replica
/
hanchenye-llvm-project
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[X86] Don't use RCP14 and RSQRT14 for reciprocal estimations or for legacy SSE rcp/rsqrt intrinsics when AVX512 features are enabled. 

Summary:
AVX512 added RCP14 and RSQRT instructions which improve accuracy over the legacy RCP and RSQRT instruction, but not enough accuracy to remove the need for a Newton Raphson refinement.

Currently we use these new instructions for the legacy packed SSE instrinics, but not the scalar instrinsics. And we use it for fast math optimization of division and reciprocal sqrt.

I think switching the legacy instrinsics maybe surprising to the user since it changes the answer based on which processor you're using regardless of any fastmath settings. It's also weird that we did something different between scalar and packed.

As far at the reciprocal estimation, I think it creates unnecessary deltas in our output behavior (and prevents EVEX->VEX). A little playing around with gcc and icc and godbolt suggest they don't change which instructions they use here.

This patch adds new X86ISD nodes for the RCP14/RSQRT14 and uses those for the new intrinsics. Leaving the old intrinsics to use the old instructions.

Going forward I think our focus should be on
-Supporting 512-bit vectors, which will have to use the RCP14/RSQRT14.
-Using RSQRT28/RCP28 to remove the Newton Raphson step on processors with AVX512ER
-Supporting double precision.

Reviewers: zvi, DavidKreitzer, RKSimon

Reviewed By: RKSimon

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D39583

llvm-svn: 317413
This commit is contained in:
Craig Topper 2017-11-04 18:26:41 +00:00
parent be1f219050
commit 692c8efe30
12 changed files with 71 additions and 85 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
llvm/lib/Target/X86/X86ISelLowering.cpp
View File

@ -24808,9 +24808,7 @@ const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
case X86ISD::FMAXC: return "X86ISD::FMAXC";
case X86ISD::FMINC: return "X86ISD::FMINC";
case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
case X86ISD::FRSQRTS: return "X86ISD::FRSQRTS";
case X86ISD::FRCP: return "X86ISD::FRCP";
case X86ISD::FRCPS: return "X86ISD::FRCPS";
case X86ISD::EXTRQI: return "X86ISD::EXTRQI";
case X86ISD::INSERTQI: return "X86ISD::INSERTQI";
case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
@ -24988,9 +24986,13 @@ const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
case X86ISD::SELECT: return "X86ISD::SELECT";
case X86ISD::SELECTS: return "X86ISD::SELECTS";
case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
case X86ISD::RCP14: return "X86ISD::RCP14";
case X86ISD::RCP14S: return "X86ISD::RCP14S";
case X86ISD::RCP28: return "X86ISD::RCP28";
case X86ISD::RCP28S: return "X86ISD::RCP28S";
case X86ISD::EXP2: return "X86ISD::EXP2";
case X86ISD::RSQRT14: return "X86ISD::RSQRT14";
case X86ISD::RSQRT14S: return "X86ISD::RSQRT14S";
case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
case X86ISD::RSQRT28S: return "X86ISD::RSQRT28S";
case X86ISD::FADD_RND: return "X86ISD::FADD_RND";

4
llvm/lib/Target/X86/X86ISelLowering.h
View File

@ -254,7 +254,9 @@ namespace llvm {
/// Note that these typically require refinement
/// in order to obtain suitable precision.
FRSQRT, FRCP,
FRSQRTS, FRCPS,
// AVX-512 reciprocal approximations with a little more precision.
RSQRT14, RSQRT14S, RCP14, RCP14S,
// Thread Local Storage.
TLSADDR,

