Reland "[NVPTX] Use .common linkage for common globals" (#86824)

Switch from `.weak` to `.common` linkage for common global variables
where possible. The `.common` linkage is described in
[PTX ISA 11.6.4. Linking Directives: .common]
(https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#linking-directives-common)
> Declares identifier to be globally visible but “common”.
>
>Common symbols are similar to globally visible symbols. However
multiple object files may declare the same common symbol and they may
have different types and sizes and references to a symbol get resolved
against a common symbol with the largest size.
>
>Only one object file can initialize a common symbol and that must have
the largest size among all other definitions of that common symbol from
different object files.
>
>.common linking directive can be used only on variables with .global
storage. It cannot be used on function symbols or on symbols with opaque
type.

I've updated the logic and tests to only use `.common` for PTX 5.0 or
greater and verified that the new tests now pass with `ptxas`.

Author: AlexMaclean

llvm/lib/Target/NVPTX/NVPTXAsmPrinter.cpp

@@ -1022,14 +1022,16 @@ void NVPTXAsmPrinter::printModuleLevelGV(const GlobalVariable *GVar,
   const DataLayout &DL = getDataLayout();
 
   // GlobalVariables are always constant pointers themselves.
-  PointerType *PTy = GVar->getType();
   Type *ETy = GVar->getValueType();
 
   if (GVar->hasExternalLinkage()) {
     if (GVar->hasInitializer())
       O << ".visible ";
     else
       O << ".extern ";
+  } else if (STI.getPTXVersion() >= 50 && GVar->hasCommonLinkage() &&
+             GVar->getAddressSpace() == ADDRESS_SPACE_GLOBAL) {
+    O << ".common ";
   } else if (GVar->hasLinkOnceLinkage() || GVar->hasWeakLinkage() ||
              GVar->hasAvailableExternallyLinkage() ||
              GVar->hasCommonLinkage()) {
@@ -1141,7 +1143,7 @@ void NVPTXAsmPrinter::printModuleLevelGV(const GlobalVariable *GVar,
   }
 
   O << ".";
-  emitPTXAddressSpace(PTy->getAddressSpace(), O);
+  emitPTXAddressSpace(GVar->getAddressSpace(), O);
 
   if (isManaged(*GVar)) {
     if (STI.getPTXVersion() < 40 || STI.getSmVersion() < 30) {
@@ -1170,8 +1172,8 @@ void NVPTXAsmPrinter::printModuleLevelGV(const GlobalVariable *GVar,
     // Ptx allows variable initilization only for constant and global state
     // spaces.
     if (GVar->hasInitializer()) {
-      if ((PTy->getAddressSpace() == ADDRESS_SPACE_GLOBAL) ||
-          (PTy->getAddressSpace() == ADDRESS_SPACE_CONST)) {
+      if ((GVar->getAddressSpace() == ADDRESS_SPACE_GLOBAL) ||
+          (GVar->getAddressSpace() == ADDRESS_SPACE_CONST)) {
         const Constant *Initializer = GVar->getInitializer();
         // 'undef' is treated as there is no value specified.
         if (!Initializer->isNullValue() && !isa<UndefValue>(Initializer)) {
@@ -1186,7 +1188,7 @@ void NVPTXAsmPrinter::printModuleLevelGV(const GlobalVariable *GVar,
             !isa<UndefValue>(GVar->getInitializer())) {
           report_fatal_error("initial value of '" + GVar->getName() +
                              "' is not allowed in addrspace(" +
-                             Twine(PTy->getAddressSpace()) + ")");
+                             Twine(GVar->getAddressSpace()) + ")");
         }
       }
     }
@@ -1205,8 +1207,8 @@ void NVPTXAsmPrinter::printModuleLevelGV(const GlobalVariable *GVar,
       ElementSize = DL.getTypeStoreSize(ETy);
       // Ptx allows variable initilization only for constant and
       // global state spaces.
-      if (((PTy->getAddressSpace() == ADDRESS_SPACE_GLOBAL) ||
-           (PTy->getAddressSpace() == ADDRESS_SPACE_CONST)) &&
+      if (((GVar->getAddressSpace() == ADDRESS_SPACE_GLOBAL) ||
+           (GVar->getAddressSpace() == ADDRESS_SPACE_CONST)) &&
           GVar->hasInitializer()) {
         const Constant *Initializer = GVar->getInitializer();
         if (!isa<UndefValue>(Initializer) && !Initializer->isNullValue()) {

llvm/test/CodeGen/NVPTX/common-linkage.ll

@@ -0,0 +1,29 @@
+; RUN: llc < %s -march=nvptx -mcpu=sm_20 -mattr=+ptx43 | FileCheck %s --check-prefixes CHECK,PTX43
+; RUN: llc < %s -march=nvptx -mcpu=sm_20 -mattr=+ptx50 | FileCheck %s --check-prefixes CHECK,PTX50
+; RUN: %if ptxas %{ llc < %s -march=nvptx64 -mcpu=sm_20 -mattr=+ptx43 | %ptxas-verify %}
+; RUN: %if ptxas %{ llc < %s -march=nvptx64 -mcpu=sm_20 -mattr=+ptx50 | %ptxas-verify %}
+
+; PTX43: .weak .global .align 4 .u32 g
+; PTX50: .common .global .align 4 .u32 g
+@g = common addrspace(1) global i32 0, align 4
+
+; CHECK: .weak .const .align 4 .u32 c
+@c = common addrspace(4) global i32 0, align 4
+
+; CHECK: .weak .shared .align 4 .u32 s
+@s = common addrspace(3) global i32 0, align 4
+
+define i32 @f1() {
+  %1 = load i32, ptr addrspace(1) @g
+  ret i32 %1
+}
+
+define i32 @f4() {
+  %1 = load i32, ptr addrspace(4) @c
+  ret i32 %1
+}
+
+define i32 @f3() {
+  %1 = load i32, ptr addrspace(3) @s
+  ret i32 %1
+}

llvm/test/CodeGen/NVPTX/weak-global.ll

@@ -1,7 +1,10 @@
-; RUN: llc < %s -march=nvptx64 -mcpu=sm_20 | FileCheck %s
-; RUN: %if ptxas %{ llc < %s -march=nvptx64 -mcpu=sm_20 | %ptxas-verify %}
+; RUN: llc < %s -march=nvptx64 -mcpu=sm_20 -mattr=+ptx43 | FileCheck %s --check-prefix PTX43
+; RUN: llc < %s -march=nvptx64 -mcpu=sm_20 -mattr=+ptx50 | FileCheck %s --check-prefix PTX50
+; RUN: %if ptxas %{ llc < %s -march=nvptx64 -mcpu=sm_20 -mattr=+ptx43 | %ptxas-verify %}
+; RUN: %if ptxas %{ llc < %s -march=nvptx64 -mcpu=sm_20 -mattr=+ptx50 | %ptxas-verify %}
 
-; CHECK: .weak .global .align 4 .u32 g
+; PTX43: .weak .global .align 4 .u32 g
+; PTX50: .common .global .align 4 .u32 g
 @g = common addrspace(1) global i32 zeroinitializer
 
 define i32 @func0() {