maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/llvm/test/CodeGen/AMDGPU/amdgcn-ieee.ll

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

189 lines
7.6 KiB
LLVM
Raw Normal View History

AMDGPU: Change pre-gfx9 implementation of fcanonicalize to mul If f32 denormals were enabled pre-gfx9, we would still try to implement this with v_max_f32. Pre-gfx9, these instructions ignored the denormal mode and did not flush. Switch to the multiply form for f32 as a workaround which should always work in any case. This fixes conformance failures when the library implementation of fmin/fmax were accidentally not inlined, forcing the assumption of no flushing on targets where denormals are not enabled by default. This is a workaround, since really we should not be mixing code with different FP mode expectations, but prefer the lowering that will work in any mode. Now this will always use max to implement canonicalize on gfx9+. This is only really beneficial for f64. For f32/f16 it's a neutral choice (and worse in terms of code size in 1 case), but possibly worse for the compiler since it does add an extra register use operand. Leave this change for later.
2020-04-23 05:34:10 +08:00
; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
AMDGPU: Remove dx10-clamp from subtarget features Since this can be set with s_setreg*, it should not be a subtarget property. Set a default based on the calling convention, and Introduce a new amdgpu-dx10-clamp attribute to override this if desired. Also introduce a new amdgpu-ieee attribute to match. The values need to match to allow inlining. I think it is OK for the caller's dx10-clamp attribute to override the callee, but there doesn't appear to be the infrastructure to do this currently without definining the attribute in the generic Attributes.td. Eventually the calling convention lowering will need to insert a mode switch somewhere for these. llvm-svn: 357302
2019-03-30 03:14:54 +08:00
; GCN-LABEL: {{^}}kernel_ieee_mode_default:
; GCN: {{buffer|global|flat}}_load_dword [[VAL0:v[0-9]+]]
; GCN-NEXT: {{buffer|global|flat}}_load_dword [[VAL1:v[0-9]+]]
AMDGPU: Change pre-gfx9 implementation of fcanonicalize to mul If f32 denormals were enabled pre-gfx9, we would still try to implement this with v_max_f32. Pre-gfx9, these instructions ignored the denormal mode and did not flush. Switch to the multiply form for f32 as a workaround which should always work in any case. This fixes conformance failures when the library implementation of fmin/fmax were accidentally not inlined, forcing the assumption of no flushing on targets where denormals are not enabled by default. This is a workaround, since really we should not be mixing code with different FP mode expectations, but prefer the lowering that will work in any mode. Now this will always use max to implement canonicalize on gfx9+. This is only really beneficial for f64. For f32/f16 it's a neutral choice (and worse in terms of code size in 1 case), but possibly worse for the compiler since it does add an extra register use operand. Leave this change for later.
2020-04-23 05:34:10 +08:00
; GCN-DAG: v_mul_f32_e32 [[QUIET0:v[0-9]+]], 1.0, [[VAL0]]
; GCN-DAG: v_mul_f32_e32 [[QUIET1:v[0-9]+]], 1.0, [[VAL1]]
AMDGPU: Remove dx10-clamp from subtarget features Since this can be set with s_setreg*, it should not be a subtarget property. Set a default based on the calling convention, and Introduce a new amdgpu-dx10-clamp attribute to override this if desired. Also introduce a new amdgpu-ieee attribute to match. The values need to match to allow inlining. I think it is OK for the caller's dx10-clamp attribute to override the callee, but there doesn't appear to be the infrastructure to do this currently without definining the attribute in the generic Attributes.td. Eventually the calling convention lowering will need to insert a mode switch somewhere for these. llvm-svn: 357302
2019-03-30 03:14:54 +08:00
; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[QUIET0]], [[QUIET1]]
; GCN-NOT: v_mul_f32
define amdgpu_kernel void @kernel_ieee_mode_default() #0 {
%val0 = load volatile float, float addrspace(1)* undef
%val1 = load volatile float, float addrspace(1)* undef
%min = call float @llvm.minnum.f32(float %val0, float %val1)
store volatile float %min, float addrspace(1)* undef
ret void
}
; GCN-LABEL: {{^}}kernel_ieee_mode_on:
; GCN: {{buffer|global|flat}}_load_dword [[VAL0:v[0-9]+]]
; GCN-NEXT: {{buffer|global|flat}}_load_dword [[VAL1:v[0-9]+]]
AMDGPU: Change pre-gfx9 implementation of fcanonicalize to mul If f32 denormals were enabled pre-gfx9, we would still try to implement this with v_max_f32. Pre-gfx9, these instructions ignored the denormal mode and did not flush. Switch to the multiply form for f32 as a workaround which should always work in any case. This fixes conformance failures when the library implementation of fmin/fmax were accidentally not inlined, forcing the assumption of no flushing on targets where denormals are not enabled by default. This is a workaround, since really we should not be mixing code with different FP mode expectations, but prefer the lowering that will work in any mode. Now this will always use max to implement canonicalize on gfx9+. This is only really beneficial for f64. For f32/f16 it's a neutral choice (and worse in terms of code size in 1 case), but possibly worse for the compiler since it does add an extra register use operand. Leave this change for later.
