AMDGPU/GlobalISel: Legalize s64->s16 G_SITOFP/G_UITOFP

[Petar.Avramovic]

Summary

Add widenScalar for TypeIdx == 0 for G_SITOFP/G_UITOFP.
Legailize, using widenScalar, as s64->s32 G_SITOFP/G_UITOFP
followed by s32->s16 G_FPTRUNC.

[foad]

Legailize as s64->s32 G_SITOFP/G_UITOFP followed by s32->s16 G_FPTRUNC.

It seems like this would give the wrong result in some cases because of the double rounding. What does SelectionDAG do?

[arsenm]

In D83880#2153414, @foad wrote:

Legailize as s64->s32 G_SITOFP/G_UITOFP followed by s32->s16 G_FPTRUNC.

It seems like this would give the wrong result in some cases because of the double rounding. What does SelectionDAG do?

It looks like it does the same thing. AMDGPU and AArch64 both seem to do the same thing in custom lowering. I think since the source is an integer, there really isn't a whole lot of opportunity for rounding to do much.

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

647

Could you move this to widenScalar for TypeIdx == 0 in LegalizerHelper instead of making this custom?

1828

Missing spaces around ==

[foad]

In D83880#2153414, @foad wrote:

Legailize as s64->s32 G_SITOFP/G_UITOFP followed by s32->s16 G_FPTRUNC.

It seems like this would give the wrong result in some cases because of the double rounding.

The range of a half is only about +/- 65536, and any integer in that range will convert to float exactly without rounding, so there is no double rounding problem.

[arsenm]

LGTM

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-uitofp.mir

484

Can also use v2s64 to v2s16

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

case TargetOpcode::G_FPTOUI:
     if (TypeIdx == 0)
       widenScalarDst(MI, WideTy);
     else
       widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_FPEXT);
 
     Observer.changedInstr(MI);
     return Legalized;
   case TargetOpcode::G_SITOFP:
-    if (TypeIdx != 1)
-      return UnableToLegalize;
     Observer.changingInstr(MI);
+
+    if (TypeIdx == 0)
+      widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC);
+    else
       widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_SEXT);
+
     Observer.changedInstr(MI);
     return Legalized;
-
   case TargetOpcode::G_UITOFP:
-    if (TypeIdx != 1)
-      return UnableToLegalize;
     Observer.changingInstr(MI);
+
+    if (TypeIdx == 0)
+      widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC);
+    else
       widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ZEXT);
+
     Observer.changedInstr(MI);
     return Legalized;
-
   case TargetOpcode::G_LOAD:
   case TargetOpcode::G_SEXTLOAD:
   case TargetOpcode::G_ZEXTLOAD:
     Observer.changingInstr(MI);
     widenScalarDst(MI, WideTy);
     Observer.changedInstr(MI);
     return Legalized;
 

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

getActionDefinitionsBuilder({G_SEXT, G_ZEXT, G_ANYEXT})
     .clampScalar(0, S32, S64)
     .widenScalarToNextPow2(1, 32);
 
   // TODO: Split s1->s64 during regbankselect for VALU.
   auto &IToFP = getActionDefinitionsBuilder({G_SITOFP, G_UITOFP})
     .legalFor({{S32, S32}, {S64, S32}, {S16, S32}})
     .lowerFor({{S32, S64}})
     .lowerIf(typeIs(1, S1))
     .customFor({{S64, S64}});
   if (ST.has16BitInsts())
     IToFP.legalFor({{S16, S16}});
   IToFP.clampScalar(1, S32, S64)
+       .minScalar(0, S32)
        .scalarize(0)
        .widenScalarToNextPow2(1);
 
   auto &FPToI = getActionDefinitionsBuilder({G_FPTOSI, G_FPTOUI})
     .legalFor({{S32, S32}, {S32, S64}, {S32, S16}})
     .customFor({{S64, S64}});
   if (ST.has16BitInsts())
     FPToI.legalFor({{S16, S16}});
bool AMDGPULegalizerInfo::legalizeITOFP(
   Register Dst = MI.getOperand(0).getReg();
   Register Src = MI.getOperand(1).getReg();
 
   const LLT S64 = LLT::scalar(64);
   const LLT S32 = LLT::scalar(32);
 
   assert(MRI.getType(Src) == S64 && MRI.getType(Dst) == S64);
 
   auto Unmerge = B.buildUnmerge({S32, S32}, Src);
 
   auto CvtHi = Signed ?
     B.buildSITOFP(S64, Unmerge.getReg(1)) :
     B.buildUITOFP(S64, Unmerge.getReg(1));
 
   auto CvtLo = B.buildUITOFP(S64, Unmerge.getReg(0));
 
   auto ThirtyTwo = B.buildConstant(S32, 32);

