[llvm-branch-commits] [llvm] [DAGCombiner][GlobalISel] Extend allMulUsesCanBeContracted with FMA/FMAD pattern (PR #188117)

via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Fri Apr 3 08:57:18 PDT 2026


llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-llvm-selectiondag

@llvm/pr-subscribers-backend-amdgpu

Author: Adel Ejjeh (adelejjeh)

<details>
<summary>Changes</summary>

Supersedes #<!-- -->169735 (PR 5/5). Split into a stack of 5 PRs per reviewer request.

---

Add FMA/FMAD recognition to `allMulUsesCanBeContracted()`: when a multiply is used as an operand of an FMA/FMAD, the guard checks whether the FMA's consumer enables chain reassociation that can eliminate the multiply. Two sub-patterns are recognized:

1. **Direct chain reassociation** (`fmul -> fma -> {fadd, fsub}`): Contractable only if the consumer `fadd`/`fsub` has the `reassoc` flag. Chain reassociation transforms `fadd(fma(a, b, fmul(c, d)), e)` into `fma(a, b, fma(c, d, e))`, eliminating the multiply. Without `reassoc`, the basic `fadd -> fma` fold fires instead, re-computing the multiply and causing duplication.

2. **FPEXT chain reassociation** (`fmul -> fpext -> fma -> {fadd, fsub}` and `fmul -> fma -> fpext -> {fadd, fsub}`): Contractable without `reassoc` because the aggressive fpext folds (inside `if (Aggressive)`) handle these patterns directly.

Changes:

- **DAGCombiner.cpp**: Add `ISD::FMA`/`ISD::FMAD` case to `allMulUsesCanBeContracted()`. FMA/FMAD handler iterates FMA users; accepts if any user is `FP_EXTEND`, or if it is `FADD`/`FSUB` with `hasAllowReassociation()`. FPEXT -> FMA/FMAD handler iterates FMA users; accepts `FADD`/`FSUB` without `reassoc` check. Guards added to 8 fold sites in `visitFADDForFMACombine` and `visitFSUBForFMACombine`.
- **CombinerHelper.cpp**: Add `G_FMA`/`G_FMAD` case to `allMulUsesCanBeContracted()` with equivalent logic using `FmReassoc`. Guards added to 4 fold sites in `matchCombineFAddFpExtFMulToFMadOrFMAAggressive`.

### Test Changes

#### `fma-multiple-uses-contraction.ll`

The final FMA/FMAD chain section is activated: all remaining `P0-` prefixed baseline checks are replaced with real-prefix CHECK lines. The `--allow-unused-prefixes` flag is removed from all RUN lines since all check prefixes are now actively used.

- `fma_chain_fadd_no_reassoc`: Demonstrates the guard correctly blocking contraction when `reassoc` is absent (both SDAG and GISel show shared `v_mul` without duplication).
- `fma_chain_fadd_reassoc`: Demonstrates chain reassociation firing when `reassoc nsz` is present (no `v_mul`, multiply eliminated via FMA chain).

#### Other lit tests

- `dagcombine-fma-crash.ll` (`main`): CHECK lines revert to pre-stack (upstream) behavior. The fmul→FMA chain is now recognized as contractable, restoring the contraction that was temporarily blocked by PR 2/5.
- `dagcombine-fma-fmad.ll` (`_amdgpu_ps_main`): CHECK lines revert to pre-stack (upstream) behavior for the same reason.

---

_This PR was split from the original_ #<!-- -->169735 _with the assistance of_ [_Claude Code_](https://claude.ai/code)_._

---

Patch is 117.22 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/188117.diff


