG] Test ISD::PTRADD handling in various special cases #145329
• [AMDGPU][SDAG] Handle ISD::PTRADD in VOP3 patterns #143881
• [AMDGPU][SDAG] Test ISD::PTRADD handling in VOP3 patterns #143880
• [AMDGPU][SDAG] Add target-specific ISD::PTRADD combines #143673 👈 (View in Graphite)
• [AMDGPU][SDAG] Tests for target-specific ISD::PTRADD combines #143672
• [AMDGPU][SDAG] Handle ISD::PTRADD in SelectionDAGAddressAnalysis #142778
• [AMDGPU][SDAG] Add test for ISD::PTRADD handling in SelectionDAGAddressAnalysis #142777
• [AMDGPU][SDAG] Add ISD::PTRADD DAG combines #142739
• [AMDGPU][SDAG] Add tests for ISD::PTRADD DAG combines #142738
• [AMDGPU][SDAG] Initial support for ISD::PTRADD #141725
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[llvmbot]

@llvm/pr-subscribers-llvm-selectiondag

@llvm/pr-subscribers-backend-amdgpu

Author: Fabian Ritter (ritter-x2a)

Changes

This patch adds several (AMDGPU-)target-specific DAG combines for
ISD::PTRADD nodes that reproduce existing similar transforms for
ISD::ADD nodes. There is no functional change intended for the existing
target-specific PTRADD combine.

For SWDEV-516125.

---

Full diff: https://github.com/llvm/llvm-project/pull/143673.diff

3 Files Affected:

• (modified) llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp (+3-1)
• (modified) llvm/lib/Target/AMDGPU/SIISelLowering.cpp (+116-35)
• (modified) llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll (+48-103)

diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index 526a395181764..cc24585b4e4ad 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -6765,7 +6765,9 @@ SDValue SelectionDAG::FoldSymbolOffset(unsigned Opcode, EVT VT,
     return SDValue();
   int64_t Offset = C2->getSExtValue();
   switch (Opcode) {
-  case ISD::ADD: break;
+  case ISD::ADD:
+  case ISD::PTRADD:
+    break;
   case ISD::SUB: Offset = -uint64_t(Offset); break;
   default: return SDValue();
   }
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 0708c76bcb3fc..6149d7ca84354 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -33,6 +33,7 @@
 #include "llvm/CodeGen/MachineFrameInfo.h"
 #include "llvm/CodeGen/MachineFunction.h"
 #include "llvm/CodeGen/MachineLoopInfo.h"
+#include "llvm/CodeGen/SDPatternMatch.h"
 #include "llvm/IR/DiagnosticInfo.h"
 #include "llvm/IR/IRBuilder.h"
 #include "llvm/IR/IntrinsicInst.h"
@@ -46,6 +47,7 @@
 #include <optional>
 
 using namespace llvm;
+using namespace llvm::SDPatternMatch;
 
 #define DEBUG_TYPE "si-lower"
 
@@ -14320,7 +14322,7 @@ static SDValue tryFoldMADwithSRL(SelectionDAG &DAG, const SDLoc &SL,
 // instead of a tree.
 SDValue SITargetLowering::tryFoldToMad64_32(SDNode *N,
                                             DAGCombinerInfo &DCI) const {
-  assert(N->getOpcode() == ISD::ADD);
+  assert(N->isAnyAdd());
 
   SelectionDAG &DAG = DCI.DAG;
   EVT VT = N->getValueType(0);
@@ -14353,7 +14355,7 @@ SDValue SITargetLowering::tryFoldToMad64_32(SDNode *N,
     for (SDNode *User : LHS->users()) {
       // There is a use that does not feed into addition, so the multiply can't
       // be removed. We prefer MUL + ADD + ADDC over MAD + MUL.
-      if (User->getOpcode() != ISD::ADD)
+      if (!User->isAnyAdd())
         return SDValue();
 
       // We prefer 2xMAD over MUL + 2xADD + 2xADDC (code density), and prefer
@@ -14465,8 +14467,11 @@ SITargetLowering::foldAddSub64WithZeroLowBitsTo32(SDNode *N,
 