12
llvm/lib/Target/X86/X86InstrAVX512.td
View File

@ -7362,13 +7362,13 @@ multiclass avx512_fp14_s<bits<8> opc, string OpcodeStr, SDNode OpNode,
}
}
defm VRCP14SS : avx512_fp14_s<0x4D, "vrcp14ss", X86frcp14s, f32x_info>,
defm VRCP14SS : avx512_fp14_s<0x4D, "vrcp14ss", X86rcp14s, f32x_info>,
EVEX_CD8<32, CD8VT1>, T8PD, NotMemoryFoldable;
defm VRCP14SD : avx512_fp14_s<0x4D, "vrcp14sd", X86frcp14s, f64x_info>,
defm VRCP14SD : avx512_fp14_s<0x4D, "vrcp14sd", X86rcp14s, f64x_info>,
VEX_W, EVEX_CD8<64, CD8VT1>, T8PD, NotMemoryFoldable;
defm VRSQRT14SS : avx512_fp14_s<0x4F, "vrsqrt14ss", X86frsqrt14s, f32x_info>,
defm VRSQRT14SS : avx512_fp14_s<0x4F, "vrsqrt14ss", X86rsqrt14s, f32x_info>,
EVEX_CD8<32, CD8VT1>, T8PD, NotMemoryFoldable;
defm VRSQRT14SD : avx512_fp14_s<0x4F, "vrsqrt14sd", X86frsqrt14s, f64x_info>,
defm VRSQRT14SD : avx512_fp14_s<0x4F, "vrsqrt14sd", X86rsqrt14s, f64x_info>,
VEX_W, EVEX_CD8<64, CD8VT1>, T8PD, NotMemoryFoldable;
/// avx512_fp14_p rcp14ps, rcp14pd, rsqrt14ps, rsqrt14pd
@ -7414,8 +7414,8 @@ multiclass avx512_fp14_p_vl_all<bits<8> opc, string OpcodeStr, SDNode OpNode> {
}
}
defm VRSQRT14 : avx512_fp14_p_vl_all<0x4E, "vrsqrt14", X86frsqrt>;
defm VRCP14 : avx512_fp14_p_vl_all<0x4C, "vrcp14", X86frcp>;
defm VRSQRT14 : avx512_fp14_p_vl_all<0x4E, "vrsqrt14", X86rsqrt14>;
defm VRCP14 : avx512_fp14_p_vl_all<0x4C, "vrcp14", X86rcp14>;
/// avx512_fp28_s rcp28ss, rcp28sd, rsqrt28ss, rsqrt28sd
multiclass avx512_fp28_s<bits<8> opc, string OpcodeStr,X86VectorVTInfo _,

6
llvm/lib/Target/X86/X86InstrFragmentsSIMD.td
View File

@ -56,8 +56,6 @@ def X86fxor : SDNode<"X86ISD::FXOR", SDTFPBinOp,
def X86fandn : SDNode<"X86ISD::FANDN", SDTFPBinOp>;
def X86frsqrt : SDNode<"X86ISD::FRSQRT", SDTFPUnaryOp>;
def X86frcp : SDNode<"X86ISD::FRCP", SDTFPUnaryOp>;
def X86frsqrt14s: SDNode<"X86ISD::FRSQRTS", SDTFPBinOp>;
def X86frcp14s : SDNode<"X86ISD::FRCPS", SDTFPBinOp>;
def X86fhadd : SDNode<"X86ISD::FHADD", SDTFPBinOp>;
def X86fhsub : SDNode<"X86ISD::FHSUB", SDTFPBinOp>;
def X86hadd : SDNode<"X86ISD::HADD", SDTIntBinOp>;
@ -498,10 +496,14 @@ def SDTIFma : SDTypeProfile<1, 3, [SDTCisInt<0>, SDTCisSameAs<0,1>,
def x86vpmadd52l : SDNode<"X86ISD::VPMADD52L", SDTIFma, [SDNPCommutative]>;
def x86vpmadd52h : SDNode<"X86ISD::VPMADD52H", SDTIFma, [SDNPCommutative]>;
def X86rsqrt14 : SDNode<"X86ISD::RSQRT14", SDTFPUnaryOp>;
def X86rcp14 : SDNode<"X86ISD::RCP14", SDTFPUnaryOp>;
def X86rsqrt28 : SDNode<"X86ISD::RSQRT28", SDTFPUnaryOpRound>;
def X86rcp28 : SDNode<"X86ISD::RCP28", SDTFPUnaryOpRound>;
def X86exp2 : SDNode<"X86ISD::EXP2", SDTFPUnaryOpRound>;
def X86rsqrt14s : SDNode<"X86ISD::RSQRT14S", SDTFPBinOp>;
def X86rcp14s : SDNode<"X86ISD::RCP14S", SDTFPBinOp>;
def X86rsqrt28s : SDNode<"X86ISD::RSQRT28S", SDTFPBinOpRound>;
def X86rcp28s : SDNode<"X86ISD::RCP28S", SDTFPBinOpRound>;
def X86RndScales : SDNode<"X86ISD::VRNDSCALES", SDTFPBinOpImmRound>;

