AMDGPU: Constant fold rcp node

[arsenm]

Summary

When doing arcp optimization with a constant denominator,
this was leaving behind rcps with constant inputs.

[arsenm]

ping

[arsenm]

ping

[arsenm]

ping

[kzhuravl]

LGTM.

[arsenm]

r297248

lib/Target/AMDGPU/AMDGPUISelLowering.cpp

case AMDGPUISD::BFE_U32: {
     }
 
     break;
   }
   case ISD::LOAD:
     return performLoadCombine(N, DCI);
   case ISD::STORE:
     return performStoreCombine(N, DCI);
+  case AMDGPUISD::RCP: {
+    if (const auto *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0))) {
+      // XXX - Should this flush denormals?
+      const APFloat &Val = CFP->getValueAPF();
+      APFloat One(Val.getSemantics(), "1.0");
+      return DAG.getConstantFP(One / Val, SDLoc(N), N->getValueType(0));
+    }
+
+    break;
+  }
   }
   return SDValue();
 }
 
 //===----------------------------------------------------------------------===//
 // Helper functions
 //===----------------------------------------------------------------------===//
 

test/CodeGen/AMDGPU/fdiv.f16.ll

entry:
   %gep.r = getelementptr inbounds half, half addrspace(1)* %r, i64 %tid.ext
   %a.val = load volatile half, half addrspace(1)* %gep.a
   %b.val = load volatile half, half addrspace(1)* %gep.b
   %r.val = fdiv half %a.val, %b.val
   store half %r.val, half addrspace(1)* %gep.r
   ret void
 }
 
+; FUNC-LABEL: {{^}}div_arcp_2_x_pat_f16:
+; SI: v_mul_f32_e32 v{{[0-9]+}}, 0.5, v{{[0-9]+}}
+
+; VI: v_mul_f16_e32 [[MUL:v[0-9]+]], 0.5, v{{[0-9]+}}
+; VI: buffer_store_short [[MUL]]
+define void @div_arcp_2_x_pat_f16(half addrspace(1)* %out) #0 {
+  %x = load half, half addrspace(1)* undef
+  %rcp = fdiv arcp half %x, 2.0
+  store half %rcp, half addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: {{^}}div_arcp_k_x_pat_f16:
+; SI: v_mul_f32_e32 v{{[0-9]+}}, 0x3dcccccd, v{{[0-9]+}}
+
+; VI: v_mul_f16_e32 [[MUL:v[0-9]+]], 0x2e66, v{{[0-9]+}}
+; VI: buffer_store_short [[MUL]]
+define void @div_arcp_k_x_pat_f16(half addrspace(1)* %out) #0 {
+  %x = load half, half addrspace(1)* undef
+  %rcp = fdiv arcp half %x, 10.0
+  store half %rcp, half addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: {{^}}div_arcp_neg_k_x_pat_f16:
+; SI: v_mul_f32_e32 v{{[0-9]+}}, 0xbdcccccd, v{{[0-9]+}}
+
+; VI: v_mul_f16_e32 [[MUL:v[0-9]+]], 0xae66, v{{[0-9]+}}
+; VI: buffer_store_short [[MUL]]
+define void @div_arcp_neg_k_x_pat_f16(half addrspace(1)* %out) #0 {
+  %x = load half, half addrspace(1)* undef
+  %rcp = fdiv arcp half %x, -10.0
+  store half %rcp, half addrspace(1)* %out, align 4
+  ret void
+}
+
 declare i32 @llvm.amdgcn.workitem.id.x() #1
 declare half @llvm.sqrt.f16(half) #1
 declare half @llvm.fabs.f16(half) #1
 
 attributes #0 = { nounwind }
 attributes #1 = { nounwind readnone }
 attributes #2 = { nounwind "unsafe-fp-math"="true" }

test/CodeGen/AMDGPU/fdiv.f64.ll

-; RUN: llc -march=amdgcn -mcpu=hawaii -verify-machineinstrs < %s | FileCheck -check-prefix=CI -check-prefix=COMMON %s
-; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=COMMON %s
-; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=CI -check-prefix=COMMON %s
+; RUN: llc -march=amdgcn -mcpu=hawaii -verify-machineinstrs < %s | FileCheck -check-prefix=CI -check-prefix=GCN %s
+; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=GCN %s
+; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=CI -check-prefix=GCN %s
 
