3996311ee1
According to IR LangRef, the FMF flag: contract Allow floating-point contraction (e.g. fusing a multiply followed by an addition into a fused multiply-and-add). reassoc Allow reassociation transformations for floating-point instructions. This may dramatically change results in floating-point. My understanding is that these two flags shouldn't imply each other, as we might have a SDNode that can be reassociated with others, but not contractble. eg: We may want following fmul/fad/fsub to freely reassoc, but don't want fma being generated here. %F = fmul reassoc double %A, %B ; <double> [#uses=1] %G = fmul reassoc double %C, %D ; <double> [#uses=1] %H = fadd reassoc double %F, %G ; <double> [#uses=1] %I = fsub reassoc double %H, %E ; <double> [#uses=1] Before https://reviews.llvm.org/D45710, `reassoc` flag actually did not imply isContratable either. The current implementation also only check the flag in fadd node, ignoring fmul node, this patch update that as well. Reviewed By: spatel, qiucf Differential Revision: https://reviews.llvm.org/D104247
142 lines
7.6 KiB
LLVM
142 lines
7.6 KiB
LLVM
; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx900 -denormal-fp-math=preserve-sign -fp-contract=on -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX9-FLUSH %s
|
|
; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx900 -denormal-fp-math=preserve-sign -fp-contract=on -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX9-FLUSH %s
|
|
; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx900 -denormal-fp-math=preserve-sign -fp-contract=fast -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX9-FLUSH %s
|
|
; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx900 -denormal-fp-math=preserve-sign -fp-contract=fast -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX9-FLUSH %s
|
|
|
|
; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx900 -denormal-fp-math=ieee -fp-contract=on -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX9-DENORM-STRICT,GFX9-DENORM %s
|
|
; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx900 -denormal-fp-math=ieee -fp-contract=on -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX9-DENORM-STRICT,GFX9-DENORM %s
|
|
; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx900 -denormal-fp-math=ieee -fp-contract=fast -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX9-DENORM-CONTRACT,GFX9-DENORM %s
|
|
; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx900 -denormal-fp-math=ieee -fp-contract=fast -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX9-DENORM-CONTRACT,GFX9-DENORM %s
|
|
|
|
declare i32 @llvm.amdgcn.workitem.id.x() #1
|
|
declare <2 x half> @llvm.fmuladd.v2f16(<2 x half>, <2 x half>, <2 x half>) #1
|
|
declare <2 x half> @llvm.fabs.v2f16(<2 x half>) #1
|
|
|
|
; GCN-LABEL: {{^}}fmuladd_v2f16:
|
|
; GFX9-FLUSH: v_pk_mul_f16 {{v[0-9]+, v[0-9]+, v[0-9]+}}
|
|
; GFX9-FLUSH: v_pk_add_f16 {{v[0-9]+, v[0-9]+, v[0-9]+}}
|
|
|
|
; GFX9-DENORM: v_pk_fma_f16 {{v[0-9]+, v[0-9]+, v[0-9]+}}
