From 33f5f7933cea8cbcf552cee038a16d8600d8c664 Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault@amd.com>
Date: Wed, 22 Apr 2020 17:34:10 -0400
Subject: [PATCH] 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.
---
 lib/Target/AMDGPU/AMDGPU.td                   |   3 +
 lib/Target/AMDGPU/AMDGPUInstructions.td       |   3 +
 lib/Target/AMDGPU/SIInstructions.td           |  90 ++++---
 .../GlobalISel/inst-select-fcanonicalize.mir  |  50 ++--
 test/CodeGen/AMDGPU/amdgcn-ieee.ll            |  30 +--
 test/CodeGen/AMDGPU/clamp.ll                  |  13 +-
 .../AMDGPU/fcanonicalize-elimination.ll       |  49 ++--
 test/CodeGen/AMDGPU/fcanonicalize.f16.ll      |   7 +-
 test/CodeGen/AMDGPU/fcanonicalize.ll          | 235 ++++++++++--------
 test/CodeGen/AMDGPU/fminnum.f64.ll            | 123 +++++----
 10 files changed, 345 insertions(+), 258 deletions(-)

diff --git a/lib/Target/AMDGPU/AMDGPU.td b/lib/Target/AMDGPU/AMDGPU.td
index 97852523033..9f38e92c434 100644
--- a/lib/Target/AMDGPU/AMDGPU.td
+++ b/lib/Target/AMDGPU/AMDGPU.td
@@ -1063,6 +1063,9 @@ def Has16BitInsts : Predicate<"Subtarget->has16BitInsts()">,
 def HasVOP3PInsts : Predicate<"Subtarget->hasVOP3PInsts()">,
   AssemblerPredicate<(all_of FeatureVOP3P)>;
 
+def HasMinMaxDenormModes : Predicate<"Subtarget->supportsMinMaxDenormModes()">;
+def NotHasMinMaxDenormModes : Predicate<"!Subtarget->supportsMinMaxDenormModes()">;
+
 def HasSDWA : Predicate<"Subtarget->hasSDWA()">,
   AssemblerPredicate<(all_of FeatureSDWA, FeatureVolcanicIslands)>;
 
diff --git a/lib/Target/AMDGPU/AMDGPUInstructions.td b/lib/Target/AMDGPU/AMDGPUInstructions.td
index 483b1568e53..706053a4e31 100644
--- a/lib/Target/AMDGPU/AMDGPUInstructions.td
+++ b/lib/Target/AMDGPU/AMDGPUInstructions.td
@@ -77,6 +77,9 @@ class ILFormat<dag outs, dag ins, string asmstr, list<dag> pattern>
 
 def TruePredicate : Predicate<"">;
 
+// FIXME: Tablegen should specially supports this
+def FalsePredicate : Predicate<"false">;
+
 // Add a predicate to the list if does not already exist to deduplicate it.
 class PredConcat<list<Predicate> lst, Predicate pred> {
   list<Predicate> ret =
diff --git a/lib/Target/AMDGPU/SIInstructions.td b/lib/Target/AMDGPU/SIInstructions.td
index 40eebb044fd..dae8b0b7d39 100644
--- a/lib/Target/AMDGPU/SIInstructions.td
+++ b/lib/Target/AMDGPU/SIInstructions.td
@@ -1878,7 +1878,9 @@ def : GCNPat <
 
 }
 
-let OtherPredicates = [NoFP16Denormals] in {
+
+// Prefer selecting to max when legal, but using mul is always valid.
+let AddedComplexity = -5 in {
 def : GCNPat<
   (fcanonicalize (f16 (VOP3Mods f16:$src, i32:$src_mods))),
   (V_MUL_F16_e64 0, (i32 CONST.FP16_ONE), $src_mods, $src)
@@ -1893,23 +1895,7 @@ def : GCNPat<
   (fcanonicalize (v2f16 (VOP3PMods v2f16:$src, i32:$src_mods))),
   (V_PK_MUL_F16 0, (i32 CONST.FP16_ONE), $src_mods, $src, DSTCLAMP.NONE)
 >;
-}
 
-let OtherPredicates = [FP16Denormals] in {
-def : GCNPat<
-  (fcanonicalize (f16 (VOP3Mods f16:$src, i32:$src_mods))),
-  (V_MAX_F16_e64 $src_mods, $src, $src_mods, $src, 0, 0)
->;
-
-let SubtargetPredicate = HasVOP3PInsts in {
-def : GCNPat<
-  (fcanonicalize (v2f16 (VOP3PMods v2f16:$src, i32:$src_mods))),
-  (V_PK_MAX_F16 $src_mods, $src, $src_mods, $src, DSTCLAMP.NONE)
->;
-}
-}
-
-let OtherPredicates = [NoFP32Denormals] in {
 def : GCNPat<
   (fcanonicalize (f32 (VOP3Mods f32:$src, i32:$src_mods))),
   (V_MUL_F32_e64 0, (i32 CONST.FP32_ONE), $src_mods, $src)
@@ -1919,28 +1905,68 @@ def : GCNPat<
   (fcanonicalize (f32 (fneg (VOP3Mods f32:$src, i32:$src_mods)))),
   (V_MUL_F32_e64 0, (i32 CONST.FP32_NEG_ONE), $src_mods, $src)
 >;
-}
 
-let OtherPredicates = [FP32Denormals] in {
-def : GCNPat<
-  (fcanonicalize (f32 (VOP3Mods f32:$src, i32:$src_mods))),
-  (V_MAX_F32_e64 $src_mods, $src, $src_mods, $src)
->;
-}
-
-let OtherPredicates = [NoFP64Denormals] in {
+// TODO: Handle fneg like other types.
 def : GCNPat<
   (fcanonicalize (f64 (VOP3Mods f64:$src, i32:$src_mods))),
   (V_MUL_F64 0, CONST.FP64_ONE, $src_mods, $src)
 >;
+} // End AddedComplexity = -5
+
+multiclass SelectCanonicalizeAsMax<
+  list<Predicate> f32_preds = [],
+  list<Predicate> f64_preds = [],
+  list<Predicate> f16_preds = []> {
+  def : GCNPat<
+    (fcanonicalize (f32 (VOP3Mods f32:$src, i32:$src_mods))),
+    (V_MAX_F32_e64 $src_mods, $src, $src_mods, $src)> {
+    let OtherPredicates = f32_preds;
+  }
+
+  def : GCNPat<
+    (fcanonicalize (f64 (VOP3Mods f64:$src, i32:$src_mods))),
+    (V_MAX_F64 $src_mods, $src, $src_mods, $src)> {
+    let OtherPredicates = f64_preds;
+  }
+
+  def : GCNPat<
+    (fcanonicalize (f16 (VOP3Mods f16:$src, i32:$src_mods))),
+    (V_MAX_F16_e64 $src_mods, $src, $src_mods, $src, 0, 0)> {
+    // FIXME: Should have 16-bit inst subtarget predicate
+    let OtherPredicates = f16_preds;
+  }
+
+  def : GCNPat<
+    (fcanonicalize (v2f16 (VOP3PMods v2f16:$src, i32:$src_mods))),
+    (V_PK_MAX_F16 $src_mods, $src, $src_mods, $src, DSTCLAMP.NONE)> {
+    // FIXME: Should have VOP3P subtarget predicate
+    let OtherPredicates = f16_preds;
+  }
 }
 
-let OtherPredicates = [FP64Denormals] in {
-def : GCNPat<
-  (fcanonicalize (f64 (VOP3Mods f64:$src, i32:$src_mods))),
-  (V_MAX_F64 $src_mods, $src, $src_mods, $src)
->;
-}
+// On pre-gfx9 targets, v_max_*/v_min_* did not respect the denormal
+// mode, and would never flush. For f64, it's faster to do implement
+// this with a max. For f16/f32 it's a wash, but prefer max when
+// valid.
+//
+// FIXME: Lowering f32/f16 with max is worse since we can use a
+// smaller encoding if the input is fneg'd. It also adds an extra
+// register use.
+let SubtargetPredicate = HasMinMaxDenormModes in {
+  defm : SelectCanonicalizeAsMax<[], [], []>;
+} // End SubtargetPredicate = HasMinMaxDenormModes
+
+let SubtargetPredicate = NotHasMinMaxDenormModes in {
+  // Use the max lowering if we don't need to flush.
+
+  // FIXME: We don't do use this for f32 as a workaround for the
+  // library being compiled with the default ieee mode, but
+  // potentially being called from flushing kernels. Really we should
+  // not be mixing code expecting different default FP modes, but mul
+  // works in any FP environment.
+  defm : SelectCanonicalizeAsMax<[FalsePredicate], [FP64Denormals], [FP16Denormals]>;
+} // End SubtargetPredicate = NotHasMinMaxDenormModes
+
 
 let OtherPredicates = [HasDLInsts] in {
 def : GCNPat <
diff --git a/test/CodeGen/AMDGPU/GlobalISel/inst-select-fcanonicalize.mir b/test/CodeGen/AMDGPU/GlobalISel/inst-select-fcanonicalize.mir
index b62e24921eb..75086984a14 100644
--- a/test/CodeGen/AMDGPU/GlobalISel/inst-select-fcanonicalize.mir
+++ b/test/CodeGen/AMDGPU/GlobalISel/inst-select-fcanonicalize.mir
@@ -48,8 +48,8 @@ body: |
     ; GFX8: S_ENDPGM 0, implicit [[V_MUL_F16_e64_]]
     ; GFX9-LABEL: name: fcanonicalize_f16_flush
     ; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
-    ; GFX9: [[V_MUL_F16_e64_:%[0-9]+]]:vgpr_32 = V_MUL_F16_e64 0, 15360, 0, [[COPY]], 0, 0, implicit $exec
-    ; GFX9: S_ENDPGM 0, implicit [[V_MUL_F16_e64_]]
+    ; GFX9: [[V_MAX_F16_e64_:%[0-9]+]]:vgpr_32 = V_MAX_F16_e64 0, [[COPY]], 0, [[COPY]], 0, 0, implicit $exec
+    ; GFX9: S_ENDPGM 0, implicit [[V_MAX_F16_e64_]]
     %0:vgpr(s32) = COPY $vgpr0
     %1:vgpr(s16) = G_TRUNC %0
     %2:vgpr(s16) = G_FCANONICALIZE %1
@@ -72,8 +72,8 @@ body: |
 