2020-04-23 05:34:10 +08:00
; GCN-DAG: v_mul_f32_e32 [[QUIET0:v[0-9]+]], 1.0, [[VAL0]]
; GCN-DAG: v_mul_f32_e32 [[QUIET1:v[0-9]+]], 1.0, [[VAL1]]
AMDGPU: Remove dx10-clamp from subtarget features Since this can be set with s_setreg*, it should not be a subtarget property. Set a default based on the calling convention, and Introduce a new amdgpu-dx10-clamp attribute to override this if desired. Also introduce a new amdgpu-ieee attribute to match. The values need to match to allow inlining. I think it is OK for the caller's dx10-clamp attribute to override the callee, but there doesn't appear to be the infrastructure to do this currently without definining the attribute in the generic Attributes.td. Eventually the calling convention lowering will need to insert a mode switch somewhere for these. llvm-svn: 357302
2019-03-30 03:14:54 +08:00
; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[QUIET0]], [[QUIET1]]
; GCN-NOT: v_mul_f32
define amdgpu_kernel void @kernel_ieee_mode_on() #1 {
%val0 = load volatile float, float addrspace(1)* undef
%val1 = load volatile float, float addrspace(1)* undef
%min = call float @llvm.minnum.f32(float %val0, float %val1)
store volatile float %min, float addrspace(1)* undef
ret void
}
; GCN-LABEL: {{^}}kernel_ieee_mode_off:
; GCN: {{buffer|global|flat}}_load_dword [[VAL0:v[0-9]+]]
; GCN-NEXT: {{buffer|global|flat}}_load_dword [[VAL1:v[0-9]+]]
; GCN-NOT: [[VAL0]]
; GCN-NOT: [[VAL1]]
; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[VAL0]], [[VAL1]]
; GCN-NOT: v_mul_f32
define amdgpu_kernel void @kernel_ieee_mode_off() #2 {
%val0 = load volatile float, float addrspace(1)* undef
%val1 = load volatile float, float addrspace(1)* undef
%min = call float @llvm.minnum.f32(float %val0, float %val1)
store volatile float %min, float addrspace(1)* undef
ret void
}
; GCN-LABEL: {{^}}func_ieee_mode_default:
; GCN: {{buffer|global|flat}}_load_dword [[VAL0:v[0-9]+]]
; GCN-NEXT: {{buffer|global|flat}}_load_dword [[VAL1:v[0-9]+]]
AMDGPU: Change pre-gfx9 implementation of fcanonicalize to mul If f32 denormals were enabled pre-gfx9, we would still try to implement this with v_max_f32. Pre-gfx9, these instructions ignored the denormal mode and did not flush. Switch to the multiply form for f32 as a workaround which should always work in any case. This fixes conformance failures when the library implementation of fmin/fmax were accidentally not inlined, forcing the assumption of no flushing on targets where denormals are not enabled by default. This is a workaround, since really we should not be mixing code with different FP mode expectations, but prefer the lowering that will work in any mode. Now this will always use max to implement canonicalize on gfx9+. This is only really beneficial for f64. For f32/f16 it's a neutral choice (and worse in terms of code size in 1 case), but possibly worse for the compiler since it does add an extra register use operand. Leave this change for later.