 
-; COMMON-LABEL: {{^}}fdiv_f64:
-; COMMON-DAG: buffer_load_dwordx2 [[NUM:v\[[0-9]+:[0-9]+\]]], off, {{s\[[0-9]+:[0-9]+\]}}, 0
-; COMMON-DAG: buffer_load_dwordx2 [[DEN:v\[[0-9]+:[0-9]+\]]], off, {{s\[[0-9]+:[0-9]+\]}}, 0 offset:8
+; GCN-LABEL: {{^}}fdiv_f64:
+; GCN-DAG: buffer_load_dwordx2 [[NUM:v\[[0-9]+:[0-9]+\]]], off, {{s\[[0-9]+:[0-9]+\]}}, 0
+; GCN-DAG: buffer_load_dwordx2 [[DEN:v\[[0-9]+:[0-9]+\]]], off, {{s\[[0-9]+:[0-9]+\]}}, 0 offset:8
 ; CI-DAG: v_div_scale_f64 [[SCALE0:v\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, [[DEN]], [[DEN]], [[NUM]]
 ; CI-DAG: v_div_scale_f64 [[SCALE1:v\[[0-9]+:[0-9]+\]]], vcc, [[NUM]], [[DEN]], [[NUM]]
 
 ; Check for div_scale bug workaround on SI
 ; SI-DAG: v_div_scale_f64 [[SCALE0:v\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, [[DEN]], [[DEN]], [[NUM]]
 ; SI-DAG: v_div_scale_f64 [[SCALE1:v\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, [[NUM]], [[DEN]], [[NUM]]
 
-; COMMON-DAG: v_rcp_f64_e32 [[RCP_SCALE0:v\[[0-9]+:[0-9]+\]]], [[SCALE0]]
+; GCN-DAG: v_rcp_f64_e32 [[RCP_SCALE0:v\[[0-9]+:[0-9]+\]]], [[SCALE0]]
 
 ; SI-DAG: v_cmp_eq_u32_e32 vcc, {{v[0-9]+}}, {{v[0-9]+}}
 ; SI-DAG: v_cmp_eq_u32_e64 [[CMP0:s\[[0-9]+:[0-9]+\]]], {{v[0-9]+}}, {{v[0-9]+}}
 ; SI-DAG: s_xor_b64 vcc, [[CMP0]], vcc
 