     ; GFX8-LABEL: name: fcanonicalize_f32_denorm
     ; GFX8: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
-    ; GFX8: [[V_MAX_F32_e64_:%[0-9]+]]:vgpr_32 = V_MAX_F32_e64 0, [[COPY]], 0, [[COPY]], 0, 0, implicit $exec
-    ; GFX8: S_ENDPGM 0, implicit [[V_MAX_F32_e64_]]
+    ; GFX8: [[V_MUL_F32_e64_:%[0-9]+]]:vgpr_32 = V_MUL_F32_e64 0, 1065353216, 0, [[COPY]], 0, 0, implicit $exec
+    ; GFX8: S_ENDPGM 0, implicit [[V_MUL_F32_e64_]]
     ; GFX9-LABEL: name: fcanonicalize_f32_denorm
     ; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
     ; GFX9: [[V_MAX_F32_e64_:%[0-9]+]]:vgpr_32 = V_MAX_F32_e64 0, [[COPY]], 0, [[COPY]], 0, 0, implicit $exec
@@ -103,8 +103,8 @@ body: |
     ; GFX8: S_ENDPGM 0, implicit [[V_MUL_F32_e64_]]
     ; GFX9-LABEL: name: fcanonicalize_f32_flush
     ; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
-    ; GFX9: [[V_MUL_F32_e64_:%[0-9]+]]:vgpr_32 = V_MUL_F32_e64 0, 1065353216, 0, [[COPY]], 0, 0, implicit $exec
-    ; GFX9: S_ENDPGM 0, implicit [[V_MUL_F32_e64_]]
+    ; GFX9: [[V_MAX_F32_e64_:%[0-9]+]]:vgpr_32 = V_MAX_F32_e64 0, [[COPY]], 0, [[COPY]], 0, 0, implicit $exec
+    ; GFX9: S_ENDPGM 0, implicit [[V_MAX_F32_e64_]]
     %0:vgpr(s32) = COPY $vgpr0
     %1:vgpr(s32) = G_FCANONICALIZE %0
     S_ENDPGM 0, implicit %1
@@ -125,9 +125,9 @@ body: |
     liveins: $vgpr0
 
     ; GFX8-LABEL: name: fcanonicalize_v2f16_denorm
-    ; GFX8: [[COPY:%[0-9]+]]:vgpr(<2 x s16>) = COPY $vgpr0
-    ; GFX8: [[FCANONICALIZE:%[0-9]+]]:vgpr(<2 x s16>) = G_FCANONICALIZE [[COPY]]
-    ; GFX8: S_ENDPGM 0, implicit [[FCANONICALIZE]](<2 x s16>)
+    ; GFX8: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+    ; GFX8: [[V_PK_MAX_F16_:%[0-9]+]]:vgpr_32 = V_PK_MAX_F16 8, [[COPY]], 8, [[COPY]], 0, 0, 0, 0, 0, implicit $exec
+    ; GFX8: S_ENDPGM 0, implicit [[V_PK_MAX_F16_]]
     ; GFX9-LABEL: name: fcanonicalize_v2f16_denorm
     ; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
     ; GFX9: [[V_PK_MAX_F16_:%[0-9]+]]:vgpr_32 = V_PK_MAX_F16 8, [[COPY]], 8, [[COPY]], 0, 0, 0, 0, 0, implicit $exec
@@ -157,8 +157,8 @@ body: |
     ; GFX8: S_ENDPGM 0, implicit [[V_PK_MUL_F16_]]
     ; GFX9-LABEL: name: fcanonicalize_v2f16_flush
     ; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
-    ; GFX9: [[V_PK_MUL_F16_:%[0-9]+]]:vgpr_32 = V_PK_MUL_F16 0, 15360, 8, [[COPY]], 0, 0, 0, 0, 0, implicit $exec
-    ; GFX9: S_ENDPGM 0, implicit [[V_PK_MUL_F16_]]
+    ; GFX9: [[V_PK_MAX_F16_:%[0-9]+]]:vgpr_32 = V_PK_MAX_F16 8, [[COPY]], 8, [[COPY]], 0, 0, 0, 0, 0, implicit $exec
+    ; GFX9: S_ENDPGM 0, implicit [[V_PK_MAX_F16_]]
     %0:vgpr(<2 x s16>) = COPY $vgpr0
     %1:vgpr(<2 x s16>) = G_FCANONICALIZE %0
     S_ENDPGM 0, implicit %1
@@ -211,8 +211,8 @@ body: |
     ; GFX8: S_ENDPGM 0, implicit [[V_MUL_F64_]]
     ; GFX9-LABEL: name: fcanonicalize_f64_flush
     ; GFX9: [[COPY:%[0-9]+]]:vreg_64 = COPY $vgpr0_vgpr1
-    ; GFX9: [[V_MUL_F64_:%[0-9]+]]:vreg_64 = V_MUL_F64 0, 4607182418800017408, 0, [[COPY]], 0, 0, implicit $exec
-    ; GFX9: S_ENDPGM 0, implicit [[V_MUL_F64_]]
+    ; GFX9: [[V_MAX_F64_:%[0-9]+]]:vreg_64 = V_MAX_F64 0, [[COPY]], 0, [[COPY]], 0, 0, implicit $exec
+    ; GFX9: S_ENDPGM 0, implicit [[V_MAX_F64_]]
     %0:vgpr(s64) = COPY $vgpr0_vgpr1
     %1:vgpr(s64) = G_FCANONICALIZE %0
     S_ENDPGM 0, implicit %1
@@ -233,8 +233,8 @@ body: |
     liveins: $vgpr0
     ; GFX8-LABEL: name: fcanonicalize_fabs_f32_denorm
     ; GFX8: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
-    ; GFX8: [[V_MAX_F32_e64_:%[0-9]+]]:vgpr_32 = V_MAX_F32_e64 2, [[COPY]], 2, [[COPY]], 0, 0, implicit $exec
-    ; GFX8: S_ENDPGM 0, implicit [[V_MAX_F32_e64_]]
+    ; GFX8: [[V_MUL_F32_e64_:%[0-9]+]]:vgpr_32 = V_MUL_F32_e64 0, 1065353216, 2, [[COPY]], 0, 0, implicit $exec
+    ; GFX8: S_ENDPGM 0, implicit [[V_MUL_F32_e64_]]
     ; GFX9-LABEL: name: fcanonicalize_fabs_f32_denorm
     ; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
     ; GFX9: [[V_MAX_F32_e64_:%[0-9]+]]:vgpr_32 = V_MAX_F32_e64 2, [[COPY]], 2, [[COPY]], 0, 0, implicit $exec
@@ -265,8 +265,8 @@ body: |
     ; GFX8: S_ENDPGM 0, implicit [[V_MUL_F32_e64_]]
     ; GFX9-LABEL: name: fcanonicalize_fabs_f32_flush
     ; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
-    ; GFX9: [[V_MUL_F32_e64_:%[0-9]+]]:vgpr_32 = V_MUL_F32_e64 0, 1065353216, 2, [[COPY]], 0, 0, implicit $exec
-    ; GFX9: S_ENDPGM 0, implicit [[V_MUL_F32_e64_]]
+    ; GFX9: [[V_MAX_F32_e64_:%[0-9]+]]:vgpr_32 = V_MAX_F32_e64 2, [[COPY]], 2, [[COPY]], 0, 0, implicit $exec
+    ; GFX9: S_ENDPGM 0, implicit [[V_MAX_F32_e64_]]
     %0:vgpr(s32) = COPY $vgpr0
     %1:vgpr(s32) = G_FABS %0
     %2:vgpr(s32) = G_FCANONICALIZE %1
@@ -288,8 +288,8 @@ body: |
     liveins: $vgpr0
     ; GFX8-LABEL: name: fcanonicalize_fneg_f32_denorm
     ; GFX8: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
-    ; GFX8: [[V_MAX_F32_e64_:%[0-9]+]]:vgpr_32 = V_MAX_F32_e64 1, [[COPY]], 1, [[COPY]], 0, 0, implicit $exec
-    ; GFX8: S_ENDPGM 0, implicit [[V_MAX_F32_e64_]]
+    ; GFX8: [[V_MUL_F32_e64_:%[0-9]+]]:vgpr_32 = V_MUL_F32_e64 0, 3212836864, 0, [[COPY]], 0, 0, implicit $exec
+    ; GFX8: S_ENDPGM 0, implicit [[V_MUL_F32_e64_]]
     ; GFX9-LABEL: name: fcanonicalize_fneg_f32_denorm
     ; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
     ; GFX9: [[V_MAX_F32_e64_:%[0-9]+]]:vgpr_32 = V_MAX_F32_e64 1, [[COPY]], 1, [[COPY]], 0, 0, implicit $exec
@@ -319,8 +319,8 @@ body: |
     ; GFX8: S_ENDPGM 0, implicit [[V_MUL_F32_e64_]]
     ; GFX9-LABEL: name: fcanonicalize_fneg_f32_flush
     ; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
-    ; GFX9: [[V_MUL_F32_e64_:%[0-9]+]]:vgpr_32 = V_MUL_F32_e64 0, 3212836864, 0, [[COPY]], 0, 0, implicit $exec
-    ; GFX9: S_ENDPGM 0, implicit [[V_MUL_F32_e64_]]
+    ; GFX9: [[V_MAX_F32_e64_:%[0-9]+]]:vgpr_32 = V_MAX_F32_e64 1, [[COPY]], 1, [[COPY]], 0, 0, implicit $exec
+    ; GFX9: S_ENDPGM 0, implicit [[V_MAX_F32_e64_]]
     %0:vgpr(s32) = COPY $vgpr0
     %1:vgpr(s32) = G_FNEG %0
     %2:vgpr(s32) = G_FCANONICALIZE %1
@@ -344,8 +344,8 @@ body: |
     ; GFX8: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
     ; GFX8: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 2147483648
     ; GFX8: [[V_XOR_B32_e32_:%[0-9]+]]:vgpr_32 = V_XOR_B32_e32 [[S_MOV_B32_]], [[COPY]], implicit $exec
-    ; GFX8: [[V_MAX_F32_e64_:%[0-9]+]]:vgpr_32 = V_MAX_F32_e64 2, [[V_XOR_B32_e32_]], 2, [[V_XOR_B32_e32_]], 0, 0, implicit $exec
-    ; GFX8: S_ENDPGM 0, implicit [[V_MAX_F32_e64_]]
+    ; GFX8: [[V_MUL_F32_e64_:%[0-9]+]]:vgpr_32 = V_MUL_F32_e64 0, 1065353216, 2, [[V_XOR_B32_e32_]], 0, 0, implicit $exec
+    ; GFX8: S_ENDPGM 0, implicit [[V_MUL_F32_e64_]]
     ; GFX9-LABEL: name: fcanonicalize_fneg_fabs_f32_denorm
     ; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
     ; GFX9: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 2147483648
@@ -382,8 +382,8 @@ body: |
     ; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
     ; GFX9: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 2147483648
     ; GFX9: [[V_XOR_B32_e32_:%[0-9]+]]:vgpr_32 = V_XOR_B32_e32 [[S_MOV_B32_]], [[COPY]], implicit $exec
-    ; GFX9: [[V_MUL_F32_e64_:%[0-9]+]]:vgpr_32 = V_MUL_F32_e64 0, 1065353216, 2, [[V_XOR_B32_e32_]], 0, 0, implicit $exec
-    ; GFX9: S_ENDPGM 0, implicit [[V_MUL_F32_e64_]]
+    ; GFX9: [[V_MAX_F32_e64_:%[0-9]+]]:vgpr_32 = V_MAX_F32_e64 2, [[V_XOR_B32_e32_]], 2, [[V_XOR_B32_e32_]], 0, 0, implicit $exec
+    ; GFX9: S_ENDPGM 0, implicit [[V_MAX_F32_e64_]]
     %0:vgpr(s32) = COPY $vgpr0
     %1:vgpr(s32) = G_FNEG %0
     %2:vgpr(s32) = G_FABS %1
diff --git a/test/CodeGen/AMDGPU/amdgcn-ieee.ll b/test/CodeGen/AMDGPU/amdgcn-ieee.ll
index 2ad2ee3502d..d306268ecb5 100644
--- a/test/CodeGen/AMDGPU/amdgcn-ieee.ll
+++ b/test/CodeGen/AMDGPU/amdgcn-ieee.ll
@@ -1,10 +1,10 @@
-; RUN: llc -mtriple=amdgcn-mesa-mesa3d -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
+; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
 
