[X86] Fix a bug in WIN_FTOL_32/64 handling.

Duplicating an FP register "as itself" is a bad idea, since it violates the
invariant that every FP register is mapped to at most one FPU stack slot.
Use the scratch FP register instead.

This fixes PR23957.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@241069 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86FloatingPoint.cpp b/lib/Target/X86/X86FloatingPoint.cpp
index 3b0bd03095a9f467995440b0c48268ba8207f379..40b9c8a863a347720c605adf319b773d60727671 100644 (file)
--- a/lib/Target/X86/X86FloatingPoint.cpp
+++ b/lib/Target/X86/X86FloatingPoint.cpp
@@ -1530,7 +1530,7 @@ void FPS::handleSpecialFP(MachineBasicBlock::iterator &Inst) {
     if (Op.isKill())
       moveToTop(FPReg, Inst);
     else
     if (Op.isKill())
       moveToTop(FPReg, Inst);
     else

-      duplicateToTop(FPReg, FPReg, Inst);
+      duplicateToTop(FPReg, ScratchFPReg, Inst);

 
     // Emit the call. This will pop the operand.
     BuildMI(*MBB, Inst, MI->getDebugLoc(), TII->get(X86::CALLpcrel32))
 
     // Emit the call. This will pop the operand.
     BuildMI(*MBB, Inst, MI->getDebugLoc(), TII->get(X86::CALLpcrel32))

diff --git a/test/CodeGen/X86/win_ftol2.ll b/test/CodeGen/X86/win_ftol2.ll
index 14591248f354ee5c37139451dbe518c0c0597602..dfa6e3aa76bdd416ee6fc7928d132637d73f9281 100644 (file)
--- a/test/CodeGen/X86/win_ftol2.ll
+++ b/test/CodeGen/X86/win_ftol2.ll
@@ -142,3 +142,25 @@ define i64 @double_ui64_5(double %X) {
   %tmp.1 = fptoui double %X to i64
   ret i64 %tmp.1
 }
   %tmp.1 = fptoui double %X to i64
   ret i64 %tmp.1
 }

+
+define double @pr23957_32(double %A) {
+; FTOL-LABEL: @pr23957_32
+; FTOL: fldl
+; FTOL-NEXT: fld %st(0)
+; FTOL-NEXT: calll __ftol2
+  %B = fptoui double %A to i32
+  %C = uitofp i32 %B to double
+  %D = fsub double %C, %A
+  ret double %D
+}
+
+define double @pr23957_64(double %A) {
+; FTOL-LABEL: @pr23957_64
+; FTOL: fldl
+; FTOL-NEXT: fld %st(0)
+; FTOL-NEXT: calll __ftol2
+  %B = fptoui double %A to i64
+  %C = uitofp i64 %B to double
+  %D = fsub double %C, %A
+  ret double %D
+}