-; COMMON-DAG: v_fma_f64 [[FMA0:v\[[0-9]+:[0-9]+\]]], -[[SCALE0]], [[RCP_SCALE0]], 1.0
-; COMMON-DAG: v_fma_f64 [[FMA1:v\[[0-9]+:[0-9]+\]]], [[RCP_SCALE0]], [[FMA0]], [[RCP_SCALE0]]
-; COMMON-DAG: v_fma_f64 [[FMA2:v\[[0-9]+:[0-9]+\]]], -[[SCALE0]], [[FMA1]], 1.0
-; COMMON-DAG: v_fma_f64 [[FMA3:v\[[0-9]+:[0-9]+\]]], [[FMA1]], [[FMA2]], [[FMA1]]
-; COMMON-DAG: v_mul_f64 [[MUL:v\[[0-9]+:[0-9]+\]]], [[SCALE1]], [[FMA3]]
-; COMMON-DAG: v_fma_f64 [[FMA4:v\[[0-9]+:[0-9]+\]]], -[[SCALE0]], [[MUL]], [[SCALE1]]
-; COMMON: v_div_fmas_f64 [[FMAS:v\[[0-9]+:[0-9]+\]]], [[FMA4]], [[FMA3]], [[MUL]]
-; COMMON: v_div_fixup_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[FMAS]], [[DEN]], [[NUM]]
-; COMMON: buffer_store_dwordx2 [[RESULT]]
-; COMMON: s_endpgm
-define void @fdiv_f64(double addrspace(1)* %out, double addrspace(1)* %in) nounwind {
+; GCN-DAG: v_fma_f64 [[FMA0:v\[[0-9]+:[0-9]+\]]], -[[SCALE0]], [[RCP_SCALE0]], 1.0
+; GCN-DAG: v_fma_f64 [[FMA1:v\[[0-9]+:[0-9]+\]]], [[RCP_SCALE0]], [[FMA0]], [[RCP_SCALE0]]
+; GCN-DAG: v_fma_f64 [[FMA2:v\[[0-9]+:[0-9]+\]]], -[[SCALE0]], [[FMA1]], 1.0
+; GCN-DAG: v_fma_f64 [[FMA3:v\[[0-9]+:[0-9]+\]]], [[FMA1]], [[FMA2]], [[FMA1]]
+; GCN-DAG: v_mul_f64 [[MUL:v\[[0-9]+:[0-9]+\]]], [[SCALE1]], [[FMA3]]
+; GCN-DAG: v_fma_f64 [[FMA4:v\[[0-9]+:[0-9]+\]]], -[[SCALE0]], [[MUL]], [[SCALE1]]
+; GCN: v_div_fmas_f64 [[FMAS:v\[[0-9]+:[0-9]+\]]], [[FMA4]], [[FMA3]], [[MUL]]
+; GCN: v_div_fixup_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[FMAS]], [[DEN]], [[NUM]]
+; GCN: buffer_store_dwordx2 [[RESULT]]
+; GCN: s_endpgm
+define void @fdiv_f64(double addrspace(1)* %out, double addrspace(1)* %in) #0 {
   %gep.1 = getelementptr double, double addrspace(1)* %in, i32 1
   %num = load volatile double, double addrspace(1)* %in
   %den = load volatile double, double addrspace(1)* %gep.1
   %result = fdiv double %num, %den
   store double %result, double addrspace(1)* %out
   ret void
 }
 
-; COMMON-LABEL: {{^}}fdiv_f64_s_v:
-define void @fdiv_f64_s_v(double addrspace(1)* %out, double addrspace(1)* %in, double %num) nounwind {
+; GCN-LABEL: {{^}}fdiv_f64_s_v:
+define void @fdiv_f64_s_v(double addrspace(1)* %out, double addrspace(1)* %in, double %num) #0 {
   %den = load double, double addrspace(1)* %in
   %result = fdiv double %num, %den
   store double %result, double addrspace(1)* %out
   ret void
 }
 
-; COMMON-LABEL: {{^}}fdiv_f64_v_s:
-define void @fdiv_f64_v_s(double addrspace(1)* %out, double addrspace(1)* %in, double %den) nounwind {
+; GCN-LABEL: {{^}}fdiv_f64_v_s:
+define void @fdiv_f64_v_s(double addrspace(1)* %out, double addrspace(1)* %in, double %den) #0 {
   %num = load double, double addrspace(1)* %in
   %result = fdiv double %num, %den
   store double %result, double addrspace(1)* %out
   ret void
 }
 
-; COMMON-LABEL: {{^}}fdiv_f64_s_s:
-define void @fdiv_f64_s_s(double addrspace(1)* %out, double %num, double %den) nounwind {
+; GCN-LABEL: {{^}}fdiv_f64_s_s:
+define void @fdiv_f64_s_s(double addrspace(1)* %out, double %num, double %den) #0 {
   %result = fdiv double %num, %den
   store double %result, double addrspace(1)* %out
   ret void
 }
 
-; COMMON-LABEL: {{^}}v_fdiv_v2f64:
-define void @v_fdiv_v2f64(<2 x double> addrspace(1)* %out, <2 x double> addrspace(1)* %in) nounwind {
+; GCN-LABEL: {{^}}v_fdiv_v2f64:
+define void @v_fdiv_v2f64(<2 x double> addrspace(1)* %out, <2 x double> addrspace(1)* %in) #0 {
   %gep.1 = getelementptr <2 x double>, <2 x double> addrspace(1)* %in, i32 1
   %num = load <2 x double>, <2 x double> addrspace(1)* %in
   %den = load <2 x double>, <2 x double> addrspace(1)* %gep.1
   %result = fdiv <2 x double> %num, %den
   store <2 x double> %result, <2 x double> addrspace(1)* %out
   ret void
 }
 