4
llvm/lib/Target/X86/X86InstrSSE.td
View File

@ -3252,9 +3252,9 @@ defm SQRT : sse1_fp_unop_s<0x51, "sqrt", fsqrt, SSE_SQRTSS>,
// Reciprocal approximations. Note that these typically require refinement
// in order to obtain suitable precision.
defm RSQRT : sse1_fp_unop_s<0x52, "rsqrt", X86frsqrt, SSE_RSQRTSS>,
sse1_fp_unop_p<0x52, "rsqrt", X86frsqrt, SSE_RSQRTPS, [HasAVX, NoVLX] >;
sse1_fp_unop_p<0x52, "rsqrt", X86frsqrt, SSE_RSQRTPS, [HasAVX]>;
defm RCP : sse1_fp_unop_s<0x53, "rcp", X86frcp, SSE_RCPS>,
sse1_fp_unop_p<0x53, "rcp", X86frcp, SSE_RCPP, [HasAVX, NoVLX]>;
sse1_fp_unop_p<0x53, "rcp", X86frcp, SSE_RCPP, [HasAVX]>;
// There is no f64 version of the reciprocal approximation instructions.

32
llvm/lib/Target/X86/X86IntrinsicsInfo.h
View File

@ -1428,26 +1428,26 @@ static const IntrinsicData IntrinsicsWithoutChain[] = {
X86_INTRINSIC_DATA(avx512_ptestnm_w_128, CMP_MASK, X86ISD::TESTNM, 0),
X86_INTRINSIC_DATA(avx512_ptestnm_w_256, CMP_MASK, X86ISD::TESTNM, 0),
X86_INTRINSIC_DATA(avx512_ptestnm_w_512, CMP_MASK, X86ISD::TESTNM, 0),
X86_INTRINSIC_DATA(avx512_rcp14_pd_128, INTR_TYPE_1OP_MASK, X86ISD::FRCP, 0),
X86_INTRINSIC_DATA(avx512_rcp14_pd_256, INTR_TYPE_1OP_MASK, X86ISD::FRCP, 0),
X86_INTRINSIC_DATA(avx512_rcp14_pd_512, INTR_TYPE_1OP_MASK, X86ISD::FRCP, 0),
X86_INTRINSIC_DATA(avx512_rcp14_ps_128, INTR_TYPE_1OP_MASK, X86ISD::FRCP, 0),
X86_INTRINSIC_DATA(avx512_rcp14_ps_256, INTR_TYPE_1OP_MASK, X86ISD::FRCP, 0),
X86_INTRINSIC_DATA(avx512_rcp14_ps_512, INTR_TYPE_1OP_MASK, X86ISD::FRCP, 0),
X86_INTRINSIC_DATA(avx512_rcp14_sd, INTR_TYPE_SCALAR_MASK, X86ISD::FRCPS, 0),
X86_INTRINSIC_DATA(avx512_rcp14_ss, INTR_TYPE_SCALAR_MASK, X86ISD::FRCPS, 0),
X86_INTRINSIC_DATA(avx512_rcp14_pd_128, INTR_TYPE_1OP_MASK, X86ISD::RCP14, 0),
X86_INTRINSIC_DATA(avx512_rcp14_pd_256, INTR_TYPE_1OP_MASK, X86ISD::RCP14, 0),
X86_INTRINSIC_DATA(avx512_rcp14_pd_512, INTR_TYPE_1OP_MASK, X86ISD::RCP14, 0),
X86_INTRINSIC_DATA(avx512_rcp14_ps_128, INTR_TYPE_1OP_MASK, X86ISD::RCP14, 0),
X86_INTRINSIC_DATA(avx512_rcp14_ps_256, INTR_TYPE_1OP_MASK, X86ISD::RCP14, 0),