5 Files Affected:

- (modified) llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp (+57-1) 
- (modified) llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp (+55-1) 
- (modified) llvm/test/CodeGen/AMDGPU/dagcombine-fma-crash.ll (+10-12) 
- (modified) llvm/test/CodeGen/AMDGPU/dagcombine-fma-fmad.ll (+80-89) 
- (modified) llvm/test/CodeGen/AMDGPU/fma-multiple-uses-contraction.ll (+706-822) 


``````````diff
diff --git a/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp b/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp
index 75ead4c775c6c..45c446eba41dc 100644
--- a/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp
+++ b/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp
@@ -6321,7 +6321,10 @@ static bool hasMoreUses(const MachineInstr &MI0, const MachineInstr &MI1,
 ///   - fmul --> fneg --> fsub: Contraction through fneg
 ///   - fmul --> fneg --> fpext --> fsub: FNEG then FPEXT folds if foldable
 ///   - fmul --> fpext --> {fadd, fsub}: FPEXT folds if foldable
+///   - fmul --> fpext --> {fma, fmad} --> {fadd, fsub}: FPEXT chain reassoc
 ///   - fmul --> fpext --> fneg --> fsub: FPEXT then FNEG to FSUB
+///   - fmul --> {fma, fmad} --> {fadd, fsub} (reassoc): Direct chain reassoc
+///   - fmul --> {fma, fmad} --> fpext --> {fadd, fsub}: FPEXT outer chain
 bool CombinerHelper::allMulUsesCanBeContracted(
     const MachineInstr &MI, unsigned PreferredFusedOpcode) const {
   const auto &TLI = getTargetLowering();
@@ -6367,7 +6370,8 @@ bool CombinerHelper::allMulUsesCanBeContracted(
       continue;
     }
 
-    // FP_EXTEND - check if ALL users are FADD, FSUB, or FNEG --> FSUB
+    // FP_EXTEND - check if ALL users are FADD, FSUB, FNEG --> FSUB, or
+    // FMA/FMAD --> {FADD, FSUB}
     if (Opcode == TargetOpcode::G_FPEXT) {
       Register FPExtReg = UseMI.getOperand(0).getReg();
 
@@ -6383,6 +6387,26 @@ bool CombinerHelper::allMulUsesCanBeContracted(
             ExtUseOpcode == TargetOpcode::G_FSUB) {
           continue;
         }
+        if (ExtUseOpcode == TargetOpcode::G_FMA ||
+            ExtUseOpcode == TargetOpcode::G_FMAD) {
+          // FPEXT --> FMA/FMAD is only contractable if the FMA/FMAD is
+          // used by FADD or FSUB (chain reassociation can fire to
+          // eliminate the multiply).
+          Register FMAReg = FPExtUseMI.getOperand(0).getReg();
+          bool FMAUsedByAddSub = false;
+          for (const MachineInstr &FMAUseMI :
+               MRI.use_nodbg_instructions(FMAReg)) {
+            unsigned FMAUseOp = FMAUseMI.getOpcode();
+            if (FMAUseOp == TargetOpcode::G_FADD ||
+                FMAUseOp == TargetOpcode::G_FSUB) {
+              FMAUsedByAddSub = true;
+              break;
+            }
+          }
+          if (FMAUsedByAddSub)
+            continue;
+          return false;
+        }
         if (ExtUseOpcode == TargetOpcode::G_FNEG) {
           // FP_EXTEND --> FNEG --> FSUB
           Register FPExtFNegReg = FPExtUseMI.getOperand(0).getReg();