     SDValue Hi = getHiHalf64(LHS, DAG);
     SDValue ConstHi32 = DAG.getConstant(Hi_32(Val), SL, MVT::i32);
+    unsigned Opcode = N->getOpcode();
+    if (Opcode == ISD::PTRADD)
+      Opcode = ISD::ADD;
     SDValue AddHi =
-        DAG.getNode(N->getOpcode(), SL, MVT::i32, Hi, ConstHi32, N->getFlags());
+        DAG.getNode(Opcode, SL, MVT::i32, Hi, ConstHi32, N->getFlags());
 
     SDValue Lo = DAG.getNode(ISD::TRUNCATE, SL, MVT::i32, LHS);
     return DAG.getNode(ISD::BUILD_PAIR, SL, MVT::i64, Lo, AddHi);
@@ -14940,44 +14945,120 @@ SDValue SITargetLowering::performPtrAddCombine(SDNode *N,
                                                DAGCombinerInfo &DCI) const {
   SelectionDAG &DAG = DCI.DAG;
   SDLoc DL(N);
+  EVT VT = N->getValueType(0);
   SDValue N0 = N->getOperand(0);
   SDValue N1 = N->getOperand(1);
 
-  if (N1.getOpcode() == ISD::ADD) {
-    // (ptradd x, (add y, z)) -> (ptradd (ptradd x, y), z) if z is a constant,
-    //    y is not, and (add y, z) is used only once.
-    // (ptradd x, (add y, z)) -> (ptradd (ptradd x, z), y) if y is a constant,
-    //    z is not, and (add y, z) is used only once.
-    // The goal is to move constant offsets to the outermost ptradd, to create
-    // more opportunities to fold offsets into memory instructions.
-    // Together with the generic combines in DAGCombiner.cpp, this also
-    // implements (ptradd (ptradd x, y), z) -> (ptradd (ptradd x, z), y)).
-    //
-    // This transform is here instead of in the general DAGCombiner as it can
-    // turn in-bounds pointer arithmetic out-of-bounds, which is problematic for
-    // AArch64's CPA.
-    SDValue X = N0;
-    SDValue Y = N1.getOperand(0);
-    SDValue Z = N1.getOperand(1);
-    bool N1OneUse = N1.hasOneUse();
-    bool YIsConstant = DAG.isConstantIntBuildVectorOrConstantInt(Y);
-    bool ZIsConstant = DAG.isConstantIntBuildVectorOrConstantInt(Z);
-    if ((ZIsConstant != YIsConstant) && N1OneUse) {
-      SDNodeFlags Flags;
-      // If both additions in the original were NUW, the new ones are as well.
-      if (N->getFlags().hasNoUnsignedWrap() &&
-          N1->getFlags().hasNoUnsignedWrap())
-        Flags |= SDNodeFlags::NoUnsignedWrap;
-
-      if (YIsConstant)
-        std::swap(Y, Z);
-
-      SDValue Inner = DAG.getMemBasePlusOffset(X, Y, DL, Flags);
-      DCI.AddToWorklist(Inner.getNode());
-      return DAG.getMemBasePlusOffset(Inner, Z, DL, Flags);
+  // The following folds transform PTRADDs into regular arithmetic in cases
+  // where the PTRADD wouldn't be folded as an immediate offset into memory
+  // instructions anyway. They are target-specific in that other targets might
+  // prefer to not lose information about the pointer arithmetic.
+
+  // Fold (ptradd x, shl(0 - v, k)) -> sub(x, shl(v, k)).
+  // Adapted from DAGCombiner::visitADDLikeCommutative.
+  SDValue V, K;
+  if (sd_match(N1, m_Shl(m_Neg(m_Value(V)), m_Value(K)))) {
+    SDValue Inner = DAG.getNode(ISD::SHL, DL, VT, V, K);
+    DCI.AddToWorklist(Inner.getNode());
+    return DAG.getNode(ISD::SUB, DL, VT, N0, Inner);
+  }
+
+  // Fold into Mad64 if the right-hand side is a MUL. Analogous to a fold in
+  // performAddCombine.
+  if (N1.getOpcode() == ISD::MUL) {
+    if (Subtarget->hasMad64_32()) {
+      if (SDValue Folded = tryFoldToMad64_32(N, DCI))
+        return Folded;
+    }
+  }
+
+  // If the 32 low bits of the constant are all zero, there is nothing to fold
+  // into an immediate offset, so it's better to eliminate the unnecessary
+  // addition for the lower 32 bits than to preserve the PTRADD.
+  // Analogous to a fold in performAddCombine.
+  if (VT == MVT::i64) {
+    if (SDValue Folded = foldAddSub64WithZeroLowBitsTo32(N, DCI))
+      return Folded;
+  }
+
+  if (N0.getOpcode() == ISD::PTRADD && N1.getOpcode() == ISD::Constant) {
+    // Fold (ptradd (ptradd GA, v), c) -> (ptradd (ptradd GA, c) v) with
+    // global address GA and constant c, such that c can be folded into GA.
+    SDValue GAValue = N0.getOperand(0);
+    if (const GlobalAddressSDNode *GA =
+            dyn_cast<GlobalAddressSDNode>(GAValue)) {
+      const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+      if (DCI.isBeforeLegalizeOps() && TLI.isOffsetFoldingLegal(GA)) {
+        SDNodeFlags Flags;
+        // If both additions in the original were NUW, reassociation preserves
+        // that.
+        if (N->getFlags().hasNoUnsignedWrap() &&
+            N0->getFlags().hasNoUnsignedWrap())
+          Flags |= SDNodeFlags::NoUnsignedWrap;
+        SDValue Inner = DAG.getMemBasePlusOffset(GAValue, N1, DL, Flags);
+        DCI.AddToWorklist(Inner.getNode());
+        return DAG.getMemBasePlusOffset(Inner, N0.getOperand(1), DL, Flags);
+      }
     }
   }
 
