llvmbot wrote:

<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-clang

Author: Erich Keane (erichkeane)

<details>
<summary>Changes</summary>

Like the previous patch, this upstreams inline
assembly support re #<!-- -->153267

This patch extends this to Input Operands.  With this done, we now correctly 
compile every test in the incubator for inline assembly.

NOTE: When doing the IR differences between CIR and Classic Codegen, I noticed 
a sext/zext difference, but I'm reasonably convinced CIR is correct.  Also, we 
keep around the 'ret-val' temporary for a function,
  so that difference is also in the tests.

---

Patch is 50.08 KiB, truncated to 20.00 KiB below, full version: 
https://github.com/llvm/llvm-project/pull/176239.diff


4 Files Affected:

- (modified) clang/include/clang/CIR/MissingFeatures.h (-1) 
- (modified) clang/lib/CIR/CodeGen/CIRGenAsm.cpp (+192-15) 
- (modified) clang/lib/CIR/CodeGen/CIRGenFunction.h (+7) 
- (modified) clang/test/CIR/CodeGen/inline-asm.c (+751-2) 


``````````diff
diff --git a/clang/include/clang/CIR/MissingFeatures.h 
b/clang/include/clang/CIR/MissingFeatures.h
index 39818417fc3d0..2f3814eac0ce7 100644
--- a/clang/include/clang/CIR/MissingFeatures.h
+++ b/clang/include/clang/CIR/MissingFeatures.h
@@ -214,7 +214,6 @@ struct MissingFeatures {
   static bool appleKext() { return false; }
   static bool armComputeVolatileBitfields() { return false; }
   static bool asmGoto() { return false; }
-  static bool asmInputOperands() { return false; }
   static bool asmLabelAttr() { return false; }
   static bool asmLLVMAssume() { return false; }
   static bool asmMemoryEffects() { return false; }
diff --git a/clang/lib/CIR/CodeGen/CIRGenAsm.cpp 
b/clang/lib/CIR/CodeGen/CIRGenAsm.cpp
index 7032781dc97f8..ae60d1a6c4a2a 100644
--- a/clang/lib/CIR/CodeGen/CIRGenAsm.cpp
+++ b/clang/lib/CIR/CodeGen/CIRGenAsm.cpp
@@ -117,6 +117,67 @@ collectInOutConstraintInfos(const CIRGenFunction &cgf, 
const AsmStmt &s,
   }
 }
 
+std::pair<mlir::Value, mlir::Type> CIRGenFunction::emitAsmInputLValue(
+    const TargetInfo::ConstraintInfo &info, LValue inputValue,
+    QualType inputType, std::string &constraintString, SourceLocation loc) {
+
+  if (info.allowsRegister() || !info.allowsMemory()) {
+    if (hasScalarEvaluationKind(inputType))
+      return {emitLoadOfLValue(inputValue, loc).getValue(), mlir::Type()};
+
+    mlir::Type ty = convertType(inputType);
+    uint64_t size = cgm.getDataLayout().getTypeSizeInBits(ty);
+    if ((size <= 64 && llvm::isPowerOf2_64(size)) ||
+        getTargetHooks().isScalarizableAsmOperand(*this, ty)) {
+      ty = cir::IntType::get(&getMLIRContext(), size, false);
+
+      return {builder.createLoad(
+                  getLoc(loc),
+                  inputValue.getAddress().withElementType(builder, ty)),
+              mlir::Type()};
+    }
+  }
+
+  Address addr = inputValue.getAddress();
+  constraintString += '*';
+  return {addr.getPointer(), addr.getElementType()};
+}
+
+std::pair<mlir::Value, mlir::Type>
+CIRGenFunction::emitAsmInput(const TargetInfo::ConstraintInfo &info,