2020-04-23 05:34:10 +08:00
; GCN-DAG: v_mul_f32_e32 [[QUIET0:v[0-9]+]], 1.0, [[VAL0]]
; GCN-DAG: v_mul_f32_e32 [[QUIET1:v[0-9]+]], 1.0, [[VAL1]]
AMDGPU: Remove dx10-clamp from subtarget features Since this can be set with s_setreg*, it should not be a subtarget property. Set a default based on the calling convention, and Introduce a new amdgpu-dx10-clamp attribute to override this if desired. Also introduce a new amdgpu-ieee attribute to match. The values need to match to allow inlining. I think it is OK for the caller's dx10-clamp attribute to override the callee, but there doesn't appear to be the infrastructure to do this currently without definining the attribute in the generic Attributes.td. Eventually the calling convention lowering will need to insert a mode switch somewhere for these. llvm-svn: 357302
2019-03-30 03:14:54 +08:00
; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[QUIET0]], [[QUIET1]]
; GCN-NOT: v_mul_f32
define void @func_ieee_mode_default() #0 {
%val0 = load volatile float, float addrspace(1)* undef
%val1 = load volatile float, float addrspace(1)* undef
%min = call float @llvm.minnum.f32(float %val0, float %val1)
store volatile float %min, float addrspace(1)* undef
ret void
}
; GCN-LABEL: {{^}}func_ieee_mode_on:
; GCN: {{buffer|global|flat}}_load_dword [[VAL0:v[0-9]+]]
; GCN-NEXT: {{buffer|global|flat}}_load_dword [[VAL1:v[0-9]+]]
AMDGPU: Change pre-gfx9 implementation of fcanonicalize to mul If f32 denormals were enabled pre-gfx9, we would still try to implement this with v_max_f32. Pre-gfx9, these instructions ignored the denormal mode and did not flush. Switch to the multiply form for f32 as a workaround which should always work in any case. This fixes conformance failures when the library implementation of fmin/fmax were accidentally not inlined, forcing the assumption of no flushing on targets where denormals are not enabled by default. This is a workaround, since really we should not be mixing code with different FP mode expectations, but prefer the lowering that will work in any mode. Now this will always use max to implement canonicalize on gfx9+. This is only really beneficial for f64. For f32/f16 it's a neutral choice (and worse in terms of code size in 1 case), but possibly worse for the compiler since it does add an extra register use operand. Leave this change for later.
2020-04-23 05:34:10 +08:00
; GCN-DAG: v_mul_f32_e32 [[QUIET0:v[0-9]+]], 1.0, [[VAL0]]
; GCN-DAG: v_mul_f32_e32 [[QUIET1:v[0-9]+]], 1.0, [[VAL1]]
AMDGPU: Remove dx10-clamp from subtarget features Since this can be set with s_setreg*, it should not be a subtarget property. Set a default based on the calling convention, and Introduce a new amdgpu-dx10-clamp attribute to override this if desired. Also introduce a new amdgpu-ieee attribute to match. The values need to match to allow inlining. I think it is OK for the caller's dx10-clamp attribute to override the callee, but there doesn't appear to be the infrastructure to do this currently without definining the attribute in the generic Attributes.td. Eventually the calling convention lowering will need to insert a mode switch somewhere for these. llvm-svn: 357302
2019-03-30 03:14:54 +08:00
; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[QUIET0]], [[QUIET1]]
; GCN-NOT: v_mul_f32
define void @func_ieee_mode_on() #1 {
%val0 = load volatile float, float addrspace(1)* undef
%val1 = load volatile float, float addrspace(1)* undef
%min = call float @llvm.minnum.f32(float %val0, float %val1)
store volatile float %min, float addrspace(1)* undef
ret void
}
; GCN-LABEL: {{^}}func_ieee_mode_off:
; GCN: {{buffer|global|flat}}_load_dword [[VAL0:v[0-9]+]]
; GCN-NEXT: {{buffer|global|flat}}_load_dword [[VAL1:v[0-9]+]]
; GCN-NOT: [[VAL0]]
; GCN-NOT: [[VAL1]]
; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[VAL0]], [[VAL1]]
; GCN-NOT: v_mul_f32
define void @func_ieee_mode_off() #2 {
%val0 = load volatile float, float addrspace(1)* undef
%val1 = load volatile float, float addrspace(1)* undef
%min = call float @llvm.minnum.f32(float %val0, float %val1)
store volatile float %min, float addrspace(1)* undef
ret void
}
; GCN-LABEL: {{^}}cs_ieee_mode_default:
; GCN: {{buffer|global|flat}}_load_dword [[VAL0:v[0-9]+]]
; GCN-NEXT: {{buffer|global|flat}}_load_dword [[VAL1:v[0-9]+]]
AMDGPU: Change pre-gfx9 implementation of fcanonicalize to mul If f32 denormals were enabled pre-gfx9, we would still try to implement this with v_max_f32. Pre-gfx9, these instructions ignored the denormal mode and did not flush. Switch to the multiply form for f32 as a workaround which should always work in any case. This fixes conformance failures when the library implementation of fmin/fmax were accidentally not inlined, forcing the assumption of no flushing on targets where denormals are not enabled by default. This is a workaround, since really we should not be mixing code with different FP mode expectations, but prefer the lowering that will work in any mode. Now this will always use max to implement canonicalize on gfx9+. This is only really beneficial for f64. For f32/f16 it's a neutral choice (and worse in terms of code size in 1 case), but possibly worse for the compiler since it does add an extra register use operand. Leave this change for later.