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-sitofp.mir

bb.0:
     ; GFX8: [[ICMP3:%[0-9]+]]:_(s1) = G_ICMP intpred(ne), [[ASHR]](s64), [[C2]]
     ; GFX8: [[SELECT3:%[0-9]+]]:_(s32) = G_SELECT [[ICMP3]](s1), [[FNEG]], [[ADD]]
     ; GFX8: $vgpr0 = COPY [[SELECT3]](s32)
     %0:_(s64) = COPY $vgpr0_vgpr1
     %1:_(s33) = G_TRUNC %0
     %2:_(s32) = G_SITOFP %1
     $vgpr0 = COPY %2
 ...
+
+---
+name: test_sitofp_s64_to_s16
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; GFX6-LABEL: name: test_sitofp_s64_to_s16
+    ; GFX6: [[COPY:%[0-9]+]]:_(s64) = COPY $vgpr0_vgpr1
+    ; GFX6: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 63
+    ; GFX6: [[ASHR:%[0-9]+]]:_(s64) = G_ASHR [[COPY]], [[C]](s32)
+    ; GFX6: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[COPY]](s64)
+    ; GFX6: [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[ASHR]](s64)
+    ; GFX6: [[UADDO:%[0-9]+]]:_(s32), [[UADDO1:%[0-9]+]]:_(s1) = G_UADDO [[UV]], [[UV2]]
+    ; GFX6: [[UADDE:%[0-9]+]]:_(s32), [[UADDE1:%[0-9]+]]:_(s1) = G_UADDE [[UV1]], [[UV3]], [[UADDO1]]
+    ; GFX6: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[UADDO]](s32), [[UADDE]](s32)
+    ; GFX6: [[XOR:%[0-9]+]]:_(s64) = G_XOR [[MV]], [[ASHR]]
+    ; GFX6: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; GFX6: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+    ; GFX6: [[CTLZ_ZERO_UNDEF:%[0-9]+]]:_(s32) = G_CTLZ_ZERO_UNDEF [[XOR]](s64)
+    ; GFX6: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 190
+    ; GFX6: [[SUB:%[0-9]+]]:_(s32) = G_SUB [[C3]], [[CTLZ_ZERO_UNDEF]]
+    ; GFX6: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(ne), [[XOR]](s64), [[C2]]
+    ; GFX6: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[ICMP]](s1), [[SUB]], [[C1]]
+    ; GFX6: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 9223372036854775807
+    ; GFX6: [[SHL:%[0-9]+]]:_(s64) = G_SHL [[XOR]], [[CTLZ_ZERO_UNDEF]](s32)
+    ; GFX6: [[AND:%[0-9]+]]:_(s64) = G_AND [[SHL]], [[C4]]
+    ; GFX6: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 1099511627775
+    ; GFX6: [[AND1:%[0-9]+]]:_(s64) = G_AND [[AND]], [[C5]]
+    ; GFX6: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 40
+    ; GFX6: [[LSHR:%[0-9]+]]:_(s64) = G_LSHR [[AND]], [[C6]](s32)
+    ; GFX6: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 23
+    ; GFX6: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[SELECT]], [[C7]](s32)
+    ; GFX6: [[TRUNC:%[0-9]+]]:_(s32) = G_TRUNC [[LSHR]](s64)
+    ; GFX6: [[OR:%[0-9]+]]:_(s32) = G_OR [[SHL1]], [[TRUNC]]
+    ; GFX6: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 549755813888
+    ; GFX6: [[ICMP1:%[0-9]+]]:_(s1) = G_ICMP intpred(ugt), [[AND1]](s64), [[C8]]
+    ; GFX6: [[ICMP2:%[0-9]+]]:_(s1) = G_ICMP intpred(eq), [[AND1]](s64), [[C8]]
+    ; GFX6: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX6: [[AND2:%[0-9]+]]:_(s32) = G_AND [[OR]], [[C9]]
+    ; GFX6: [[SELECT1:%[0-9]+]]:_(s32) = G_SELECT [[ICMP2]](s1), [[AND2]], [[C1]]
+    ; GFX6: [[SELECT2:%[0-9]+]]:_(s32) = G_SELECT [[ICMP1]](s1), [[C9]], [[SELECT1]]
+    ; GFX6: [[ADD:%[0-9]+]]:_(s32) = G_ADD [[OR]], [[SELECT2]]
+    ; GFX6: [[FNEG:%[0-9]+]]:_(s32) = G_FNEG [[ADD]]
+    ; GFX6: [[ICMP3:%[0-9]+]]:_(s1) = G_ICMP intpred(ne), [[ASHR]](s64), [[C2]]
+    ; GFX6: [[SELECT3:%[0-9]+]]:_(s32) = G_SELECT [[ICMP3]](s1), [[FNEG]], [[ADD]]