|
|
define amdgpu_kernel void @fmuladd_v2f16(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %in1,
|
|
<2 x half> addrspace(1)* %in2, <2 x half> addrspace(1)* %in3) #0 {
|
|
%r0 = load <2 x half>, <2 x half> addrspace(1)* %in1
|
|
%r1 = load <2 x half>, <2 x half> addrspace(1)* %in2
|
|
%r2 = load <2 x half>, <2 x half> addrspace(1)* %in3
|
|
%r3 = tail call <2 x half> @llvm.fmuladd.v2f16(<2 x half> %r0, <2 x half> %r1, <2 x half> %r2)
|
|
store <2 x half> %r3, <2 x half> addrspace(1)* %out
|
|
ret void
|
|
}
|
|
|
|
; GCN-LABEL: {{^}}fmul_fadd_v2f16:
|
|
; GFX9-DENORM-STRICT: v_pk_mul_f16 {{v[0-9]+, v[0-9]+, v[0-9]+}}
|
|
; GFX9-DENORM-STRICT: v_pk_add_f16 {{v[0-9]+, v[0-9]+, v[0-9]+}}
|
|
|
|
; GFX9-DENORM-CONTRACT: v_pk_fma_f16 {{v[0-9]+, v[0-9]+, v[0-9]+}}
|
|
define amdgpu_kernel void @fmul_fadd_v2f16(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %in1,
|
|
<2 x half> addrspace(1)* %in2, <2 x half> addrspace(1)* %in3) #0 {
|
|
%r0 = load <2 x half>, <2 x half> addrspace(1)* %in1
|
|
%r1 = load <2 x half>, <2 x half> addrspace(1)* %in2
|
|
%r2 = load <2 x half>, <2 x half> addrspace(1)* %in3
|
|
%r3 = fmul <2 x half> %r0, %r1
|
|
%r4 = fadd <2 x half> %r3, %r2
|
|
store <2 x half> %r4, <2 x half> addrspace(1)* %out
|
|
ret void
|
|
}
|
|
|
|
; GCN-LABEL: {{^}}fmul_fadd_contract_v2f16:
|
|
; GFX9-FLUSH: v_pk_mul_f16 {{v[0-9]+, v[0-9]+, v[0-9]+}}
|
|
; GFX9-FLUSH: v_pk_add_f16 {{v[0-9]+, v[0-9]+, v[0-9]+}}
|
|
|
|
; GFX9-DENORM: v_pk_fma_f16 {{v[0-9]+, v[0-9]+, v[0-9]+}}
|
|
define amdgpu_kernel void @fmul_fadd_contract_v2f16(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %in1,
|
|
<2 x half> addrspace(1)* %in2, <2 x half> addrspace(1)* %in3) #0 {
|
|
%r0 = load <2 x half>, <2 x half> addrspace(1)* %in1
|
|
%r1 = load <2 x half>, <2 x half> addrspace(1)* %in2
|
|
%r2 = load <2 x half>, <2 x half> addrspace(1)* %in3
|
|
%r3 = fmul contract <2 x half> %r0, %r1
|
|
%r4 = fadd contract <2 x half> %r3, %r2
|
|
store <2 x half> %r4, <2 x half> addrspace(1)* %out
|
|
ret void
|
|
}
|
|
|
|
|
|
; GCN-LABEL: {{^}}fmuladd_2.0_a_b_v2f16:
|
|
; GCN: {{buffer|flat|global}}_load_dword [[R1:v[0-9]+]],
|
|
; GCN: {{buffer|flat|global}}_load_dword [[R2:v[0-9]+]],
|
|
; GFX9-FLUSH: v_pk_add_f16 [[ADD0:v[0-9]+]], [[R1]], [[R1]]
|
|
; GFX9-FLUSH: v_pk_add_f16 [[RESULT:v[0-9]+]], [[ADD0]], [[R2]]
|
|
|
|
; GFX9-FLUSH: global_store_dword v{{[0-9]+}}, [[RESULT]], s{{\[[0-9]+:[0-9]+\]}}
|
|
|
|
; GFX9-DENORM: v_pk_fma_f16 [[RESULT:v[0-9]+]], [[R1]], 2.0, [[R2]]
|
|
; GFX9-DENORM: global_store_dword v{{[0-9]+}}, [[RESULT]], s{{\[[0-9]+:[0-9]+\]}}
|
|
define amdgpu_kernel void @fmuladd_2.0_a_b_v2f16(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %in) #0 {
|
|
%tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
%gep.0 = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid
|
|
%gep.1 = getelementptr <2 x half>, <2 x half> addrspace(1)* %gep.0, i32 1
|
|
%gep.out = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid
|
|
|
|
