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/GlobalISel/inst-select-fcmp.s16.mir

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

442 lines
17 KiB
Plaintext
Raw Normal View History

GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
# RUN: llc -march=amdgcn -mcpu=fiji -run-pass=instruction-select -verify-machineinstrs -global-isel-abort=0 -o - %s | FileCheck -check-prefix=WAVE64 %s
# RUN: llc -march=amdgcn -mcpu=gfx1010 -run-pass=instruction-select -verify-machineinstrs -global-isel-abort=0 -o - %s | FileCheck -check-prefix=WAVE32 %s
---
name: fcmp_false_s16_vv
legalized: true
regBankSelected: true
body: |
bb.0:
liveins: $vgpr0, $vgpr1
; WAVE64-LABEL: name: fcmp_false_s16_vv
; WAVE64: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; WAVE64: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; WAVE64: [[TRUNC:%[0-9]+]]:vgpr(s16) = G_TRUNC [[COPY]](s32)
; WAVE64: [[TRUNC1:%[0-9]+]]:vgpr(s16) = G_TRUNC [[COPY1]](s32)
; WAVE64: [[FCMP:%[0-9]+]]:vcc(s1) = G_FCMP floatpred(false), [[TRUNC]](s16), [[TRUNC1]]
; WAVE64: S_ENDPGM 0, implicit [[FCMP]](s1)
; WAVE32-LABEL: name: fcmp_false_s16_vv
; WAVE32: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; WAVE32: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; WAVE32: [[TRUNC:%[0-9]+]]:vgpr(s16) = G_TRUNC [[COPY]](s32)
; WAVE32: [[TRUNC1:%[0-9]+]]:vgpr(s16) = G_TRUNC [[COPY1]](s32)
; WAVE32: [[FCMP:%[0-9]+]]:vcc(s1) = G_FCMP floatpred(false), [[TRUNC]](s16), [[TRUNC1]]
; WAVE32: S_ENDPGM 0, implicit [[FCMP]](s1)
%0:vgpr(s32) = COPY $vgpr0
%1:vgpr(s32) = COPY $vgpr1
%2:vgpr(s16) = G_TRUNC %0
%3:vgpr(s16) = G_TRUNC %1
%4:vcc(s1) = G_FCMP floatpred(false), %2, %3
S_ENDPGM 0, implicit %4
...
---
name: fcmp_oeq_s16_vv
legalized: true
regBankSelected: true
body: |
bb.0:
liveins: $vgpr0, $vgpr1
; WAVE64-LABEL: name: fcmp_oeq_s16_vv
; WAVE64: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE64: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE64: [[V_CMP_EQ_F16_e64_:%[0-9]+]]:sreg_64 = V_CMP_EQ_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE64: S_ENDPGM 0, implicit [[V_CMP_EQ_F16_e64_]]
; WAVE32-LABEL: name: fcmp_oeq_s16_vv
; WAVE32: $vcc_hi = IMPLICIT_DEF
; WAVE32: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE32: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE32: [[V_CMP_EQ_F16_e64_:%[0-9]+]]:sreg_32 = V_CMP_EQ_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE32: S_ENDPGM 0, implicit [[V_CMP_EQ_F16_e64_]]
%0:vgpr(s32) = COPY $vgpr0
%1:vgpr(s32) = COPY $vgpr1
%2:vgpr(s16) = G_TRUNC %0
%3:vgpr(s16) = G_TRUNC %1
%4:vcc(s1) = G_FCMP floatpred(oeq), %2, %3
S_ENDPGM 0, implicit %4
...
---
name: fcmp_ogt_s16_vv
legalized: true
regBankSelected: true
body: |
bb.0:
liveins: $vgpr0, $vgpr1
; WAVE64-LABEL: name: fcmp_ogt_s16_vv
; WAVE64: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE64: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE64: [[V_CMP_GT_F16_e64_:%[0-9]+]]:sreg_64 = V_CMP_GT_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE64: S_ENDPGM 0, implicit [[V_CMP_GT_F16_e64_]]
; WAVE32-LABEL: name: fcmp_ogt_s16_vv
; WAVE32: $vcc_hi = IMPLICIT_DEF
; WAVE32: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE32: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE32: [[V_CMP_GT_F16_e64_:%[0-9]+]]:sreg_32 = V_CMP_GT_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE32: S_ENDPGM 0, implicit [[V_CMP_GT_F16_e64_]]
%0:vgpr(s32) = COPY $vgpr0
%1:vgpr(s32) = COPY $vgpr1
%2:vgpr(s16) = G_TRUNC %0
%3:vgpr(s16) = G_TRUNC %1
%4:vcc(s1) = G_FCMP floatpred(ogt), %2, %3
S_ENDPGM 0, implicit %4
...