 ; 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]+]]
-; GCN-DAG: v_max_f32_e32 [[QUIET0:v[0-9]+]], [[VAL0]], [[VAL0]]
-; GCN-DAG: v_max_f32_e32 [[QUIET1:v[0-9]+]], [[VAL1]], [[VAL1]]
+; 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]]
 ; 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 {
@@ -18,8 +18,8 @@ define amdgpu_kernel void @kernel_ieee_mode_default() #0 {
 ; 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]+]]
-; GCN-DAG: v_max_f32_e32 [[QUIET0:v[0-9]+]], [[VAL0]], [[VAL0]]
-; GCN-DAG: v_max_f32_e32 [[QUIET1:v[0-9]+]], [[VAL1]], [[VAL1]]
+; 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]]
 ; 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 {
@@ -48,8 +48,8 @@ define amdgpu_kernel void @kernel_ieee_mode_off() #2 {
 ; 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]+]]
-; GCN-DAG: v_max_f32_e32 [[QUIET0:v[0-9]+]], [[VAL0]], [[VAL0]]
-; GCN-DAG: v_max_f32_e32 [[QUIET1:v[0-9]+]], [[VAL1]], [[VAL1]]
+; 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]]
 ; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[QUIET0]], [[QUIET1]]
 ; GCN-NOT: v_mul_f32
 define void @func_ieee_mode_default() #0 {
@@ -63,8 +63,8 @@ define void @func_ieee_mode_default() #0 {
 ; 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]+]]
-; GCN-DAG: v_max_f32_e32 [[QUIET0:v[0-9]+]], [[VAL0]], [[VAL0]]
-; GCN-DAG: v_max_f32_e32 [[QUIET1:v[0-9]+]], [[VAL1]], [[VAL1]]
+; 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]]
 ; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[QUIET0]], [[QUIET1]]
 ; GCN-NOT: v_mul_f32
 define void @func_ieee_mode_on() #1 {
@@ -93,8 +93,8 @@ define void @func_ieee_mode_off() #2 {
 ; 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]+]]
-; GCN-DAG: v_max_f32_e32 [[QUIET0:v[0-9]+]], [[VAL0]], [[VAL0]]
-; GCN-DAG: v_max_f32_e32 [[QUIET1:v[0-9]+]], [[VAL1]], [[VAL1]]
+; 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]]
 ; 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 {
@@ -108,8 +108,8 @@ define amdgpu_cs void @cs_ieee_mode_default() #0 {
 ; 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]+]]
-; GCN-DAG: v_max_f32_e32 [[QUIET0:v[0-9]+]], [[VAL0]], [[VAL0]]
-; GCN-DAG: v_max_f32_e32 [[QUIET1:v[0-9]+]], [[VAL1]], [[VAL1]]
+; 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]]
 ; 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 {
@@ -153,8 +153,8 @@ define amdgpu_ps void @ps_ieee_mode_default() #0 {
 ; 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]+]]
-; GCN-DAG: v_max_f32_e32 [[QUIET0:v[0-9]+]], [[VAL0]], [[VAL0]]
-; GCN-DAG: v_max_f32_e32 [[QUIET1:v[0-9]+]], [[VAL1]], [[VAL1]]
+; 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]]
 ; 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 {
diff --git a/test/CodeGen/AMDGPU/clamp.ll b/test/CodeGen/AMDGPU/clamp.ll
index df48096a368..67ff2d0452a 100644
--- a/test/CodeGen/AMDGPU/clamp.ll
+++ b/test/CodeGen/AMDGPU/clamp.ll
@@ -1,5 +1,5 @@
-; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,SI %s
-; RUN: llc -march=amdgcn -mcpu=fiji -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX89,VI %s
+; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,SI,GFX678 %s
+; RUN: llc -march=amdgcn -mcpu=fiji -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX89,VI,GFX678 %s
 ; RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX89,GFX9 %s
 
 ; GCN-LABEL: {{^}}v_clamp_f32:
@@ -74,7 +74,8 @@ define amdgpu_kernel void @v_clamp_negzero_f32(float addrspace(1)* %out, float a
 
 ; GCN-LABEL: {{^}}v_clamp_negzero_maybe_snan_f32:
 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
-; GCN: v_mul_f32_e32 [[QUIET:v[0-9]+]], 1.0, [[A]]
+; GFX678: v_mul_f32_e32 [[QUIET:v[0-9]+]], 1.0, [[A]]
+; GFX9: v_max_f32_e32 [[QUIET:v[0-9]+]], [[A]], [[A]]
 ; GCN: v_max_f32_e32 [[MAX:v[0-9]+]], 0x80000000, [[QUIET]]
 ; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], 1.0, [[MAX]]
 define amdgpu_kernel void @v_clamp_negzero_maybe_snan_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 {
@@ -91,7 +92,8 @@ define amdgpu_kernel void @v_clamp_negzero_maybe_snan_f32(float addrspace(1)* %o
 
 ; GCN-LABEL: {{^}}v_clamp_multi_use_max_f32:
 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
-; GCN: v_mul_f32_e32 [[QUIET_A:v[0-9]+]], 1.0, [[A]]
+; GFX678: v_mul_f32_e32 [[QUIET_A:v[0-9]+]], 1.0, [[A]]
+; GFX9: v_max_f32_e32 [[QUIET_A:v[0-9]+]], [[A]], [[A]]
 ; GCN: v_max_f32_e32 [[MAX:v[0-9]+]], 0, [[QUIET_A]]
 ; GCN: v_min_f32_e32 [[MED:v[0-9]+]], 1.0, [[QUIET_A]]
 ; GCN-NOT: [[MAX]]
@@ -416,7 +418,8 @@ define amdgpu_kernel void @v_clamp_f32_snan_dx10clamp(float addrspace(1)* %out,
 
 ; GCN-LABEL: {{^}}v_clamp_f32_snan_no_dx10clamp:
 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
-; GCN: v_mul_f32_e32 [[QUIET_A:v[0-9]+]], 1.0, [[A]]
+; GFX678: v_mul_f32_e32 [[QUIET_A:v[0-9]+]], 1.0, [[A]]
+; GFX9: v_max_f32_e32 [[QUIET_A:v[0-9]+]], [[A]], [[A]]
 ; GCN: v_med3_f32 {{v[0-9]+}}, [[QUIET_A]], 0, 1.0
 define amdgpu_kernel void @v_clamp_f32_snan_no_dx10clamp(float addrspace(1)* %out, float addrspace(1)* %aptr) #4 {
   %tid = call i32 @llvm.amdgcn.workitem.id.x()
diff --git a/test/CodeGen/AMDGPU/fcanonicalize-elimination.ll b/test/CodeGen/AMDGPU/fcanonicalize-elimination.ll
index 3d8a1c7d745..b3fae8b5411 100644
--- a/test/CodeGen/AMDGPU/fcanonicalize-elimination.ll
+++ b/test/CodeGen/AMDGPU/fcanonicalize-elimination.ll
@@ -5,8 +5,8 @@
 ; RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs -denormal-fp-math-f32=preserve-sign < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX9,GFX9-FLUSH,GCN-FLUSH,GCN-NOEXCEPT %s
 
 ; GCN-LABEL: {{^}}test_no_fold_canonicalize_loaded_value_f32:
-; GCN-FLUSH:   v_mul_f32_e32 v{{[0-9]+}}, 1.0, v{{[0-9]+}}
-; GFX9-DENORM: v_max_f32_e32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
+; VI: v_mul_f32_e32 v{{[0-9]+}}, 1.0, v{{[0-9]+}}
+; GFX9: v_max_f32_e32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
 define amdgpu_kernel void @test_no_fold_canonicalize_loaded_value_f32(float addrspace(1)* %arg) {
   %id = tail call i32 @llvm.amdgcn.workitem.id.x()
   %gep = getelementptr inbounds float, float addrspace(1)* %arg, i32 %id
@@ -172,8 +172,9 @@ define amdgpu_kernel void @test_fold_canonicalize_fmuladd_value_f32(float addrsp
 
 ; GCN-LABEL: test_fold_canonicalize_canonicalize_value_f32:
 ; GCN: {{flat|global}}_load_dword [[LOAD:v[0-9]+]],
-; GCN-FLUSH:  v_mul_f32_e32 [[V:v[0-9]+]], 1.0, [[LOAD]]
-; GCN-DENORM: v_max_f32_e32 [[V:v[0-9]+]], [[LOAD]], [[LOAD]]
+; VI:  v_mul_f32_e32 [[V:v[0-9]+]], 1.0, [[LOAD]]
+; GFX9: v_max_f32_e32 [[V:v[0-9]+]], [[LOAD]], [[LOAD]]
+
 ; GCN-NOT: v_mul
 ; GCN-NOT: v_max
 ; GCN: {{flat|global}}_store_dword v[{{[0-9:]+}}], [[V]]
@@ -305,8 +306,8 @@ define amdgpu_kernel void @test_fold_canonicalize_fpround_value_v2f16_v2f32(<2 x
 }
 
 ; GCN-LABEL: test_no_fold_canonicalize_fneg_value_f32:
-; GCN-FLUSH:  v_mul_f32_e32 v{{[0-9]+}}, -1.0, v{{[0-9]+}}
-; GCN-DENORM: v_max_f32_e64 v{{[0-9]+}}, -v{{[0-9]+}}, -v{{[0-9]+}}
+; VI:  v_mul_f32_e32 v{{[0-9]+}}, -1.0, v{{[0-9]+}}
+; GFX9: v_max_f32_e64 v{{[0-9]+}}, -v{{[0-9]+}}, -v{{[0-9]+}}
 define amdgpu_kernel void @test_no_fold_canonicalize_fneg_value_f32(float addrspace(1)* %arg) {
   %id = tail call i32 @llvm.amdgcn.workitem.id.x()
   %gep = getelementptr inbounds float, float addrspace(1)* %arg, i32 %id
@@ -334,8 +335,8 @@ define amdgpu_kernel void @test_fold_canonicalize_fneg_value_f32(float addrspace
 }
 