@@ -6398,6 +6422,34 @@ bool CombinerHelper::allMulUsesCanBeContracted(
       continue;
     }
 
+    // FMA/FMAD - the multiply is used as an operand of an FMA. This is
+    // contractable only if chain reassociation can fire to eliminate the
+    // multiply. Chain reassociation transforms:
+    //   fadd/fsub(fma(a, b, fmul(c, d)), e) -> fma(a, b, fma(c, d, e))
+    // For direct fmul -> fma -> fadd/fsub, this requires the consumer
+    // fadd/fsub to have the reassoc flag. For fmul -> fma -> fpext ->
+    // fadd/fsub, the aggressive fpext folds handle it without reassoc.
+    if (Opcode == TargetOpcode::G_FMA || Opcode == TargetOpcode::G_FMAD) {
+      Register FMAReg = UseMI.getOperand(0).getReg();
+      bool FMAIsContractable = false;
+      for (const MachineInstr &FMAUseMI : MRI.use_nodbg_instructions(FMAReg)) {
+        unsigned FMAUseOp = FMAUseMI.getOpcode();
+        if (FMAUseOp == TargetOpcode::G_FPEXT) {
+          FMAIsContractable = true;
+          break;
+        }
+        if ((FMAUseOp == TargetOpcode::G_FADD ||
+             FMAUseOp == TargetOpcode::G_FSUB) &&
+            FMAUseMI.getFlag(MachineInstr::MIFlag::FmReassoc)) {
+          FMAIsContractable = true;
+          break;
+        }
+      }
+      if (FMAIsContractable)
+        continue;
+      return false;
+    }
+
     // Any other use type is not currently recognized as contractable.
     return false;
   }
@@ -6666,6 +6718,7 @@ bool CombinerHelper::matchCombineFAddFpExtFMulToFMadOrFMAAggressive(
       mi_match(LHS.MI->getOperand(3).getReg(), MRI,
                m_GFPExt(m_MInstr(FMulMI))) &&
       isContractableFMul(*FMulMI, AllowFusionGlobally) &&
+      allMulUsesCanBeContracted(*FMulMI, PreferredFusedOpcode) &&
       TLI.isFPExtFoldable(MI, PreferredFusedOpcode, DstType,
                           MRI.getType(FMulMI->getOperand(0).getReg()))) {
     MatchInfo = [=](MachineIRBuilder &B) {
@@ -6686,6 +6739,7 @@ bool CombinerHelper::matchCombineFAddFpExtFMulToFMadOrFMAAggressive(
       FMAMI->getOpcode() == PreferredFusedOpcode) {
     MachineInstr *FMulMI = MRI.getVRegDef(FMAMI->getOperand(3).getReg());
     if (isContractableFMul(*FMulMI, AllowFusionGlobally) &&
+        allMulUsesCanBeContracted(*FMulMI, PreferredFusedOpcode) &&
         TLI.isFPExtFoldable(MI, PreferredFusedOpcode, DstType,
                             MRI.getType(FMAMI->getOperand(0).getReg()))) {
       MatchInfo = [=](MachineIRBuilder &B) {
@@ -6707,6 +6761,7 @@ bool CombinerHelper::matchCombineFAddFpExtFMulToFMadOrFMAAggressive(
       mi_match(RHS.MI->getOperand(3).getReg(), MRI,
                m_GFPExt(m_MInstr(FMulMI))) &&
       isContractableFMul(*FMulMI, AllowFusionGlobally) &&
+      allMulUsesCanBeContracted(*FMulMI, PreferredFusedOpcode) &&
       TLI.isFPExtFoldable(MI, PreferredFusedOpcode, DstType,
                           MRI.getType(FMulMI->getOperand(0).getReg()))) {
     MatchInfo = [=](MachineIRBuilder &B) {
@@ -6727,6 +6782,7 @@ bool CombinerHelper::matchCombineFAddFpExtFMulToFMadOrFMAAggressive(