---
name: fcmp_oge_s16_vv
legalized: true
regBankSelected: true
body: |
bb.0:
liveins: $vgpr0, $vgpr1
; WAVE64-LABEL: name: fcmp_oge_s16_vv
; WAVE64: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE64: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE64: [[V_CMP_GE_F16_e64_:%[0-9]+]]:sreg_64 = V_CMP_GE_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE64: S_ENDPGM 0, implicit [[V_CMP_GE_F16_e64_]]
; WAVE32-LABEL: name: fcmp_oge_s16_vv
; WAVE32: $vcc_hi = IMPLICIT_DEF
; WAVE32: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE32: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE32: [[V_CMP_GE_F16_e64_:%[0-9]+]]:sreg_32 = V_CMP_GE_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE32: S_ENDPGM 0, implicit [[V_CMP_GE_F16_e64_]]
%0:vgpr(s32) = COPY $vgpr0
%1:vgpr(s32) = COPY $vgpr1
%2:vgpr(s16) = G_TRUNC %0
%3:vgpr(s16) = G_TRUNC %1
%4:vcc(s1) = G_FCMP floatpred(oge), %2, %3
S_ENDPGM 0, implicit %4
...
---
name: fcmp_olt_s16_vv
legalized: true
regBankSelected: true
body: |
bb.0:
liveins: $vgpr0, $vgpr1
; WAVE64-LABEL: name: fcmp_olt_s16_vv
; WAVE64: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE64: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE64: [[V_CMP_LT_F16_e64_:%[0-9]+]]:sreg_64 = V_CMP_LT_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE64: S_ENDPGM 0, implicit [[V_CMP_LT_F16_e64_]]
; WAVE32-LABEL: name: fcmp_olt_s16_vv
; WAVE32: $vcc_hi = IMPLICIT_DEF
; WAVE32: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE32: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE32: [[V_CMP_LT_F16_e64_:%[0-9]+]]:sreg_32 = V_CMP_LT_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE32: S_ENDPGM 0, implicit [[V_CMP_LT_F16_e64_]]
%0:vgpr(s32) = COPY $vgpr0
%1:vgpr(s32) = COPY $vgpr1
%2:vgpr(s16) = G_TRUNC %0
%3:vgpr(s16) = G_TRUNC %1
%4:vcc(s1) = G_FCMP floatpred(olt), %2, %3
S_ENDPGM 0, implicit %4
...
---
name: fcmp_ole_s16_vv
legalized: true
regBankSelected: true
body: |
bb.0:
liveins: $vgpr0, $vgpr1
; WAVE64-LABEL: name: fcmp_ole_s16_vv
; WAVE64: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE64: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE64: [[V_CMP_LE_F16_e64_:%[0-9]+]]:sreg_64 = V_CMP_LE_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE64: S_ENDPGM 0, implicit [[V_CMP_LE_F16_e64_]]
; WAVE32-LABEL: name: fcmp_ole_s16_vv
; WAVE32: $vcc_hi = IMPLICIT_DEF
; WAVE32: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE32: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE32: [[V_CMP_LE_F16_e64_:%[0-9]+]]:sreg_32 = V_CMP_LE_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE32: S_ENDPGM 0, implicit [[V_CMP_LE_F16_e64_]]
%0:vgpr(s32) = COPY $vgpr0
%1:vgpr(s32) = COPY $vgpr1
%2:vgpr(s16) = G_TRUNC %0
%3:vgpr(s16) = G_TRUNC %1
%4:vcc(s1) = G_FCMP floatpred(ole), %2, %3
S_ENDPGM 0, implicit %4
...