 ; GCN-LABEL: test_no_fold_canonicalize_fabs_value_f32:
-; GCN-FLUSH:  v_mul_f32_e64 v{{[0-9]+}}, 1.0, |v{{[0-9]+}}|
-; GCN-DENORM: v_max_f32_e64 v{{[0-9]+}}, |v{{[0-9]+}}|, |v{{[0-9]+}}|
+; VI:  v_mul_f32_e64 v{{[0-9]+}}, 1.0, |v{{[0-9]+}}|
+; GFX9: v_max_f32_e64 v{{[0-9]+}}, |v{{[0-9]+}}|, |v{{[0-9]+}}|
 define amdgpu_kernel void @test_no_fold_canonicalize_fabs_value_f32(float addrspace(1)* %arg) {
   %id = tail call i32 @llvm.amdgcn.workitem.id.x()
   %gep = getelementptr inbounds float, float addrspace(1)* %arg, i32 %id
@@ -347,8 +348,9 @@ define amdgpu_kernel void @test_no_fold_canonicalize_fabs_value_f32(float addrsp
 }
 
 ; GCN-LABEL: test_no_fold_canonicalize_fcopysign_value_f32:
-; GCN-FLUSH:  v_mul_f32_e64 v{{[0-9]+}}, 1.0, |v{{[0-9]+}}|
-; GCN-DENORM: v_max_f32_e64 v{{[0-9]+}}, |v{{[0-9]+}}|, |v{{[0-9]+}}|
+; VI:  v_mul_f32_e64 v{{[0-9]+}}, 1.0, |v{{[0-9]+}}|
+; GFX9: v_max_f32_e64 v{{[0-9]+}}, |v{{[0-9]+}}|, |v{{[0-9]+}}|
+
 ; GCN-NOT: v_mul_
 ; GCN-NOT: v_max_
 define amdgpu_kernel void @test_no_fold_canonicalize_fcopysign_value_f32(float addrspace(1)* %arg, float %sign) {
@@ -454,10 +456,10 @@ define amdgpu_kernel void @test_fold_canonicalize_qNaN_value_f32(float addrspace
 
 ; GCN-LABEL: test_fold_canonicalize_minnum_value_from_load_f32_ieee_mode:
 ; GCN: {{flat|global}}_load_dword [[VAL:v[0-9]+]]
-; GCN-FLUSH: v_mul_f32_e32 [[QUIET:v[0-9]+]], 1.0, [[VAL]]
-; GCN-DENORM: v_max_f32_e32 [[QUIET:v[0-9]+]], [[VAL]], [[VAL]]
-; GCN: v_min_f32_e32 [[V:v[0-9]+]], 0, [[QUIET]]
+; VI: v_mul_f32_e32 [[QUIET:v[0-9]+]], 1.0, [[VAL]]
+; GFX9: v_max_f32_e32 [[QUIET:v[0-9]+]], [[VAL]], [[VAL]]
 
+; GCN: v_min_f32_e32 [[V:v[0-9]+]], 0, [[QUIET]]
 ; GCN-NOT: v_max
 ; GCN-NOT: v_mul
 
@@ -512,8 +514,8 @@ define amdgpu_kernel void @test_fold_canonicalize_minnum_value_f32(float addrspa
 
 ; GCN-LABEL: test_fold_canonicalize_sNaN_value_f32:
 ; GCN: {{flat|global}}_load_dword [[LOAD:v[0-9]+]]
-; GCN-FLUSH: v_mul_f32_e32 v{{[0-9]+}}, 1.0, [[LOAD]]
-; GCN-DENORM: v_max_f32_e32 v{{[0-9]+}}, [[LOAD]], [[LOAD]]
+; VI: v_mul_f32_e32 v{{[0-9]+}}, 1.0, [[LOAD]]
+; GFX9: v_max_f32_e32 v{{[0-9]+}}, [[LOAD]], [[LOAD]]
 define amdgpu_kernel void @test_fold_canonicalize_sNaN_value_f32(float addrspace(1)* %arg) {
   %id = tail call i32 @llvm.amdgcn.workitem.id.x()
   %gep = getelementptr inbounds float, float addrspace(1)* %arg, i32 %id
@@ -530,10 +532,9 @@ define amdgpu_kernel void @test_fold_canonicalize_sNaN_value_f32(float addrspace
 ; GFX9-DENORM: v_max_f32_e32 [[QUIET:v[0-9]+]], [[VAL]], [[VAL]]
 ; GFX9-DENORM: v_min_f32_e32 [[RESULT:v[0-9]+]], 0x7fffff, [[QUIET]]
 
-; GFX9-FLUSH: v_mul_f32_e32 [[QUIET:v[0-9]+]], 1.0, [[VAL]]
+; GFX9-FLUSH: v_max_f32_e32 [[QUIET:v[0-9]+]], [[VAL]], [[VAL]]
 ; GFX9-FLUSH: v_min_f32_e32 [[RESULT:v[0-9]+]], 0, [[QUIET]]
 
-
 ; VI-FLUSH: v_mul_f32_e32 [[QUIET_V0:v[0-9]+]], 1.0, [[VAL]]
 ; VI-FLUSH: v_min_f32_e32 [[RESULT:v[0-9]+]], 0, [[QUIET_V0]]
 
@@ -647,7 +648,9 @@ entry:
 ; GFX9-DENORM: global_load_dword [[V:v[0-9]+]],
 ; GFX9-DENORM: global_store_dword v[{{[0-9:]+}}], [[V]]
 ; GFX9-DENORM-NOT: 1.0
-; GCN-FLUSH: v_mul_f32_e32 v{{[0-9]+}}, 1.0, v{{[0-9]+}}
+; GFX9-DENORM-NOT: v_max
+; VI-FLUSH: v_mul_f32_e32 v{{[0-9]+}}, 1.0, v{{[0-9]+}}
+; GFX9-FLUSH: v_max_f32_e32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
 define amdgpu_kernel void @test_fold_canonicalize_load_nnan_value_f32(float addrspace(1)* %arg, float addrspace(1)* %out) #1 {
   %id = tail call i32 @llvm.amdgcn.workitem.id.x()
   %gep = getelementptr inbounds float, float addrspace(1)* %arg, i32 %id
@@ -736,13 +739,9 @@ define amdgpu_ps float @test_fold_canonicalize_minnum_value_no_ieee_mode(float %
 ; GFX9: v_min_f32_e32 v0, v0, v1
 ; GFX9-NEXT: s_setpc_b64
 
-; VI-FLUSH-DAG: v_mul_f32_e32 v0, 1.0, v0
-; VI-FLUSH-DAG: v_mul_f32_e32 v1, 1.0, v1
-; VI-FLUSH: v_min_f32_e32 v0, v0, v1
-
-; VI-DENORM-DAG: v_max_f32_e32 v0, v0, v0
-; VI-DENORM-DAG: v_max_f32_e32 v1, v1, v1
-; VI-DENORM: v_min_f32_e32 v0, v0, v1
+; VI-DAG: v_mul_f32_e32 v0, 1.0, v0
+; VI-DAG: v_mul_f32_e32 v1, 1.0, v1
+; VI: v_min_f32_e32 v0, v0, v1
 
 ; VI-NEXT: s_setpc_b64
 define float @test_fold_canonicalize_minnum_value_ieee_mode(float %arg0, float %arg1) {
diff --git a/test/CodeGen/AMDGPU/fcanonicalize.f16.ll b/test/CodeGen/AMDGPU/fcanonicalize.f16.ll
index e179ef37b3f..7a44d11ad09 100644
--- a/test/CodeGen/AMDGPU/fcanonicalize.f16.ll
+++ b/test/CodeGen/AMDGPU/fcanonicalize.f16.ll
@@ -98,7 +98,8 @@ define amdgpu_kernel void @v_test_canonicalize_fneg_var_f16(half addrspace(1)* %
 }
 
 ; GCN-LABEL: {{^}}v_test_no_denormals_canonicalize_fneg_var_f16:
-; GFX89: v_mul_f16_e32 [[REG:v[0-9]+]], -1.0, v{{[0-9]+}}
+; VI: v_mul_f16_e32 [[REG:v[0-9]+]], -1.0, v{{[0-9]+}}
+; GFX9: v_max_f16_e64 [[REG:v[0-9]+]], -v{{[0-9]+}}, -v{{[0-9]+}}
 ; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @v_test_no_denormals_canonicalize_fneg_var_f16(half addrspace(1)* %out) #2 {
   %val = load half, half addrspace(1)* %out
@@ -109,7 +110,9 @@ define amdgpu_kernel void @v_test_no_denormals_canonicalize_fneg_var_f16(half ad
 }
 
 ; GCN-LABEL: {{^}}v_test_no_denormals_canonicalize_fneg_fabs_var_f16:
-; GFX89: v_mul_f16_e64 [[REG:v[0-9]+]], -1.0, |v{{[0-9]+}}|
+; VI: v_mul_f16_e64 [[REG:v[0-9]+]], -1.0, |v{{[0-9]+}}|
+; GFX9: v_max_f16_e64 [[REG:v[0-9]+]], -|v{{[0-9]+}}|, -|v{{[0-9]+}}|
+
 ; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 
 ; CI: v_cvt_f32_f16_e64 {{v[0-9]+}}, -|{{v[0-9]+}}|
diff --git a/test/CodeGen/AMDGPU/fcanonicalize.ll b/test/CodeGen/AMDGPU/fcanonicalize.ll
index ad472a92f86..e302d59b1ac 100644
--- a/test/CodeGen/AMDGPU/fcanonicalize.ll
+++ b/test/CodeGen/AMDGPU/fcanonicalize.ll
@@ -1,4 +1,6 @@
-; RUN:  llc -amdgpu-scalarize-global-loads=false -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
+; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn-amd-amdhsa -mcpu=hawaii -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX6,GFX678 %s
+; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn-amd-amdhsa -mcpu=fiji -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX8,GFX678 %s
+; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX9 %s
 
 declare float @llvm.fabs.f32(float) #0
 declare float @llvm.canonicalize.f32(float) #0
@@ -16,8 +18,9 @@ declare <2 x half> @llvm.canonicalize.v2f16(<2 x half>) #0
 declare i32 @llvm.amdgcn.workitem.id.x() #0
 
 ; GCN-LABEL: {{^}}v_test_canonicalize_var_f32:
-; GCN: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
-; GCN: buffer_store_dword [[REG]]
+; GFX678: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
+; GFX9: v_max_f32_e32 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}}
+; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @v_test_canonicalize_var_f32(float addrspace(1)* %out) #1 {
   %val = load float, float addrspace(1)* %out
   %canonicalized = call float @llvm.canonicalize.f32(float %val)
@@ -26,8 +29,9 @@ define amdgpu_kernel void @v_test_canonicalize_var_f32(float addrspace(1)* %out)
 }
 
