From: Matt Arsenault <Matthew.Arsenault@amd.com>
Date: Sat, 22 Aug 2015 00:50:41 +0000 (+0000)
Subject: AMDGPU: Improve accuracy of instruction rates for some FP instructions
X-Git-Url: http://demsky.eecs.uci.edu/git/?a=commitdiff_plain;h=e48caeb48f4af56d1bb401951fb9d38496186050;p=oota-llvm.git

AMDGPU: Improve accuracy of instruction rates for some FP instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@245774 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/Target/AMDGPU/CIInstructions.td b/lib/Target/AMDGPU/CIInstructions.td
index fd569097ae8..9ec6fd12499 100644
--- a/lib/Target/AMDGPU/CIInstructions.td
+++ b/lib/Target/AMDGPU/CIInstructions.td
@@ -41,6 +41,7 @@ def HasFlatAddressSpace : Predicate<"Subtarget->hasFlatAddressSpace()">;
 
 let SubtargetPredicate = isCIVI in {
 
+let SchedRW = [WriteDoubleAdd] in {
 defm V_TRUNC_F64 : VOP1Inst <vop1<0x17>, "v_trunc_f64",
   VOP_F64_F64, ftrunc
 >;
@@ -53,12 +54,16 @@ defm V_FLOOR_F64 : VOP1Inst <vop1<0x1A>, "v_floor_f64",
 defm V_RNDNE_F64 : VOP1Inst <vop1<0x19>, "v_rndne_f64",
   VOP_F64_F64, frint
 >;
+} // End SchedRW = [WriteDoubleAdd]
+
+let SchedRW = [WriteQuarterRate32] in {
 defm V_LOG_LEGACY_F32 : VOP1Inst <vop1<0x45, 0x4c>, "v_log_legacy_f32",
   VOP_F32_F32
 >;
 defm V_EXP_LEGACY_F32 : VOP1Inst <vop1<0x46, 0x4b>, "v_exp_legacy_f32",
   VOP_F32_F32
 >;
+} // End SchedRW = [WriteQuarterRate32]
 
 //===----------------------------------------------------------------------===//
 // VOP3 Instructions
diff --git a/lib/Target/AMDGPU/SIInstructions.td b/lib/Target/AMDGPU/SIInstructions.td
index 8ad9d53831c..833a19f2d43 100644
--- a/lib/Target/AMDGPU/SIInstructions.td
+++ b/lib/Target/AMDGPU/SIInstructions.td
@@ -1286,7 +1286,9 @@ defm V_SQRT_F64 : VOP1Inst <vop1<0x34, 0x28>, "v_sqrt_f64",
   VOP_F64_F64, fsqrt
 >;
 
-} // let SchedRW = [WriteDouble]
+} // End SchedRW = [WriteDouble]
+
+let SchedRW = [WriteQuarterRate32] in {
 
 defm V_SIN_F32 : VOP1Inst <vop1<0x35, 0x29>, "v_sin_f32",
   VOP_F32_F32, AMDGPUsin
@@ -1294,6 +1296,9 @@ defm V_SIN_F32 : VOP1Inst <vop1<0x35, 0x29>, "v_sin_f32",
 defm V_COS_F32 : VOP1Inst <vop1<0x36, 0x2a>, "v_cos_f32",
   VOP_F32_F32, AMDGPUcos
 >;
+
+} // End SchedRW = [WriteQuarterRate32]
+
 defm V_NOT_B32 : VOP1Inst <vop1<0x37, 0x2b>, "v_not_b32", VOP_I32_I32>;
 defm V_BFREV_B32 : VOP1Inst <vop1<0x38, 0x2c>, "v_bfrev_b32", VOP_I32_I32>;
 defm V_FFBH_U32 : VOP1Inst <vop1<0x39, 0x2d>, "v_ffbh_u32", VOP_I32_I32>;
@@ -1302,10 +1307,18 @@ defm V_FFBH_I32 : VOP1Inst <vop1<0x3b, 0x2f>, "v_ffbh_i32", VOP_I32_I32>;
 defm V_FREXP_EXP_I32_F64 : VOP1Inst <vop1<0x3c,0x30>, "v_frexp_exp_i32_f64",
   VOP_I32_F64
 >;
+
+let SchedRW = [WriteDoubleAdd] in {
 defm V_FREXP_MANT_F64 : VOP1Inst <vop1<0x3d, 0x31>, "v_frexp_mant_f64",
   VOP_F64_F64
 >;
-defm V_FRACT_F64 : VOP1Inst <vop1<0x3e, 0x32>, "v_fract_f64", VOP_F64_F64>;
+
+defm V_FRACT_F64 : VOP1Inst <vop1<0x3e, 0x32>, "v_fract_f64",
+  VOP_F64_F64
+>;
+} // End SchedRW = [WriteDoubleAdd]
+
+
 defm V_FREXP_EXP_I32_F32 : VOP1Inst <vop1<0x3f, 0x33>, "v_frexp_exp_i32_f32",
   VOP_I32_F32
 >;
@@ -1337,7 +1350,7 @@ defm V_RSQ_LEGACY_F32 : VOP1InstSI <vop1<0x2d>, "v_rsq_legacy_f32",
   VOP_F32_F32, AMDGPUrsq_legacy
 >;
 
-} // End let SchedRW = [WriteQuarterRate32]
+} // End SchedRW = [WriteQuarterRate32]
 
 let SchedRW = [WriteDouble] in {
 
@@ -1698,15 +1711,15 @@ defm V_DIV_FIXUP_F32 : VOP3Inst <
   vop3<0x15f, 0x1de>, "v_div_fixup_f32", VOP_F32_F32_F32_F32, AMDGPUdiv_fixup
 >;
 