+    ; GFX6: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[SELECT3]](s32)
+    ; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FPTRUNC]](s16)
+    ; GFX6: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; GFX8-LABEL: name: test_sitofp_s64_to_s16
+    ; GFX8: [[COPY:%[0-9]+]]:_(s64) = COPY $vgpr0_vgpr1
+    ; GFX8: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 63
+    ; GFX8: [[ASHR:%[0-9]+]]:_(s64) = G_ASHR [[COPY]], [[C]](s32)
+    ; GFX8: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[COPY]](s64)
+    ; GFX8: [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[ASHR]](s64)
+    ; GFX8: [[UADDO:%[0-9]+]]:_(s32), [[UADDO1:%[0-9]+]]:_(s1) = G_UADDO [[UV]], [[UV2]]
+    ; GFX8: [[UADDE:%[0-9]+]]:_(s32), [[UADDE1:%[0-9]+]]:_(s1) = G_UADDE [[UV1]], [[UV3]], [[UADDO1]]
+    ; GFX8: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[UADDO]](s32), [[UADDE]](s32)
+    ; GFX8: [[XOR:%[0-9]+]]:_(s64) = G_XOR [[MV]], [[ASHR]]
+    ; GFX8: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; GFX8: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+    ; GFX8: [[CTLZ_ZERO_UNDEF:%[0-9]+]]:_(s32) = G_CTLZ_ZERO_UNDEF [[XOR]](s64)
+    ; GFX8: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 190
+    ; GFX8: [[SUB:%[0-9]+]]:_(s32) = G_SUB [[C3]], [[CTLZ_ZERO_UNDEF]]
+    ; GFX8: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(ne), [[XOR]](s64), [[C2]]
+    ; GFX8: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[ICMP]](s1), [[SUB]], [[C1]]
+    ; GFX8: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 9223372036854775807
+    ; GFX8: [[SHL:%[0-9]+]]:_(s64) = G_SHL [[XOR]], [[CTLZ_ZERO_UNDEF]](s32)
+    ; GFX8: [[AND:%[0-9]+]]:_(s64) = G_AND [[SHL]], [[C4]]
+    ; GFX8: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 1099511627775
+    ; GFX8: [[AND1:%[0-9]+]]:_(s64) = G_AND [[AND]], [[C5]]
+    ; GFX8: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 40
+    ; GFX8: [[LSHR:%[0-9]+]]:_(s64) = G_LSHR [[AND]], [[C6]](s32)
+    ; GFX8: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 23
+    ; GFX8: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[SELECT]], [[C7]](s32)
+    ; GFX8: [[TRUNC:%[0-9]+]]:_(s32) = G_TRUNC [[LSHR]](s64)
+    ; GFX8: [[OR:%[0-9]+]]:_(s32) = G_OR [[SHL1]], [[TRUNC]]
+    ; GFX8: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 549755813888
+    ; GFX8: [[ICMP1:%[0-9]+]]:_(s1) = G_ICMP intpred(ugt), [[AND1]](s64), [[C8]]
+    ; GFX8: [[ICMP2:%[0-9]+]]:_(s1) = G_ICMP intpred(eq), [[AND1]](s64), [[C8]]
+    ; GFX8: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX8: [[AND2:%[0-9]+]]:_(s32) = G_AND [[OR]], [[C9]]
+    ; GFX8: [[SELECT1:%[0-9]+]]:_(s32) = G_SELECT [[ICMP2]](s1), [[AND2]], [[C1]]
+    ; GFX8: [[SELECT2:%[0-9]+]]:_(s32) = G_SELECT [[ICMP1]](s1), [[C9]], [[SELECT1]]
+    ; GFX8: [[ADD:%[0-9]+]]:_(s32) = G_ADD [[OR]], [[SELECT2]]
+    ; GFX8: [[FNEG:%[0-9]+]]:_(s32) = G_FNEG [[ADD]]
+    ; GFX8: [[ICMP3:%[0-9]+]]:_(s1) = G_ICMP intpred(ne), [[ASHR]](s64), [[C2]]
+    ; GFX8: [[SELECT3:%[0-9]+]]:_(s32) = G_SELECT [[ICMP3]](s1), [[FNEG]], [[ADD]]
+    ; GFX8: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[SELECT3]](s32)
+    ; GFX8: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FPTRUNC]](s16)
+    ; GFX8: $vgpr0 = COPY [[ANYEXT]](s32)
+    %0:_(s64) = COPY $vgpr0_vgpr1
+    %1:_(s16) = G_SITOFP %0
+    %2:_(s32) = G_ANYEXT %1
+    $vgpr0 = COPY %2
+...