 ; GCN-LABEL: {{^}}s_test_canonicalize_var_f32:
-; GCN: v_mul_f32_e64 [[REG:v[0-9]+]], 1.0, {{s[0-9]+}}
-; GCN: buffer_store_dword [[REG]]
+; GFX678: v_mul_f32_e64 [[REG:v[0-9]+]], 1.0, {{s[0-9]+}}
+; GFX9: v_max_f32_e64 [[REG:v[0-9]+]], {{s[0-9]+}}, {{s[0-9]+}}
+; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @s_test_canonicalize_var_f32(float addrspace(1)* %out, float %val) #1 {
   %canonicalized = call float @llvm.canonicalize.f32(float %val)
   store float %canonicalized, float addrspace(1)* %out
@@ -35,8 +39,9 @@ define amdgpu_kernel void @s_test_canonicalize_var_f32(float addrspace(1)* %out,
 }
 
 ; GCN-LABEL: {{^}}v_test_canonicalize_fabs_var_f32:
-; GCN: v_mul_f32_e64 [[REG:v[0-9]+]], 1.0, |{{v[0-9]+}}|
-; GCN: buffer_store_dword [[REG]]
+; GFX678: v_mul_f32_e64 [[REG:v[0-9]+]], 1.0, |{{v[0-9]+}}|
+; GFX9: v_max_f32_e64 [[REG:v[0-9]+]], |{{v[0-9]+}}|, |{{v[0-9]+}}|
+; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @v_test_canonicalize_fabs_var_f32(float addrspace(1)* %out) #1 {
   %val = load float, float addrspace(1)* %out
   %val.fabs = call float @llvm.fabs.f32(float %val)
@@ -46,8 +51,9 @@ define amdgpu_kernel void @v_test_canonicalize_fabs_var_f32(float addrspace(1)*
 }
 
 ; GCN-LABEL: {{^}}v_test_canonicalize_fneg_fabs_var_f32:
-; GCN: v_mul_f32_e64 [[REG:v[0-9]+]], -1.0, |{{v[0-9]+}}|
-; GCN: buffer_store_dword [[REG]]
+; GFX678: v_mul_f32_e64 [[REG:v[0-9]+]], -1.0, |{{v[0-9]+}}|
+; GFX9: v_max_f32_e64 [[REG:v[0-9]+]], -|{{v[0-9]+}}|, -|{{v[0-9]+}}|
+; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @v_test_canonicalize_fneg_fabs_var_f32(float addrspace(1)* %out) #1 {
   %val = load float, float addrspace(1)* %out
   %val.fabs = call float @llvm.fabs.f32(float %val)
@@ -58,8 +64,9 @@ define amdgpu_kernel void @v_test_canonicalize_fneg_fabs_var_f32(float addrspace
 }
 
 ; GCN-LABEL: {{^}}v_test_canonicalize_fneg_var_f32:
-; GCN: v_mul_f32_e32 [[REG:v[0-9]+]], -1.0, {{v[0-9]+}}
-; GCN: buffer_store_dword [[REG]]
+; GFX678: v_mul_f32_e32 [[REG:v[0-9]+]], -1.0, {{v[0-9]+}}
+; GFX9: v_max_f32_e64 [[REG:v[0-9]+]], -{{v[0-9]+}}, -{{v[0-9]+}}
+; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @v_test_canonicalize_fneg_var_f32(float addrspace(1)* %out) #1 {
   %val = load float, float addrspace(1)* %out
   %val.fneg = fsub float -0.0, %val
@@ -70,7 +77,7 @@ define amdgpu_kernel void @v_test_canonicalize_fneg_var_f32(float addrspace(1)*
 
 ; GCN-LABEL: {{^}}test_fold_canonicalize_undef_f32:
 ; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7fc00000{{$}}
-; GCN: buffer_store_dword [[REG]]
+; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @test_fold_canonicalize_undef_f32(float addrspace(1)* %out) #1 {
   %canonicalized = call float @llvm.canonicalize.f32(float undef)
   store float %canonicalized, float addrspace(1)* %out
@@ -79,7 +86,7 @@ define amdgpu_kernel void @test_fold_canonicalize_undef_f32(float addrspace(1)*
 
 ; GCN-LABEL: {{^}}test_fold_canonicalize_p0_f32:
 ; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0{{$}}
-; GCN: buffer_store_dword [[REG]]
+; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @test_fold_canonicalize_p0_f32(float addrspace(1)* %out) #1 {
   %canonicalized = call float @llvm.canonicalize.f32(float 0.0)
   store float %canonicalized, float addrspace(1)* %out
@@ -88,7 +95,7 @@ define amdgpu_kernel void @test_fold_canonicalize_p0_f32(float addrspace(1)* %ou
 
 ; GCN-LABEL: {{^}}test_fold_canonicalize_n0_f32:
 ; GCN: v_bfrev_b32_e32 [[REG:v[0-9]+]], 1{{$}}
-; GCN: buffer_store_dword [[REG]]
+; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @test_fold_canonicalize_n0_f32(float addrspace(1)* %out) #1 {
   %canonicalized = call float @llvm.canonicalize.f32(float -0.0)
   store float %canonicalized, float addrspace(1)* %out
@@ -97,7 +104,7 @@ define amdgpu_kernel void @test_fold_canonicalize_n0_f32(float addrspace(1)* %ou
 
 ; GCN-LABEL: {{^}}test_fold_canonicalize_p1_f32:
 ; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 1.0{{$}}
-; GCN: buffer_store_dword [[REG]]
+; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @test_fold_canonicalize_p1_f32(float addrspace(1)* %out) #1 {
   %canonicalized = call float @llvm.canonicalize.f32(float 1.0)
   store float %canonicalized, float addrspace(1)* %out
@@ -106,7 +113,7 @@ define amdgpu_kernel void @test_fold_canonicalize_p1_f32(float addrspace(1)* %ou
 
 ; GCN-LABEL: {{^}}test_fold_canonicalize_n1_f32:
 ; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], -1.0{{$}}
-; GCN: buffer_store_dword [[REG]]
+; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @test_fold_canonicalize_n1_f32(float addrspace(1)* %out) #1 {
   %canonicalized = call float @llvm.canonicalize.f32(float -1.0)
   store float %canonicalized, float addrspace(1)* %out
@@ -115,7 +122,7 @@ define amdgpu_kernel void @test_fold_canonicalize_n1_f32(float addrspace(1)* %ou
 
 ; GCN-LABEL: {{^}}test_fold_canonicalize_literal_f32:
 ; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x41800000{{$}}
-; GCN: buffer_store_dword [[REG]]
+; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @test_fold_canonicalize_literal_f32(float addrspace(1)* %out) #1 {
   %canonicalized = call float @llvm.canonicalize.f32(float 16.0)
   store float %canonicalized, float addrspace(1)* %out
@@ -124,7 +131,7 @@ define amdgpu_kernel void @test_fold_canonicalize_literal_f32(float addrspace(1)
 
 ; GCN-LABEL: {{^}}test_no_denormals_fold_canonicalize_denormal0_f32:
 ; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0{{$}}
-; GCN: buffer_store_dword [[REG]]
+; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @test_no_denormals_fold_canonicalize_denormal0_f32(float addrspace(1)* %out) #1 {
   %canonicalized = call float @llvm.canonicalize.f32(float bitcast (i32 8388607 to float))
   store float %canonicalized, float addrspace(1)* %out
@@ -133,7 +140,7 @@ define amdgpu_kernel void @test_no_denormals_fold_canonicalize_denormal0_f32(flo
 
 ; GCN-LABEL: {{^}}test_denormals_fold_canonicalize_denormal0_f32:
 ; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7fffff{{$}}
-; GCN: buffer_store_dword [[REG]]
+; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @test_denormals_fold_canonicalize_denormal0_f32(float addrspace(1)* %out) #3 {
   %canonicalized = call float @llvm.canonicalize.f32(float bitcast (i32 8388607 to float))
   store float %canonicalized, float addrspace(1)* %out
@@ -142,7 +149,7 @@ define amdgpu_kernel void @test_denormals_fold_canonicalize_denormal0_f32(float
 
 ; GCN-LABEL: {{^}}test_no_denormals_fold_canonicalize_denormal1_f32:
 ; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0{{$}}
-; GCN: buffer_store_dword [[REG]]
+; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @test_no_denormals_fold_canonicalize_denormal1_f32(float addrspace(1)* %out) #1 {
   %canonicalized = call float @llvm.canonicalize.f32(float bitcast (i32 2155872255 to float))
   store float %canonicalized, float addrspace(1)* %out
@@ -151,7 +158,7 @@ define amdgpu_kernel void @test_no_denormals_fold_canonicalize_denormal1_f32(flo
 
 ; GCN-LABEL: {{^}}test_denormals_fold_canonicalize_denormal1_f32:
 ; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x807fffff{{$}}
-; GCN: buffer_store_dword [[REG]]
+; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @test_denormals_fold_canonicalize_denormal1_f32(float addrspace(1)* %out) #3 {
   %canonicalized = call float @llvm.canonicalize.f32(float bitcast (i32 2155872255 to float))
   store float %canonicalized, float addrspace(1)* %out
@@ -160,7 +167,7 @@ define amdgpu_kernel void @test_denormals_fold_canonicalize_denormal1_f32(float
 
 ; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_f32:
 ; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7fc00000{{$}}
-; GCN: buffer_store_dword [[REG]]
+; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @test_fold_canonicalize_qnan_f32(float addrspace(1)* %out) #1 {
   %canonicalized = call float @llvm.canonicalize.f32(float 0x7FF8000000000000)
   store float %canonicalized, float addrspace(1)* %out
@@ -169,7 +176,7 @@ define amdgpu_kernel void @test_fold_canonicalize_qnan_f32(float addrspace(1)* %
 
 ; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_value_neg1_f32:
 ; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7fc00000{{$}}
-; GCN: buffer_store_dword [[REG]]
+; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @test_fold_canonicalize_qnan_value_neg1_f32(float addrspace(1)* %out) #1 {
   %canonicalized = call float @llvm.canonicalize.f32(float bitcast (i32 -1 to float))
   store float %canonicalized, float addrspace(1)* %out
@@ -178,7 +185,7 @@ define amdgpu_kernel void @test_fold_canonicalize_qnan_value_neg1_f32(float addr
 
 ; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_value_neg2_f32:
 ; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7fc00000{{$}}
-; GCN: buffer_store_dword [[REG]]
+; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @test_fold_canonicalize_qnan_value_neg2_f32(float addrspace(1)* %out) #1 {
   %canonicalized = call float @llvm.canonicalize.f32(float bitcast (i32 -2 to float))
   store float %canonicalized, float addrspace(1)* %out
@@ -187,7 +194,7 @@ define amdgpu_kernel void @test_fold_canonicalize_qnan_value_neg2_f32(float addr
 