+  if (N1.getOpcode() != ISD::ADD || !N1.hasOneUse())
+    return SDValue();
+
+  // (ptradd x, (add y, z)) -> (ptradd (ptradd x, y), z) if z is a constant,
+  //    y is not, and (add y, z) is used only once.
+  // (ptradd x, (add y, z)) -> (ptradd (ptradd x, z), y) if y is a constant,
+  //    z is not, and (add y, z) is used only once.
+  // The goal is to move constant offsets to the outermost ptradd, to create
+  // more opportunities to fold offsets into memory instructions.
+  // Together with the generic combines in DAGCombiner.cpp, this also
+  // implements (ptradd (ptradd x, y), z) -> (ptradd (ptradd x, z), y)).
+  //
+  // This transform is here instead of in the general DAGCombiner as it can
+  // turn in-bounds pointer arithmetic out-of-bounds, which is problematic for
+  // AArch64's CPA.
+  SDValue X = N0;
+  SDValue Y = N1.getOperand(0);
+  SDValue Z = N1.getOperand(1);
+  bool YIsConstant = DAG.isConstantIntBuildVectorOrConstantInt(Y);
+  bool ZIsConstant = DAG.isConstantIntBuildVectorOrConstantInt(Z);
+
+  SDNodeFlags ReassocFlags;
+  // If both additions in the original were NUW, reassociation preserves that.
+  if (N->getFlags().hasNoUnsignedWrap() && N1->getFlags().hasNoUnsignedWrap())
+    ReassocFlags |= SDNodeFlags::NoUnsignedWrap;
+  if (ZIsConstant != YIsConstant) {
+
+    if (YIsConstant)
+      std::swap(Y, Z);
+
+    SDValue Inner = DAG.getMemBasePlusOffset(X, Y, DL, ReassocFlags);
+    DCI.AddToWorklist(Inner.getNode());
+    return DAG.getMemBasePlusOffset(Inner, Z, DL, ReassocFlags);
+  }
+
+  // If one of Y and Z is constant, they have been handled above. If both were
+  // constant, the addition would have been folded in SelectionDAG::getNode
+  // already. This ensures that the generic DAG combines won't undo the
+  // following reassociation.
+  assert(!YIsConstant && !ZIsConstant);
+
+  if (!X->isDivergent() && Y->isDivergent() != Z->isDivergent()) {
+    // Reassociate (ptradd x, (add y, z)) -> (ptradd (ptradd x, y), z) if x and
+    // y are uniform and z isn't.
+    // Reassociate (ptradd x, (add y, z)) -> (ptradd (ptradd x, z), y) if x and
+    // z are uniform and y isn't.
+    // The goal is to push uniform operands up in the computation, so that they
+    // can be handled with scalar operations. We can't use reassociateScalarOps
+    // for this since it requires two identical commutative operations to
+    // reassociate.
+    if (Y->isDivergent())
+      std::swap(Y, Z);
+    SDValue UniformInner = DAG.getMemBasePlusOffset(X, Y, DL, ReassocFlags);
+    DCI.AddToWorklist(UniformInner.getNode());
+    return DAG.getMemBasePlusOffset(UniformInner, Z, DL, ReassocFlags);
+  }
+
   return SDValue();
 }
 
diff --git a/llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll b/llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll
index 1ec94162951a6..c00bccdbce6b7 100644
--- a/llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll
+++ b/llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll
@@ -145,49 +145,29 @@ entry:
 