X86_INTRINSIC_DATA(avx512_rcp14_ps_512, INTR_TYPE_1OP_MASK, X86ISD::RCP14, 0),
X86_INTRINSIC_DATA(avx512_rcp14_sd, INTR_TYPE_SCALAR_MASK, X86ISD::RCP14S, 0),
X86_INTRINSIC_DATA(avx512_rcp14_ss, INTR_TYPE_SCALAR_MASK, X86ISD::RCP14S, 0),
X86_INTRINSIC_DATA(avx512_rcp28_pd, INTR_TYPE_1OP_MASK_RM, X86ISD::RCP28, 0),
X86_INTRINSIC_DATA(avx512_rcp28_ps, INTR_TYPE_1OP_MASK_RM, X86ISD::RCP28, 0),
X86_INTRINSIC_DATA(avx512_rcp28_sd, INTR_TYPE_SCALAR_MASK_RM, X86ISD::RCP28S, 0),
X86_INTRINSIC_DATA(avx512_rcp28_ss, INTR_TYPE_SCALAR_MASK_RM, X86ISD::RCP28S, 0),
X86_INTRINSIC_DATA(avx512_rsqrt14_pd_128, INTR_TYPE_1OP_MASK, X86ISD::FRSQRT, 0),
X86_INTRINSIC_DATA(avx512_rsqrt14_pd_256, INTR_TYPE_1OP_MASK, X86ISD::FRSQRT, 0),
X86_INTRINSIC_DATA(avx512_rsqrt14_pd_512, INTR_TYPE_1OP_MASK, X86ISD::FRSQRT, 0),
X86_INTRINSIC_DATA(avx512_rsqrt14_ps_128, INTR_TYPE_1OP_MASK, X86ISD::FRSQRT, 0),
X86_INTRINSIC_DATA(avx512_rsqrt14_ps_256, INTR_TYPE_1OP_MASK, X86ISD::FRSQRT, 0),
X86_INTRINSIC_DATA(avx512_rsqrt14_ps_512, INTR_TYPE_1OP_MASK, X86ISD::FRSQRT, 0),
X86_INTRINSIC_DATA(avx512_rsqrt14_sd, INTR_TYPE_SCALAR_MASK, X86ISD::FRSQRTS, 0),
X86_INTRINSIC_DATA(avx512_rsqrt14_ss, INTR_TYPE_SCALAR_MASK, X86ISD::FRSQRTS, 0),
X86_INTRINSIC_DATA(avx512_rsqrt14_pd_128, INTR_TYPE_1OP_MASK, X86ISD::RSQRT14, 0),
X86_INTRINSIC_DATA(avx512_rsqrt14_pd_256, INTR_TYPE_1OP_MASK, X86ISD::RSQRT14, 0),
X86_INTRINSIC_DATA(avx512_rsqrt14_pd_512, INTR_TYPE_1OP_MASK, X86ISD::RSQRT14, 0),
X86_INTRINSIC_DATA(avx512_rsqrt14_ps_128, INTR_TYPE_1OP_MASK, X86ISD::RSQRT14, 0),
X86_INTRINSIC_DATA(avx512_rsqrt14_ps_256, INTR_TYPE_1OP_MASK, X86ISD::RSQRT14, 0),
X86_INTRINSIC_DATA(avx512_rsqrt14_ps_512, INTR_TYPE_1OP_MASK, X86ISD::RSQRT14, 0),
X86_INTRINSIC_DATA(avx512_rsqrt14_sd, INTR_TYPE_SCALAR_MASK, X86ISD::RSQRT14S, 0),
X86_INTRINSIC_DATA(avx512_rsqrt14_ss, INTR_TYPE_SCALAR_MASK, X86ISD::RSQRT14S, 0),
X86_INTRINSIC_DATA(avx512_rsqrt28_pd, INTR_TYPE_1OP_MASK_RM,X86ISD::RSQRT28, 0),
X86_INTRINSIC_DATA(avx512_rsqrt28_ps, INTR_TYPE_1OP_MASK_RM,X86ISD::RSQRT28, 0),
X86_INTRINSIC_DATA(avx512_rsqrt28_sd, INTR_TYPE_SCALAR_MASK_RM,X86ISD::RSQRT28S, 0),