       FMAMI->getOpcode() == PreferredFusedOpcode) {
     MachineInstr *FMulMI = MRI.getVRegDef(FMAMI->getOperand(3).getReg());
     if (isContractableFMul(*FMulMI, AllowFusionGlobally) &&
+        allMulUsesCanBeContracted(*FMulMI, PreferredFusedOpcode) &&
         TLI.isFPExtFoldable(MI, PreferredFusedOpcode, DstType,
                             MRI.getType(FMAMI->getOperand(0).getReg()))) {
       MatchInfo = [=](MachineIRBuilder &B) {
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 6593ff2f3f52b..a8be683e5285d 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -17690,7 +17690,10 @@ static bool isFusedOp(const MatchContextClass &Matcher, SDValue N) {
 ///   - fmul --> fneg --> fsub: Contraction through fneg
 ///   - fmul --> fneg --> fpext --> fsub: FNEG then FPEXT folds if foldable
 ///   - fmul --> fpext --> {fadd, fsub}: FPEXT folds if foldable
+///   - fmul --> fpext --> {fma, fmad} --> {fadd, fsub}: FPEXT chain reassoc
 ///   - fmul --> fpext --> fneg --> fsub: FPEXT then FNEG to FSUB
+///   - fmul --> {fma, fmad} --> {fadd, fsub} (reassoc): Direct chain reassoc
+///   - fmul --> {fma, fmad} --> fpext --> {fadd, fsub}: FPEXT outer chain
 static bool allMulUsesCanBeContracted(SDValue Mul,
                                       const unsigned PreferredFusedOpcode,
                                       const TargetLowering &TLI,
@@ -17729,7 +17732,8 @@ static bool allMulUsesCanBeContracted(SDValue Mul,
       continue; // All FNEG uses are contractable
     }
 
-    // FP_EXTEND - check if ALL users are FADD, FSUB, or FNEG --> FSUB
+    // FP_EXTEND - check if ALL users are FADD, FSUB, FNEG --> FSUB, or
+    // FMA/FMAD --> {FADD, FSUB}
     if (Opcode == ISD::FP_EXTEND) {
       EVT SrcVT = Mul.getValueType();
 
@@ -17742,6 +17746,22 @@ static bool allMulUsesCanBeContracted(SDValue Mul,
         if (ExtUserOp == ISD::FADD || ExtUserOp == ISD::FSUB) {
           continue; // FPEXT --> {FADD, FSUB} is contractable
         }
+        if (ExtUserOp == ISD::FMA || ExtUserOp == ISD::FMAD) {
+          // FPEXT --> FMA/FMAD is only contractable if the FMA/FMAD is
+          // used by FADD or FSUB (chain reassociation can fire to
+          // eliminate the multiply).
+          bool FMAUsedByAddSub = false;
+          for (const auto *FMAUser : FPExtUser->users()) {
+            unsigned FMAUserOp = FMAUser->getOpcode();
+            if (FMAUserOp == ISD::FADD || FMAUserOp == ISD::FSUB) {
+              FMAUsedByAddSub = true;
+              break;
+            }
+          }
+          if (FMAUsedByAddSub)
+            continue;
+          return false;
+        }
         if (ExtUserOp == ISD::FNEG) {
           // FP_EXTEND --> FNEG --> FSUB
           for (const auto *FPExtFNegUser : FPExtUser->users()) {
@@ -17756,6 +17776,32 @@ static bool allMulUsesCanBeContracted(SDValue Mul,
       continue; // All FPEXT uses are contractable
     }
 
+    // FMA/FMAD - the multiply is used as an operand of an FMA. This is