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-uitofp.mir

bb.0:
     ; GFX8: [[SELECT2:%[0-9]+]]:_(s32) = G_SELECT [[ICMP1]](s1), [[C9]], [[SELECT1]]
     ; GFX8: [[ADD:%[0-9]+]]:_(s32) = G_ADD [[OR]], [[SELECT2]]
     ; GFX8: $vgpr0 = COPY [[ADD]](s32)
     %0:_(s64) = COPY $vgpr0_vgpr1
     %1:_(s33) = G_TRUNC %0
     %2:_(s32) = G_UITOFP %1
     $vgpr0 = COPY %2
 ...
+
+---
+name: test_uitofp_s64_to_s16
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; GFX6-LABEL: name: test_uitofp_s64_to_s16
+    ; GFX6: [[COPY:%[0-9]+]]:_(s64) = COPY $vgpr0_vgpr1
+    ; GFX6: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; GFX6: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+    ; GFX6: [[CTLZ_ZERO_UNDEF:%[0-9]+]]:_(s32) = G_CTLZ_ZERO_UNDEF [[COPY]](s64)
+    ; GFX6: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 190
+    ; GFX6: [[SUB:%[0-9]+]]:_(s32) = G_SUB [[C2]], [[CTLZ_ZERO_UNDEF]]
+    ; GFX6: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(ne), [[COPY]](s64), [[C1]]
+    ; GFX6: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[ICMP]](s1), [[SUB]], [[C]]
+    ; GFX6: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 9223372036854775807
+    ; GFX6: [[SHL:%[0-9]+]]:_(s64) = G_SHL [[COPY]], [[CTLZ_ZERO_UNDEF]](s32)
+    ; GFX6: [[AND:%[0-9]+]]:_(s64) = G_AND [[SHL]], [[C3]]
+    ; GFX6: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 1099511627775
+    ; GFX6: [[AND1:%[0-9]+]]:_(s64) = G_AND [[AND]], [[C4]]
+    ; GFX6: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 40
+    ; GFX6: [[LSHR:%[0-9]+]]:_(s64) = G_LSHR [[AND]], [[C5]](s32)
+    ; GFX6: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 23
+    ; GFX6: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[SELECT]], [[C6]](s32)
+    ; GFX6: [[TRUNC:%[0-9]+]]:_(s32) = G_TRUNC [[LSHR]](s64)
+    ; GFX6: [[OR:%[0-9]+]]:_(s32) = G_OR [[SHL1]], [[TRUNC]]
+    ; GFX6: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 549755813888
+    ; GFX6: [[ICMP1:%[0-9]+]]:_(s1) = G_ICMP intpred(ugt), [[AND1]](s64), [[C7]]
+    ; GFX6: [[ICMP2:%[0-9]+]]:_(s1) = G_ICMP intpred(eq), [[AND1]](s64), [[C7]]
+    ; GFX6: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX6: [[AND2:%[0-9]+]]:_(s32) = G_AND [[OR]], [[C8]]
+    ; GFX6: [[SELECT1:%[0-9]+]]:_(s32) = G_SELECT [[ICMP2]](s1), [[AND2]], [[C]]
+    ; GFX6: [[SELECT2:%[0-9]+]]:_(s32) = G_SELECT [[ICMP1]](s1), [[C8]], [[SELECT1]]
+    ; GFX6: [[ADD:%[0-9]+]]:_(s32) = G_ADD [[OR]], [[SELECT2]]