 ; GCN-LABEL: {{^}}test_fold_canonicalize_snan0_value_f32:
 ; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7fc00000{{$}}
-; GCN: buffer_store_dword [[REG]]
+; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @test_fold_canonicalize_snan0_value_f32(float addrspace(1)* %out) #1 {
   %canonicalized = call float @llvm.canonicalize.f32(float bitcast (i32 2139095041 to float))
   store float %canonicalized, float addrspace(1)* %out
@@ -196,7 +203,7 @@ define amdgpu_kernel void @test_fold_canonicalize_snan0_value_f32(float addrspac
 
 ; GCN-LABEL: {{^}}test_fold_canonicalize_snan1_value_f32:
 ; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7fc00000{{$}}
-; GCN: buffer_store_dword [[REG]]
+; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @test_fold_canonicalize_snan1_value_f32(float addrspace(1)* %out) #1 {
   %canonicalized = call float @llvm.canonicalize.f32(float bitcast (i32 2143289343 to float))
   store float %canonicalized, float addrspace(1)* %out
@@ -205,7 +212,7 @@ define amdgpu_kernel void @test_fold_canonicalize_snan1_value_f32(float addrspac
 
 ; GCN-LABEL: {{^}}test_fold_canonicalize_snan2_value_f32:
 ; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7fc00000{{$}}
-; GCN: buffer_store_dword [[REG]]
+; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @test_fold_canonicalize_snan2_value_f32(float addrspace(1)* %out) #1 {
   %canonicalized = call float @llvm.canonicalize.f32(float bitcast (i32 4286578689 to float))
   store float %canonicalized, float addrspace(1)* %out
@@ -214,7 +221,7 @@ define amdgpu_kernel void @test_fold_canonicalize_snan2_value_f32(float addrspac
 
 ; GCN-LABEL: {{^}}test_fold_canonicalize_snan3_value_f32:
 ; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7fc00000{{$}}
-; GCN: buffer_store_dword [[REG]]
+; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @test_fold_canonicalize_snan3_value_f32(float addrspace(1)* %out) #1 {
   %canonicalized = call float @llvm.canonicalize.f32(float bitcast (i32 4290772991 to float))
   store float %canonicalized, float addrspace(1)* %out
@@ -223,7 +230,7 @@ define amdgpu_kernel void @test_fold_canonicalize_snan3_value_f32(float addrspac
 
 ; GCN-LABEL: {{^}}v_test_canonicalize_var_f64:
 ; GCN: v_max_f64 [[REG:v\[[0-9]+:[0-9]+\]]], {{v\[[0-9]+:[0-9]+\]}}, {{v\[[0-9]+:[0-9]+\]}}
-; GCN: buffer_store_dwordx2 [[REG]]
+; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @v_test_canonicalize_var_f64(double addrspace(1)* %out) #1 {
   %val = load double, double addrspace(1)* %out
   %canonicalized = call double @llvm.canonicalize.f64(double %val)
@@ -233,7 +240,7 @@ define amdgpu_kernel void @v_test_canonicalize_var_f64(double addrspace(1)* %out
 
 ; GCN-LABEL: {{^}}s_test_canonicalize_var_f64:
 ; GCN: v_max_f64 [[REG:v\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}
-; GCN: buffer_store_dwordx2 [[REG]]
+; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @s_test_canonicalize_var_f64(double addrspace(1)* %out, double %val) #1 {
   %canonicalized = call double @llvm.canonicalize.f64(double %val)
   store double %canonicalized, double addrspace(1)* %out
@@ -242,7 +249,7 @@ define amdgpu_kernel void @s_test_canonicalize_var_f64(double addrspace(1)* %out
 
 ; GCN-LABEL: {{^}}v_test_canonicalize_fabs_var_f64:
 ; GCN: v_max_f64 [[REG:v\[[0-9]+:[0-9]+\]]], |{{v\[[0-9]+:[0-9]+\]}}|, |{{v\[[0-9]+:[0-9]+\]}}|
-; GCN: buffer_store_dwordx2 [[REG]]
+; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @v_test_canonicalize_fabs_var_f64(double addrspace(1)* %out) #1 {
   %val = load double, double addrspace(1)* %out
   %val.fabs = call double @llvm.fabs.f64(double %val)
@@ -253,7 +260,7 @@ define amdgpu_kernel void @v_test_canonicalize_fabs_var_f64(double addrspace(1)*
 
 ; GCN-LABEL: {{^}}v_test_canonicalize_fneg_fabs_var_f64:
 ; GCN: v_max_f64 [[REG:v\[[0-9]+:[0-9]\]]], -|{{v\[[0-9]+:[0-9]+\]}}|, -|{{v\[[0-9]+:[0-9]+\]}}|
-; GCN: buffer_store_dwordx2 [[REG]]
+; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @v_test_canonicalize_fneg_fabs_var_f64(double addrspace(1)* %out) #1 {
   %val = load double, double addrspace(1)* %out
   %val.fabs = call double @llvm.fabs.f64(double %val)
@@ -265,7 +272,7 @@ define amdgpu_kernel void @v_test_canonicalize_fneg_fabs_var_f64(double addrspac
 
 ; GCN-LABEL: {{^}}v_test_canonicalize_fneg_var_f64:
 ; GCN: v_max_f64 [[REG:v\[[0-9]+:[0-9]+\]]], -{{v\[[0-9]+:[0-9]+\]}}, -{{v\[[0-9]+:[0-9]+\]}}
-; GCN: buffer_store_dwordx2 [[REG]]
+; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @v_test_canonicalize_fneg_var_f64(double addrspace(1)* %out) #1 {
   %val = load double, double addrspace(1)* %out
   %val.fneg = fsub double -0.0, %val
@@ -277,7 +284,7 @@ define amdgpu_kernel void @v_test_canonicalize_fneg_var_f64(double addrspace(1)*
 ; GCN-LABEL: {{^}}test_fold_canonicalize_p0_f64:
 ; GCN: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}}
 ; GCN: v_mov_b32_e32 v[[HI:[0-9]+]], v[[LO]]{{$}}
-; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
+; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[LO]]:[[HI]]{{\]}}
 define amdgpu_kernel void @test_fold_canonicalize_p0_f64(double addrspace(1)* %out) #1 {
   %canonicalized = call double @llvm.canonicalize.f64(double 0.0)
   store double %canonicalized, double addrspace(1)* %out
@@ -287,7 +294,7 @@ define amdgpu_kernel void @test_fold_canonicalize_p0_f64(double addrspace(1)* %o
 ; GCN-LABEL: {{^}}test_fold_canonicalize_n0_f64:
 ; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}}
 ; GCN-DAG: v_bfrev_b32_e32 v[[HI:[0-9]+]], 1{{$}}
-; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
+; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[LO]]:[[HI]]{{\]}}
 define amdgpu_kernel void @test_fold_canonicalize_n0_f64(double addrspace(1)* %out) #1 {
   %canonicalized = call double @llvm.canonicalize.f64(double -0.0)
   store double %canonicalized, double addrspace(1)* %out
@@ -297,7 +304,7 @@ define amdgpu_kernel void @test_fold_canonicalize_n0_f64(double addrspace(1)* %o
 ; GCN-LABEL: {{^}}test_fold_canonicalize_p1_f64:
 ; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}}
 ; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0x3ff00000{{$}}
-; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
+; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[LO]]:[[HI]]{{\]}}
 define amdgpu_kernel void @test_fold_canonicalize_p1_f64(double addrspace(1)* %out) #1 {
   %canonicalized = call double @llvm.canonicalize.f64(double 1.0)
   store double %canonicalized, double addrspace(1)* %out
@@ -307,7 +314,7 @@ define amdgpu_kernel void @test_fold_canonicalize_p1_f64(double addrspace(1)* %o
 ; GCN-LABEL: {{^}}test_fold_canonicalize_n1_f64:
 ; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}}
 ; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0xbff00000{{$}}
-; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
+; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[LO]]:[[HI]]{{\]}}
 define amdgpu_kernel void @test_fold_canonicalize_n1_f64(double addrspace(1)* %out) #1 {
   %canonicalized = call double @llvm.canonicalize.f64(double -1.0)
   store double %canonicalized, double addrspace(1)* %out
@@ -317,7 +324,7 @@ define amdgpu_kernel void @test_fold_canonicalize_n1_f64(double addrspace(1)* %o
 ; GCN-LABEL: {{^}}test_fold_canonicalize_literal_f64:
 ; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}}
 ; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0x40300000{{$}}
-; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
+; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[LO]]:[[HI]]{{\]}}
 define amdgpu_kernel void @test_fold_canonicalize_literal_f64(double addrspace(1)* %out) #1 {
   %canonicalized = call double @llvm.canonicalize.f64(double 16.0)
   store double %canonicalized, double addrspace(1)* %out
@@ -327,7 +334,7 @@ define amdgpu_kernel void @test_fold_canonicalize_literal_f64(double addrspace(1
 ; GCN-LABEL: {{^}}test_no_denormals_fold_canonicalize_denormal0_f64:
 ; GCN: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}}
 ; GCN: v_mov_b32_e32 v[[HI:[0-9]+]], v[[LO]]{{$}}
-; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
+; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[LO]]:[[HI]]{{\]}}
 define amdgpu_kernel void @test_no_denormals_fold_canonicalize_denormal0_f64(double addrspace(1)* %out) #2 {
   %canonicalized = call double @llvm.canonicalize.f64(double bitcast (i64 4503599627370495 to double))
   store double %canonicalized, double addrspace(1)* %out
@@ -337,7 +344,7 @@ define amdgpu_kernel void @test_no_denormals_fold_canonicalize_denormal0_f64(dou
 ; GCN-LABEL: {{^}}test_denormals_fold_canonicalize_denormal0_f64:
 ; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], -1{{$}}
 ; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0xfffff{{$}}
-; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
+; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[LO]]:[[HI]]{{\]}}
 define amdgpu_kernel void @test_denormals_fold_canonicalize_denormal0_f64(double addrspace(1)* %out) #3 {
   %canonicalized = call double @llvm.canonicalize.f64(double bitcast (i64 4503599627370495 to double))
   store double %canonicalized, double addrspace(1)* %out
@@ -347,7 +354,7 @@ define amdgpu_kernel void @test_denormals_fold_canonicalize_denormal0_f64(double
 ; GCN-LABEL: {{^}}test_no_denormals_fold_canonicalize_denormal1_f64:
 ; GCN: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}}
 ; GCN: v_mov_b32_e32 v[[HI:[0-9]+]], v[[LO]]{{$}}
-; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
+; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[LO]]:[[HI]]{{\]}}
 define amdgpu_kernel void @test_no_denormals_fold_canonicalize_denormal1_f64(double addrspace(1)* %out) #2 {
   %canonicalized = call double @llvm.canonicalize.f64(double bitcast (i64 9227875636482146303 to double))
   store double %canonicalized, double addrspace(1)* %out
@@ -357,7 +364,7 @@ define amdgpu_kernel void @test_no_denormals_fold_canonicalize_denormal1_f64(dou
 ; GCN-LABEL: {{^}}test_denormals_fold_canonicalize_denormal1_f64:
 ; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], -1{{$}}
 ; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0x800fffff{{$}}
-; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
+; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[LO]]:[[HI]]{{\]}}
 define amdgpu_kernel void @test_denormals_fold_canonicalize_denormal1_f64(double addrspace(1)* %out) #3 {
   %canonicalized = call double @llvm.canonicalize.f64(double bitcast (i64 9227875636482146303 to double))
   store double %canonicalized, double addrspace(1)* %out
@@ -367,7 +374,7 @@ define amdgpu_kernel void @test_denormals_fold_canonicalize_denormal1_f64(double
 ; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_f64:
 ; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0x7ff80000{{$}}
 ; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}}
-; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
+; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[LO]]:[[HI]]{{\]}}
 define amdgpu_kernel void @test_fold_canonicalize_qnan_f64(double addrspace(1)* %out) #1 {
   %canonicalized = call double @llvm.canonicalize.f64(double 0x7FF8000000000000)
   store double %canonicalized, double addrspace(1)* %out
@@ -377,7 +384,7 @@ define amdgpu_kernel void @test_fold_canonicalize_qnan_f64(double addrspace(1)*
 ; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_value_neg1_f64:
 ; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0x7ff80000{{$}}
 ; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}}
-; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
+; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[LO]]:[[HI]]{{\]}}
 define amdgpu_kernel void @test_fold_canonicalize_qnan_value_neg1_f64(double addrspace(1)* %out) #1 {
   %canonicalized = call double @llvm.canonicalize.f64(double bitcast (i64 -1 to double))
   store double %canonicalized, double addrspace(1)* %out
@@ -387,7 +394,7 @@ define amdgpu_kernel void @test_fold_canonicalize_qnan_value_neg1_f64(double add
 ; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_value_neg2_f64:
 ; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0x7ff80000{{$}}
 ; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}}
-; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
+; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[LO]]:[[HI]]{{\]}}
 define amdgpu_kernel void @test_fold_canonicalize_qnan_value_neg2_f64(double addrspace(1)* %out) #1 {
   %canonicalized = call double @llvm.canonicalize.f64(double bitcast (i64 -2 to double))
   store double %canonicalized, double addrspace(1)* %out
@@ -397,7 +404,7 @@ define amdgpu_kernel void @test_fold_canonicalize_qnan_value_neg2_f64(double add
 ; GCN-LABEL: {{^}}test_fold_canonicalize_snan0_value_f64:
 ; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0x7ff80000{{$}}
 ; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}}
-; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
+; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[LO]]:[[HI]]{{\]}}
 define amdgpu_kernel void @test_fold_canonicalize_snan0_value_f64(double addrspace(1)* %out) #1 {
   %canonicalized = call double @llvm.canonicalize.f64(double bitcast (i64 9218868437227405313 to double))
   store double %canonicalized, double addrspace(1)* %out
@@ -407,7 +414,7 @@ define amdgpu_kernel void @test_fold_canonicalize_snan0_value_f64(double addrspa
 ; GCN-LABEL: {{^}}test_fold_canonicalize_snan1_value_f64:
 ; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0x7ff80000{{$}}
 ; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}}
-; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
+; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[LO]]:[[HI]]{{\]}}
 define amdgpu_kernel void @test_fold_canonicalize_snan1_value_f64(double addrspace(1)* %out) #1 {
   %canonicalized = call double @llvm.canonicalize.f64(double bitcast (i64 9223372036854775807 to double))
   store double %canonicalized, double addrspace(1)* %out
@@ -417,7 +424,7 @@ define amdgpu_kernel void @test_fold_canonicalize_snan1_value_f64(double addrspa
 ; GCN-LABEL: {{^}}test_fold_canonicalize_snan2_value_f64:
 ; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0x7ff80000{{$}}
 ; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}}
-; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
+; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[LO]]:[[HI]]{{\]}}
 define amdgpu_kernel void @test_fold_canonicalize_snan2_value_f64(double addrspace(1)* %out) #1 {
   %canonicalized = call double @llvm.canonicalize.f64(double bitcast (i64 18442240474082181121 to double))
   store double %canonicalized, double addrspace(1)* %out
@@ -427,7 +434,7 @@ define amdgpu_kernel void @test_fold_canonicalize_snan2_value_f64(double addrspa
 ; GCN-LABEL: {{^}}test_fold_canonicalize_snan3_value_f64:
 ; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0x7ff80000{{$}}
 ; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}}
-; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
+; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[LO]]:[[HI]]{{\]}}
 define amdgpu_kernel void @test_fold_canonicalize_snan3_value_f64(double addrspace(1)* %out) #1 {
   %canonicalized = call double @llvm.canonicalize.f64(double bitcast (i64 18446744073709551615 to double))
   store double %canonicalized, double addrspace(1)* %out
@@ -435,7 +442,8 @@ define amdgpu_kernel void @test_fold_canonicalize_snan3_value_f64(double addrspa
 }
 