+    // contractable only if chain reassociation can fire to eliminate the
+    // multiply. Chain reassociation transforms:
+    //   fadd/fsub(fma(a, b, fmul(c, d)), e) -> fma(a, b, fma(c, d, e))
+    // For direct fmul -> fma -> fadd/fsub, this requires the consumer
+    // fadd/fsub to have the reassoc flag. For fmul -> fma -> fpext ->
+    // fadd/fsub, the aggressive fpext folds handle it without reassoc.
+    if (Opcode == ISD::FMA || Opcode == ISD::FMAD) {
+      bool FMAIsContractable = false;
+      for (const auto *FMAUser : UserNode->users()) {
+        unsigned FMAUserOp = FMAUser->getOpcode();
+        if (FMAUserOp == ISD::FP_EXTEND) {
+          FMAIsContractable = true;
+          break;
+        }
+        if ((FMAUserOp == ISD::FADD || FMAUserOp == ISD::FSUB) &&
+            FMAUser->getFlags().hasAllowReassociation()) {
+          FMAIsContractable = true;
+          break;
+        }
+      }
+      if (FMAIsContractable)
+        continue;
+      return false;
+    }
+
     // Any other use type is not currently recognized as contractable.
     return false;
   }
@@ -17926,6 +17972,7 @@ SDValue DAGCombiner::visitFADDForFMACombine(SDNode *N) {
       if (matcher.match(N02, ISD::FP_EXTEND)) {
         SDValue N020 = N02.getOperand(0);
         if (isContractableFMUL(N020) &&
+            allMulUsesCanBeContracted(N020, PreferredFusedOpcode, TLI, DAG) &&
             TLI.isFPExtFoldable(DAG, PreferredFusedOpcode, VT,
                                 N020.getValueType())) {
           return FoldFAddFMAFPExtFMul(N0.getOperand(0), N0.getOperand(1),
@@ -17955,6 +18002,7 @@ SDValue DAGCombiner::visitFADDForFMACombine(SDNode *N) {
       if (isFusedOp(matcher, N00)) {
         SDValue N002 = N00.getOperand(2);
         if (isContractableFMUL(N002) &&
+            allMulUsesCanBeContracted(N002, PreferredFusedOpcode, TLI, DAG) &&
             TLI.isFPExtFoldable(DAG, PreferredFusedOpcode, VT,
                                 N00.getValueType())) {
           return FoldFAddFPExtFMAFMul(N00.getOperand(0), N00.getOperand(1),
@@ -17971,6 +18019,7 @@ SDValue DAGCombiner::visitFADDForFMACombine(SDNode *N) {
       if (N12.getOpcode() == ISD::FP_EXTEND) {
         SDValue N120 = N12.getOperand(0);
         if (isContractableFMUL(N120) &&
+            allMulUsesCanBeContracted(N120, PreferredFusedOpcode, TLI, DAG) &&
             TLI.isFPExtFoldable(DAG, PreferredFusedOpcode, VT,
                                 N120.getValueType())) {
           return FoldFAddFMAFPExtFMul(N1.getOperand(0), N1.getOperand(1),
@@ -17990,6 +18039,7 @@ SDValue DAGCombiner::visitFADDForFMACombine(SDNode *N) {
       if (isFusedOp(matcher, N10)) {
         SDValue N102 = N10.getOperand(2);
         if (isContractableFMUL(N102) &&
+            allMulUsesCanBeContracted(N102, PreferredFusedOpcode, TLI, DAG) &&
             TLI.isFPExtFoldable(DAG, PreferredFusedOpcode, VT,
                                 N10.getValueType())) {
