Don't clobber pending ST regs when FP regs are killed.

X86FloatingPoint keeps track of pending ST registers for an upcoming
inline asm instruction with fixed stack register constraints.  It does
this by remembering which FP register holds the value that should appear
at a fixed stack position for the inline asm.

When that FP register is killed before the inline asm, make sure to
duplicate it to a scratch register, so the ST register still has a live
FP reference.

This could happen when the same FP register was copied to two ST
registers, or when a spill instruction is inserted between the ST copy
and the inline asm.

This fixes PR10602.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@137050 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86FloatingPoint.cpp b/lib/Target/X86/X86FloatingPoint.cpp
index 6f619a9e567cc68a4e627b0e8fbec70da90d3261..a9b6fdd140855a25801ac8e8ac18f6b82761c224 100644 (file)
--- a/lib/Target/X86/X86FloatingPoint.cpp
+++ b/lib/Target/X86/X86FloatingPoint.cpp
@@ -260,6 +260,21 @@ namespace {
       BuildMI(*MBB, I, dl, TII->get(X86::LD_Frr)).addReg(STReg);
     }
 
+    /// duplicatePendingSTBeforeKill - The instruction at I is about to kill
+    /// RegNo. If any PendingST registers still need the RegNo value, duplicate
+    /// them to new scratch registers.
+    void duplicatePendingSTBeforeKill(unsigned RegNo, MachineInstr *I) {
+      for (unsigned i = 0; i != NumPendingSTs; ++i) {
+        if (PendingST[i] != RegNo)
+          continue;
+        unsigned SR = getScratchReg();
+        DEBUG(dbgs() << "Duplicating pending ST" << i
+                     << " in FP" << RegNo << " to FP" << SR << '\n');
+        duplicateToTop(RegNo, SR, I);
+        PendingST[i] = SR;
+      }
+    }
+
     /// popStackAfter - Pop the current value off of the top of the FP stack
     /// after the specified instruction.
     void popStackAfter(MachineBasicBlock::iterator &I);
@@ -973,6 +988,9 @@ void FPS::handleOneArgFP(MachineBasicBlock::iterator &I) {
   unsigned Reg = getFPReg(MI->getOperand(NumOps-1));
   bool KillsSrc = MI->killsRegister(X86::FP0+Reg);
 
+  if (KillsSrc)
+    duplicatePendingSTBeforeKill(Reg, I);
+
   // FISTP64m is strange because there isn't a non-popping versions.
   // If we have one _and_ we don't want to pop the operand, duplicate the value
   // on the stack instead of moving it.  This ensure that popping the value is
@@ -1036,6 +1054,7 @@ void FPS::handleOneArgFPRW(MachineBasicBlock::iterator &I) {
   bool KillsSrc = MI->killsRegister(X86::FP0+Reg);
 
   if (KillsSrc) {
+    duplicatePendingSTBeforeKill(Reg, I);
     // If this is the last use of the source register, just make sure it's on
     // the top of the stack.
     moveToTop(Reg, I);
@@ -1322,6 +1341,7 @@ void FPS::handleSpecialFP(MachineBasicBlock::iterator &I) {
 
       // When the source is killed, allocate a scratch FP register.
       if (KillsSrc) {
+        duplicatePendingSTBeforeKill(SrcFP, I);
         unsigned Slot = getSlot(SrcFP);
         unsigned SR = getScratchReg();
         PendingST[DstST] = SR;

diff --git a/test/CodeGen/X86/inline-asm-fpstack.ll b/test/CodeGen/X86/inline-asm-fpstack.ll
index 8e48bbec8e3a556771b6a5669546d1ed5675100d..c9a1c1c38c5b7ded83d2691869c7238a7bc986b9 100644 (file)
--- a/test/CodeGen/X86/inline-asm-fpstack.ll
+++ b/test/CodeGen/X86/inline-asm-fpstack.ll
@@ -329,3 +329,14 @@ entry:
   %asmresult = extractvalue %complex %0, 0
   ret float %asmresult
 }
+
+; Pass the same value in two fixed stack slots.
+; CHECK: PR10602
+; CHECK: flds LCPI
+; CHECK: fld %st(0)
+; CHECK: fcomi %st(1), %st(0)
+define i32 @PR10602() nounwind ssp {
+entry:
+  %0 = tail call i32 asm "fcomi $2, $1; pushf; pop $0", "=r,{st},{st(1)},~{dirflag},~{fpsr},~{flags}"(double 2.000000e+00, double 2.000000e+00) nounwind
+  ret i32 %0
+}