%r1 = load volatile <2 x half>, <2 x half> addrspace(1)* %gep.0
|
|
%r2 = load volatile <2 x half>, <2 x half> addrspace(1)* %gep.1
|
|
|
|
%r3 = tail call <2 x half> @llvm.fmuladd.v2f16(<2 x half> <half 2.0, half 2.0>, <2 x half> %r1, <2 x half> %r2)
|
|
store <2 x half> %r3, <2 x half> addrspace(1)* %gep.out
|
|
ret void
|
|
}
|
|
|
|
; GCN-LABEL: {{^}}fmuladd_a_2.0_b_v2f16:
|
|
; GCN: {{buffer|flat|global}}_load_dword [[R1:v[0-9]+]],
|
|
; GCN: {{buffer|flat|global}}_load_dword [[R2:v[0-9]+]],
|
|
; GFX9-FLUSH: v_pk_add_f16 [[ADD0:v[0-9]+]], [[R1]], [[R1]]
|
|
; GFX9-FLUSH: v_pk_add_f16 [[RESULT:v[0-9]+]], [[ADD0]], [[R2]]
|
|
|
|
; GFX9-FLUSH: global_store_dword v{{[0-9]+}}, [[RESULT]], s{{\[[0-9]+:[0-9]+\]}}
|
|
|
|
; GFX9-DENORM: v_pk_fma_f16 [[RESULT:v[0-9]+]], [[R1]], 2.0, [[R2]]
|
|
; GFX9-DENORM: global_store_dword v{{[0-9]+}}, [[RESULT]], s{{\[[0-9]+:[0-9]+\]}}
|
|
define amdgpu_kernel void @fmuladd_a_2.0_b_v2f16(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %in) #0 {
|
|
%tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
%gep.0 = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid
|
|
%gep.1 = getelementptr <2 x half>, <2 x half> addrspace(1)* %gep.0, i32 1
|
|
%gep.out = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid
|
|
|
|
%r1 = load volatile <2 x half>, <2 x half> addrspace(1)* %gep.0
|
|
%r2 = load volatile <2 x half>, <2 x half> addrspace(1)* %gep.1
|
|
|
|
%r3 = tail call <2 x half> @llvm.fmuladd.v2f16(<2 x half> %r1, <2 x half> <half 2.0, half 2.0>, <2 x half> %r2)
|
|
store <2 x half> %r3, <2 x half> addrspace(1)* %gep.out
|
|
ret void
|
|
}
|
|
|
|
; GCN-LABEL: {{^}}fadd_a_a_b_v2f16:
|
|
; GCN: {{buffer|flat|global}}_load_dword [[R1:v[0-9]+]],
|
|
; GCN: {{buffer|flat|global}}_load_dword [[R2:v[0-9]+]],
|
|
; GFX9-FLUSH: v_pk_add_f16 [[ADD0:v[0-9]+]], [[R1]], [[R1]]
|
|
; GFX9-FLUSH: v_pk_add_f16 [[RESULT:v[0-9]+]], [[ADD0]], [[R2]]
|
|
|
|
; GFX9-DENORM-STRICT: v_pk_add_f16 [[ADD0:v[0-9]+]], [[R1]], [[R1]]
|
|
; GFX9-DENORM-STRICT: v_pk_add_f16 [[RESULT:v[0-9]+]], [[ADD0]], [[R2]]
|
|
|
|
; GFX9-DENORM-CONTRACT: v_pk_fma_f16 [[RESULT:v[0-9]+]], [[R1]], 2.0, [[R2]]
|
|
|
|
; GCN: {{flat|global}}_store_dword v{{.+}}, [[RESULT]]
|
|
define amdgpu_kernel void @fadd_a_a_b_v2f16(<2 x half> addrspace(1)* %out,
|
|
<2 x half> addrspace(1)* %in1,
|
|
<2 x half> addrspace(1)* %in2) #0 {
|
|
%tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
%gep.0 = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid
|
|
%gep.1 = getelementptr <2 x half>, <2 x half> addrspace(1)* %gep.0, i32 1
|
|
%gep.out = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid
|
|
|
|
%r0 = load volatile <2 x half>, <2 x half> addrspace(1)* %gep.0
|
|
%r1 = load volatile <2 x half>, <2 x half> addrspace(1)* %gep.1
|
|
|
|
%add.0 = fadd <2 x half> %r0, %r0
|
|
%add.1 = fadd <2 x half> %add.0, %r1
|
|
store <2 x half> %add.1, <2 x half> addrspace(1)* %gep.out
|
|
ret void
|
|
}
|
|
|
|
attributes #0 = { nounwind }
|
|
attributes #1 = { nounwind readnone }
|