[CodeGen] Round [SU]INT_TO_FP result when promoting from f16.

If we don't, values that aren't precisely representable in f16 could
be used as-is in a promoted f32 operation, which would produce
incorrect results.

AArch64 had the correct behavior; add a focused test.

Fixes http://llvm.org/PR26871

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@268700 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp b/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp
index 6c0193a76732a2635077e4d2eeba8e50e734483e..38639a2a3a29745929d5f4e5d71cedd9aad52661 100644 (file)
--- a/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp
+++ b/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp
@@ -2102,9 +2102,14 @@ SDValue DAGTypeLegalizer::PromoteFloatRes_SELECT_CC(SDNode *N) {
 // Construct a SDNode that transforms the SINT or UINT operand to the promoted
 // float type.
 SDValue DAGTypeLegalizer::PromoteFloatRes_XINT_TO_FP(SDNode *N) {
+  SDLoc DL(N);
   EVT VT = N->getValueType(0);
   EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT);
-  return DAG.getNode(N->getOpcode(), SDLoc(N), NVT, N->getOperand(0));
+  SDValue NV = DAG.getNode(N->getOpcode(), DL, NVT, N->getOperand(0));
+  // Round the value to the desired precision (that of the source type).
+  return DAG.getNode(
+      ISD::FP_EXTEND, DL, NVT,
+      DAG.getNode(ISD::FP_ROUND, DL, VT, NV, DAG.getIntPtrConstant(0, DL)));
 }
 
 SDValue DAGTypeLegalizer::PromoteFloatRes_UNDEF(SDNode *N) {

diff --git a/test/CodeGen/AArch64/f16-instructions.ll b/test/CodeGen/AArch64/f16-instructions.ll
index 8ba4e363741ebf0a92a434de33f8f0c71d6f1acb..df26cefa3284fe2754fdc04c5c3e81713aaed51a 100644 (file)
--- a/test/CodeGen/AArch64/f16-instructions.ll
+++ b/test/CodeGen/AArch64/f16-instructions.ll
@@ -446,6 +446,34 @@ define half @test_sitofp_i64(i64 %a) #0 {
   ret half %r
 }
 
+; CHECK-LABEL: test_uitofp_i32_fadd:
+; CHECK-NEXT: ucvtf s1, w0
+; CHECK-NEXT: fcvt h1, s1
+; CHECK-NEXT: fcvt s0, h0
+; CHECK-NEXT: fcvt s1, h1
+; CHECK-NEXT: fadd s0, s0, s1
+; CHECK-NEXT: fcvt h0, s0
+; CHECK-NEXT: ret
+define half @test_uitofp_i32_fadd(i32 %a, half %b) #0 {
+  %c = uitofp i32 %a to half
+  %r = fadd half %b, %c
+  ret half %r
+}
+
+; CHECK-LABEL: test_sitofp_i32_fadd:
+; CHECK-NEXT: scvtf s1, w0
+; CHECK-NEXT: fcvt h1, s1
+; CHECK-NEXT: fcvt s0, h0
+; CHECK-NEXT: fcvt s1, h1
+; CHECK-NEXT: fadd s0, s0, s1
+; CHECK-NEXT: fcvt h0, s0
+; CHECK-NEXT: ret
+define half @test_sitofp_i32_fadd(i32 %a, half %b) #0 {
+  %c = sitofp i32 %a to half
+  %r = fadd half %b, %c
+  ret half %r
+}
+
 ; CHECK-LABEL: test_fptrunc_float:
 ; CHECK-NEXT: fcvt h0, s0
 ; CHECK-NEXT: ret

diff --git a/test/CodeGen/ARM/fp16-promote.ll b/test/CodeGen/ARM/fp16-promote.ll
index 2a2eb8d2b6ba71767372183fe2a7452687c81512..4da56b7ef357ce20da985730178559c35c9c31f2 100644 (file)
--- a/test/CodeGen/ARM/fp16-promote.ll
+++ b/test/CodeGen/ARM/fp16-promote.ll
@@ -889,4 +889,44 @@ define half @test_struct_arg(%struct.dummy %p) {
   ret half %a
 }
 
+; CHECK-LABEL: test_uitofp_i32_fadd:
+; CHECK-VFP-DAG: vcvt.f32.u32
+; CHECK-NOVFP-DAG: bl __aeabi_ui2f
+
+; CHECK-FP16-DAG: vcvtb.f16.f32
+; CHECK-FP16-DAG: vcvtb.f32.f16
+; CHECK-LIBCALL-DAG: bl __aeabi_h2f
+; CHECK-LIBCALL-DAG: bl __aeabi_h2f
+
+; CHECK-VFP-DAG: vadd.f32
+; CHECK-NOVFP-DAG: bl __aeabi_fadd
+
+; CHECK-FP16-DAG: vcvtb.f16.f32
+; CHECK-LIBCALL-DAG: bl __aeabi_f2h
+define half @test_uitofp_i32_fadd(i32 %a, half %b) #0 {
+  %c = uitofp i32 %a to half
+  %r = fadd half %b, %c
+  ret half %r
+}
+
+; CHECK-LABEL: test_sitofp_i32_fadd:
+; CHECK-VFP-DAG: vcvt.f32.s32
+; CHECK-NOVFP-DAG: bl __aeabi_i2f
+
+; CHECK-FP16-DAG: vcvtb.f16.f32
+; CHECK-FP16-DAG: vcvtb.f32.f16
+; CHECK-LIBCALL-DAG: bl __aeabi_h2f
+; CHECK-LIBCALL-DAG: bl __aeabi_h2f
+
+; CHECK-VFP-DAG: vadd.f32
+; CHECK-NOVFP-DAG: bl __aeabi_fadd
+
+; CHECK-FP16-DAG: vcvtb.f16.f32
+; CHECK-LIBCALL-DAG: bl __aeabi_f2h
+define half @test_sitofp_i32_fadd(i32 %a, half %b) #0 {
+  %c = sitofp i32 %a to half
+  %r = fadd half %b, %c
+  ret half %r
+}
+
 attributes #0 = { nounwind }