+                             const Expr *inputExpr,
+                             std::string &constraintString) {
+  auto loc = getLoc(inputExpr->getExprLoc());
+
+  // If this can't be a register or memory, i.e., has to be a constant
+  // (immediate or symbolic), try to emit it as such.
+  if (!info.allowsRegister() && !info.allowsMemory()) {
+    if (info.requiresImmediateConstant()) {
+      Expr::EvalResult evalResult;
+      inputExpr->EvaluateAsRValue(evalResult, getContext(), true);
+
+      llvm::APSInt intResult;
+      if (evalResult.Val.toIntegralConstant(intResult, inputExpr->getType(),
+                                            getContext()))
+        return {builder.getConstInt(loc, intResult), mlir::Type()};
+    }
+
+    Expr::EvalResult result;
+    if (inputExpr->EvaluateAsInt(result, getContext()))
+      return {builder.getConstInt(loc, result.Val.getInt()), mlir::Type()};
+  }
+
+  if (info.allowsRegister() || !info.allowsMemory())
+    if (CIRGenFunction::hasScalarEvaluationKind(inputExpr->getType()))
+      return {emitScalarExpr(inputExpr), mlir::Type()};
+  if (inputExpr->getStmtClass() == Expr::CXXThisExprClass)
+    return {emitScalarExpr(inputExpr), mlir::Type()};
+  inputExpr = inputExpr->IgnoreParenNoopCasts(getContext());
+  LValue dest = emitLValue(inputExpr);
+  return emitAsmInputLValue(info, dest, inputExpr->getType(), constraintString,
+                            inputExpr->getExprLoc());
+}
+
 static void emitAsmStores(CIRGenFunction &cgf, const AsmStmt &s,
                           const llvm::ArrayRef<mlir::Value> regResults,
                           const llvm::ArrayRef<mlir::Type> resultRegTypes,
@@ -227,6 +288,11 @@ mlir::LogicalResult CIRGenFunction::emitAsmStmt(const 
AsmStmt &s) {
   llvm::BitVector resultTypeRequiresCast;
   llvm::BitVector resultRegIsFlagReg;
 
+  // Keep track of input constraints.
+  std::string inOutConstraints;
+  std::vector<mlir::Type> inOutArgTypes;
+  std::vector<mlir::Type> inOutArgElemTypes;
+
   // Keep track of out constraints for tied input operand.
   SmallVector<std::string> outputConstraints;
 
@@ -241,11 +307,6 @@ mlir::LogicalResult CIRGenFunction::emitAsmStmt(const 
AsmStmt &s) {
   // in addition to meeting the conditions listed above.
   bool readOnly = true, readNone = true;
 
-  if (s.getNumInputs() != 0) {
-    assert(!cir::MissingFeatures::asmInputOperands());
-    cgm.errorNYI(srcLoc, "asm with input operands");
-  }
-
   std::string outputConstraint;
   for (unsigned i = 0, e = s.getNumOutputs(); i != e; ++i) {
     TargetInfo::ConstraintInfo &info = outputConstraintInfos[i];
@@ -310,15 +371,31 @@ mlir::LogicalResult CIRGenFunction::emitAsmStmt(const 
AsmStmt &s) {
       // If this output is tied to an input, and if the input is larger, then
       // we need to set the actual result type of the inline asm node to be the
       // same as the input type.
-      if (info.hasMatchingInput())
-        assert(!cir::MissingFeatures::asmInputOperands());
+      if (info.hasMatchingInput()) {
+        unsigned inputNo;
+        for (inputNo = 0; inputNo != s.getNumInputs(); ++inputNo) {