26
llvm/test/CodeGen/X86/avx-intrinsics-x86.ll
View File

@ -581,15 +581,10 @@ declare i32 @llvm.x86.avx.ptestz.256(<4 x i64>, <4 x i64>) nounwind readnone
define <8 x float> @test_x86_avx_rcp_ps_256(<8 x float> %a0) {
; AVX-LABEL: test_x86_avx_rcp_ps_256:
; AVX: # BB#0:
; AVX-NEXT: vrcpps %ymm0, %ymm0 # encoding: [0xc5,0xfc,0x53,0xc0]
; AVX-NEXT: ret{{[l|q]}} # encoding: [0xc3]
;
; AVX512VL-LABEL: test_x86_avx_rcp_ps_256:
; AVX512VL: # BB#0:
; AVX512VL-NEXT: vrcp14ps %ymm0, %ymm0 # encoding: [0x62,0xf2,0x7d,0x28,0x4c,0xc0]
; AVX512VL-NEXT: ret{{[l|q]}} # encoding: [0xc3]
; CHECK-LABEL: test_x86_avx_rcp_ps_256:
; CHECK: # BB#0:
; CHECK-NEXT: vrcpps %ymm0, %ymm0 # encoding: [0xc5,0xfc,0x53,0xc0]
; CHECK-NEXT: ret{{[l|q]}} # encoding: [0xc3]
%res = call <8 x float> @llvm.x86.avx.rcp.ps.256(<8 x float> %a0) ; <<8 x float>> [#uses=1]
ret <8 x float> %res
}
@ -619,15 +614,10 @@ declare <8 x float> @llvm.x86.avx.round.ps.256(<8 x float>, i32) nounwind readno
define <8 x float> @test_x86_avx_rsqrt_ps_256(<8 x float> %a0) {
; AVX-LABEL: test_x86_avx_rsqrt_ps_256:
; AVX: # BB#0:
; AVX-NEXT: vrsqrtps %ymm0, %ymm0 # encoding: [0xc5,0xfc,0x52,0xc0]
; AVX-NEXT: ret{{[l|q]}} # encoding: [0xc3]
;
; AVX512VL-LABEL: test_x86_avx_rsqrt_ps_256:
; AVX512VL: # BB#0:
; AVX512VL-NEXT: vrsqrt14ps %ymm0, %ymm0 # encoding: [0x62,0xf2,0x7d,0x28,0x4e,0xc0]
; AVX512VL-NEXT: ret{{[l|q]}} # encoding: [0xc3]
; CHECK-LABEL: test_x86_avx_rsqrt_ps_256:
; CHECK: # BB#0:
; CHECK-NEXT: vrsqrtps %ymm0, %ymm0 # encoding: [0xc5,0xfc,0x52,0xc0]
; CHECK-NEXT: ret{{[l|q]}} # encoding: [0xc3]
%res = call <8 x float> @llvm.x86.avx.rsqrt.ps.256(<8 x float> %a0) ; <<8 x float>> [#uses=1]
ret <8 x float> %res
}

8
llvm/test/CodeGen/X86/avx-schedule.ll
View File

@ -3982,8 +3982,8 @@ define <8 x float> @test_rcpps(<8 x float> %a0, <8 x float> *%a1) {
;
; SKX-LABEL: test_rcpps:
; SKX: # BB#0:
; SKX-NEXT: vrcp14ps %ymm0, %ymm0 # sched: [4:1.00]
; SKX-NEXT: vrcp14ps (%rdi), %ymm1 # sched: [11:1.00]
; SKX-NEXT: vrcpps %ymm0, %ymm0 # sched: [4:1.00]
; SKX-NEXT: vrcpps (%rdi), %ymm1 # sched: [11:1.00]
; SKX-NEXT: vaddps %ymm1, %ymm0, %ymm0 # sched: [4:0.33]
; SKX-NEXT: retq # sched: [7:1.00]
;
@ -4174,8 +4174,8 @@ define <8 x float> @test_rsqrtps(<8 x float> %a0, <8 x float> *%a1) {
;
; SKX-LABEL: test_rsqrtps:
; SKX: # BB#0:
; SKX-NEXT: vrsqrt14ps %ymm0, %ymm0 # sched: [4:1.00]
; SKX-NEXT: vrsqrt14ps (%rdi), %ymm1 # sched: [11:1.00]
; SKX-NEXT: vrsqrtps %ymm0, %ymm0 # sched: [4:1.00]
; SKX-NEXT: vrsqrtps (%rdi), %ymm1 # sched: [11:1.00]
; SKX-NEXT: vaddps %ymm1, %ymm0, %ymm0 # sched: [4:0.33]
; SKX-NEXT: retq # sched: [7:1.00]
;