diff --git a/test/CodeGen/ARM/fp16-v3.ll b/test/CodeGen/ARM/fp16-v3.ll
index 4a120723542aff1ae078a0b7b7a0ecd762474fbc..e26455e61e7f07135214e4c59d49df4697d505e6 100644 (file)
--- a/test/CodeGen/ARM/fp16-v3.ll
+++ b/test/CodeGen/ARM/fp16-v3.ll
@@ -4,11 +4,13 @@ target datalayout = "e-m:e-p:32:32-i64:64-v128:64:128-a:0:32-n32-S64"
 target triple = "armv7a--none-eabi"
 
 ; CHECK-LABEL: test_vec3:
-; CHECK: vcvtb.f32.f16
-; CHECK: vcvt.f32.s32
-; CHECK: vadd.f32
-; CHECK-NEXT: vcvtb.f16.f32 [[SREG:s[0-9]+]], {{.*}}
-; CHECK-NEXT: vmov [[RREG1:r[0-9]+]], [[SREG]]
+; CHECK-DAG: vcvtb.f32.f16 [[SREG1:s[0-9]+]],
+; CHECK-DAG: vcvt.f32.s32 [[SREG2:s[0-9]+]],
+; CHECK-DAG: vcvtb.f16.f32 [[SREG3:s[0-9]+]], [[SREG2]]
+; CHECK-DAG: vcvtb.f32.f16 [[SREG4:s[0-9]+]], [[SREG3]]
+; CHECK: vadd.f32 [[SREG5:s[0-9]+]], [[SREG4]], [[SREG1]]
+; CHECK-NEXT: vcvtb.f16.f32 [[SREG6:s[0-9]+]], [[SREG5]]
+; CHECK-NEXT: vmov [[RREG1:r[0-9]+]], [[SREG6]]
 ; CHECK-NEXT: uxth [[RREG2:r[0-9]+]], [[RREG1]]
 ; CHECK-NEXT: pkhbt [[RREG3:r[0-9]+]], [[RREG1]], [[RREG1]], lsl #16
 ; CHECK-DAG: strh [[RREG1]], [r0, #4]

diff --git a/test/CodeGen/X86/half.ll b/test/CodeGen/X86/half.ll
index 9a9fb5820554b483c5d225bb45257832f2ade401..717ddbfa6fdc8693212340ddacfb3f52ef1c6396 100644 (file)
--- a/test/CodeGen/X86/half.ll
+++ b/test/CodeGen/X86/half.ll
@@ -279,4 +279,38 @@ define half @test_f80trunc_nodagcombine() #0 {
   ret half %2
 }
 
+; CHECK-LABEL: test_sitofp_fadd_i32:
+
+; CHECK-LIBCALL-NEXT: pushq %rbx
+; CHECK-LIBCALL-NEXT: subq $16, %rsp
+; CHECK-LIBCALL-NEXT: movl %edi, %ebx
+; CHECK-LIBCALL-NEXT: movzwl (%rsi), %edi
+; CHECK-LIBCALL-NEXT: callq __gnu_h2f_ieee
+; CHECK-LIBCALL-NEXT: movss %xmm0, 12(%rsp)
+; CHECK-LIBCALL-NEXT: cvtsi2ssl %ebx, %xmm0
+; CHECK-LIBCALL-NEXT: callq __gnu_f2h_ieee
+; CHECK-LIBCALL-NEXT: movzwl %ax, %edi
+; CHECK-LIBCALL-NEXT: callq __gnu_h2f_ieee
+; CHECK-LIBCALL-NEXT: addss 12(%rsp), %xmm0
+; CHECK-LIBCALL-NEXT: addq $16, %rsp
+; CHECK-LIBCALL-NEXT: popq %rbx
+; CHECK-LIBCALL-NEXT: retq
+
+; CHECK-F16C-NEXT: movswl (%rsi), %eax
+; CHECK-F16C-NEXT: vmovd %eax, %xmm0
+; CHECK-F16C-NEXT: vcvtph2ps %xmm0, %xmm0
+; CHECK-F16C-NEXT: vcvtsi2ssl %edi, %xmm0, %xmm1
+; CHECK-F16C-NEXT: vcvtps2ph $4, %xmm1, %xmm1
+; CHECK-F16C-NEXT: vcvtph2ps %xmm1, %xmm1
+; CHECK-F16C-NEXT: vaddss %xmm1, %xmm0, %xmm0
+; CHECK-F16C-NEXT: retq
+
+define float @test_sitofp_fadd_i32(i32 %a, half* %b) #0 {
+  %tmp0 = load half, half* %b
+  %tmp1 = sitofp i32 %a to half
+  %tmp2 = fadd half %tmp0, %tmp1
+  %tmp3 = fpext half %tmp2 to float
+  ret float %tmp3
+}
+
 attributes #0 = { nounwind }