+          TargetInfo::ConstraintInfo &input = inputConstraintInfos[inputNo];
+          if (input.hasTiedOperand() && input.getTiedOperand() == i)
+            break;
+        }
+        assert(inputNo != s.getNumInputs() && "Didn't find matching input!");
 
-      if (mlir::Type adjTy = cgm.getTargetCIRGenInfo().adjustInlineAsmType(
-              *this, outputConstraint, resultRegTypes.back()))
-        resultRegTypes.back() = adjTy;
-      else
-        cgm.getDiags().Report(srcLoc, diag::err_asm_invalid_type_in_input)
-            << outExpr->getType() << outputConstraint;
+        QualType inputTy = s.getInputExpr(inputNo)->getType();
+        QualType outputType = outExpr->getType();
+
+        uint64_t inputSize = getContext().getTypeSize(inputTy);
+        if (getContext().getTypeSize(outputType) < inputSize) {
+          // Form the asm to return the value as a larger integer or fp type.
+          resultRegTypes.back() = convertType(inputTy);
+        }
+
+        if (mlir::Type adjTy = cgm.getTargetCIRGenInfo().adjustInlineAsmType(
+                *this, outputConstraint, resultRegTypes.back()))
+          resultRegTypes.back() = adjTy;
+        else
+          cgm.getDiags().Report(srcLoc, diag::err_asm_invalid_type_in_input)
+              << outExpr->getType() << outputConstraint;
+      }
 
       // Update largest vector width for any vector types.
       assert(!cir::MissingFeatures::asmVectorType());
@@ -343,11 +420,111 @@ mlir::LogicalResult CIRGenFunction::emitAsmStmt(const 
AsmStmt &s) {
       readOnly = readNone = false;
     }
 
-    if (info.isReadWrite())
-      assert(!cir::MissingFeatures::asmInputOperands());
+    if (info.isReadWrite()) {
+      inOutConstraints += ',';
+      const Expr *inputExpr = s.getOutputExpr(i);
+
+      auto [arg, argElementType] =
+          emitAsmInputLValue(info, dest, inputExpr->getType(), 
inOutConstraints,
+                             inputExpr->getExprLoc());
+
+      if (mlir::Type adjTy = getTargetHooks().adjustInlineAsmType(
+              *this, outputConstraint, arg.getType()))
+        arg = builder.createBitcast(arg, adjTy);
+
+      // Update largest vector width for any vector types.
+      assert(!cir::MissingFeatures::asmVectorType());
+
+      // Only tie earlyclobber physregs.
+      if (info.allowsRegister() && (gccReg.empty() || info.earlyClobber()))
+        inOutConstraints += llvm::utostr(i);
+      else
+        inOutConstraints += outputConstraint;
+
+      inOutArgTypes.push_back(arg.getType());
+      inOutArgElemTypes.push_back(argElementType);
+      inOutArgs.push_back(arg);
+    }
 
   } // iterate over output operands
 
+  for (unsigned i = 0, e = s.getNumInputs(); i != e; ++i) {
+    TargetInfo::ConstraintInfo &info = inputConstraintInfos[i];
+    const Expr *inputExpr = s.getInputExpr(i);
+
+    if (info.allowsMemory())
+      readNone = false;
+
+    if (!constraints.empty())
+      constraints += ',';
+
+    std::string inputConstraint(s.getInputConstraint(i));
+    inputConstraint =
+        getTarget().simplifyConstraint(inputConstraint, 
