[llvm] d4569d4 - [AMDGPU] Let LowerModuleLDS run twice on the same module (#81729)

[via llvm-commits]

Author: Pierre van Houtryve
Date: 2024-03-11T09:20:01+01:00
New Revision: d4569d42b5cb8ba076f0115d3d21d89f68e6ce9d

URL: https://github.com/llvm/llvm-project/commit/d4569d42b5cb8ba076f0115d3d21d89f68e6ce9d
DIFF: https://github.com/llvm/llvm-project/commit/d4569d42b5cb8ba076f0115d3d21d89f68e6ce9d.diff

LOG: [AMDGPU] Let LowerModuleLDS run twice on the same module (#81729)

If all variables in the module are absolute, this means we're running
the pass again on an already lowered module, and that works.
If none of them are absolute, lowering can proceed as usual.
Only diagnose cases where we have a mix of absolute/non-absolute GVs,
which means we added LDS GVs after lowering, which is broken.

See #81491
Split from #75333

Added: 
    llvm/test/CodeGen/AMDGPU/lds-reject-mixed-absolute-addresses.ll
    llvm/test/CodeGen/AMDGPU/lds-run-twice-absolute-md.ll
    llvm/test/CodeGen/AMDGPU/lds-run-twice.ll

Modified: 
    llvm/lib/Target/AMDGPU/AMDGPULowerModuleLDSPass.cpp

Removed: 
    llvm/test/CodeGen/AMDGPU/lds-reject-absolute-addresses.ll


################################################################################
diff  --git a/llvm/lib/Target/AMDGPU/AMDGPULowerModuleLDSPass.cpp b/llvm/lib/Target/AMDGPU/AMDGPULowerModuleLDSPass.cpp
index 5762f1906a16d3..b85cb26fdc9565 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPULowerModuleLDSPass.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPULowerModuleLDSPass.cpp
@@ -340,15 +340,25 @@ class AMDGPULowerModuleLDS {
 
     // Get uses from the current function, excluding uses by called functions
     // Two output variables to avoid walking the globals list twice
+    std::optional<bool> HasAbsoluteGVs;
     for (auto &GV : M.globals()) {
       if (!AMDGPU::isLDSVariableToLower(GV)) {
         continue;
       }
 
-      if (GV.isAbsoluteSymbolRef()) {
-        report_fatal_error(
-            "LDS variables with absolute addresses are unimplemented.");
-      }
+      // Check if the module is consistent: either all GVs are absolute (happens
+      // when we run the pass more than once), or none are.
+      const bool IsAbsolute = GV.isAbsoluteSymbolRef();
+      if (HasAbsoluteGVs.has_value()) {
+        if (*HasAbsoluteGVs != IsAbsolute) {
+          report_fatal_error(
+              "Module cannot mix absolute and non-absolute LDS GVs");
+        }
+      } else
+        HasAbsoluteGVs = IsAbsolute;
+
+      if (IsAbsolute)
+        continue;
 
       for (User *V : GV.users()) {
         if (auto *I = dyn_cast<Instruction>(V)) {

diff  --git a/llvm/test/CodeGen/AMDGPU/lds-reject-absolute-addresses.ll b/llvm/test/CodeGen/AMDGPU/lds-reject-mixed-absolute-addresses.ll
similarity index 81%
rename from llvm/test/CodeGen/AMDGPU/lds-reject-absolute-addresses.ll
rename to llvm/test/CodeGen/AMDGPU/lds-reject-mixed-absolute-addresses.ll
index 659cdb55ded2fe..b512a43aa10222 100644
--- a/llvm/test/CodeGen/AMDGPU/lds-reject-absolute-addresses.ll
+++ b/llvm/test/CodeGen/AMDGPU/lds-reject-mixed-absolute-addresses.ll
@@ -2,8 +2,9 @@
 ; RUN: not --crash opt -S -mtriple=amdgcn-- -passes=amdgpu-lower-module-lds < %s 2>&1 | FileCheck %s
 
 @var1 = addrspace(3) global i32 undef, !absolute_symbol !0
+ at var2 = addrspace(3) global i32 undef
 
-; CHECK: LLVM ERROR: LDS variables with absolute addresses are unimplemented.
+; CHECK: Module cannot mix absolute and non-absolute LDS GVs
 define amdgpu_kernel void @kern() {
   %val0 = load i32, ptr addrspace(3) @var1
   %val1 = add i32 %val0, 4
@@ -12,4 +13,3 @@ define amdgpu_kernel void @kern() {
 }
 
 !0 = !{i32 0, i32 1}
-

diff  --git a/llvm/test/CodeGen/AMDGPU/lds-run-twice-absolute-md.ll b/llvm/test/CodeGen/AMDGPU/lds-run-twice-absolute-md.ll
new file mode 100644
index 00000000000000..52b44eea35c827
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/lds-run-twice-absolute-md.ll
@@ -0,0 +1,16 @@
+; RUN: opt -S -mtriple=amdgcn-- -amdgpu-lower-module-lds %s -o %t.ll
+; RUN: opt -S -mtriple=amdgcn-- -amdgpu-lower-module-lds %t.ll -o %t.second.ll
+; RUN: 
diff  -ub %t.ll %t.second.ll -I ".*ModuleID.*"
+
+; Check AMDGPULowerModuleLDS can run more than once on the same module, and that
+; the second run is a no-op.
+
+ at lds = internal unnamed_addr addrspace(3) global i32 undef, align 4, !absolute_symbol !0
+
+define amdgpu_kernel void @test() {
+entry:
+  store i32 1, ptr addrspace(3) @lds
+  ret void
+}
+
+!0 = !{i32 0, i32 1}

diff  --git a/llvm/test/CodeGen/AMDGPU/lds-run-twice.ll b/llvm/test/CodeGen/AMDGPU/lds-run-twice.ll
new file mode 100644
index 00000000000000..b830ccb944a282
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/lds-run-twice.ll
@@ -0,0 +1,14 @@
+; RUN: opt -S -mtriple=amdgcn-- -amdgpu-lower-module-lds %s -o %t.ll
+; RUN: opt -S -mtriple=amdgcn-- -amdgpu-lower-module-lds %t.ll -o %t.second.ll
+; RUN: 
diff  -ub %t.ll %t.second.ll -I ".*ModuleID.*"
+
+; Check AMDGPULowerModuleLDS can run more than once on the same module, and that
+; the second run is a no-op.
+
+ at lds = internal unnamed_addr addrspace(3) global i32 undef, align 4
+
+define amdgpu_kernel void @test() {
+entry:
+  store i32 1, ptr addrspace(3) @lds
+  ret void
+}