 ; GCN-LABEL:  {{^}}test_canonicalize_value_f64_flush:
-; GCN: v_mul_f64 v[{{[0-9:]+}}], 1.0, v[{{[0-9:]+}}]
+; GFX678: v_mul_f64 v[{{[0-9:]+}}], 1.0, v[{{[0-9:]+}}]
+; GCN9: v_max_f64 v[{{[0-9:]+}}], v[{{[0-9:]+}}], v[{{[0-9:]+}}]
 define amdgpu_kernel void @test_canonicalize_value_f64_flush(double addrspace(1)* %arg, double addrspace(1)* %out) #4 {
   %id = tail call i32 @llvm.amdgcn.workitem.id.x()
   %gep = getelementptr inbounds double, double addrspace(1)* %arg, i32 %id
@@ -447,7 +455,8 @@ define amdgpu_kernel void @test_canonicalize_value_f64_flush(double addrspace(1)
 }
 
 ; GCN-LABEL:  {{^}}test_canonicalize_value_f32_flush:
-; GCN: v_mul_f32_e32 {{v[0-9]+}}, 1.0, {{v[0-9]+}}
+; GFX6: v_mul_f32_e32 {{v[0-9]+}}, 1.0, {{v[0-9]+}}
+; GFX9: v_max_f32_e32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
 define amdgpu_kernel void @test_canonicalize_value_f32_flush(float addrspace(1)* %arg, float addrspace(1)* %out) #4 {
   %id = tail call i32 @llvm.amdgcn.workitem.id.x()
   %gep = getelementptr inbounds float, float addrspace(1)* %arg, i32 %id
@@ -459,7 +468,8 @@ define amdgpu_kernel void @test_canonicalize_value_f32_flush(float addrspace(1)*
 }
 
 ; GCN-LABEL:  {{^}}test_canonicalize_value_f16_flush:
-; GCN: v_mul_f16_e32 {{v[0-9]+}}, 1.0, {{v[0-9]+}}
+; GFX8: v_mul_f16_e32 {{v[0-9]+}}, 1.0, {{v[0-9]+}}
+; GFX9: v_max_f16_e32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
 define amdgpu_kernel void @test_canonicalize_value_f16_flush(half addrspace(1)* %arg, half addrspace(1)* %out) #4 {
   %id = tail call i32 @llvm.amdgcn.workitem.id.x()
   %gep = getelementptr inbounds half, half addrspace(1)* %arg, i32 %id
@@ -470,23 +480,13 @@ define amdgpu_kernel void @test_canonicalize_value_f16_flush(half addrspace(1)*
   ret void
 }
 
-; GCN-LABEL:  {{^}}test_canonicalize_value_v2f16_flush_gfx8:
-; GCN:     v_mov_b32_e32 [[ONE:v[0-9]+]], 0x3c00
-; GCN-DAG: v_mul_f16_sdwa v{{[0-9]+}}, [[ONE]], v{{[0-9]+}} dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-; GCN-DAG: v_mul_f16_e32 v{{[0-9]+}}, 1.0, v{{[0-9]+}}
-define amdgpu_kernel void @test_canonicalize_value_v2f16_flush_gfx8(<2 x half> addrspace(1)* %arg, <2 x half> addrspace(1)* %out) #4 {
-  %id = tail call i32 @llvm.amdgcn.workitem.id.x()
-  %gep = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %arg, i32 %id
-  %v = load <2 x half>, <2 x half> addrspace(1)* %gep, align 4
-  %canonicalized = tail call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %v)
-  %gep2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %out, i32 %id
-  store <2 x half> %canonicalized, <2 x half> addrspace(1)* %gep2, align 2
-  ret void
-}
+; GCN-LABEL:  {{^}}test_canonicalize_value_v2f16_flush:
+; GFX8:     v_mov_b32_e32 [[ONE:v[0-9]+]], 0x3c00
+; GFX8-DAG: v_mul_f16_sdwa v{{[0-9]+}}, [[ONE]], v{{[0-9]+}} dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-DAG: v_mul_f16_e32 v{{[0-9]+}}, 1.0, v{{[0-9]+}}
 
-; GCN-LABEL:  {{^}}test_canonicalize_value_v2f16_flush_gfx9:
-; GCN-DAG: v_pk_mul_f16 v{{[0-9]+}}, 1.0, v{{[0-9]+}} op_sel_hi:[0,1]{{$}}
-define amdgpu_kernel void @test_canonicalize_value_v2f16_flush_gfx9(<2 x half> addrspace(1)* %arg, <2 x half> addrspace(1)* %out) #6 {
+; GFX9: v_pk_max_f16 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}{{$}}
+define amdgpu_kernel void @test_canonicalize_value_v2f16_flush(<2 x half> addrspace(1)* %arg, <2 x half> addrspace(1)* %out) #4 {
   %id = tail call i32 @llvm.amdgcn.workitem.id.x()
   %gep = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %arg, i32 %id
   %v = load <2 x half>, <2 x half> addrspace(1)* %gep, align 4
@@ -498,7 +498,7 @@ define amdgpu_kernel void @test_canonicalize_value_v2f16_flush_gfx9(<2 x half> a
 
 ; GCN-LABEL:  {{^}}test_canonicalize_value_f64_denorm:
 ; GCN: v_max_f64 v[{{[0-9:]+}}], v[{{[0-9:]+}}], v[{{[0-9:]+}}]
-define amdgpu_kernel void @test_canonicalize_value_f64_denorm(double addrspace(1)* %arg, double addrspace(1)* %out) #5 {
+define amdgpu_kernel void @test_canonicalize_value_f64_denorm(double addrspace(1)* %arg, double addrspace(1)* %out) #3 {
   %id = tail call i32 @llvm.amdgcn.workitem.id.x()
   %gep = getelementptr inbounds double, double addrspace(1)* %arg, i32 %id
   %v = load double, double addrspace(1)* %gep, align 8
@@ -509,8 +509,9 @@ define amdgpu_kernel void @test_canonicalize_value_f64_denorm(double addrspace(1
 }
 
