[X86] Don't call LowerUINT_TO_FP_i32 for i32->f80 on 32-bit targets with sse2.

We were performing an emulated i32->f64 in the SSE registers, then
storing that value to memory and doing a extload into the X87
domain.

After this patch we'll now just store the i32 to memory along
with an i32 0. Then do a 64-bit FILD to f80 completely in the X87
unit. This matches what we do without SSE.

2 files changed, 14 insertions, 46 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index d7593f52573..0f152968ddf 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -19331,7 +19331,7 @@ SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
 
   if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
     return LowerUINT_TO_FP_i64(Op, DAG, Subtarget);
-  if (SrcVT == MVT::i32 && X86ScalarSSEf64)
+  if (SrcVT == MVT::i32 && X86ScalarSSEf64 && DstVT != MVT::f80)
     return LowerUINT_TO_FP_i32(Op, DAG, Subtarget);
   if (Subtarget.is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
     return SDValue();
diff --git a/llvm/test/CodeGen/X86/scalar-int-to-fp.ll b/llvm/test/CodeGen/X86/scalar-int-to-fp.ll
index 61771044897..29e8df668b5 100644
--- a/llvm/test/CodeGen/X86/scalar-int-to-fp.ll
+++ b/llvm/test/CodeGen/X86/scalar-int-to-fp.ll
@@ -214,21 +214,19 @@ define double @s32_to_d(i32 %a) nounwind {
 }
 
 define x86_fp80 @u32_to_x(i32 %a) nounwind {
-; AVX512_32-LABEL: u32_to_x:
-; AVX512_32:       # %bb.0:
-; AVX512_32-NEXT:    pushl %ebp
-; AVX512_32-NEXT:    movl %esp, %ebp
-; AVX512_32-NEXT:    andl $-8, %esp
-; AVX512_32-NEXT:    subl $8, %esp
-; AVX512_32-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
-; AVX512_32-NEXT:    vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero
-; AVX512_32-NEXT:    vorpd %xmm0, %xmm1, %xmm1
-; AVX512_32-NEXT:    vsubsd %xmm0, %xmm1, %xmm0
-; AVX512_32-NEXT:    vmovsd %xmm0, (%esp)
-; AVX512_32-NEXT:    fldl (%esp)
-; AVX512_32-NEXT:    movl %ebp, %esp
-; AVX512_32-NEXT:    popl %ebp
-; AVX512_32-NEXT:    retl
+; CHECK32-LABEL: u32_to_x:
+; CHECK32:       # %bb.0:
+; CHECK32-NEXT:    pushl %ebp
+; CHECK32-NEXT:    movl %esp, %ebp
+; CHECK32-NEXT:    andl $-8, %esp
+; CHECK32-NEXT:    subl $8, %esp
+; CHECK32-NEXT:    movl 8(%ebp), %eax
+; CHECK32-NEXT:    movl %eax, (%esp)
+; CHECK32-NEXT:    movl $0, {{[0-9]+}}(%esp)
+; CHECK32-NEXT:    fildll (%esp)
+; CHECK32-NEXT:    movl %ebp, %esp
+; CHECK32-NEXT:    popl %ebp
+; CHECK32-NEXT:    retl
 ;
 ; CHECK64-LABEL: u32_to_x:
 ; CHECK64:       # %bb.0:
@@ -236,36 +234,6 @@ define x86_fp80 @u32_to_x(i32 %a) nounwind {
 ; CHECK64-NEXT:    movq %rax, -{{[0-9]+}}(%rsp)
 ; CHECK64-NEXT:    fildll -{{[0-9]+}}(%rsp)
 ; CHECK64-NEXT:    retq
-;
-; SSE2_32-LABEL: u32_to_x:
-; SSE2_32:       # %bb.0:
-; SSE2_32-NEXT:    pushl %ebp
-; SSE2_32-NEXT:    movl %esp, %ebp
-; SSE2_32-NEXT:    andl $-8, %esp
-; SSE2_32-NEXT:    subl $8, %esp
-; SSE2_32-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
-; SSE2_32-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
-; SSE2_32-NEXT:    orpd %xmm0, %xmm1
-; SSE2_32-NEXT:    subsd %xmm0, %xmm1
-; SSE2_32-NEXT:    movsd %xmm1, (%esp)
-; SSE2_32-NEXT:    fldl (%esp)
-; SSE2_32-NEXT:    movl %ebp, %esp
-; SSE2_32-NEXT:    popl %ebp
-; SSE2_32-NEXT:    retl
-;
-; X87-LABEL: u32_to_x:
-; X87:       # %bb.0:
-; X87-NEXT:    pushl %ebp
-; X87-NEXT:    movl %esp, %ebp
-; X87-NEXT:    andl $-8, %esp
-; X87-NEXT:    subl $8, %esp
-; X87-NEXT:    movl 8(%ebp), %eax
-; X87-NEXT:    movl %eax, (%esp)
-; X87-NEXT:    movl $0, {{[0-9]+}}(%esp)
-; X87-NEXT:    fildll (%esp)
-; X87-NEXT:    movl %ebp, %esp
-; X87-NEXT:    popl %ebp
-; X87-NEXT:    retl
   %r = uitofp i32 %a to x86_fp80
   ret x86_fp80 %r
 }