[X86][SchedModel] SSE reciprocal square root instruction latencies.

The SSE rsqrt instruction (a fast reciprocal square root estimate) was grouped in the same scheduling IIC_SSE_SQRT* class as the accurate (but very slow) SSE sqrt instruction. For code which uses rsqrt (possibly with newton-raphson iterations) this poor scheduling was affecting performances. This patch splits off the rsqrt instruction from the sqrt instruction scheduling classes and creates new IIC_SSE_RSQER* classes with latency values based on Agner's table. Differential Revision: http://reviews.llvm.org/D5370 Patch by Simon Pilgrim. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@218517 91177308-0d34-0410-b5e6-96231b3b80d8
luqmana · Sep 26, 2014 · a5ab9ba · a5ab9ba
1 parent a0d5d7a
commit a5ab9ba
Show file tree

Hide file tree

Showing 7 changed files with 39 additions and 15 deletions.
diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
@@ -3344,6 +3344,16 @@ def SSE_SQRTSD : OpndItins<
 >;
 }
 
+let Sched = WriteFRsqrt in {
+def SSE_RSQRTPS : OpndItins<
+  IIC_SSE_RSQRTPS_RR, IIC_SSE_RSQRTPS_RM
+>;
+
+def SSE_RSQRTSS : OpndItins<
+  IIC_SSE_RSQRTSS_RR, IIC_SSE_RSQRTSS_RM
+>;
+}
+
 let Sched = WriteFRcp in {
 def SSE_RCPP : OpndItins<
   IIC_SSE_RCPP_RR, IIC_SSE_RCPP_RM
@@ -3622,10 +3632,10 @@ defm SQRT  : sse1_fp_unop_s<0x51, "sqrt",  fsqrt, int_x86_sse_sqrt_ss,
 
 // Reciprocal approximations. Note that these typically require refinement
 // in order to obtain suitable precision.
-defm RSQRT : sse1_fp_unop_rw<0x52, "rsqrt", X86frsqrt, SSE_SQRTSS>,
-             sse1_fp_unop_p<0x52, "rsqrt", X86frsqrt, SSE_SQRTPS>,
+defm RSQRT : sse1_fp_unop_rw<0x52, "rsqrt", X86frsqrt, SSE_RSQRTSS>,
+             sse1_fp_unop_p<0x52, "rsqrt", X86frsqrt, SSE_RSQRTPS>,
              sse1_fp_unop_p_int<0x52, "rsqrt", int_x86_sse_rsqrt_ps,
-                                int_x86_avx_rsqrt_ps_256, SSE_SQRTPS>;
+                                int_x86_avx_rsqrt_ps_256, SSE_RSQRTPS>;
 defm RCP   : sse1_fp_unop_rw<0x53, "rcp", X86frcp, SSE_RCPS>,
              sse1_fp_unop_p<0x53, "rcp", X86frcp, SSE_RCPP>,
              sse1_fp_unop_p_int<0x53, "rcp", int_x86_sse_rcp_ps,

diff --git a/lib/Target/X86/X86SchedHaswell.td b/lib/Target/X86/X86SchedHaswell.td
@@ -129,6 +129,7 @@ defm : HWWriteResPair<WriteFAdd,   HWPort1, 3>;
 defm : HWWriteResPair<WriteFMul,   HWPort0, 5>;
 defm : HWWriteResPair<WriteFDiv,   HWPort0, 12>; // 10-14 cycles.
 defm : HWWriteResPair<WriteFRcp,   HWPort0, 5>;
+defm : HWWriteResPair<WriteFRsqrt, HWPort0, 5>;
 defm : HWWriteResPair<WriteFSqrt,  HWPort0, 15>;
 defm : HWWriteResPair<WriteCvtF2I, HWPort1, 3>;
 defm : HWWriteResPair<WriteCvtI2F, HWPort1, 4>;

diff --git a/lib/Target/X86/X86SchedSandyBridge.td b/lib/Target/X86/X86SchedSandyBridge.td
@@ -117,6 +117,7 @@ defm : SBWriteResPair<WriteFAdd,   SBPort1, 3>;
 defm : SBWriteResPair<WriteFMul,   SBPort0, 5>;
 defm : SBWriteResPair<WriteFDiv,   SBPort0, 12>; // 10-14 cycles.
 defm : SBWriteResPair<WriteFRcp,   SBPort0, 5>;
+defm : SBWriteResPair<WriteFRsqrt, SBPort0, 5>;
 defm : SBWriteResPair<WriteFSqrt,  SBPort0, 15>;
 defm : SBWriteResPair<WriteCvtF2I, SBPort1, 3>;
 defm : SBWriteResPair<WriteCvtI2F, SBPort1, 4>;

diff --git a/lib/Target/X86/X86Schedule.td b/lib/Target/X86/X86Schedule.td
@@ -63,12 +63,13 @@ def WriteZero : SchedWrite;
 defm WriteJump : X86SchedWritePair;
 
