[llvm] r237004 - [X86] Updates to X86 backend for f16 promotion
Pirama Arumuga Nainar
pirama at google.com
Mon May 11 10:14:39 PDT 2015
Author: pirama
Date: Mon May 11 12:14:39 2015
New Revision: 237004
URL: http://llvm.org/viewvc/llvm-project?rev=237004&view=rev
Log:
[X86] Updates to X86 backend for f16 promotion
Summary:
r235215 adds support for f16 to be considered as a load/store type and
promote f16 operations to f32.
This patch has miscellaneous fixes for the X86 backend so all f16
operations are handled:
1. Set loadextaction for f16 vectors to expand.
2. Handle FP_EXTEND in a switch statement when handling v2f32
3. Do not fold (FP_TO_SINT (load f16)) into FP_TO_INT*_IN_MEM or
(store (SINT_TO_FP )) to a FILD.
Tests included.
Reviewers: ab, srhines, delena
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D9092
Modified:
llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
llvm/trunk/test/CodeGen/X86/half.ll
Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=237004&r1=237003&r2=237004&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Mon May 11 12:14:39 2015
@@ -750,6 +750,11 @@ X86TargetLowering::X86TargetLowering(con
// them legal.
if (VT.getVectorElementType() == MVT::i1)
setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
+
+ // EXTLOAD for MVT::f16 vectors is not legal because f16 vectors are
+ // split/scalarized right now.
+ if (VT.getVectorElementType() == MVT::f16)
+ setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
}
}
@@ -17623,6 +17628,11 @@ void X86TargetLowering::ReplaceNodeResul
return;
}
case ISD::FP_TO_SINT:
+ // FP_TO_INT*_IN_MEM is not legal for f16 inputs. Do not convert
+ // (FP_TO_SINT (load f16)) to FP_TO_INT*.
+ if (N->getOperand(0).getValueType() == MVT::f16)
+ break;
+ // fallthrough
case ISD::FP_TO_UINT: {
bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
@@ -17668,6 +17678,13 @@ void X86TargetLowering::ReplaceNodeResul
Results.push_back(V);
return;
}
+ case ISD::FP_EXTEND: {
+ // Right now, only MVT::v2f32 has OperationAction for FP_EXTEND.
+ // No other ValueType for FP_EXTEND should reach this point.
+ assert(N->getValueType(0) == MVT::v2f32 &&
+ "Do not know how to legalize this Node");
+ return;
+ }
case ISD::INTRINSIC_W_CHAIN: {
unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
switch (IntNo) {
@@ -24021,6 +24038,11 @@ static SDValue PerformSINT_TO_FPCombine(
if (Op0.getOpcode() == ISD::LOAD) {
LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
EVT VT = Ld->getValueType(0);
+
+ // This transformation is not supported if the result type is f16
+ if (N->getValueType(0) == MVT::f16)
+ return SDValue();
+
if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
!Subtarget->is64Bit() && VT == MVT::i64) {