-; COMMON-LABEL: {{^}}s_fdiv_v2f64:
+; GCN-LABEL: {{^}}s_fdiv_v2f64:
 define void @s_fdiv_v2f64(<2 x double> addrspace(1)* %out, <2 x double> %num, <2 x double> %den) {
   %result = fdiv <2 x double> %num, %den
   store <2 x double> %result, <2 x double> addrspace(1)* %out
   ret void
 }
 
-; COMMON-LABEL: {{^}}v_fdiv_v4f64:
-define void @v_fdiv_v4f64(<4 x double> addrspace(1)* %out, <4 x double> addrspace(1)* %in) nounwind {
+; GCN-LABEL: {{^}}v_fdiv_v4f64:
+define void @v_fdiv_v4f64(<4 x double> addrspace(1)* %out, <4 x double> addrspace(1)* %in) #0 {
   %gep.1 = getelementptr <4 x double>, <4 x double> addrspace(1)* %in, i32 1
   %num = load <4 x double>, <4 x double> addrspace(1)* %in
   %den = load <4 x double>, <4 x double> addrspace(1)* %gep.1
   %result = fdiv <4 x double> %num, %den
   store <4 x double> %result, <4 x double> addrspace(1)* %out
   ret void
 }
 
-; COMMON-LABEL: {{^}}s_fdiv_v4f64:
-define void @s_fdiv_v4f64(<4 x double> addrspace(1)* %out, <4 x double> %num, <4 x double> %den) {
+; GCN-LABEL: {{^}}s_fdiv_v4f64:
+define void @s_fdiv_v4f64(<4 x double> addrspace(1)* %out, <4 x double> %num, <4 x double> %den) #0 {
   %result = fdiv <4 x double> %num, %den
   store <4 x double> %result, <4 x double> addrspace(1)* %out
   ret void
 }
+
+; GCN-LABEL: {{^}}div_fast_2_x_pat_f64:
+; GCN: v_mul_f64 [[MUL:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, 0.5
+; GCN: buffer_store_dwordx2 [[MUL]]
+define void @div_fast_2_x_pat_f64(double addrspace(1)* %out) #1 {
+  %x = load double, double addrspace(1)* undef
+  %rcp = fdiv fast double %x, 2.0
+  store double %rcp, double addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}div_fast_k_x_pat_f64:
+; GCN-DAG: s_mov_b32 s[[K_LO:[0-9]+]], 0x9999999a
+; GCN-DAG: s_mov_b32 s[[K_HI:[0-9]+]], 0x3fb99999
+; GCN: v_mul_f64 [[MUL:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[}}[[K_LO]]:[[K_HI]]{{\]}}
+; GCN: buffer_store_dwordx2 [[MUL]]
+define void @div_fast_k_x_pat_f64(double addrspace(1)* %out) #1 {
+  %x = load double, double addrspace(1)* undef
+  %rcp = fdiv fast double %x, 10.0
+  store double %rcp, double addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}div_fast_neg_k_x_pat_f64:
+; GCN-DAG: s_mov_b32 s[[K_LO:[0-9]+]], 0x9999999a
+; GCN-DAG: s_mov_b32 s[[K_HI:[0-9]+]], 0xbfb99999
+; GCN: v_mul_f64 [[MUL:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[}}[[K_LO]]:[[K_HI]]{{\]}}
+; GCN: buffer_store_dwordx2 [[MUL]]
+define void @div_fast_neg_k_x_pat_f64(double addrspace(1)* %out) #1 {
+  %x = load double, double addrspace(1)* undef
+  %rcp = fdiv fast double %x, -10.0
+  store double %rcp, double addrspace(1)* %out, align 4
+  ret void
+}
+
+attributes #0 = { nounwind }
+attributes #1 = { nounwind "unsafe-fp-math"="true" }

test/CodeGen/AMDGPU/llvm.amdgcn.rcp.ll

 ; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
 
 declare float @llvm.amdgcn.rcp.f32(float) #0
 declare double @llvm.amdgcn.rcp.f64(double) #0
 
 declare double @llvm.sqrt.f64(double) #0
 declare float @llvm.sqrt.f32(float) #0
 
 ; FUNC-LABEL: {{^}}rcp_undef_f32:
 ; SI-NOT: v_rcp_f32
 define void @rcp_undef_f32(float addrspace(1)* %out) #1 {
   %rcp = call float @llvm.amdgcn.rcp.f32(float undef)
   store float %rcp, float addrspace(1)* %out, align 4
   ret void
 }
 