           return FoldFAddFPExtFMAFMul(N10.getOperand(0), N10.getOperand(1),
@@ -18244,6 +18294,7 @@ SDValue DAGCombiner::visitFSUBForFMACombine(SDNode *N) {
       if (matcher.match(N02, ISD::FP_EXTEND)) {
         SDValue N020 = N02.getOperand(0);
         if (isContractableAndReassociableFMUL(N020) &&
+            allMulUsesCanBeContracted(N020, PreferredFusedOpcode, TLI, DAG) &&
             TLI.isFPExtFoldable(DAG, PreferredFusedOpcode, VT,
                                 N020.getValueType())) {
           return matcher.getNode(
@@ -18268,6 +18319,7 @@ SDValue DAGCombiner::visitFSUBForFMACombine(SDNode *N) {
       if (isFusedOp(matcher, N00)) {
         SDValue N002 = N00.getOperand(2);
         if (isContractableAndReassociableFMUL(N002) &&
+            allMulUsesCanBeContracted(N002, PreferredFusedOpcode, TLI, DAG) &&
             TLI.isFPExtFoldable(DAG, PreferredFusedOpcode, VT,
                                 N00.getValueType())) {
           return matcher.getNode(
@@ -18289,6 +18341,7 @@ SDValue DAGCombiner::visitFSUBForFMACombine(SDNode *N) {
         matcher.match(N1.getOperand(2), ISD::FP_EXTEND) && N1->hasOneUse()) {
       SDValue N120 = N1.getOperand(2).getOperand(0);
       if (isContractableAndReassociableFMUL(N120) &&
+          allMulUsesCanBeContracted(N120, PreferredFusedOpcode, TLI, DAG) &&
           TLI.isFPExtFoldable(DAG, PreferredFusedOpcode, VT,
                               N120.getValueType())) {
         SDValue N1200 = N120.getOperand(0);
@@ -18318,6 +18371,7 @@ SDValue DAGCombiner::visitFSUBForFMACombine(SDNode *N) {
       SDValue N101 = CvtSrc.getOperand(1);
       SDValue N102 = CvtSrc.getOperand(2);
       if (isContractableAndReassociableFMUL(N102) &&
+          allMulUsesCanBeContracted(N102, PreferredFusedOpcode, TLI, DAG) &&
           TLI.isFPExtFoldable(DAG, PreferredFusedOpcode, VT,
                               CvtSrc.getValueType())) {
         SDValue N1020 = N102.getOperand(0);
diff --git a/llvm/test/CodeGen/AMDGPU/dagcombine-fma-crash.ll b/llvm/test/CodeGen/AMDGPU/dagcombine-fma-crash.ll
index 57070e763e79b..142494a803755 100644
--- a/llvm/test/CodeGen/AMDGPU/dagcombine-fma-crash.ll
+++ b/llvm/test/CodeGen/AMDGPU/dagcombine-fma-crash.ll
@@ -20,24 +20,22 @@ define void @main(float %arg) {
   ; CHECK-NEXT: bb.1.bb2:
   ; CHECK-NEXT:   successors: %bb.2(0x80000000)
   ; CHECK-NEXT: {{  $}}
-  ; CHECK-NEXT:   [[V_MUL_F32_e64_:%[0-9]+]]:vgpr_32 = contract reassoc nofpexcept V_MUL_F32_e64 0, [[S_MOV_B32_]], 0, [[S_MOV_B32_]], 0, 0, implicit $mode, implicit $exec
-  ; CHECK-NEXT:   [[S_MOV_B32_2:%[0-9]+]]:sgpr_32 = S_MOV_B32 1065353216
-  ; CHECK-NEXT:   [[V_ADD_F32_e64_:%[0-9]+]]:vgpr_32 = contract reassoc nofpexcept V_ADD_F32_e64 0, [[V_MUL_F32_e64_]], 0, [[S_MOV_B32_2]], 0, 0, implicit $mode, implicit $exec
-  ; CHECK-NEXT:   [[V_FMAC_F32_e64_:%[0-9]+]]:vgpr_32 = contract reassoc nofpexcept V_FMAC_F32_e64 0, [[COPY]], 0, [[COPY]], 0, [[V_ADD_F32_e64_]], 0, 0, implicit $mode, implicit $exec