 ; Test skipping the lower-32-bit addition if it is unnecessary.
 define ptr @huge_offset_low_32_unused(ptr %p) {
-; GFX942_PTRADD-LABEL: huge_offset_low_32_unused:
-; GFX942_PTRADD:       ; %bb.0:
-; GFX942_PTRADD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX942_PTRADD-NEXT:    s_mov_b32 s0, 0
-; GFX942_PTRADD-NEXT:    s_mov_b32 s1, 1
-; GFX942_PTRADD-NEXT:    v_lshl_add_u64 v[0:1], v[0:1], 0, s[0:1]
-; GFX942_PTRADD-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX942_LEGACY-LABEL: huge_offset_low_32_unused:
-; GFX942_LEGACY:       ; %bb.0:
-; GFX942_LEGACY-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX942_LEGACY-NEXT:    v_add_u32_e32 v1, 1, v1
-; GFX942_LEGACY-NEXT:    s_setpc_b64 s[30:31]
+; GFX942-LABEL: huge_offset_low_32_unused:
+; GFX942:       ; %bb.0:
+; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX942-NEXT:    v_add_u32_e32 v1, 1, v1
+; GFX942-NEXT:    s_setpc_b64 s[30:31]
   %gep = getelementptr inbounds i8, ptr %p, i64 u0x100000000
   ret ptr %gep
 }
 
 ; Reassociate address computation if it leads to more scalar operations.
 define amdgpu_kernel void @reassoc_scalar_r(ptr addrspace(1) %out, ptr addrspace(1) %p, i64 %soffset) {
-; GFX942_PTRADD-LABEL: reassoc_scalar_r:
-; GFX942_PTRADD:       ; %bb.0: ; %entry
-; GFX942_PTRADD-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x10
-; GFX942_PTRADD-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
-; GFX942_PTRADD-NEXT:    v_mov_b32_e32 v1, 0
-; GFX942_PTRADD-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
-; GFX942_PTRADD-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX942_PTRADD-NEXT:    v_lshl_add_u64 v[2:3], v[0:1], 0, s[6:7]
-; GFX942_PTRADD-NEXT:    v_lshl_add_u64 v[2:3], s[2:3], 0, v[2:3]
-; GFX942_PTRADD-NEXT:    global_store_dwordx2 v1, v[2:3], s[0:1]
-; GFX942_PTRADD-NEXT:    s_endpgm
-;
-; GFX942_LEGACY-LABEL: reassoc_scalar_r:
-; GFX942_LEGACY:       ; %bb.0: ; %entry
-; GFX942_LEGACY-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
-; GFX942_LEGACY-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x10
-; GFX942_LEGACY-NEXT:    v_mov_b32_e32 v1, 0
-; GFX942_LEGACY-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
-; GFX942_LEGACY-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX942_LEGACY-NEXT:    s_add_u32 s2, s2, s6
-; GFX942_LEGACY-NEXT:    s_addc_u32 s3, s3, s7
-; GFX942_LEGACY-NEXT:    v_lshl_add_u64 v[2:3], s[2:3], 0, v[0:1]
-; GFX942_LEGACY-NEXT:    global_store_dwordx2 v1, v[2:3], s[0:1]
-; GFX942_LEGACY-NEXT:    s_endpgm
+; GFX942-LABEL: reassoc_scalar_r:
+; GFX942:       ; %bb.0: ; %entry
+; GFX942-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX942-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x10
+; GFX942-NEXT:    v_mov_b32_e32 v1, 0
+; GFX942-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
+; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NEXT:    s_add_u32 s2, s2, s6
+; GFX942-NEXT:    s_addc_u32 s3, s3, s7
+; GFX942-NEXT:    v_lshl_add_u64 v[2:3], s[2:3], 0, v[0:1]
+; GFX942-NEXT:    global_store_dwordx2 v1, v[2:3], s[0:1]
+; GFX942-NEXT:    s_endpgm
 entry:
   %voffset32 = call i32 @llvm.amdgcn.workitem.id.x()
   %voffset = zext i32 %voffset32 to i64
@@ -198,30 +178,18 @@ entry:
 }
 