Modified: llvm/trunk/test/CodeGen/X86/half.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/half.ll?rev=237004&r1=237003&r2=237004&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/half.ll (original)
+++ llvm/trunk/test/CodeGen/X86/half.ll Mon May 11 12:14:39 2015
@@ -1,5 +1,7 @@
-; RUN: llc < %s -march=x86-64 -mtriple=x86_64-unknown-linux-gnu -mcpu=corei7 -mattr=-f16c | FileCheck %s -check-prefix=CHECK -check-prefix=CHECK-LIBCALL
-; RUN: llc < %s -march=x86-64 -mtriple=x86_64-unknown-linux-gnu -mcpu=corei7 -mattr=+f16c | FileCheck %s -check-prefix=CHECK -check-prefix=CHECK-F16C
+; RUN: llc < %s -march=x86-64 -mtriple=x86_64-unknown-linux-gnu -mcpu=corei7 -mattr=-f16c -asm-verbose=false \
+; RUN: | FileCheck %s -check-prefix=CHECK -check-prefix=CHECK-LIBCALL
+; RUN: llc < %s -march=x86-64 -mtriple=x86_64-unknown-linux-gnu -mcpu=corei7 -mattr=+f16c -asm-verbose=false \
+; RUN: | FileCheck %s -check-prefix=CHECK -check-prefix=CHECK-F16C
define void @test_load_store(half* %in, half* %out) {
; CHECK-LABEL: test_load_store:
@@ -30,7 +32,7 @@ define float @test_extend32(half* %addr)
; CHECK-LABEL: test_extend32:
; CHECK-LIBCALL: jmp __gnu_h2f_ieee
-; CHECK-FP16: vcvtph2ps
+; CHECK-F16C: vcvtph2ps
%val16 = load half, half* %addr
%val32 = fpext half %val16 to float
ret float %val32
@@ -41,8 +43,8 @@ define double @test_extend64(half* %addr
; CHECK-LIBCALL: callq __gnu_h2f_ieee
; CHECK-LIBCALL: cvtss2sd
-; CHECK-FP16: vcvtph2ps
-; CHECK-FP16: vcvtss2sd
+; CHECK-F16C: vcvtph2ps
+; CHECK-F16C: vcvtss2sd
%val16 = load half, half* %addr
%val32 = fpext half %val16 to double
ret double %val32
@@ -52,7 +54,7 @@ define void @test_trunc32(float %in, hal
; CHECK-LABEL: test_trunc32:
; CHECK-LIBCALL: callq __gnu_f2h_ieee
-; CHECK-FP16: vcvtps2ph
+; CHECK-F16C: vcvtps2ph
%val16 = fptrunc float %in to half
store half %val16, half* %addr
ret void
@@ -62,8 +64,200 @@ define void @test_trunc64(double %in, ha
; CHECK-LABEL: test_trunc64:
; CHECK-LIBCALL: callq __truncdfhf2
-; CHECK-FP16: callq __truncdfhf2
+; CHECK-F16C: callq __truncdfhf2
%val16 = fptrunc double %in to half
store half %val16, half* %addr
ret void
}
+
+define i64 @test_fptosi_i64(half* %p) #0 {
+; CHECK-LABEL: test_fptosi_i64:
+
+; CHECK-LIBCALL-NEXT: pushq %rax
+; CHECK-LIBCALL-NEXT: movzwl (%rdi), %edi
+; CHECK-LIBCALL-NEXT: callq __gnu_h2f_ieee
+; CHECK-LIBCALL-NEXT: cvttss2si %xmm0, %rax
+; CHECK-LIBCALL-NEXT: popq %rdx
+; CHECK-LIBCALL-NEXT: retq
+
+; CHECK-F16C-NEXT: movswl (%rdi), [[REG0:%[a-z0-9]+]]
+; CHECK-F16C-NEXT: vmovd [[REG0]], [[REG1:%[a-z0-9]+]]
+; CHECK-F16C-NEXT: vcvtph2ps [[REG1]], [[REG2:%[a-z0-9]+]]
+; CHECK-F16C-NEXT: vcvttss2si [[REG2]], %rax
+; CHECK-F16C-NEXT: retq
+ %a = load half, half* %p, align 2
+ %r = fptosi half %a to i64
+ ret i64 %r
+}
+
+define void @test_sitofp_i64(i64 %a, half* %p) #0 {
+; CHECK-LABEL: test_sitofp_i64:
+
+; CHECK-LIBCALL-NEXT: pushq [[ADDR:%[a-z]+]]