&outputConstraintInfos);
+
+    inputConstraint = s.addVariableConstraints(
+        inputConstraint, *inputExpr->IgnoreParenNoopCasts(getContext()),
+        getTarget(), /*EarlyClobber=*/false,
+        [&](const Stmt *unspStmt, StringRef msg) {
+          cgm.errorUnsupported(unspStmt, msg);
+        });
+
+    std::string replaceConstraint(inputConstraint);
+    auto [arg, argElemType] = emitAsmInput(info, inputExpr, constraints);
+
+    // If this input argument is tied to a larger output result, extend the
+    // input to be the same size as the output.  The LLVM backend wants to see
+    // the input and output of a matching constraint be the same size.  Note
+    // that GCC does not define what the top bits are here.  We use zext 
because
+    // that is usually cheaper, but LLVM IR should really get an anyext 
someday.
+    if (info.hasTiedOperand()) {
+      unsigned output = info.getTiedOperand();
+      QualType outputType = s.getOutputExpr(output)->getType();
+      QualType inputTy = inputExpr->getType();
+
+      if (getContext().getTypeSize(outputType) >
+          getContext().getTypeSize(inputTy)) {
+        // Use ptrtoint as appropriate so that we can do our extension.
+        if (isa<cir::PointerType>(arg.getType()))
+          arg = builder.createPtrToInt(arg, uIntPtrTy);
+        mlir::Type outputTy = convertType(outputType);
+        if (isa<cir::IntType>(outputTy))
+          arg = builder.createIntCast(arg, outputTy);
+        else if (isa<cir::PointerType>(outputTy))
+          arg = builder.createIntCast(arg, uIntPtrTy);
+        else if (isa<cir::FPTypeInterface>(outputTy))
+          arg = builder.createFloatingCast(arg, outputTy);
+      }
+
+      // Deal with the tied operands' constraint code in adjustInlineAsmType.
+      replaceConstraint = outputConstraints[output];
+    }
+
+    if (mlir::Type adjTy = getTargetHooks().adjustInlineAsmType(
+            *this, replaceConstraint, arg.getType()))
+      arg = builder.createBitcast(arg, adjTy);
+    else
+      cgm.getDiags().Report(s.getAsmLoc(), diag::err_asm_invalid_type_in_input)
+          << inputExpr->getType() << inputConstraint;
+
+    // Update largest vector width for any vector types.
+    assert(!cir::MissingFeatures::asmVectorType());
+
+    argTypes.push_back(arg.getType());
+    argElemTypes.push_back(argElemType);
+    inArgs.push_back(arg);
+    args.push_back(arg);
+    constraints += inputConstraint;
+  } // iterate over input operands
+
+  // Append the "input" part of inout constraints.
+  for (unsigned i = 0, e = inOutArgs.size(); i != e; ++i) {
+    args.push_back(inOutArgs[i]);
+    argTypes.push_back(inOutArgTypes[i]);
+    argElemTypes.push_back(inOutArgElemTypes[i]);
+  }
+  constraints += inOutConstraints;
+
   bool hasUnwindClobber = false;
   collectClobbers(*this, s, constraints, hasUnwindClobber, readOnly, readNone);
 
diff --git a/clang/lib/CIR/CodeGen/CIRGenFunction.h 
b/clang/lib/CIR/CodeGen/CIRGenFunction.h
index 98b3f8384fb75..27af920ea7703 100644
--- a/clang/lib/CIR/CodeGen/CIRGenFunction.h