---
name: fcmp_one_s16_vv
legalized: true
regBankSelected: true
body: |
bb.0:
liveins: $vgpr0, $vgpr1
; WAVE64-LABEL: name: fcmp_one_s16_vv
; WAVE64: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE64: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE64: [[V_CMP_LG_F16_e64_:%[0-9]+]]:sreg_64 = V_CMP_LG_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE64: S_ENDPGM 0, implicit [[V_CMP_LG_F16_e64_]]
; WAVE32-LABEL: name: fcmp_one_s16_vv
; WAVE32: $vcc_hi = IMPLICIT_DEF
; WAVE32: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE32: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE32: [[V_CMP_LG_F16_e64_:%[0-9]+]]:sreg_32 = V_CMP_LG_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE32: S_ENDPGM 0, implicit [[V_CMP_LG_F16_e64_]]
%0:vgpr(s32) = COPY $vgpr0
%1:vgpr(s32) = COPY $vgpr1
%2:vgpr(s16) = G_TRUNC %0
%3:vgpr(s16) = G_TRUNC %1
%4:vcc(s1) = G_FCMP floatpred(one), %2, %3
S_ENDPGM 0, implicit %4
...
---
name: fcmp_ord_s16_vv
legalized: true
regBankSelected: true
body: |
bb.0:
liveins: $vgpr0, $vgpr1
; WAVE64-LABEL: name: fcmp_ord_s16_vv
; WAVE64: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE64: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE64: [[V_CMP_LG_F16_e64_:%[0-9]+]]:sreg_64 = V_CMP_LG_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE64: S_ENDPGM 0, implicit [[V_CMP_LG_F16_e64_]]
; WAVE32-LABEL: name: fcmp_ord_s16_vv
; WAVE32: $vcc_hi = IMPLICIT_DEF
; WAVE32: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE32: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE32: [[V_CMP_LG_F16_e64_:%[0-9]+]]:sreg_32 = V_CMP_LG_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE32: S_ENDPGM 0, implicit [[V_CMP_LG_F16_e64_]]
%0:vgpr(s32) = COPY $vgpr0
%1:vgpr(s32) = COPY $vgpr1
%2:vgpr(s16) = G_TRUNC %0
%3:vgpr(s16) = G_TRUNC %1
%4:vcc(s1) = G_FCMP floatpred(one), %2, %3
S_ENDPGM 0, implicit %4
...
---
name: fcmp_uno_s16_vv
legalized: true
regBankSelected: true
body: |
bb.0:
liveins: $vgpr0, $vgpr1
; WAVE64-LABEL: name: fcmp_uno_s16_vv
; WAVE64: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE64: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE64: [[V_CMP_U_F16_e64_:%[0-9]+]]:sreg_64 = V_CMP_U_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE64: S_ENDPGM 0, implicit [[V_CMP_U_F16_e64_]]
; WAVE32-LABEL: name: fcmp_uno_s16_vv
; WAVE32: $vcc_hi = IMPLICIT_DEF
; WAVE32: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE32: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE32: [[V_CMP_U_F16_e64_:%[0-9]+]]:sreg_32 = V_CMP_U_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE32: S_ENDPGM 0, implicit [[V_CMP_U_F16_e64_]]
%0:vgpr(s32) = COPY $vgpr0
%1:vgpr(s32) = COPY $vgpr1
%2:vgpr(s16) = G_TRUNC %0
%3:vgpr(s16) = G_TRUNC %1
%4:vcc(s1) = G_FCMP floatpred(uno), %2, %3
S_ENDPGM 0, implicit %4
...
---
name: fcmp_ueq_s16_vv
legalized: true
regBankSelected: true
body: |
bb.0:
liveins: $vgpr0, $vgpr1
; WAVE64-LABEL: name: fcmp_ueq_s16_vv
; WAVE64: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE64: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE64: [[V_CMP_NLG_F16_e64_:%[0-9]+]]:sreg_64 = V_CMP_NLG_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE64: S_ENDPGM 0, implicit [[V_CMP_NLG_F16_e64_]]
; WAVE32-LABEL: name: fcmp_ueq_s16_vv
; WAVE32: $vcc_hi = IMPLICIT_DEF
; WAVE32: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE32: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE32: [[V_CMP_NLG_F16_e64_:%[0-9]+]]:sreg_32 = V_CMP_NLG_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE32: S_ENDPGM 0, implicit [[V_CMP_NLG_F16_e64_]]
%0:vgpr(s32) = COPY $vgpr0
%1:vgpr(s32) = COPY $vgpr1
%2:vgpr(s16) = G_TRUNC %0
%3:vgpr(s16) = G_TRUNC %1
%4:vcc(s1) = G_FCMP floatpred(ueq), %2, %3
S_ENDPGM 0, implicit %4
...