8
llvm/test/CodeGen/X86/recip-fastmath.ll
View File

@ -416,7 +416,7 @@ define <4 x float> @v4f32_one_step(<4 x float> %x) #1 {
;
; SKX-LABEL: v4f32_one_step:
; SKX: # BB#0:
; SKX-NEXT: vrcp14ps %xmm0, %xmm1 # sched: [4:1.00]
; SKX-NEXT: vrcpps %xmm0, %xmm1 # sched: [4:1.00]
; SKX-NEXT: vfnmadd213ps {{.*}}(%rip){1to4}, %xmm1, %xmm0 # sched: [10:0.50]
; SKX-NEXT: vfmadd132ps %xmm1, %xmm1, %xmm0 # sched: [4:0.33]
; SKX-NEXT: retq # sched: [7:1.00]
@ -533,7 +533,7 @@ define <4 x float> @v4f32_two_step(<4 x float> %x) #2 {
;
; SKX-LABEL: v4f32_two_step:
; SKX: # BB#0:
; SKX-NEXT: vrcp14ps %xmm0, %xmm1 # sched: [4:1.00]
; SKX-NEXT: vrcpps %xmm0, %xmm1 # sched: [4:1.00]
; SKX-NEXT: vbroadcastss {{.*#+}} xmm2 = [1,1,1,1] sched: [6:0.50]
; SKX-NEXT: vmovaps %xmm1, %xmm3 # sched: [1:1.00]
; SKX-NEXT: vfnmadd213ps %xmm2, %xmm0, %xmm3 # sched: [4:0.33]
@ -691,7 +691,7 @@ define <8 x float> @v8f32_one_step(<8 x float> %x) #1 {
;
; SKX-LABEL: v8f32_one_step:
; SKX: # BB#0:
; SKX-NEXT: vrcp14ps %ymm0, %ymm1 # sched: [4:1.00]
; SKX-NEXT: vrcpps %ymm0, %ymm1 # sched: [4:1.00]
; SKX-NEXT: vfnmadd213ps {{.*}}(%rip){1to8}, %ymm1, %ymm0 # sched: [11:0.50]
; SKX-NEXT: vfmadd132ps %ymm1, %ymm1, %ymm0 # sched: [4:0.33]
; SKX-NEXT: retq # sched: [7:1.00]
@ -821,7 +821,7 @@ define <8 x float> @v8f32_two_step(<8 x float> %x) #2 {
;
; SKX-LABEL: v8f32_two_step:
; SKX: # BB#0:
; SKX-NEXT: vrcp14ps %ymm0, %ymm1 # sched: [4:1.00]
; SKX-NEXT: vrcpps %ymm0, %ymm1 # sched: [4:1.00]
; SKX-NEXT: vbroadcastss {{.*#+}} ymm2 = [1,1,1,1,1,1,1,1] sched: [7:0.50]
; SKX-NEXT: vmovaps %ymm1, %ymm3 # sched: [1:1.00]
; SKX-NEXT: vfnmadd213ps %ymm2, %ymm0, %ymm3 # sched: [4:0.33]