+; CHECK-LIBCALL-NEXT: movq %rsi, [[ADDR]]
+; CHECK-LIBCALL-NEXT: cvtsi2ssq %rdi, %xmm0
+; CHECK-LIBCALL-NEXT: callq __gnu_f2h_ieee
+; CHECK-LIBCALL-NEXT: movw %ax, ([[ADDR]])
+; CHECK_LIBCALL-NEXT: popq [[ADDR]]
+; CHECK_LIBCALL-NEXT: retq
+
+; CHECK-F16C-NEXT: vcvtsi2ssq %rdi, [[REG0:%[a-z0-9]+]], [[REG0]]
+; CHECK-F16C-NEXT: vcvtps2ph $0, [[REG0]], [[REG0]]
+; CHECK-F16C-NEXT: vmovd [[REG0]], %eax
+; CHECK-F16C-NEXT: movw %ax, (%rsi)
+; CHECK-F16C-NEXT: retq
+ %r = sitofp i64 %a to half
+ store half %r, half* %p
+ ret void
+}
+
+define i64 @test_fptoui_i64(half* %p) #0 {
+; CHECK-LABEL: test_fptoui_i64:
+
+; FP_TO_UINT is expanded using FP_TO_SINT
+; CHECK-LIBCALL-NEXT: pushq %rax
+; CHECK-LIBCALL-NEXT: movzwl (%rdi), %edi
+; CHECK-LIBCALL-NEXT: callq __gnu_h2f_ieee
+; CHECK-LIBCALL-NEXT: movss {{.[A-Z_0-9]+}}(%rip), [[REG1:%[a-z0-9]+]]
+; CHECK-LIBCALL-NEXT: movaps %xmm0, [[REG2:%[a-z0-9]+]]
+; CHECK-LIBCALL-NEXT: subss [[REG1]], [[REG2]]
+; CHECK-LIBCALL-NEXT: cvttss2si [[REG2]], [[REG3:%[a-z0-9]+]]
+; CHECK-LIBCALL-NEXT: movabsq $-9223372036854775808, [[REG4:%[a-z0-9]+]]
+; CHECK-LIBCALL-NEXT: xorq [[REG3]], [[REG4]]
+; CHECK-LIBCALL-NEXT: cvttss2si %xmm0, [[REG5:%[a-z0-9]+]]
+; CHECK-LIBCALL-NEXT: ucomiss [[REG1]], %xmm0
+; CHECK-LIBCALL-NEXT: cmovaeq [[REG4]], [[REG5]]
+; CHECK-LIBCALL-NEXT: popq %rdx
+; CHECK-LIBCALL-NEXT: retq
+
+; CHECK-F16C-NEXT: movswl (%rdi), [[REG0:%[a-z0-9]+]]
+; CHECK-F16C-NEXT: vmovd [[REG0]], [[REG1:%[a-z0-9]+]]
+; CHECK-F16C-NEXT: vcvtph2ps [[REG1]], [[REG2:%[a-z0-9]+]]
+; CHECK-F16C-NEXT: vmovss {{.[A-Z_0-9]+}}(%rip), [[REG3:%[a-z0-9]+]]
+; CHECK-F16C-NEXT: vsubss [[REG3]], [[REG2]], [[REG4:%[a-z0-9]+]]
+; CHECK-F16C-NEXT: vcvttss2si [[REG4]], [[REG5:%[a-z0-9]+]]
+; CHECK-F16C-NEXT: movabsq $-9223372036854775808, [[REG6:%[a-z0-9]+]]
+; CHECK-F16C-NEXT: xorq [[REG5]], [[REG6:%[a-z0-9]+]]
+; CHECK-F16C-NEXT: vcvttss2si [[REG2]], [[REG7:%[a-z0-9]+]]
+; CHECK-F16C-NEXT: vucomiss [[REG3]], [[REG2]]
+; CHECK-F16C-NEXT: cmovaeq [[REG6]], %rax
+; CHECK-F16C-NEXT: retq
+ %a = load half, half* %p, align 2
+ %r = fptoui half %a to i64
+ ret i64 %r
+}
+
+define void @test_uitofp_i64(i64 %a, half* %p) #0 {
+; CHECK-LABEL: test_uitofp_i64:
+; CHECK-LIBCALL-NEXT: pushq [[ADDR:%[a-z0-9]+]]
+; CHECK-LIBCALL-NEXT: movq %rsi, [[ADDR]]
+; CHECK-NEXT: movl %edi, [[REG0:%[a-z0-9]+]]
+; CHECK-NEXT: andl $1, [[REG0]]
+; CHECK-NEXT: testq %rdi, %rdi
+; CHECK-NEXT: js [[LABEL1:.LBB[0-9_]+]]
+
+; simple conversion to float if non-negative
+; CHECK-LIBCALL-NEXT: cvtsi2ssq %rdi, [[REG1:%[a-z0-9]+]]
+; CHECK-F16C-NEXT: vcvtsi2ssq %rdi, [[REG1:%[a-z0-9]+]], [[REG1]]
+; CHECK-NEXT: jmp [[LABEL2:.LBB[0-9_]+]]
+
+; convert using shift+or if negative
+; CHECK-NEXT: [[LABEL1]]:
+; CHECK-NEXT: shrq %rdi
+; CHECK-NEXT: orq %rdi, [[REG2:%[a-z0-9]+]]
