[llvm] [AMDGPU] Disable dpp src1 sgpr on gfx11 (PR #164241)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-amdgpu

Author: Paul Trojahn (ptrojahn)

<details>
<summary>Changes</summary>

https://github.com/llvm/llvm-project/pull/67461 enabled SGPRs as src1 by default for all opcodes with manual checks for targets where this is not supported. https://github.com/llvm/llvm-project/pull/155595 disabled this check for hasDPPSrc1SGPR, which resulted in SGPRs being used as operands for src1 on gfx11. This PR reenables this check and fixes the lit test.

---
Full diff: https://github.com/llvm/llvm-project/pull/164241.diff


2 Files Affected:

- (modified) llvm/lib/Target/AMDGPU/GCNDPPCombine.cpp (+5-3) 
- (modified) llvm/test/CodeGen/AMDGPU/dpp_combine_gfx11.mir (+7-5) 


``````````diff
diff --git a/llvm/lib/Target/AMDGPU/GCNDPPCombine.cpp b/llvm/lib/Target/AMDGPU/GCNDPPCombine.cpp
index 464cbec6c46bc..e394096b9fe09 100644
--- a/llvm/lib/Target/AMDGPU/GCNDPPCombine.cpp
+++ b/llvm/lib/Target/AMDGPU/GCNDPPCombine.cpp
@@ -307,10 +307,12 @@ MachineInstr *GCNDPPCombine::createDPPInst(MachineInstr &OrigMI,
     if (Src1) {
       assert(AMDGPU::hasNamedOperand(DPPOp, AMDGPU::OpName::src1) &&
              "dpp version of instruction missing src1");
-      // If subtarget does not support SGPRs for src1 operand then the
-      // requirements are the same as for src0. We check src0 instead because
-      // pseudos are shared between subtargets and allow SGPR for src1 on all.
       if (!ST->hasDPPSrc1SGPR()) {
+        if (Src1->isReg() && ST->getRegisterInfo()->isSGPRClass(
+                                 MRI->getRegClass(Src1->getReg()))) {
+          Fail = true;
+          break;
+        }
         assert(TII->getOpSize(*DPPInst, Src0Idx) ==
                    TII->getOpSize(*DPPInst, NumOperands) &&
                "Src0 and Src1 operands should have the same size");
diff --git a/llvm/test/CodeGen/AMDGPU/dpp_combine_gfx11.mir b/llvm/test/CodeGen/AMDGPU/dpp_combine_gfx11.mir
index 3725384e885ee..87ec1eabc2033 100644
--- a/llvm/test/CodeGen/AMDGPU/dpp_combine_gfx11.mir
+++ b/llvm/test/CodeGen/AMDGPU/dpp_combine_gfx11.mir
@@ -1,6 +1,6 @@
-# RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -run-pass=gcn-dpp-combine -verify-machineinstrs -o - %s | FileCheck %s -check-prefixes=GCN
-# RUN: llc -mtriple=amdgcn -mcpu=gfx1150 -mattr=-real-true16 -run-pass=gcn-dpp-combine -verify-machineinstrs -o - %s | FileCheck %s -check-prefixes=GCN
-# RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -mattr=-real-true16 -run-pass=gcn-dpp-combine -verify-machineinstrs -o - %s | FileCheck %s -check-prefixes=GCN
+# RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -run-pass=gcn-dpp-combine -verify-machineinstrs -o - %s | FileCheck %s -check-prefixes=GCN,GFX1100
+# RUN: llc -mtriple=amdgcn -mcpu=gfx1150 -mattr=-real-true16 -run-pass=gcn-dpp-combine -verify-machineinstrs -o - %s | FileCheck %s -check-prefixes=GCN,GFX1150
+# RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -mattr=-real-true16 -run-pass=gcn-dpp-combine -verify-machineinstrs -o - %s | FileCheck %s -check-prefixes=GCN,GFX1150
 
 ---
 
@@ -38,8 +38,10 @@ body:             |
 
 # GCN-LABEL: name: vop3_sgpr_src1
 # GCN: %6:vgpr_32 = V_MED3_F32_e64_dpp %4, 0, %0, 0, %1, 0, %2, 0, 0, 1, 15, 15, 1, implicit $mode, implicit $exec
-# GCN: %8:vgpr_32 = V_MED3_F32_e64_dpp %4, 0, %0, 0, %2, 0, %1, 0, 0, 1, 15, 15, 1, implicit $mode, implicit $exec
-# GCN: %10:vgpr_32 = V_MED3_F32_e64_dpp %4, 0, %0, 0, %2, 0, %3, 0, 0, 1, 15, 15, 1, implicit $mode, implicit $exec
+# GFX1100: %8:vgpr_32 = V_MED3_F32_e64 0, %7, 0, %2, 0, %1, 0, 0, implicit $mode, implicit $exec
+# GFX1150: %8:vgpr_32 = V_MED3_F32_e64_dpp %4, 0, %0, 0, %2, 0, %1, 0, 0, 1, 15, 15, 1, implicit $mode, implicit $exec
+# GFX1100: %10:vgpr_32 = V_MED3_F32_e64 0, %9, 0, %2, 0, %3, 0, 0, implicit $mode, implicit $exec
+# GFX1150: %10:vgpr_32 = V_MED3_F32_e64_dpp %4, 0, %0, 0, %2, 0, %3, 0, 0, 1, 15, 15, 1, implicit $mode, implicit $exec
 # GCN: %12:vgpr_32 = V_MED3_F32_e64_dpp %4, 0, %0, 0, 42, 0, %2, 0, 0, 1, 15, 15, 1, implicit $mode, implicit $exec
 # GCN: %14:vgpr_32 = V_MED3_F32_e64 0, %13, 0, 4242, 0, %2, 0, 0, implicit $mode, implicit $exec
 name: vop3_sgpr_src1

``````````

</details>


https://github.com/llvm/llvm-project/pull/164241