16
llvm/test/CodeGen/X86/recip-fastmath2.ll
View File

@ -478,7 +478,7 @@ define <4 x float> @v4f32_one_step2(<4 x float> %x) #1 {
;
; SKX-LABEL: v4f32_one_step2:
; SKX: # BB#0:
; SKX-NEXT: vrcp14ps %xmm0, %xmm1 # sched: [4:1.00]
; SKX-NEXT: vrcpps %xmm0, %xmm1 # sched: [4:1.00]
; SKX-NEXT: vfnmadd213ps {{.*}}(%rip){1to4}, %xmm1, %xmm0 # sched: [10:0.50]
; SKX-NEXT: vfmadd132ps %xmm1, %xmm1, %xmm0 # sched: [4:0.33]
; SKX-NEXT: vmulps {{.*}}(%rip), %xmm0, %xmm0 # sched: [10:0.50]
@ -580,7 +580,7 @@ define <4 x float> @v4f32_one_step_2_divs(<4 x float> %x) #1 {
;
; SKX-LABEL: v4f32_one_step_2_divs:
; SKX: # BB#0:
; SKX-NEXT: vrcp14ps %xmm0, %xmm1 # sched: [4:1.00]
; SKX-NEXT: vrcpps %xmm0, %xmm1 # sched: [4:1.00]
; SKX-NEXT: vfnmadd213ps {{.*}}(%rip){1to4}, %xmm1, %xmm0 # sched: [10:0.50]
; SKX-NEXT: vfmadd132ps %xmm1, %xmm1, %xmm0 # sched: [4:0.33]
; SKX-NEXT: vmulps {{.*}}(%rip), %xmm0, %xmm1 # sched: [10:0.50]
@ -708,7 +708,7 @@ define <4 x float> @v4f32_two_step2(<4 x float> %x) #2 {
;
; SKX-LABEL: v4f32_two_step2:
; SKX: # BB#0:
; SKX-NEXT: vrcp14ps %xmm0, %xmm1 # sched: [4:1.00]
; SKX-NEXT: vrcpps %xmm0, %xmm1 # sched: [4:1.00]
; SKX-NEXT: vbroadcastss {{.*#+}} xmm2 = [1,1,1,1] sched: [6:0.50]
; SKX-NEXT: vmovaps %xmm1, %xmm3 # sched: [1:1.00]
; SKX-NEXT: vfnmadd213ps %xmm2, %xmm0, %xmm3 # sched: [4:0.33]
@ -814,7 +814,7 @@ define <8 x float> @v8f32_one_step2(<8 x float> %x) #1 {
;
; SKX-LABEL: v8f32_one_step2:
; SKX: # BB#0:
; SKX-NEXT: vrcp14ps %ymm0, %ymm1 # sched: [4:1.00]
; SKX-NEXT: vrcpps %ymm0, %ymm1 # sched: [4:1.00]
; SKX-NEXT: vfnmadd213ps {{.*}}(%rip){1to8}, %ymm1, %ymm0 # sched: [11:0.50]
; SKX-NEXT: vfmadd132ps %ymm1, %ymm1, %ymm0 # sched: [4:0.33]
; SKX-NEXT: vmulps {{.*}}(%rip), %ymm0, %ymm0 # sched: [11:0.50]
@ -925,7 +925,7 @@ define <8 x float> @v8f32_one_step_2_divs(<8 x float> %x) #1 {
;
; SKX-LABEL: v8f32_one_step_2_divs:
; SKX: # BB#0:
; SKX-NEXT: vrcp14ps %ymm0, %ymm1 # sched: [4:1.00]
; SKX-NEXT: vrcpps %ymm0, %ymm1 # sched: [4:1.00]
; SKX-NEXT: vfnmadd213ps {{.*}}(%rip){1to8}, %ymm1, %ymm0 # sched: [11:0.50]
; SKX-NEXT: vfmadd132ps %ymm1, %ymm1, %ymm0 # sched: [4:0.33]
; SKX-NEXT: vmulps {{.*}}(%rip), %ymm0, %ymm1 # sched: [11:0.50]
@ -1067,7 +1067,7 @@ define <8 x float> @v8f32_two_step2(<8 x float> %x) #2 {
;
; SKX-LABEL: v8f32_two_step2:
; SKX: # BB#0:
; SKX-NEXT: vrcp14ps %ymm0, %ymm1 # sched: [4:1.00]
; SKX-NEXT: vrcpps %ymm0, %ymm1 # sched: [4:1.00]
; SKX-NEXT: vbroadcastss {{.*#+}} ymm2 = [1,1,1,1,1,1,1,1] sched: [7:0.50]
; SKX-NEXT: vmovaps %ymm1, %ymm3 # sched: [1:1.00]
; SKX-NEXT: vfnmadd213ps %ymm2, %ymm0, %ymm3 # sched: [4:0.33]
@ -1124,7 +1124,7 @@ define <8 x float> @v8f32_no_step(<8 x float> %x) #3 {
;
; SKX-LABEL: v8f32_no_step:
; SKX: # BB#0:
; SKX-NEXT: vrcp14ps %ymm0, %ymm0 # sched: [4:1.00]
; SKX-NEXT: vrcpps %ymm0, %ymm0 # sched: [4:1.00]
; SKX-NEXT: retq # sched: [7:1.00]
%div = fdiv fast <8 x float> <float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0>, %x
ret <8 x float> %div
@ -1183,7 +1183,7 @@ define <8 x float> @v8f32_no_step2(<8 x float> %x) #3 {
;
; SKX-LABEL: v8f32_no_step2:
; SKX: # BB#0:
; SKX-NEXT: vrcp14ps %ymm0, %ymm0 # sched: [4:1.00]
; SKX-NEXT: vrcpps %ymm0, %ymm0 # sched: [4:1.00]
; SKX-NEXT: vmulps {{.*}}(%rip), %ymm0, %ymm0 # sched: [11:0.50]
; SKX-NEXT: retq # sched: [7:1.00]
%div = fdiv fast <8 x float> <float 1.0, float 2.0, float 3.0, float 4.0, float 5.0, float 6.0, float 7.0, float 8.0>, %x