 define amdgpu_kernel void @reassoc_scalar_l(ptr addrspace(1) %out, ptr addrspace(1) %p, i64 %soffset) {
-; GFX942_PTRADD-LABEL: reassoc_scalar_l:
-; GFX942_PTRADD:       ; %bb.0: ; %entry
-; GFX942_PTRADD-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x10
-; GFX942_PTRADD-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
-; GFX942_PTRADD-NEXT:    v_mov_b32_e32 v1, 0
-; GFX942_PTRADD-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
-; GFX942_PTRADD-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX942_PTRADD-NEXT:    v_lshl_add_u64 v[2:3], s[6:7], 0, v[0:1]
-; GFX942_PTRADD-NEXT:    v_lshl_add_u64 v[2:3], s[2:3], 0, v[2:3]
-; GFX942_PTRADD-NEXT:    global_store_dwordx2 v1, v[2:3], s[0:1]
-; GFX942_PTRADD-NEXT:    s_endpgm
-;
-; GFX942_LEGACY-LABEL: reassoc_scalar_l:
-; GFX942_LEGACY:       ; %bb.0: ; %entry
-; GFX942_LEGACY-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
-; GFX942_LEGACY-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x10
-; GFX942_LEGACY-NEXT:    v_mov_b32_e32 v1, 0
-; GFX942_LEGACY-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
-; GFX942_LEGACY-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX942_LEGACY-NEXT:    s_add_u32 s2, s2, s6
-; GFX942_LEGACY-NEXT:    s_addc_u32 s3, s3, s7
-; GFX942_LEGACY-NEXT:    v_lshl_add_u64 v[2:3], s[2:3], 0, v[0:1]
-; GFX942_LEGACY-NEXT:    global_store_dwordx2 v1, v[2:3], s[0:1]
-; GFX942_LEGACY-NEXT:    s_endpgm
+; GFX942-LABEL: reassoc_scalar_l:
+; GFX942:       ; %bb.0: ; %entry
+; GFX942-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX942-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x10
+; GFX942-NEXT:    v_mov_b32_e32 v1, 0
+; GFX942-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
+; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NEXT:    s_add_u32 s2, s2, s6
+; GFX942-NEXT:    s_addc_u32 s3, s3, s7
+; GFX942-NEXT:    v_lshl_add_u64 v[2:3], s[2:3], 0, v[0:1]
+; GFX942-NEXT:    global_store_dwordx2 v1, v[2:3], s[0:1]
+; GFX942-NEXT:    s_endpgm
 entry:
   %voffset32 = call i32 @llvm.amdgcn.workitem.id.x()
   %voffset = zext i32 %voffset32 to i64
@@ -233,24 +201,14 @@ entry:
 