-let SchedRW = [WriteDouble] in {
+let SchedRW = [WriteDoubleAdd] in {
 
 defm V_DIV_FIXUP_F64 : VOP3Inst <
   vop3<0x160, 0x1df>, "v_div_fixup_f64", VOP_F64_F64_F64_F64, AMDGPUdiv_fixup
 >;
 
-} // let SchedRW = [WriteDouble]
+} // End SchedRW = [WriteDouble]
 
-let SchedRW = [WriteDouble] in {
+let SchedRW = [WriteDoubleAdd] in {
 let isCommutable = 1 in {
 
 defm V_ADD_F64 : VOP3Inst <vop3<0x164, 0x280>, "v_add_f64",
@@ -1729,7 +1742,7 @@ defm V_LDEXP_F64 : VOP3Inst <vop3<0x168, 0x284>, "v_ldexp_f64",
   VOP_F64_F64_I32, AMDGPUldexp
 >;
 
-} // let SchedRW = [WriteDouble]
+} // let SchedRW = [WriteDoubleAdd]
 
 let isCommutable = 1, SchedRW = [WriteQuarterRate32] in {
 
@@ -1760,6 +1773,7 @@ defm V_DIV_SCALE_F64 : VOP3b_64 <vop3<0x16e, 0x1e1>, "v_div_scale_f64", []>;
 
 let isCommutable = 1, Uses = [VCC] in {
 
+let SchedRW = [WriteFloatFMA] in {
 // v_div_fmas_f32:
 //   result = src0 * src1 + src2
 //   if (vcc)
@@ -1768,6 +1782,7 @@ let isCommutable = 1, Uses = [VCC] in {
 defm V_DIV_FMAS_F32 : VOP3_VCC_Inst <vop3<0x16f, 0x1e2>, "v_div_fmas_f32",
   VOP_F32_F32_F32_F32, AMDGPUdiv_fmas
 >;
+}
 
 let SchedRW = [WriteDouble] in {
 // v_div_fmas_f64:
diff --git a/test/CodeGen/AMDGPU/sint_to_fp.f64.ll b/test/CodeGen/AMDGPU/sint_to_fp.f64.ll
index 0db7cdc171b..a94ccc32e61 100644
--- a/test/CodeGen/AMDGPU/sint_to_fp.f64.ll
+++ b/test/CodeGen/AMDGPU/sint_to_fp.f64.ll
@@ -46,9 +46,9 @@ define void @s_sint_to_fp_i64_to_f64(double addrspace(1)* %out, i64 %in) {
 
 ; SI-LABEL: @v_sint_to_fp_i64_to_f64
 ; SI: buffer_load_dwordx2 v{{\[}}[[LO:[0-9]+]]:[[HI:[0-9]+]]{{\]}}
-; SI: v_cvt_f64_i32_e32 [[HI_CONV:v\[[0-9]+:[0-9]+\]]], v[[HI]]
-; SI: v_ldexp_f64 [[LDEXP:v\[[0-9]+:[0-9]+\]]], [[HI_CONV]], 32
-; SI: v_cvt_f64_u32_e32 [[LO_CONV:v\[[0-9]+:[0-9]+\]]], v[[LO]]
+; SI-DAG: v_cvt_f64_i32_e32 [[HI_CONV:v\[[0-9]+:[0-9]+\]]], v[[HI]]
+; SI-DAG: v_cvt_f64_u32_e32 [[LO_CONV:v\[[0-9]+:[0-9]+\]]], v[[LO]]
+; SI-DAG: v_ldexp_f64 [[LDEXP:v\[[0-9]+:[0-9]+\]]], [[HI_CONV]], 32
 ; SI: v_add_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[LDEXP]], [[LO_CONV]]
 ; SI: buffer_store_dwordx2 [[RESULT]]
 define void @v_sint_to_fp_i64_to_f64(double addrspace(1)* %out, i64 addrspace(1)* %in) {
diff --git a/test/CodeGen/AMDGPU/uint_to_fp.f64.ll b/test/CodeGen/AMDGPU/uint_to_fp.f64.ll
index 6f608df5e9f..65fe580792a 100644
--- a/test/CodeGen/AMDGPU/uint_to_fp.f64.ll
+++ b/test/CodeGen/AMDGPU/uint_to_fp.f64.ll
@@ -4,9 +4,9 @@ declare i32 @llvm.r600.read.tidig.x() nounwind readnone
 
 ; SI-LABEL: {{^}}v_uint_to_fp_i64_to_f64
 ; SI: buffer_load_dwordx2 v{{\[}}[[LO:[0-9]+]]:[[HI:[0-9]+]]{{\]}}
-; SI: v_cvt_f64_u32_e32 [[HI_CONV:v\[[0-9]+:[0-9]+\]]], v[[HI]]
-; SI: v_ldexp_f64 [[LDEXP:v\[[0-9]+:[0-9]+\]]], [[HI_CONV]], 32
-; SI: v_cvt_f64_u32_e32 [[LO_CONV:v\[[0-9]+:[0-9]+\]]], v[[LO]]
+; SI-DAG: v_cvt_f64_u32_e32 [[HI_CONV:v\[[0-9]+:[0-9]+\]]], v[[HI]]
+; SI-DAG: v_cvt_f64_u32_e32 [[LO_CONV:v\[[0-9]+:[0-9]+\]]], v[[LO]]
+; SI-DAG: v_ldexp_f64 [[LDEXP:v\[[0-9]+:[0-9]+\]]], [[HI_CONV]], 32
 ; SI: v_add_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[LDEXP]], [[LO_CONV]]
 ; SI: buffer_store_dwordx2 [[RESULT]]
 define void @v_uint_to_fp_i64_to_f64(double addrspace(1)* %out, i64 addrspace(1)* %in) {