Check and allow floating point registers to select the size of the

register for inline asm. This conforms to how gcc allows for effective
casting of inputs into gprs (fprs is already handled).

llvm-svn: 174008

2 files changed, 44 insertions, 2 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 36d1ad42677..c34010c9787 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -18170,7 +18170,7 @@ X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
         Res.first = DestReg;
         Res.second = &X86::GR8RegClass;
       }
-    } else if (VT == MVT::i32) {
+    } else if (VT == MVT::i32 || VT == MVT::f32) {
       unsigned DestReg = 0;
       switch (Res.first) {
       default: break;
@@ -18187,7 +18187,7 @@ X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
         Res.first = DestReg;
         Res.second = &X86::GR32RegClass;
       }
-    } else if (VT == MVT::i64) {
+    } else if (VT == MVT::i64 || VT == MVT::f64) {
       unsigned DestReg = 0;
       switch (Res.first) {
       default: break;
diff --git a/llvm/test/CodeGen/X86/cas.ll b/llvm/test/CodeGen/X86/cas.ll
new file mode 100644
index 00000000000..12e49f94704
--- /dev/null
+++ b/llvm/test/CodeGen/X86/cas.ll
@@ -0,0 +1,42 @@
+; RUN: llc -mtriple=x86_64-pc-linux-gnu %s -o - | FileCheck %s
+
+; C code this came from
+;bool cas(float volatile *p, float *expected, float desired) {
+;  bool success;
+;  __asm__ __volatile__("lock; cmpxchg %[desired], %[mem]; "
+;                       "mov %[expected], %[expected_out]; "
+;                       "sete %[success]"
+;                       : [success] "=a" (success),
+;                         [expected_out] "=rm" (*expected)
+;                       : [expected] "a" (*expected),
+;                         [desired] "q" (desired),
+;                         [mem] "m" (*p)
+;                       : "memory", "cc");
+;  return success;
+;}
+
+define zeroext i1 @cas(float* %p, float* %expected, float %desired) nounwind {
+entry:
+  %p.addr = alloca float*, align 8
+  %expected.addr = alloca float*, align 8
+  %desired.addr = alloca float, align 4
+  %success = alloca i8, align 1
+  store float* %p, float** %p.addr, align 8
+  store float* %expected, float** %expected.addr, align 8
+  store float %desired, float* %desired.addr, align 4
+  %0 = load float** %expected.addr, align 8
+  %1 = load float** %expected.addr, align 8
+  %2 = load float* %1, align 4
+  %3 = load float* %desired.addr, align 4
+  %4 = load float** %p.addr, align 8
+  %5 = call i8 asm sideeffect "lock; cmpxchg $3, $4; mov $2, $1; sete $0", "={ax},=*rm,{ax},q,*m,~{memory},~{cc},~{dirflag},~{fpsr},~{flags}"(float* %0, float %2, float %3, float* %4) nounwind
+  store i8 %5, i8* %success, align 1
+  %6 = load i8* %success, align 1
+  %tobool = trunc i8 %6 to i1
+  ret i1 %tobool
+}
+
+; Make sure we're emitting a move from 
+; CHECK: #APP
+; CHECK-NEXT: lock;{{.*}}mov %eax,{{.*}}
+; CHECK-NEXT: #NO_APP