+++ b/clang/lib/CIR/CodeGen/CIRGenFunction.h
@@ -1360,6 +1360,13 @@ class CIRGenFunction : public CIRGenTypeCache {
   Address emitArrayToPointerDecay(const Expr *e,
                                   LValueBaseInfo *baseInfo = nullptr);
 
+  std::pair<mlir::Value, mlir::Type>
+  emitAsmInputLValue(const TargetInfo::ConstraintInfo &info, LValue inputValue,
+                     QualType inputType, std::string &constraintString,
+                     SourceLocation loc);
+  std::pair<mlir::Value, mlir::Type>
+  emitAsmInput(const TargetInfo::ConstraintInfo &info, const Expr *inputExpr,
+               std::string &constraintString);
   mlir::LogicalResult emitAsmStmt(const clang::AsmStmt &s);
 
   RValue emitAtomicExpr(AtomicExpr *e);
diff --git a/clang/test/CIR/CodeGen/inline-asm.c 
b/clang/test/CIR/CodeGen/inline-asm.c
index adcd0c96905a0..55570e1e05550 100644
--- a/clang/test/CIR/CodeGen/inline-asm.c
+++ b/clang/test/CIR/CodeGen/inline-asm.c
@@ -1,9 +1,9 @@
 // RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -fclangir -emit-cir %s -o 
%t.cir
 // RUN: FileCheck --input-file=%t.cir %s -check-prefix=CIR
 // RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -fclangir -emit-llvm %s -o 
%t-cir.ll
-// RUN: FileCheck --input-file=%t-cir.ll %s --check-prefixes=LLVM
+// RUN: FileCheck --input-file=%t-cir.ll %s --check-prefixes=LLVM,CIRLLVMONLY
 // RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -emit-llvm %s -o %t-cir.ll
-// RUN: FileCheck --input-file=%t-cir.ll %s --check-prefixes=LLVM
+// RUN: FileCheck --input-file=%t-cir.ll %s --check-prefixes=LLVM,LLVMONLY
 
 __asm__ ("foo1");
 __asm__ ("foo2");
@@ -37,6 +37,34 @@ void empty2() {
   __asm__ volatile("nop" : : : );
 }
 
+//      CIR: cir.func{{.*}}@empty3
+//      CIR: %[[X:.*]] = cir.alloca !s32i, !cir.ptr<!s32i>, ["x", init]
+//      CIR: cir.asm(x86_att, 
+// CIR-NEXT:   out = [%[[X]] : !cir.ptr<!s32i> (maybe_memory)],
+// CIR-NEXT:   in = [],
+// CIR-NEXT:   in_out = [%[[X]] : !cir.ptr<!s32i> (maybe_memory)],
+// CIR-NEXT:   {"" "=*m,*m,~{dirflag},~{fpsr},~{flags}"}) side_effects
+// LLVM: define{{.*}}@empty3
+// LLVM: %[[X:.*]] = alloca i32
+// LLVM:   call void asm sideeffect "", 
"=*m,*m,~{dirflag},~{fpsr},~{flags}"(ptr elementtype(i32) %[[X]], ptr 
elementtype(i32) %[[X]])
+void empty3(int x) {
+  __asm__ volatile("" : "+m"(x));
+}
+
+//      CIR: cir.func{{.*}}@empty4
+//      CIR: %[[X:.*]] = cir.alloca !s32i, !cir.ptr<!s32i>, ["x", init]
+//      CIR: cir.asm(x86_att, 
+// CIR-NEXT:   out = [],
+// CIR-NEXT:   in = [%[[X]] : !cir.ptr<!s32i> (maybe_memory)],
+// CIR-NEXT:   in_out = [],
+// CIR-NEXT:   {"" "*m,~{dirflag},~{fpsr},~{flags}"}) side_effects
+// LLVM: define{{.*}}@empty4
+// LLVM: %[[X:.*]] = alloca i32
+// LLVM: call void asm sideeffect "", "*m,~{dirflag},~{fpsr},~{flags}"(ptr 
elementtype(i32) %[[X]])
+void empty4(int x) {
+  __asm__ volatile("" : : "m"(x));
+}
+
 //      CIR: cir.func{{.*}}@empty5
 //      CIR: %[[X:.*]] = cir.alloca !s32i, !cir.ptr<!s32i>, ["x", init]
 //      CIR: cir.asm(x86_att, 
@@ -51,6 +79,95 @@ void empty5(int x) {
   __asm__ volatile("" : "=m"(x));
 }
 