2020-04-23 05:34:10 +08:00
; GCN-DAG: v_mul_f32_e32 [[QUIET0:v[0-9]+]], 1.0, [[VAL0]]
; GCN-DAG: v_mul_f32_e32 [[QUIET1:v[0-9]+]], 1.0, [[VAL1]]
AMDGPU: Remove dx10-clamp from subtarget features Since this can be set with s_setreg*, it should not be a subtarget property. Set a default based on the calling convention, and Introduce a new amdgpu-dx10-clamp attribute to override this if desired. Also introduce a new amdgpu-ieee attribute to match. The values need to match to allow inlining. I think it is OK for the caller's dx10-clamp attribute to override the callee, but there doesn't appear to be the infrastructure to do this currently without definining the attribute in the generic Attributes.td. Eventually the calling convention lowering will need to insert a mode switch somewhere for these. llvm-svn: 357302
2019-03-30 03:14:54 +08:00
; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[QUIET0]], [[QUIET1]]
; GCN-NOT: v_mul_f32
define amdgpu_cs void @cs_ieee_mode_default() #0 {
%val0 = load volatile float, float addrspace(1)* undef
%val1 = load volatile float, float addrspace(1)* undef
%min = call float @llvm.minnum.f32(float %val0, float %val1)
store volatile float %min, float addrspace(1)* undef
ret void
}
; GCN-LABEL: {{^}}cs_ieee_mode_on:
; GCN: {{buffer|global|flat}}_load_dword [[VAL0:v[0-9]+]]
; GCN-NEXT: {{buffer|global|flat}}_load_dword [[VAL1:v[0-9]+]]
AMDGPU: Change pre-gfx9 implementation of fcanonicalize to mul If f32 denormals were enabled pre-gfx9, we would still try to implement this with v_max_f32. Pre-gfx9, these instructions ignored the denormal mode and did not flush. Switch to the multiply form for f32 as a workaround which should always work in any case. This fixes conformance failures when the library implementation of fmin/fmax were accidentally not inlined, forcing the assumption of no flushing on targets where denormals are not enabled by default. This is a workaround, since really we should not be mixing code with different FP mode expectations, but prefer the lowering that will work in any mode. Now this will always use max to implement canonicalize on gfx9+. This is only really beneficial for f64. For f32/f16 it's a neutral choice (and worse in terms of code size in 1 case), but possibly worse for the compiler since it does add an extra register use operand. Leave this change for later.
2020-04-23 05:34:10 +08:00
; GCN-DAG: v_mul_f32_e32 [[QUIET0:v[0-9]+]], 1.0, [[VAL0]]
; GCN-DAG: v_mul_f32_e32 [[QUIET1:v[0-9]+]], 1.0, [[VAL1]]
AMDGPU: Remove dx10-clamp from subtarget features Since this can be set with s_setreg*, it should not be a subtarget property. Set a default based on the calling convention, and Introduce a new amdgpu-dx10-clamp attribute to override this if desired. Also introduce a new amdgpu-ieee attribute to match. The values need to match to allow inlining. I think it is OK for the caller's dx10-clamp attribute to override the callee, but there doesn't appear to be the infrastructure to do this currently without definining the attribute in the generic Attributes.td. Eventually the calling convention lowering will need to insert a mode switch somewhere for these. llvm-svn: 357302
2019-03-30 03:14:54 +08:00
; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[QUIET0]], [[QUIET1]]
; GCN-NOT: v_mul_f32
define amdgpu_cs void @cs_ieee_mode_on() #1 {
%val0 = load volatile float, float addrspace(1)* undef
%val1 = load volatile float, float addrspace(1)* undef
%min = call float @llvm.minnum.f32(float %val0, float %val1)
store volatile float %min, float addrspace(1)* undef
ret void
}
; GCN-LABEL: {{^}}cs_ieee_mode_off:
; GCN: {{buffer|global|flat}}_load_dword [[VAL0:v[0-9]+]]
; GCN-NEXT: {{buffer|global|flat}}_load_dword [[VAL1:v[0-9]+]]
; GCN-NOT: [[VAL0]]
; GCN-NOT: [[VAL1]]
; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[VAL0]], [[VAL1]]
; GCN-NOT: v_mul_f32
define amdgpu_cs void @cs_ieee_mode_off() #2 {