+; CHECK-LIBCALL-NEXT: cvtsi2ssq [[REG2]], [[REG3:%[a-z0-9]+]]
+; CHECK-LIBCALL-NEXT: addss [[REG3]], [[REG1]]
+; CHECK-F16C-NEXT: vcvtsi2ssq [[REG2]], [[REG3:%[a-z0-9]+]], [[REG3]]
+; CHECK-F16C-NEXT: vaddss [[REG3]], [[REG3]], [[REG1:[%a-z0-9]+]]
+
+; convert float to half
+; CHECK-NEXT: [[LABEL2]]:
+; CHECK-LIBCALL-NEXT: callq __gnu_f2h_ieee
+; CHECK-LIBCALL-NEXT: movw %ax, ([[ADDR]])
+; CHECK-LIBCALL-NEXT: popq [[ADDR]]
+; CHECK-F16C-NEXT: vcvtps2ph $0, [[REG1]], [[REG4:%[a-z0-9]+]]
+; CHECK-F16C-NEXT: vmovd [[REG4]], %eax
+; CHECK-F16C-NEXT: movw %ax, (%rsi)
+; CHECK-NEXT: retq
+
+ %r = uitofp i64 %a to half
+ store half %r, half* %p
+ ret void
+}
+
+define <4 x float> @test_extend32_vec4(<4 x half>* %p) #0 {
+; CHECK-LABEL: test_extend32_vec4:
+
+; CHECK-LIBCALL: callq __gnu_h2f_ieee
+; CHECK-LIBCALL: callq __gnu_h2f_ieee
+; CHECK-LIBCALL: callq __gnu_h2f_ieee
+; CHECK-LIBCALL: callq __gnu_h2f_ieee
+; CHECK-F16C: vcvtph2ps
+; CHECK-F16C: vcvtph2ps
+; CHECK-F16C: vcvtph2ps
+; CHECK-F16C: vcvtph2ps
+ %a = load <4 x half>, <4 x half>* %p, align 8
+ %b = fpext <4 x half> %a to <4 x float>
+ ret <4 x float> %b
+}
+
+define <4 x double> @test_extend64_vec4(<4 x half>* %p) #0 {
+; CHECK-LABEL: test_extend64_vec4
+
+; CHECK-LIBCALL: callq __gnu_h2f_ieee
+; CHECK-LIBCALL-DAG: callq __gnu_h2f_ieee
+; CHECK-LIBCALL-DAG: callq __gnu_h2f_ieee
+; CHECK-LIBCALL-DAG: callq __gnu_h2f_ieee
+; CHECK-LIBCALL-DAG: cvtss2sd
+; CHECK-LIBCALL-DAG: cvtss2sd
+; CHECK-LIBCALL-DAG: cvtss2sd
+; CHECK-LIBCALL: cvtss2sd
+; CHECK-F16C: vcvtph2ps
+; CHECK-F16C-DAG: vcvtph2ps
+; CHECK-F16C-DAG: vcvtph2ps
+; CHECK-F16C-DAG: vcvtph2ps
+; CHECK-F16C-DAG: vcvtss2sd
+; CHECK-F16C-DAG: vcvtss2sd
+; CHECK-F16C-DAG: vcvtss2sd
+; CHECK-F16C: vcvtss2sd
+ %a = load <4 x half>, <4 x half>* %p, align 8
+ %b = fpext <4 x half> %a to <4 x double>
+ ret <4 x double> %b
+}
+
+define void @test_trunc32_vec4(<4 x float> %a, <4 x half>* %p) {
+; CHECK-LABEL: test_trunc32_vec4:
+
+; CHECK-LIBCALL: callq __gnu_f2h_ieee
+; CHECK-LIBCALL: callq __gnu_f2h_ieee
+; CHECK-LIBCALL: callq __gnu_f2h_ieee
+; CHECK-LIBCALL: callq __gnu_f2h_ieee
+; CHECK-F16C: vcvtps2ph
+; CHECK-F16C: vcvtps2ph
+; CHECK-F16C: vcvtps2ph
+; CHECK-F16C: vcvtps2ph
+; CHECK: movw
+; CHECK: movw
+; CHECK: movw
+; CHECK: movw
+ %v = fptrunc <4 x float> %a to <4 x half>
+ store <4 x half> %v, <4 x half>* %p
+ ret void
+}
+
+define void @test_trunc64_vec4(<4 x double> %a, <4 x half>* %p) {
+; CHECK-LABEL: test_trunc64_vec4:
+; CHECK: callq __truncdfhf2
+; CHECK: callq __truncdfhf2
+; CHECK: callq __truncdfhf2
+; CHECK: callq __truncdfhf2
+; CHECK: movw
+; CHECK: movw
+; CHECK: movw
+; CHECK: movw
+ %v = fptrunc <4 x double> %a to <4 x half>
+ store <4 x half> %v, <4 x half>* %p
+ ret void
+}
+
+attributes #0 = { nounwind }
More information about the llvm-commits
mailing list