 // Floating point. This covers both scalar and vector operations.
-defm WriteFAdd  : X86SchedWritePair; // Floating point add/sub/compare.
-defm WriteFMul  : X86SchedWritePair; // Floating point multiplication.
-defm WriteFDiv  : X86SchedWritePair; // Floating point division.
-defm WriteFSqrt : X86SchedWritePair; // Floating point square root.
-defm WriteFRcp  : X86SchedWritePair; // Floating point reciprocal.
-defm WriteFMA   : X86SchedWritePair; // Fused Multiply Add.
+defm WriteFAdd   : X86SchedWritePair; // Floating point add/sub/compare.
+defm WriteFMul   : X86SchedWritePair; // Floating point multiplication.
+defm WriteFDiv   : X86SchedWritePair; // Floating point division.
+defm WriteFSqrt  : X86SchedWritePair; // Floating point square root.
+defm WriteFRcp   : X86SchedWritePair; // Floating point reciprocal estimate.
+defm WriteFRsqrt : X86SchedWritePair; // Floating point reciprocal square root estimate.
+defm WriteFMA    : X86SchedWritePair; // Fused Multiply Add.
 defm WriteFShuffle  : X86SchedWritePair; // Floating point vector shuffles.
 defm WriteFBlend  : X86SchedWritePair; // Floating point vector blends.
 defm WriteFVarBlend  : X86SchedWritePair; // Fp vector variable blends.
@@ -314,6 +315,11 @@ def IIC_SSE_SQRTPD_RM : InstrItinClass;
 def IIC_SSE_SQRTSD_RR : InstrItinClass;
 def IIC_SSE_SQRTSD_RM : InstrItinClass;
 
+def IIC_SSE_RSQRTPS_RR : InstrItinClass;
+def IIC_SSE_RSQRTPS_RM : InstrItinClass;
+def IIC_SSE_RSQRTSS_RR : InstrItinClass;
+def IIC_SSE_RSQRTSS_RM : InstrItinClass;
+
 def IIC_SSE_RCPP_RR : InstrItinClass;
 def IIC_SSE_RCPP_RM : InstrItinClass;
 def IIC_SSE_RCPS_RR : InstrItinClass;

diff --git a/lib/Target/X86/X86ScheduleAtom.td b/lib/Target/X86/X86ScheduleAtom.td
@@ -224,6 +224,11 @@ def AtomItineraries : ProcessorItineraries<
   InstrItinData<IIC_SSE_SQRTSD_RR, [InstrStage<62, [Port0, Port1]>] >,
   InstrItinData<IIC_SSE_SQRTSD_RM, [InstrStage<62, [Port0, Port1]>] >,
 
+  InstrItinData<IIC_SSE_RSQRTPS_RR, [InstrStage<9, [Port0, Port1]>] >,
+  InstrItinData<IIC_SSE_RSQRTPS_RM, [InstrStage<10, [Port0, Port1]>] >,
+  InstrItinData<IIC_SSE_RSQRTSS_RR, [InstrStage<4, [Port0]>] >,
+  InstrItinData<IIC_SSE_RSQRTSS_RM, [InstrStage<4, [Port0]>] >,
+
   InstrItinData<IIC_SSE_RCPP_RR, [InstrStage<9, [Port0, Port1]>] >,
   InstrItinData<IIC_SSE_RCPP_RM, [InstrStage<10, [Port0, Port1]>] >,
   InstrItinData<IIC_SSE_RCPS_RR, [InstrStage<4, [Port0]>] >,

diff --git a/lib/Target/X86/X86ScheduleBtVer2.td b/lib/Target/X86/X86ScheduleBtVer2.td
@@ -163,15 +163,15 @@ defm : JWriteResIntPair<WriteJump,  JALU01, 1>;
 // FIXME: should we bother splitting JFPU pipe + unit stages for fast instructions?
 // FIXME: Double precision latencies
 // FIXME: SS vs PS latencies
-// FIXME: RSQRT latencies
 // FIXME: ymm latencies
 ////////////////////////////////////////////////////////////////////////////////
 
-defm : JWriteResFpuPair<WriteFAdd,      JFPU0,  3>;
-defm : JWriteResFpuPair<WriteFMul,      JFPU1,  2>;
-defm : JWriteResFpuPair<WriteFRcp,      JFPU1,  2>;
-defm : JWriteResFpuPair<WriteFShuffle, JFPU01,  1>;
-defm : JWriteResFpuPair<WriteFBlend,   JFPU01,  1>;
+defm : JWriteResFpuPair<WriteFAdd,        JFPU0,  3>;
+defm : JWriteResFpuPair<WriteFMul,        JFPU1,  2>;
+defm : JWriteResFpuPair<WriteFRcp,        JFPU1,  2>;
+defm : JWriteResFpuPair<WriteFRsqrt,      JFPU1,  2>;
+defm : JWriteResFpuPair<WriteFShuffle,   JFPU01,  1>;
+defm : JWriteResFpuPair<WriteFBlend,     JFPU01,  1>;
 defm : JWriteResFpuPair<WriteFShuffle256, JFPU01, 1>;
 
 def : WriteRes<WriteFSqrt, [JFPU1, JLAGU, JFPM]> {

diff --git a/lib/Target/X86/X86ScheduleSLM.td b/lib/Target/X86/X86ScheduleSLM.td
@@ -101,6 +101,7 @@ def : WriteRes<WriteIDivLd, [MEC_RSV, IEC_RSV01, SMDivider]> {
 // Scalar and vector floating point.
 defm : SMWriteResPair<WriteFAdd,   FPC_RSV1, 3>;
 defm : SMWriteResPair<WriteFRcp,   FPC_RSV0, 5>;
+defm : SMWriteResPair<WriteFRsqrt, FPC_RSV0, 5>;
 defm : SMWriteResPair<WriteFSqrt,  FPC_RSV0, 15>;
 defm : SMWriteResPair<WriteCvtF2I, FPC_RSV01, 4>;
 defm : SMWriteResPair<WriteCvtI2F, FPC_RSV01, 4>;