26
llvm/test/CodeGen/X86/sse-intrinsics-x86.ll
View File

@ -401,15 +401,10 @@ define <4 x float> @test_x86_sse_rcp_ps(<4 x float> %a0) {
; SSE-NEXT: rcpps %xmm0, %xmm0 ## encoding: [0x0f,0x53,0xc0]
; SSE-NEXT: retl ## encoding: [0xc3]
;
; AVX2-LABEL: test_x86_sse_rcp_ps:
; AVX2: ## BB#0:
; AVX2-NEXT: vrcpps %xmm0, %xmm0 ## encoding: [0xc5,0xf8,0x53,0xc0]
; AVX2-NEXT: retl ## encoding: [0xc3]
;
; SKX-LABEL: test_x86_sse_rcp_ps:
; SKX: ## BB#0:
; SKX-NEXT: vrcp14ps %xmm0, %xmm0 ## encoding: [0x62,0xf2,0x7d,0x08,0x4c,0xc0]
; SKX-NEXT: retl ## encoding: [0xc3]
; VCHECK-LABEL: test_x86_sse_rcp_ps:
; VCHECK: ## BB#0:
; VCHECK-NEXT: vrcpps %xmm0, %xmm0 ## encoding: [0xc5,0xf8,0x53,0xc0]
; VCHECK-NEXT: retl ## encoding: [0xc3]
%res = call <4 x float> @llvm.x86.sse.rcp.ps(<4 x float> %a0) ; <<4 x float>> [#uses=1]
ret <4 x float> %res
}
@ -438,15 +433,10 @@ define <4 x float> @test_x86_sse_rsqrt_ps(<4 x float> %a0) {
; SSE-NEXT: rsqrtps %xmm0, %xmm0 ## encoding: [0x0f,0x52,0xc0]
; SSE-NEXT: retl ## encoding: [0xc3]
;
; AVX2-LABEL: test_x86_sse_rsqrt_ps:
; AVX2: ## BB#0:
; AVX2-NEXT: vrsqrtps %xmm0, %xmm0 ## encoding: [0xc5,0xf8,0x52,0xc0]
; AVX2-NEXT: retl ## encoding: [0xc3]
;
; SKX-LABEL: test_x86_sse_rsqrt_ps:
; SKX: ## BB#0:
; SKX-NEXT: vrsqrt14ps %xmm0, %xmm0 ## encoding: [0x62,0xf2,0x7d,0x08,0x4e,0xc0]
; SKX-NEXT: retl ## encoding: [0xc3]
; VCHECK-LABEL: test_x86_sse_rsqrt_ps:
; VCHECK: ## BB#0:
; VCHECK-NEXT: vrsqrtps %xmm0, %xmm0 ## encoding: [0xc5,0xf8,0x52,0xc0]
; VCHECK-NEXT: retl ## encoding: [0xc3]
%res = call <4 x float> @llvm.x86.sse.rsqrt.ps(<4 x float> %a0) ; <<4 x float>> [#uses=1]
ret <4 x float> %res
}

8
llvm/test/CodeGen/X86/sse-schedule.ll
View File

@ -2547,8 +2547,8 @@ define <4 x float> @test_rcpps(<4 x float> %a0, <4 x float> *%a1) {
;
; SKX-LABEL: test_rcpps:
; SKX: # BB#0:
; SKX-NEXT: vrcp14ps %xmm0, %xmm0 # sched: [4:1.00]
; SKX-NEXT: vrcp14ps (%rdi), %xmm1 # sched: [10:1.00]
; SKX-NEXT: vrcpps %xmm0, %xmm0 # sched: [4:1.00]
; SKX-NEXT: vrcpps (%rdi), %xmm1 # sched: [10:1.00]
; SKX-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.33]
; SKX-NEXT: retq # sched: [7:1.00]
;
@ -2719,8 +2719,8 @@ define <4 x float> @test_rsqrtps(<4 x float> %a0, <4 x float> *%a1) {
;
; SKX-LABEL: test_rsqrtps:
; SKX: # BB#0:
; SKX-NEXT: vrsqrt14ps %xmm0, %xmm0 # sched: [4:1.00]
; SKX-NEXT: vrsqrt14ps (%rdi), %xmm1 # sched: [10:1.00]
; SKX-NEXT: vrsqrtps %xmm0, %xmm0 # sched: [4:1.00]
; SKX-NEXT: vrsqrtps (%rdi), %xmm1 # sched: [10:1.00]
; SKX-NEXT: vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.33]
; SKX-NEXT: retq # sched: [7:1.00]
;