---
name: fcmp_ugt_s16_vv
legalized: true
regBankSelected: true
body: |
bb.0:
liveins: $vgpr0, $vgpr1
; WAVE64-LABEL: name: fcmp_ugt_s16_vv
; WAVE64: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE64: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE64: [[V_CMP_NLE_F16_e64_:%[0-9]+]]:sreg_64 = V_CMP_NLE_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE64: S_ENDPGM 0, implicit [[V_CMP_NLE_F16_e64_]]
; WAVE32-LABEL: name: fcmp_ugt_s16_vv
; WAVE32: $vcc_hi = IMPLICIT_DEF
; WAVE32: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE32: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE32: [[V_CMP_NLE_F16_e64_:%[0-9]+]]:sreg_32 = V_CMP_NLE_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE32: S_ENDPGM 0, implicit [[V_CMP_NLE_F16_e64_]]
%0:vgpr(s32) = COPY $vgpr0
%1:vgpr(s32) = COPY $vgpr1
%2:vgpr(s16) = G_TRUNC %0
%3:vgpr(s16) = G_TRUNC %1
%4:vcc(s1) = G_FCMP floatpred(ugt), %2, %3
S_ENDPGM 0, implicit %4
...
---
name: fcmp_uge_s16_vv
legalized: true
regBankSelected: true
body: |
bb.0:
liveins: $vgpr0, $vgpr1
; WAVE64-LABEL: name: fcmp_uge_s16_vv
; WAVE64: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE64: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE64: [[V_CMP_NLT_F16_e64_:%[0-9]+]]:sreg_64 = V_CMP_NLT_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE64: S_ENDPGM 0, implicit [[V_CMP_NLT_F16_e64_]]
; WAVE32-LABEL: name: fcmp_uge_s16_vv
; WAVE32: $vcc_hi = IMPLICIT_DEF
; WAVE32: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE32: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE32: [[V_CMP_NLT_F16_e64_:%[0-9]+]]:sreg_32 = V_CMP_NLT_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE32: S_ENDPGM 0, implicit [[V_CMP_NLT_F16_e64_]]
%0:vgpr(s32) = COPY $vgpr0
%1:vgpr(s32) = COPY $vgpr1
%2:vgpr(s16) = G_TRUNC %0
%3:vgpr(s16) = G_TRUNC %1
%4:vcc(s1) = G_FCMP floatpred(uge), %2, %3
S_ENDPGM 0, implicit %4
...
---
name: fcmp_ult_s16_vv
legalized: true
regBankSelected: true
body: |
bb.0:
liveins: $vgpr0, $vgpr1
; WAVE64-LABEL: name: fcmp_ult_s16_vv
; WAVE64: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE64: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE64: [[V_CMP_NGE_F16_e64_:%[0-9]+]]:sreg_64 = V_CMP_NGE_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE64: S_ENDPGM 0, implicit [[V_CMP_NGE_F16_e64_]]
; WAVE32-LABEL: name: fcmp_ult_s16_vv
; WAVE32: $vcc_hi = IMPLICIT_DEF
; WAVE32: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE32: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE32: [[V_CMP_NGE_F16_e64_:%[0-9]+]]:sreg_32 = V_CMP_NGE_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE32: S_ENDPGM 0, implicit [[V_CMP_NGE_F16_e64_]]
%0:vgpr(s32) = COPY $vgpr0
%1:vgpr(s32) = COPY $vgpr1
%2:vgpr(s16) = G_TRUNC %0
%3:vgpr(s16) = G_TRUNC %1
%4:vcc(s1) = G_FCMP floatpred(ult), %2, %3
S_ENDPGM 0, implicit %4
...
---
name: fcmp_ule_s16_vv
legalized: true
regBankSelected: true
body: |
bb.0:
liveins: $vgpr0, $vgpr1
; WAVE64-LABEL: name: fcmp_ule_s16_vv
; WAVE64: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE64: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE64: [[V_CMP_NGT_F16_e64_:%[0-9]+]]:sreg_64 = V_CMP_NGT_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE64: S_ENDPGM 0, implicit [[V_CMP_NGT_F16_e64_]]
; WAVE32-LABEL: name: fcmp_ule_s16_vv
; WAVE32: $vcc_hi = IMPLICIT_DEF
; WAVE32: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE32: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE32: [[V_CMP_NGT_F16_e64_:%[0-9]+]]:sreg_32 = V_CMP_NGT_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE32: S_ENDPGM 0, implicit [[V_CMP_NGT_F16_e64_]]
%0:vgpr(s32) = COPY $vgpr0
%1:vgpr(s32) = COPY $vgpr1
%2:vgpr(s16) = G_TRUNC %0
%3:vgpr(s16) = G_TRUNC %1
%4:vcc(s1) = G_FCMP floatpred(ule), %2, %3
S_ENDPGM 0, implicit %4
...