+//      CIR: cir.func{{.*}}@empty6
+//      CIR: %[[X:.*]] = cir.alloca !s32i, !cir.ptr<!s32i>, ["x", init]
+//      CIR: %[[X_LOAD:.*]] = cir.load align(4) %[[X]] : !cir.ptr<!s32i>, !s32i
+//      CIR: cir.asm(x86_att, 
+// CIR-NEXT:   out = [],
+// CIR-NEXT:   in = [],
+// CIR-NEXT:   in_out = [%[[X_LOAD]] : !s32i],
+// CIR-NEXT:   {"" "=&r,=&r,1,~{dirflag},~{fpsr},~{flags}"}) side_effects -> 
!rec_anon_struct
+// LLVM: define{{.*}}@empty6
+// LLVM: %[[X:.*]] = alloca i32
+// LLVM: %[[X_LOAD:.*]] = load i32, ptr %[[X]]
+// LLVM: call { i32, i32 } asm sideeffect "", 
"=&r,=&r,1,~{dirflag},~{fpsr},~{flags}"(i32 %[[X_LOAD]])
+//
+void empty6(int x) {
+  __asm__ volatile("" : "=&r"(x), "+&r"(x));
+}
+
+//      CIR: cir.func{{.*}}@add1
+//      CIR: %[[X:.*]] = cir.alloca !u32i, !cir.ptr<!u32i>, ["x", init]
+//      CIR: %[[A:.*]] = cir.alloca !s32i, !cir.ptr<!s32i>, ["a"]
+//      CIR: %[[X_LOAD:.*]] = cir.load align(4) %[[X]] : !cir.ptr<!u32i>, !u32i
+//      CIR: %[[ASM_RES:.*]] = cir.asm(x86_att, 
+// CIR-NEXT:                     out = [],
+// CIR-NEXT:                     in = [%[[X_LOAD]] : !u32i],
+// CIR-NEXT:                     in_out = [],
+// CIR-NEXT:                     {"addl $$42, $1" 
"=r,r,~{dirflag},~{fpsr},~{flags}"}) -> !s32i
+// CIR-NEXT: cir.store{{.*}} %[[ASM_RES]], %[[A]] : !s32i, !cir.ptr<!s32i>
+// LLVM: define{{.*}}@add1
+// LLVM: %[[X:.*]] = alloca i32
+// Classic codegen doesn't alloca for a return slot.
+// CIRLLVMONLY: alloca i32
+// LLVM: %[[A:.*]] = alloca i32
+// LLVM: %[[X_LOAD:.*]] = load i32, ptr %[[X]]
+// LLVM: %[[ASM_RES:.*]] = call i32 asm "addl $$42, $1", 
"=r,r,~{dirflag},~{fpsr},~{flags}"(i32 %[[X_LOAD]])
+// LLVM: store i32 %[[ASM_RES]], ptr %[[A]]
+unsigned add1(unsigned int x) {
+  int a;
+  __asm__("addl $42, %[val]"
+      : "=r" (a)
+      : [val] "r" (x)
+      );
+
+  return a;
+}
+
+//      CIR: cir.func{{.*}}@add2
+//      CIR: %[[X:.*]] = cir.alloca !u32i, !cir.ptr<!u32i>, ["x", init]
+//      CIR: %[[X_LOAD:.*]] = cir.load align(4) %[[X]] : !cir.ptr<!u32i>, !u32i
+//      CIR: %[[ASM_RES:.*]] = cir.asm(x86_att, 
+// CIR-NEXT:                     out = [],
+// CIR-NEXT:                     in = [],
+// CIR-NEXT:                     in_out = [%[[X_LOAD]] : !u32i],
+// CIR-NEXT:                     {"addl $$42, $0" 
"=r,0,~{dirflag},~{fpsr},~{flags}"}) -> !u32i
+// CIR-NEXT: cir.store{{.*}} %[[ASM_RES]], %[[X]] : !u32i, !cir.ptr<!u32i>
+// LLVM: define{{.*}}@add2
+// LLVM: %[[X:.*]] = alloca i32
+// LLVM: %[[X_LOAD:.*]] = load i32, ptr %[[X]]
+// LLVM: %[[ASM_RES:.*]] = call i32 asm "addl $$42, $0", 
"=r,0,~{dirflag},~{fpsr},~{flags}"(i32 %[[X_LOAD]])
+// LLVM: store i32 %[[ASM_RES]], ptr %[[X]]
+unsigned add2(unsigned int x) {
+  __asm__("addl $42, %[val]"
+      : [val] "+r" (x)
+      );
+  return x;
+}
+
+//      CIR: cir.func{{.*}}@add3
+//      CIR: %[[X:.*]] = cir.alloca !u32i, !cir.ptr<!u32i>, ["x", init]
+//      CIR: %[[X_LOAD:.*]] = cir.load{{.*}} %[[X]] : !cir.ptr<!u32i>, !u32i
+//      CIR: %[[ASM_RES:.*]] = cir.asm(x86_att, 
+// CIR-NEXT:                     out = [],
