[PATCH] D138937: [AMDGPU] Update InstrCost calculation

[Yashwant Singh via Phabricator via llvm-commits]

yassingh created this revision.
yassingh added reviewers: rampitec, arsenm, foad.
Herald added subscribers: kosarev, kerbowa, hiraditya, tpr, dstuttard, yaxunl, jvesely, kzhuravl.
Herald added a project: All.
yassingh requested review of this revision.
Herald added subscribers: llvm-commits, wdng.
Herald added a project: LLVM.

Fixes inconsistency in PerfHintAnalysis while using opaque 
pointers. Opaque pointers got rid of bitcasts that affected
both MemoryBound and WaveLimiterHint calculation as 
InstCost will be different.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D138937

Files:
  llvm/lib/Target/AMDGPU/AMDGPUPerfHintAnalysis.cpp
  llvm/test/CodeGen/AMDGPU/perfhint-instr-cost.ll
  llvm/test/CodeGen/AMDGPU/perfhint.ll


Index: llvm/test/CodeGen/AMDGPU/perfhint.ll
===================================================================
--- llvm/test/CodeGen/AMDGPU/perfhint.ll
+++ llvm/test/CodeGen/AMDGPU/perfhint.ll
@@ -1,4 +1,5 @@
 ; RUN: llc -march=amdgcn < %s | FileCheck -check-prefix=GCN %s
+; RUN: llc -march=amdgcn -opaque-pointers=1 < %s | FileCheck -check-prefix=GCN %s
 
 ; GCN-LABEL: {{^}}test_membound:
 ; GCN: MemoryBound: 1
@@ -21,6 +22,7 @@
 
 ; GCN-LABEL: {{^}}test_membound_1:
 ; GCN: MemoryBound: 1
+; GCN: WaveLimiterHint : 1
 define amdgpu_kernel void @test_membound_1(<2 x double> addrspace(1)* nocapture readonly %ptr.0,
                                            <2 x double> addrspace(1)* nocapture %ptr.1,
                                            <2 x double> %arg.0, i32 %arg.1, <4 x double> %arg.2) {
@@ -112,7 +114,7 @@
 }
 
 ; GCN-LABEL: {{^}}test_indirect:
-; GCN: MemoryBound: 0
+; GCN: MemoryBound: 1
 ; GCN: WaveLimiterHint : 1
 define amdgpu_kernel void @test_indirect(i32 addrspace(1)* nocapture %arg) {
 bb:
Index: llvm/test/CodeGen/AMDGPU/perfhint-instr-cost.ll
===================================================================
--- /dev/null
+++ llvm/test/CodeGen/AMDGPU/perfhint-instr-cost.ll
@@ -0,0 +1,15 @@
+; RUN: llc -march=amdgcn -debug-only=amdgpu-perf-hint < %s 2>&1 | FileCheck %s
+; RUN: llc -march=amdgcn -debug-only=amdgpu-perf-hint -opaque-pointers=1 < %s 2>&1 | FileCheck %s
+define i32 @perfHintInstrCost(i8 addrspace(4)* %p1, i8 addrspace(4)* %p2, i8 addrspace(4)* %p3) #0 {
+; CHECK: MemInst cost: 3
+; CHECK: TotalInst cost: 6 
+  %x.cast = bitcast i8 addrspace(4)* %p1 to i32 addrspace(4)*
+  %y.cast = bitcast i8 addrspace(4)* %p2 to i32 addrspace(4)*
+  %z.cast = bitcast i8 addrspace(4)* %p3 to i32 addrspace(4)*
+  %x = load volatile i32, i32 addrspace(4)* %x.cast
+  %y = load volatile i32, i32 addrspace(4)* %y.cast
+  %z = load volatile i32, i32 addrspace(4)* %z.cast
+  %sum = add i32 %x, %y
+  %sum2 = add i32 %sum, %z
+  ret i32 %sum2
+}
\ No newline at end of file
Index: llvm/lib/Target/AMDGPU/AMDGPUPerfHintAnalysis.cpp
===================================================================
--- llvm/lib/Target/AMDGPU/AMDGPUPerfHintAnalysis.cpp
+++ llvm/lib/Target/AMDGPU/AMDGPUPerfHintAnalysis.cpp
@@ -224,6 +224,9 @@
     LastAccess = MemAccessInfo();
     unsigned UsedGlobalLoadsInBB = 0;
     for (auto &I : B) {
+      // skip bitcast for cost calculation
+      if (I.isCast())
+        continue;
       if (const Type *Ty = getMemoryInstrPtrAndType(&I).second) {
         unsigned Size = divideCeil(Ty->getPrimitiveSizeInBits(), 32);
         // TODO: Check if the global load and its user are close to each other


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D138937.478663.patch
Type: text/x-patch
Size: 2673 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20221129/ee48f609/attachment.bin>