-  ; CHECK-NEXT:   [[V_FMAC_F32_e64_1:%[0-9]+]]:vgpr_32 = contract reassoc nofpexcept V_FMAC_F32_e64 0, [[COPY]], 0, [[COPY]], 0, [[V_MUL_F32_e64_]], 0, 0, implicit $mode, implicit $exec
-  ; CHECK-NEXT:   [[V_ADD_F32_e64_1:%[0-9]+]]:vgpr_32 = contract reassoc nofpexcept V_ADD_F32_e64 0, killed [[V_FMAC_F32_e64_1]], 0, [[S_MOV_B32_2]], 0, 0, implicit $mode, implicit $exec
-  ; CHECK-NEXT:   [[S_MOV_B32_3:%[0-9]+]]:sreg_32 = S_MOV_B32 0
+  ; CHECK-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 1065353216, implicit $exec
+  ; CHECK-NEXT:   [[V_FMAC_F32_e64_:%[0-9]+]]:vgpr_32 = contract reassoc nofpexcept V_FMAC_F32_e64 0, [[S_MOV_B32_]], 0, [[S_MOV_B32_]], 0, [[V_MOV_B32_e32_]], 0, 0, implicit $mode, implicit $exec
+  ; CHECK-NEXT:   [[V_FMAC_F32_e64_1:%[0-9]+]]:vgpr_32 = contract reassoc nofpexcept V_FMAC_F32_e64 0, [[COPY]], 0, [[COPY]], 0, [[V_FMAC_F32_e64_]], 0, 0, implicit $mode, implicit $exec
+  ; CHECK-NEXT:   [[V_ADD_F32_e64_:%[0-9]+]]:vgpr_32 = contract reassoc nofpexcept V_ADD_F32_e64 0, [[V_FMAC_F32_e64_1]], 0, [[V_MOV_B32_e32_]], 0, 0, implicit $mode, implicit $exec
+  ; CHECK-NEXT:   [[S_MOV_B32_2:%[0-9]+]]:sreg_32 = S_MOV_B32 0
   ; CHECK-NEXT: {{  $}}
   ; CHECK-NEXT: bb.2.bb11:
   ; CHECK-NEXT:   successors: %bb.3(0x40000000), %bb.4(0x40000000)
   ; CHECK-NEXT: {{  $}}
-  ; CHECK-NEXT:   [[PHI:%[0-9]+]]:vgpr_32 = PHI [[DEF]], %bb.0, [[V_FMAC_F32_e64_]], %bb.1
-  ; CHECK-NEXT:   [[PHI1:%[0-9]+]]:vgpr_32 = PHI [[DEF]], %bb.0, [[V_ADD_F32_e64_1]], %bb.1
-  ; CHECK-NEXT:   [[PHI2:%[0-9]+]]:sreg_32_xm0_xexec = PHI [[S_MOV_B32_1]], %bb.0, [[S_MOV_B32_3]], %bb.1
+  ; CHECK-NEXT:   [[PHI:%[0-9]+]]:vgpr_32 = PHI [[DEF]], %bb.0, [[V_FMAC_F32_e64_1]], %bb.1
+  ; CHECK-NEXT:   [[PHI1:%[0-9]+]]:vgpr_32 = PHI [[DEF]], %bb.0, [[V_ADD_F32_e64_]], %bb.1
+  ; CHECK-NEXT:   [[PHI2:%[0-9]+]]:sreg_32_xm0_xexec = PHI [[S_MOV_B32_1]], %bb.0, [[S_MOV_B32_2]], %bb.1
   ; CHECK-NEXT:   [[V_CNDMASK_B32_e64_:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, 0, 0, 1, [[PHI2]], implicit $exec
-  ; CHECK-NEXT:   [[S_MOV_B32_4:%[0-9]+]]:sreg_32 = S_MOV_B32 1
+  ; CHECK-NEXT:   [[S_MOV_B32_3:%[0-9]+]]:sreg_32 = S_MOV_B32 1
   ; CHECK-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY [[V_CNDMASK_B32_e64_]]
-  ; CHECK-NEXT:   S_CMP_LG_U32 killed [[COPY1]], killed [[S_MOV_B32_4]], implicit-def $scc
+  ; CHECK-NEXT:   S_CMP_LG_U32 killed [[COPY1]], killed [[S_MOV_B32_3]], implicit-def $scc
   ; CHECK-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $scc
   ; CHECK-NEXT:   [[S_AND_B32_1:%[0-9]+]]:sreg_32 = S_AND_B32 $exec_lo, killed [[COPY2]], implicit-def dead $scc
   ; CHECK-NEXT:   $vcc_lo = COPY [[S_AND_B32_1]]
diff --git a/llvm/test/CodeGen/AMDGPU/dagcombine-fma-fmad.ll b/llvm/test/CodeGen/AMDGPU/dagcombine-fma-fmad.ll
index e95f19a19d134..8e0e7d2a735f8 100644
--- a/llvm/test/CodeGen/AMDGPU/dagcombine-fma-fmad.ll
+++ b/llvm/test/CodeGen/AMDGPU/dagcombine-fma-fmad.ll
@@ -17,57 +17,54 @@ define amdgpu_ps float @_amdgpu_ps_main() #0 {