+// CIR-NEXT:                     in = [],
+// CIR-NEXT:                     in_out = [%[[X_LOAD]] : !u32i],
+// CIR-NEXT:                     {"addl $$42, $0  \0A\09          subl $$1, $0 
   \0A\09          imul $$2, $0" "=r,0,~{dirflag},~{fpsr},~{flags}"}) -> !u32i
+// CIR-NEXT: cir.store{{.*}} %[[ASM_RES]], %[[X]]  : !u32i, !cir.ptr<!u32i>
+// LLVM: define{{.*}}@add3
+// LLVM: %[[X:.*]] = alloca i32
+// LLVM: %[[X_LOAD:.*]] = load i32, ptr %[[X]]
+// LLVM: %[[ASM_RES:.*]] = call i32 asm "addl $$42, $0  \0A\09          subl 
$$1, $0    \0A\09          imul $$2, $0", 
"=r,0,~{dirflag},~{fpsr},~{flags}"(i32 %[[X_LOAD]])
+// LLVM: store i32 %[[ASM_RES]], ptr %[[X]]
+unsigned add3(unsigned int x) { // ((42 + x) - 1) * 2
+  __asm__("addl $42, %[val]  \n\t\
+          subl $1, %[val]    \n\t\
+          imul $2, %[val]"
+      : [val] "+r" (x)
+      );
+  return x;
+}
+
 //      CIR: cir.func{{.*}}@add4
 //      CIR: %[[X:.*]] = cir.alloca !cir.ptr<!s32i>, 
!cir.ptr<!cir.ptr<!s32i>>, ["x", init] 
 //      CIR: %[[X_LOAD:.*]] = cir.load {{.*}}  %[[X]] : 
!cir.ptr<!cir.ptr<!s32i>>, !cir.ptr<!s32i>
@@ -68,6 +185,36 @@ void add4(int *x) {
   __asm__("addl $42, %[addr]" : [addr] "=m" (*x));
 }
 
+//      CIR: cir.func{{.*}}@add5
+//      CIR: %[[X:.*]] = cir.alloca !cir.float, !cir.ptr<!cir.float>, ["x", 
init]
+//      CIR: %[[Y:.*]] = cir.alloca !cir.float, !cir.ptr<!cir.float>, ["y", 
init]
+//      CIR: %[[R:.*]] = cir.alloca !cir.float, !cir.ptr<!cir.float>, ["r"]
+//      CIR: %[[X_LOAD:.*]] = cir.load{{.*}} %[[X]] : !cir.ptr<!cir.float>, 
!cir.float
+//      CIR: %[[Y_LOAD:.*]] = cir.load{{.*}} %[[Y]] : !cir.ptr<!cir.float>, 
!cir.float
+//      CIR: %[[ASM_RES:.*]] = cir.asm(x86_att, 
+// CIR-NEXT:                      out = [],
+// CIR-NEXT:                      in = [%[[X_LOAD]] : !cir.float, %[[Y_LOAD]] 
: !cir.float],
+// CIR-NEXT:                      in_out = [],
+// CIR-NEXT:                      {"flds $1; flds $2; faddp" 
"=&{st},imr,imr,~{dirflag},~{fpsr},~{flags}"}) -> !cir.float
+// CIR-NEXT: cir.store{{.*}} %[[ASM_RES]], %[[R]] : !cir.float, 
!cir.ptr<!cir.float>
+// LLVM: define{{.*}}@add5
+// LLVM: %[[X:.*]] = alloca float
+// LLVM: %[[Y:.*]] = alloca float
+// Classic codegen doesn't alloca for a return slot.
+// CIRLLVMONLY: alloca float
+// LLVM: %[[R:.*]] = alloca float
+// LLVM: %[[X_LOAD:.*]] = load float, ptr %[[X]]
+// LLVM: %[[Y_LOAD:.*]] = load float, ptr %[[Y]]
+// LLVM: %[[ASM_RES:.*]] = call float asm "flds $1; flds $2; faddp", 
"=&{st},imr,imr,~{dirflag},~{fpsr},~{flags}"(float %[[X_LOAD]], float 
%[[Y_LOAD]])
+// LLVM: store float %[[ASM_RES]], ptr %[[R]]
+float add5(float x, float y) {
+   float r;
+  __asm__("flds %[x]; flds %[y]; faddp"
+          : "=&t" (r)
+          : [x] "g" (x), [y] "g" (y));
+  return r;
+}
+
 //      CIR: cir.func{{.*}}@mov
 //      CIR: %[[A:.*]] = cir.alloca !s32i, !cir.ptr<!s32i>, ["a"] {alignment = 
4 : i64}
 //      CIR: %[[RES:.*]] = cir.asm(x...
[truncated]

``````````

</details>


https://github.com/llvm/llvm-project/pull/176239
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to