 ; Tests the target-specific (ptradd x, shl(0 - y, k)) -> sub(x, shl(y, k)) fold
 define ptr addrspace(1) @shl_neg_offset(ptr addrspace(1) %p, i64 %noffset, i64 %shift) {
-; GFX942_PTRADD-LABEL: shl_neg_offset:
-; GFX942_PTRADD:       ; %bb.0:
-; GFX942_PTRADD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX942_PTRADD-NEXT:    v_sub_co_u32_e32 v2, vcc, 0, v2
-; GFX942_PTRADD-NEXT:    s_nop 1
-; GFX942_PTRADD-NEXT:    v_subb_co_u32_e32 v3, vcc, 0, v3, vcc
-; GFX942_PTRADD-NEXT:    v_lshlrev_b64 v[2:3], v4, v[2:3]
-; GFX942_PTRADD-NEXT:    v_lshl_add_u64 v[0:1], v[0:1], 0, v[2:3]
-; GFX942_PTRADD-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX942_LEGACY-LABEL: shl_neg_offset:
-; GFX942_LEGACY:       ; %bb.0:
-; GFX942_LEGACY-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX942_LEGACY-NEXT:    v_lshlrev_b64 v[2:3], v4, v[2:3]
-; GFX942_LEGACY-NEXT:    v_sub_co_u32_e32 v0, vcc, v0, v2
-; GFX942_LEGACY-NEXT:    s_nop 1
-; GFX942_LEGACY-NEXT:    v_subb_co_u32_e32 v1, vcc, v1, v3, vcc
-; GFX942_LEGACY-NEXT:    s_setpc_b64 s[30:31]
+; GFX942-LABEL: shl_neg_offset:
+; GFX942:       ; %bb.0:
+; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX942-NEXT:    v_lshlrev_b64 v[2:3], v4, v[2:3]
+; GFX942-NEXT:    v_sub_co_u32_e32 v0, vcc, v0, v2
+; GFX942-NEXT:    s_nop 1
+; GFX942-NEXT:    v_subb_co_u32_e32 v1, vcc, v1, v3, vcc
+; GFX942-NEXT:    s_setpc_b64 s[30:31]
   %offset = sub i64 0, %noffset
   %x = shl i64 %offset, %shift
   %gep = getelementptr inbounds i8, ptr addrspace(1) %p, i64 %x
@@ -268,10 +226,9 @@ define ptr addrspace(1) @complextype_global_gep(i64 %offset) {
 ; GFX942_PTRADD:       ; %bb.0:
 ; GFX942_PTRADD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX942_PTRADD-NEXT:    s_getpc_b64 s[0:1]
-; GFX942_PTRADD-NEXT:    s_add_u32 s0, s0, v0@rel32@lo+4
-; GFX942_PTRADD-NEXT:    s_addc_u32 s1, s1, v0@rel32@hi+12
+; GFX942_PTRADD-NEXT:    s_add_u32 s0, s0, v0@rel32@lo+14
+; GFX942_PTRADD-NEXT:    s_addc_u32 s1, s1, v0@rel32@hi+22
 ; GFX942_PTRADD-NEXT:    v_lshl_add_u64 v[0:1], s[0:1], 0, v[0:1]
-; GFX942_PTRADD-NEXT:    v_lshl_add_u64 v[0:1], v[0:1], 0, 10
 ; GFX942_PTRADD-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX942_LEGACY-LABEL: complextype_global_gep:
@@ -291,27 +248,15 @@ define ptr addrspace(1) @complextype_global_gep(i64 %offset) {
 
 ; Tests the tryFoldToMad64_32 PTRADD combine.
 define amdgpu_kernel void @fold_mad64(ptr addrspace(1) %p) {
-; GFX942_PTRADD-LABEL: fold_mad64:
-; GFX942_PTRADD:       ; %bb.0:
-; GFX942_PTRADD-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
-; GFX942_PTRADD-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
-; GFX942_PTRADD-NEXT:    v_mul_hi_u32_u24_e32 v1, 12, v0
-; GFX942_PTRADD-NEXT:    v_mul_u32_u24_e32 v0, 12, v0
-; GFX942_PTRADD-NEXT:    v_mov_b32_e32 v2, 1.0
-; GFX942_PTRADD-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX942_PTRADD-NEXT:    v_lshl_add_u64 v[0:1], s[0:1], 0, v[0:1]
-; GFX942_PTRADD-NEXT:    global_store_dword v[0:1], v2, off
-; GFX942_PTRADD-NEXT:    s_endpgm
-;
-; GFX942_LEGACY-LABEL: fold_mad64:
-; GFX942_LEGACY:       ; %bb.0:
-; GFX942_LEGACY-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
-; GFX942_LEGACY-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
-; GFX942_LEGACY-NEXT:    v_mov_b32_e32 v2, 1.0
-; GFX942_LEGACY-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX942_LEGACY-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v0, 12, s[0:1]
-; GFX942_LEGACY-NEXT:    global_store_dword v[0:1], v2, off
-; GFX942_LEGACY-NEXT:    s_endpgm
+; GFX942-LABEL: fold_mad64:
+; GFX942:       ; %bb.0:
+; GFX942-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX942-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
+; GFX942-NEXT:    v_mov_b32_e32 v2, 1.0
+; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v0, 12, s[0:1]
+; GFX942-NEXT:    global_store_dword v[0:1], v2, off
+; GFX942-NEXT:    s_endpgm
   %voffset32 = call i32 @llvm.amdgcn.workitem.id.x()
   %voffset = zext i32 %voffset32 to i64
   %p1 = getelementptr inbounds %S, ptr addrspace(1) %p, i64 %voffset, i32 0