 ; GFX10-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    image_load_mip v4, v[2:4], s[0:7] dmask:0x4 dim:SQ_RSRC_IMG_2D unorm
-; GFX10-NEXT:    s_clause 0x2
+; GFX10-NEXT:    s_clause 0x3
 ; GFX10-NEXT:    s_buffer_load_dword s24, s[0:3], 0x5c
 ; GFX10-NEXT:    s_buffer_load_dword s25, s[0:3], 0x7c
+; GFX10-NEXT:    s_buffer_load_dword s28, s[0:3], 0xc0
 ; GFX10-NEXT:    s_waitcnt_depctr depctr_vm_vsrc(0)
 ; GFX10-NEXT:    s_nop 0
-; GFX10-NEXT:    s_buffer_load_dwordx4 s[0:3], s[0:3], 0x40
-; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-NEXT:    s_clause 0x2
-; GFX10-NEXT:    s_buffer_load_dword s26, s[0:3], 0xc0
-; GFX10-NEXT:    s_buffer_load_dwordx4 s[4:7], s[0:3], 0x50
-; GFX10-NEXT:    s_buffer_load_dwordx4 s[8:11], s[0:3], 0x60
+; GFX10-NEXT:    s_buffer_load_dwordx4 s[0:3], s[0:3], 0x70
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-NEXT:    s_clause 0x3
-; GFX10-NEXT:    s_buffer_load_dword s4, s[0:3], 0x2c
-; GFX10-NEXT:    s_buffer_load_dwordx4 s[12:15], s[0:3], 0x70
-; GFX10-NEXT:    s_buffer_load_dwordx4 s[16:19], s[0:3], 0x20
-; GFX10-NEXT:    s_buffer_load_dwordx4 s[20:23], s[0:3], 0x0
-; GFX10-NEXT:    v_max_f32_e64 v5, s0, s0 clamp
-; GFX10-NEXT:    v_sub_f32_e64 v6, s24, s25
-; GFX10-NEXT:    v_mul_f32_e32 v7, s2, v5
-; GFX10-NEXT:    s_buffer_load_dwordx4 s[0:3], s[0:3], 0x10
-; GFX10-NEXT:    v_fma_f32 v1, v1, v6, s25
+; GFX10-NEXT:    v_sub_f32_e64 v5, s24, s25
+; GFX10-NEXT:    s_buffer_load_dword s0, s[0:3], 0x2c
+; GFX10-NEXT:    v_mul_f32_e32 v2, s2, v2
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_clause 0x4
+; GFX10-NEXT:    s_buffer_load_dwordx4 s[4:7], s[0:3], 0x40
+; GFX10-NEXT:    s_buffer_load_dwordx4 s[8:11], s[0:3], 0x50
+; GFX10-NEXT:    s_buffer_load_dwordx4 s[12:15], s[0:3], 0x0
+; GFX10-NEXT:    s_buffer_load_dwordx4 s[16:19], s[0:3], 0x60
+; GFX10-NEXT:    s_buffer_load_dwordx4 s[20:23], s[0:3], 0x20
+; GFX10-NEXT:    v_fma_f32 v1, v1, v5, s25
+; GFX10-NEXT:    s_buffer_load_dwordx4 s[24:27], s[0:3], 0x10
+; GFX10-NEXT:    v_max_f32_e64 v6, s0, s0 clamp
+; GFX10-NEXT:    v_add_f32_e64 v5, s28, -1.0
+; GFX10-NEXT:    v_sub_f32_e32 v7, s0, v1
 ; GFX10-NEXT:    s_mov_b32 s0, 0x3c23d70a
-; GFX10-NEXT:    v_add_f32_e64 v6, s26, -1.0
-; GFX10-NEXT:    v_sub_f32_e32 v8, s6, v7
-; GFX10-NEXT:    v_mul_f32_e32 v0, s10, v0
-; GFX10-NEXT:    v_sub_f32_e32 v9, s4, v1
-; GFX10-NEXT:    v_mul_f32_e32 v2, s14, v2
-; GFX10-NEXT:    v_fma_f32 v6, v5, v6, 1.0
-; GFX10-NEXT:    v_fmac_f32_e32 v7, v8, v5
-; GFX10-NEXT:    v_sub_f32_e32 v8, s18, v0
-; GFX10-NEXT:    v_fmac_f32_e32 v1, v5, v9
-; GFX10-NEXT:    v_mul_f32_e32 v9, s22, v3
-; GFX10-NEXT:    v_mul_f32_e32 v2, v5, v2
-; GFX10-NEXT:    v_add_f32_e32 v7, v3, v7
-; GFX10-NEXT:    v_fmac_f32_e32 v0, v8, v5
-; GFX10-NEXT:    v_sub_f32_e32 v1, v1, v6
-; GFX10-NEXT:    v_mul_f32_e32 v8, v9, v5
-; GFX10-NEXT...
[truncated]

``````````

</details>


https://github.com/llvm/llvm-project/pull/188117


More information about the llvm-branch-commits mailing list