+; FUNC-LABEL: {{^}}rcp_2_f32:
+; SI-NOT: v_rcp_f32
+; SI: v_mov_b32_e32 v{{[0-9]+}}, 0.5
+define void @rcp_2_f32(float addrspace(1)* %out) #1 {
+  %rcp = call float @llvm.amdgcn.rcp.f32(float 2.0)
+  store float %rcp, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: {{^}}rcp_10_f32:
+; SI-NOT: v_rcp_f32
+; SI: v_mov_b32_e32 v{{[0-9]+}}, 0x3dcccccd
+define void @rcp_10_f32(float addrspace(1)* %out) #1 {
+  %rcp = call float @llvm.amdgcn.rcp.f32(float 10.0)
+  store float %rcp, float addrspace(1)* %out, align 4
+  ret void
+}
+
 ; FUNC-LABEL: {{^}}safe_no_fp32_denormals_rcp_f32:
 ; SI: v_rcp_f32_e32 [[RESULT:v[0-9]+]], s{{[0-9]+}}
 ; SI-NOT: [[RESULT]]
 ; SI: buffer_store_dword [[RESULT]]
 define void @safe_no_fp32_denormals_rcp_f32(float addrspace(1)* %out, float %src) #1 {
   %rcp = fdiv float 1.0, %src
   store float %rcp, float addrspace(1)* %out, align 4
   ret void

test/CodeGen/AMDGPU/rcp-pattern.ll

define void @rcp_fabs_fneg_pat_multi_use_f32(float addrspace(1)* %out, float %src) #0 {
   %rcp = fdiv float 1.0, %src.fabs.fneg
   store volatile float %rcp, float addrspace(1)* %out, align 4
 
   %other = fmul float %src, %src.fabs.fneg
   store volatile float %other, float addrspace(1)* %out, align 4
   ret void
 }
 
+; FUNC-LABEL: {{^}}div_arcp_2_x_pat_f32:
+; GCN: v_mul_f32_e32 [[MUL:v[0-9]+]], 0.5, v{{[0-9]+}}
+; GCN: buffer_store_dword [[MUL]]
+define void @div_arcp_2_x_pat_f32(float addrspace(1)* %out) #0 {
+  %x = load float, float addrspace(1)* undef
+  %rcp = fdiv arcp float %x, 2.0
+  store float %rcp, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: {{^}}div_arcp_k_x_pat_f32:
+; GCN: v_mul_f32_e32 [[MUL:v[0-9]+]], 0x3dcccccd, v{{[0-9]+}}
+; GCN: buffer_store_dword [[MUL]]
+define void @div_arcp_k_x_pat_f32(float addrspace(1)* %out) #0 {
+  %x = load float, float addrspace(1)* undef
+  %rcp = fdiv arcp float %x, 10.0
+  store float %rcp, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: {{^}}div_arcp_neg_k_x_pat_f32:
+; GCN: v_mul_f32_e32 [[MUL:v[0-9]+]], 0xbdcccccd, v{{[0-9]+}}
+; GCN: buffer_store_dword [[MUL]]
+define void @div_arcp_neg_k_x_pat_f32(float addrspace(1)* %out) #0 {
+  %x = load float, float addrspace(1)* undef
+  %rcp = fdiv arcp float %x, -10.0
+  store float %rcp, float addrspace(1)* %out, align 4
+  ret void
+}
 
 declare float @llvm.fabs.f32(float) #1
 declare float @llvm.sqrt.f32(float) #1
 
 attributes #0 = { nounwind "unsafe-fp-math"="false" }
 attributes #1 = { nounwind readnone }
 attributes #2 = { nounwind "unsafe-fp-math"="true" }
 
 !0 = !{float 2.500000e+00}