---
name: fcmp_une_s16_vv
legalized: true
regBankSelected: true
body: |
bb.0:
liveins: $vgpr0, $vgpr1
; WAVE64-LABEL: name: fcmp_une_s16_vv
; WAVE64: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE64: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE64: [[V_CMP_NEQ_F16_e64_:%[0-9]+]]:sreg_64 = V_CMP_NEQ_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE64: S_ENDPGM 0, implicit [[V_CMP_NEQ_F16_e64_]]
; WAVE32-LABEL: name: fcmp_une_s16_vv
; WAVE32: $vcc_hi = IMPLICIT_DEF
; WAVE32: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; WAVE32: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
AMDGPU: Start adding MODE register uses to instructions This is the groundwork required to implement strictfp. For now, this should be NFC for regular instructoins (many instructions just gain an extra use of a reserved register). Regalloc won't rematerialize instructions with reads of physical registers, but we were suffering from that anyway with the exec reads. Should add it for all the related FP uses (possibly with some extras). I did not add it to either the gpr index mode instructions (or every single VALU instruction) since it's a ridiculous feature already modeled as an arbitrary side effect. Also work towards marking instructions with FP exceptions. This doesn't actually set the bit yet since this would start to change codegen. It seems nofpexcept is currently not implied from the regular IR FP operations. Add it to some MIR tests where I think it might matter.
2020-05-28 01:25:37 +08:00
; WAVE32: [[V_CMP_NEQ_F16_e64_:%[0-9]+]]:sreg_32 = V_CMP_NEQ_F16_e64 0, [[COPY]], 0, [[COPY1]], 0, implicit $mode, implicit $exec
GlobalISel/TableGen: Handle setcc patterns This is a special case because one node maps to two different G_ instructions, and the operand order is changed. This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually selected for now since it has the SALU and VALU complication to deal with. llvm-svn: 370280
2019-08-29 09:13:41 +08:00
; WAVE32: S_ENDPGM 0, implicit [[V_CMP_NEQ_F16_e64_]]
%0:vgpr(s32) = COPY $vgpr0
%1:vgpr(s32) = COPY $vgpr1
%2:vgpr(s16) = G_TRUNC %0
%3:vgpr(s16) = G_TRUNC %1
%4:vcc(s1) = G_FCMP floatpred(une), %2, %3
S_ENDPGM 0, implicit %4
...
---
name: fcmp_true_s16_vv
legalized: true
regBankSelected: true
body: |
bb.0:
liveins: $vgpr0, $vgpr1
; WAVE64-LABEL: name: fcmp_true_s16_vv
; WAVE64: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; WAVE64: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; WAVE64: [[TRUNC:%[0-9]+]]:vgpr(s16) = G_TRUNC [[COPY]](s32)
; WAVE64: [[TRUNC1:%[0-9]+]]:vgpr(s16) = G_TRUNC [[COPY1]](s32)
; WAVE64: [[FCMP:%[0-9]+]]:vcc(s1) = G_FCMP floatpred(true), [[TRUNC]](s16), [[TRUNC1]]
; WAVE64: S_ENDPGM 0, implicit [[FCMP]](s1)
; WAVE32-LABEL: name: fcmp_true_s16_vv
; WAVE32: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
; WAVE32: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
; WAVE32: [[TRUNC:%[0-9]+]]:vgpr(s16) = G_TRUNC [[COPY]](s32)
; WAVE32: [[TRUNC1:%[0-9]+]]:vgpr(s16) = G_TRUNC [[COPY1]](s32)
; WAVE32: [[FCMP:%[0-9]+]]:vcc(s1) = G_FCMP floatpred(true), [[TRUNC]](s16), [[TRUNC1]]
; WAVE32: S_ENDPGM 0, implicit [[FCMP]](s1)
%0:vgpr(s32) = COPY $vgpr0
%1:vgpr(s32) = COPY $vgpr1
%2:vgpr(s16) = G_TRUNC %0
%3:vgpr(s16) = G_TRUNC %1
%4:vcc(s1) = G_FCMP floatpred(true), %2, %3
S_ENDPGM 0, implicit %4
...