[PowerPC] Handle FP physical register in inline asm constraint.

[sfertile]

Summary

Do not defer to the base class when the register constraint is a physical fpr. The base class will select SPILLTOVSX as the register class and register allocation will fail on subtargets without VSX registers.

[sfertile]

ping.

[ZarkoCA]

LGTM, only some question about a comment.

llvm/lib/Target/PowerPC/PPCISelLowering.cpp

15111

I think this should be SPILLTOVSRRC instead of SPILLTOVSX

llvm/lib/Target/PowerPC/PPCISelLowering.cpp

if (Constraint.size() == 1) {
     return std::make_pair(0U, &PPC::VSRCRegClass);
   } else if ((Constraint == "ws" || Constraint == "ww") && Subtarget.hasVSX()) {
     if (VT == MVT::f32 && Subtarget.hasP8Vector())
       return std::make_pair(0U, &PPC::VSSRCRegClass);
     else
       return std::make_pair(0U, &PPC::VSFRCRegClass);
   }
 
+  // Handle special cases of physical registers that are not properly handled
+  // by the base class.
+  if (Constraint[0] == '{' && Constraint[Constraint.size() - 1] == '}') {
     // If we name a VSX register, we can't defer to the base class because it
     // will not recognize the correct register (their names will be VSL{0-31}
     // and V{0-31} so they won't match). So we match them here.
     if (Constraint.size() > 3 && Constraint[1] == 'v' && Constraint[2] == 's') {
       int VSNum = atoi(Constraint.data() + 3);
       assert(VSNum >= 0 && VSNum <= 63 &&
              "Attempted to access a vsr out of range");
       if (VSNum < 32)
         return std::make_pair(PPC::VSL0 + VSNum, &PPC::VSRCRegClass);
       return std::make_pair(PPC::V0 + VSNum - 32, &PPC::VSRCRegClass);
     }
+
+    // For float registers, we can't defer to the base class as it will match
+    // the SPILLTOVSRRC class.
+    if (Constraint.size() > 3 && Constraint[1] == 'f') {
+      int RegNum = atoi(Constraint.data() + 2);
+      if (RegNum > 31 || RegNum < 0)
+        report_fatal_error("Invalid floating point register number");
+      if (VT == MVT::f32 || VT == MVT::i32)
+        return Subtarget.hasSPE()
+                   ? std::make_pair(PPC::R0 + RegNum, &PPC::GPRCRegClass)
+                   : std::make_pair(PPC::F0 + RegNum, &PPC::F4RCRegClass);
+      if (VT == MVT::f64 || VT == MVT::i64)
+        return Subtarget.hasSPE()
+                   ? std::make_pair(PPC::S0 + RegNum, &PPC::SPERCRegClass)
+                   : std::make_pair(PPC::F0 + RegNum, &PPC::F8RCRegClass);
+    }
+  }
+
   std::pair<unsigned, const TargetRegisterClass *> R =
       TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
 
   // r[0-9]+ are used, on PPC64, to refer to the corresponding 64-bit registers
   // (which we call X[0-9]+). If a 64-bit value has been requested, and a
   // 32-bit GPR has been selected, then 'upgrade' it to the 64-bit parent
   // register.
   // FIXME: If TargetLowering::getRegForInlineAsmConstraint could somehow use

llvm/test/CodeGen/PowerPC/inline-asm-physical-fpr-spe.ll

+; RUN: llc -verify-machineinstrs < %s -mtriple=powerpc-unknown-linux-gnu \
+; RUN:          -mattr=+spe |  FileCheck %s
+
+define i32 @test_f32(float %x) {
+; CHECK-LABEL: test_f32:
+; CHECK:         #APP
+; CHECK-NEXT:    efsctsi 31, 3
+; CHECK-NEXT:    #NO_APP
+entry:
+  %0 = call i32 asm sideeffect "efsctsi $0, $1", "={f31},f"(float %x)
+  ret i32 %0
+}
+
+define i32 @test_f64(double %x) {
+; CHECK-LABEL: test_f64:
+; CHECK:         #APP
+; CHECK-NEXT:    efdctsi 0, 3
+; CHECK-NEXT:    #NO_APP
+entry:
+  %0 = call i32 asm sideeffect "efdctsi $0, $1", "={f0},d"(double %x)
+  ret i32 %0
+}
+

llvm/test/CodeGen/PowerPC/inline-asm-physical-fpr.ll

+; RUN: llc -mcpu=pwr7 -mattr=-altivec -verify-machineinstrs \
+; RUN:     -mtriple=powerpc-unknown-aix < %s  | FileCheck %s
+
+; RUN: llc -mcpu=pwr7 -mattr=-altivec -verify-machineinstrs \
+; RUN:     -mtriple=powerpc64-unknown-aix < %s | FileCheck %s
+
+
+define dso_local double @test_double(double %a, double %b) {
+entry:
+  %0 = tail call double asm "fadd. $0,$1,$2\0A", "={f31},d,d,0"(double %a, double %b, double 0.000000e+00)
+  ret double %0
+}
+
+; CHECK-LABEL: test_double
+; CHECK:         #APP
+; CHECK-NEXT:    fadd. 31,1,2
+
+define dso_local signext i32 @test_int(double %a, double %b) {
+entry:
+  %0 = tail call i32 asm "fadd. $0,$1,$2\0A", "={f0},d,d,0"(double %a, double %b, i32 0)
+  ret i32 %0
+}
+
+; CHECK-LABEL: test_int
+; CHECK:         #APP
+; CHECK-NEXT:    fadd. 0,1,2