%val0 = load volatile float, float addrspace(1)* undef
%val1 = load volatile float, float addrspace(1)* undef
%min = call float @llvm.minnum.f32(float %val0, float %val1)
store volatile float %min, float addrspace(1)* undef
ret void
}
; GCN-LABEL: {{^}}ps_ieee_mode_default:
; GCN: {{buffer|global|flat}}_load_dword [[VAL0:v[0-9]+]]
; GCN-NEXT: {{buffer|global|flat}}_load_dword [[VAL1:v[0-9]+]]
; GCN-NOT: [[VAL0]]
; GCN-NOT: [[VAL1]]
; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[VAL0]], [[VAL1]]
; GCN-NOT: v_mul_f32
define amdgpu_ps void @ps_ieee_mode_default() #0 {
%val0 = load volatile float, float addrspace(1)* undef
%val1 = load volatile float, float addrspace(1)* undef
%min = call float @llvm.minnum.f32(float %val0, float %val1)
store volatile float %min, float addrspace(1)* undef
ret void
}
; GCN-LABEL: {{^}}ps_ieee_mode_on:
; GCN: {{buffer|global|flat}}_load_dword [[VAL0:v[0-9]+]]
; GCN-NEXT: {{buffer|global|flat}}_load_dword [[VAL1:v[0-9]+]]
AMDGPU: Change pre-gfx9 implementation of fcanonicalize to mul If f32 denormals were enabled pre-gfx9, we would still try to implement this with v_max_f32. Pre-gfx9, these instructions ignored the denormal mode and did not flush. Switch to the multiply form for f32 as a workaround which should always work in any case. This fixes conformance failures when the library implementation of fmin/fmax were accidentally not inlined, forcing the assumption of no flushing on targets where denormals are not enabled by default. This is a workaround, since really we should not be mixing code with different FP mode expectations, but prefer the lowering that will work in any mode. Now this will always use max to implement canonicalize on gfx9+. This is only really beneficial for f64. For f32/f16 it's a neutral choice (and worse in terms of code size in 1 case), but possibly worse for the compiler since it does add an extra register use operand. Leave this change for later.
2020-04-23 05:34:10 +08:00
; GCN-DAG: v_mul_f32_e32 [[QUIET0:v[0-9]+]], 1.0, [[VAL0]]
; GCN-DAG: v_mul_f32_e32 [[QUIET1:v[0-9]+]], 1.0, [[VAL1]]
AMDGPU: Remove dx10-clamp from subtarget features Since this can be set with s_setreg*, it should not be a subtarget property. Set a default based on the calling convention, and Introduce a new amdgpu-dx10-clamp attribute to override this if desired. Also introduce a new amdgpu-ieee attribute to match. The values need to match to allow inlining. I think it is OK for the caller's dx10-clamp attribute to override the callee, but there doesn't appear to be the infrastructure to do this currently without definining the attribute in the generic Attributes.td. Eventually the calling convention lowering will need to insert a mode switch somewhere for these. llvm-svn: 357302
2019-03-30 03:14:54 +08:00
; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[QUIET0]], [[QUIET1]]
; GCN-NOT: v_mul_f32
define amdgpu_ps void @ps_ieee_mode_on() #1 {
%val0 = load volatile float, float addrspace(1)* undef
%val1 = load volatile float, float addrspace(1)* undef
%min = call float @llvm.minnum.f32(float %val0, float %val1)
store volatile float %min, float addrspace(1)* undef
ret void
}
; GCN-LABEL: {{^}}ps_ieee_mode_off:
; GCN: {{buffer|global|flat}}_load_dword [[VAL0:v[0-9]+]]
; GCN-NEXT: {{buffer|global|flat}}_load_dword [[VAL1:v[0-9]+]]
; GCN-NOT: [[VAL0]]
; GCN-NOT: [[VAL1]]
; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[VAL0]], [[VAL1]]
; GCN-NOT: v_mul_f32
define amdgpu_ps void @ps_ieee_mode_off() #2 {
%val0 = load volatile float, float addrspace(1)* undef
%val1 = load volatile float, float addrspace(1)* undef
%min = call float @llvm.minnum.f32(float %val0, float %val1)
store volatile float %min, float addrspace(1)* undef
ret void
}
declare float @llvm.minnum.f32(float, float) #3
attributes #0 = { nounwind }
attributes #1 = { nounwind "amdgpu-ieee"="true" }
attributes #2 = { nounwind "amdgpu-ieee"="false" }
attributes #3 = { nounwind readnone speculatable }