+    ; GFX6: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[ADD]](s32)
+    ; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FPTRUNC]](s16)
+    ; GFX6: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; GFX8-LABEL: name: test_uitofp_s64_to_s16
+    ; GFX8: [[COPY:%[0-9]+]]:_(s64) = COPY $vgpr0_vgpr1
+    ; GFX8: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; GFX8: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+    ; GFX8: [[CTLZ_ZERO_UNDEF:%[0-9]+]]:_(s32) = G_CTLZ_ZERO_UNDEF [[COPY]](s64)
+    ; GFX8: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 190
+    ; GFX8: [[SUB:%[0-9]+]]:_(s32) = G_SUB [[C2]], [[CTLZ_ZERO_UNDEF]]
+    ; GFX8: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(ne), [[COPY]](s64), [[C1]]
+    ; GFX8: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[ICMP]](s1), [[SUB]], [[C]]
+    ; GFX8: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 9223372036854775807
+    ; GFX8: [[SHL:%[0-9]+]]:_(s64) = G_SHL [[COPY]], [[CTLZ_ZERO_UNDEF]](s32)
+    ; GFX8: [[AND:%[0-9]+]]:_(s64) = G_AND [[SHL]], [[C3]]
+    ; GFX8: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 1099511627775
+    ; GFX8: [[AND1:%[0-9]+]]:_(s64) = G_AND [[AND]], [[C4]]
+    ; GFX8: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 40
+    ; GFX8: [[LSHR:%[0-9]+]]:_(s64) = G_LSHR [[AND]], [[C5]](s32)
+    ; GFX8: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 23
+    ; GFX8: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[SELECT]], [[C6]](s32)
+    ; GFX8: [[TRUNC:%[0-9]+]]:_(s32) = G_TRUNC [[LSHR]](s64)
+    ; GFX8: [[OR:%[0-9]+]]:_(s32) = G_OR [[SHL1]], [[TRUNC]]
+    ; GFX8: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 549755813888
+    ; GFX8: [[ICMP1:%[0-9]+]]:_(s1) = G_ICMP intpred(ugt), [[AND1]](s64), [[C7]]
+    ; GFX8: [[ICMP2:%[0-9]+]]:_(s1) = G_ICMP intpred(eq), [[AND1]](s64), [[C7]]
+    ; GFX8: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX8: [[AND2:%[0-9]+]]:_(s32) = G_AND [[OR]], [[C8]]
+    ; GFX8: [[SELECT1:%[0-9]+]]:_(s32) = G_SELECT [[ICMP2]](s1), [[AND2]], [[C]]
+    ; GFX8: [[SELECT2:%[0-9]+]]:_(s32) = G_SELECT [[ICMP1]](s1), [[C8]], [[SELECT1]]
+    ; GFX8: [[ADD:%[0-9]+]]:_(s32) = G_ADD [[OR]], [[SELECT2]]
+    ; GFX8: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[ADD]](s32)
+    ; GFX8: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FPTRUNC]](s16)
+    ; GFX8: $vgpr0 = COPY [[ANYEXT]](s32)
+    %0:_(s64) = COPY $vgpr0_vgpr1
+    %1:_(s16) = G_UITOFP %0
+    %2:_(s32) = G_ANYEXT %1
+    $vgpr0 = COPY %2
+...