 ; GCN-LABEL:  {{^}}test_canonicalize_value_f32_denorm:
-; GCN: v_max_f32_e32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
-define amdgpu_kernel void @test_canonicalize_value_f32_denorm(float addrspace(1)* %arg, float addrspace(1)* %out) #5 {
+; GFX678: v_mul_f32_e32 {{v[0-9]+}}, 1.0, {{v[0-9]+}}
+; GFX9: v_max_f32_e32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
+define amdgpu_kernel void @test_canonicalize_value_f32_denorm(float addrspace(1)* %arg, float addrspace(1)* %out) #3 {
   %id = tail call i32 @llvm.amdgcn.workitem.id.x()
   %gep = getelementptr inbounds float, float addrspace(1)* %arg, i32 %id
   %v = load float, float addrspace(1)* %gep, align 4
@@ -520,9 +521,12 @@ define amdgpu_kernel void @test_canonicalize_value_f32_denorm(float addrspace(1)
   ret void
 }
 
+; FIXME: Conversion to float should count as the canonicalize pre-gfx8
 ; GCN-LABEL:  {{^}}test_canonicalize_value_f16_denorm:
-; GCN: v_max_f16_e32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
-define amdgpu_kernel void @test_canonicalize_value_f16_denorm(half addrspace(1)* %arg, half addrspace(1)* %out) #5 {
+; GFX6: v_mul_f32_e32 {{v[0-9]+}}, 1.0, {{v[0-9]+}}
+; GFX8: v_max_f16_e32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
+; GFX9: v_max_f16_e32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
+define amdgpu_kernel void @test_canonicalize_value_f16_denorm(half addrspace(1)* %arg, half addrspace(1)* %out) #3 {
   %id = tail call i32 @llvm.amdgcn.workitem.id.x()
   %gep = getelementptr inbounds half, half addrspace(1)* %arg, i32 %id
   %v = load half, half addrspace(1)* %gep, align 2
@@ -533,8 +537,14 @@ define amdgpu_kernel void @test_canonicalize_value_f16_denorm(half addrspace(1)*
 }
 
 ; GCN-LABEL:  {{^}}test_canonicalize_value_v2f16_denorm:
-; GCN: v_pk_max_f16 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
-define amdgpu_kernel void @test_canonicalize_value_v2f16_denorm(<2 x half> addrspace(1)* %arg, <2 x half> addrspace(1)* %out) #5 {
+; GFX6: v_mul_f32_e32 {{v[0-9]+}}, 1.0, {{v[0-9]+}}
+; GFX6: v_mul_f32_e32 {{v[0-9]+}}, 1.0, {{v[0-9]+}}
+
+; GFX8: v_max_f16_sdwa
+; GFX8: v_max_f16_e32
+
+; GFX9: v_pk_max_f16 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
+define amdgpu_kernel void @test_canonicalize_value_v2f16_denorm(<2 x half> addrspace(1)* %arg, <2 x half> addrspace(1)* %out) #3 {
   %id = tail call i32 @llvm.amdgcn.workitem.id.x()
   %gep = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %arg, i32 %id
   %v = load <2 x half>, <2 x half> addrspace(1)* %gep, align 4
@@ -556,43 +566,64 @@ define amdgpu_kernel void @v_test_canonicalize_var_v2f64(<2 x double> addrspace(
   ret void
 }
 
-; GCN-LABEL: {{^}}v_test_canonicalize_v2f32:
-; GCN: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
-; GCN: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
-define <2 x float> @v_test_canonicalize_v2f32(<2 x float> %arg) #1 {
+; GCN-LABEL: {{^}}v_test_canonicalize_v2f32_flush:
+; GFX6: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
+; GFX6: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
+
+; GFX9: v_max_f32_e32 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}}
+; GFX9: v_max_f32_e32 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}}
+define <2 x float> @v_test_canonicalize_v2f32_flush(<2 x float> %arg) #1 {
   %canon = call <2 x float> @llvm.canonicalize.v2f32(<2 x float> %arg)
   ret <2 x float> %canon
 }
 
-; GCN-LABEL: {{^}}v_test_canonicalize_v3f32:
-; GCN: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
-; GCN: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
-; GCN: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
-define <3 x float> @v_test_canonicalize_v3f32(<3 x float> %arg) #1 {
+; GCN-LABEL: {{^}}v_test_canonicalize_v3f32_flush:
+; GFX6: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
+; GFX6: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
+; GFX6: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
+
+; GFX9: v_max_f32_e32 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}}
+; GFX9: v_max_f32_e32 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}}
+; GFX9: v_max_f32_e32 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}}
+define <3 x float> @v_test_canonicalize_v3f32_flush(<3 x float> %arg) #1 {
   %canon = call <3 x float> @llvm.canonicalize.v3f32(<3 x float> %arg)
   ret <3 x float> %canon
 }
 
-; GCN-LABEL: {{^}}v_test_canonicalize_v4f32:
-; GCN: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
-; GCN: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
-; GCN: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
-; GCN: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
-define <4 x float> @v_test_canonicalize_v4f32(<4 x float> %arg) #1 {
+; GCN-LABEL: {{^}}v_test_canonicalize_v4f32_flush:
+; GFX6: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
+; GFX6: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
+; GFX6: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
+; GFX6: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
+
+; GFX9: v_max_f32_e32 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}}
+; GFX9: v_max_f32_e32 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}}
+; GFX9: v_max_f32_e32 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}}
+; GFX9: v_max_f32_e32 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}}
+define <4 x float> @v_test_canonicalize_v4f32_flush(<4 x float> %arg) #1 {
   %canon = call <4 x float> @llvm.canonicalize.v4f32(<4 x float> %arg)
   ret <4 x float> %canon
 }
 
-; GCN-LABEL: {{^}}v_test_canonicalize_v8f32:
-; GCN: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
-; GCN: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
-; GCN: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
-; GCN: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
-; GCN: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
-; GCN: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
-; GCN: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
-; GCN: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
-define <8 x float> @v_test_canonicalize_v8f32(<8 x float> %arg) #1 {
+; GCN-LABEL: {{^}}v_test_canonicalize_v8f32_flush:
+; GFX6: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
+; GFX6: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
+; GFX6: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
+; GFX6: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
+; GFX6: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
+; GFX6: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
+; GFX6: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
+; GFX6: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
+
+; GFX9: v_max_f32_e32 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}}
+; GFX9: v_max_f32_e32 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}}
+; GFX9: v_max_f32_e32 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}}
+; GFX9: v_max_f32_e32 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}}
+; GFX9: v_max_f32_e32 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}}
+; GFX9: v_max_f32_e32 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}}
+; GFX9: v_max_f32_e32 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}}
+; GFX9: v_max_f32_e32 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}}
+define <8 x float> @v_test_canonicalize_v8f32_flush(<8 x float> %arg) #1 {
   %canon = call <8 x float> @llvm.canonicalize.v8f32(<8 x float> %arg)
   ret <8 x float> %canon
 }
@@ -628,6 +659,4 @@ attributes #0 = { nounwind readnone }
 attributes #1 = { nounwind "denormal-fp-math-f32"="preserve-sign,preserve-sign" }
 attributes #2 = { nounwind "denormal-fp-math"="preserve-sign,preserve-sign" }
 attributes #3 = { nounwind "denormal-fp-math"="ieee,ieee" }
-attributes #4 = { nounwind "denormal-fp-math"="preserve-sign,preserve-sign" "target-cpu"="tonga" }
-attributes #5 = { nounwind "denormal-fp-math"="ieee,ieee" "target-cpu"="gfx900" }
-attributes #6 = { nounwind "denormal-fp-math"="preserve-sign,preserve-sign" "target-cpu"="gfx900" }
+attributes #4 = { nounwind "denormal-fp-math"="preserve-sign,preserve-sign" }
diff --git a/test/CodeGen/AMDGPU/fminnum.f64.ll b/test/CodeGen/AMDGPU/fminnum.f64.ll
index e37a1cead47..04fc0e35bee 100644
--- a/test/CodeGen/AMDGPU/fminnum.f64.ll
+++ b/test/CodeGen/AMDGPU/fminnum.f64.ll
@@ -1,5 +1,6 @@
-; RUN: llc -march=amdgcn -mcpu=tahiti < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
-; RUN: llc -march=amdgcn -mcpu=tonga < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
+; RUN: llc -march=amdgcn -mcpu=tahiti < %s | FileCheck -check-prefixes=GCN,GFX678,SI %s
+; RUN: llc -march=amdgcn -mcpu=tonga < %s | FileCheck -check-prefixes=GCN,GFX678,VI %s
+; RUN: llc -march=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefixes=GCN,GFX9 %s
 
 declare double @llvm.minnum.f64(double, double) #0
 declare <2 x double> @llvm.minnum.v2f64(<2 x double>, <2 x double>) #0
@@ -7,26 +8,44 @@ declare <4 x double> @llvm.minnum.v4f64(<4 x double>, <4 x double>) #0
 declare <8 x double> @llvm.minnum.v8f64(<8 x double>, <8 x double>) #0
 declare <16 x double> @llvm.minnum.v16f64(<16 x double>, <16 x double>) #0
 
-; FUNC-LABEL: {{^}}test_fmin_f64_ieee:
-; SI: s_load_dwordx2 [[A:s\[[0-9]+:[0-9]+\]]]
-; SI: s_load_dwordx2 [[B:s\[[0-9]+:[0-9]+\]]]
-; SI-DAG: v_max_f64 [[QUIETA:v\[[0-9]+:[0-9]+\]]], [[A]], [[A]]
-; SI-DAG: v_max_f64 [[QUIETB:v\[[0-9]+:[0-9]+\]]], [[B]], [[B]]
-; SI: v_min_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[QUIETA]], [[QUIETB]]
-define amdgpu_kernel void @test_fmin_f64_ieee([8 x i32], double %a, [8 x i32], double %b) nounwind {
+; GCN-LABEL: {{^}}test_fmin_f64_ieee_noflush:
+; GCN: s_load_dwordx2 [[A:s\[[0-9]+:[0-9]+\]]]
+; GCN: s_load_dwordx2 [[B:s\[[0-9]+:[0-9]+\]]]
+
+; GCN-DAG: v_max_f64 [[QUIETA:v\[[0-9]+:[0-9]+\]]], [[A]], [[A]]
+; GCN-DAG: v_max_f64 [[QUIETB:v\[[0-9]+:[0-9]+\]]], [[B]], [[B]]
+
+; GCN: v_min_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[QUIETA]], [[QUIETB]]
+define amdgpu_kernel void @test_fmin_f64_ieee_noflush([8 x i32], double %a, [8 x i32], double %b) #1 {
   %val = call double @llvm.minnum.f64(double %a, double %b) #0
   store double %val, double addrspace(1)* undef, align 8
   ret void
 }
 
-; FUNC-LABEL: {{^}}test_fmin_f64_no_ieee:
-; SI: ds_read_b64 [[VAL0:v\[[0-9]+:[0-9]+\]]]
-; SI: ds_read_b64 [[VAL1:v\[[0-9]+:[0-9]+\]]]
-; SI-NOT: [[VAL0]]
-; SI-NOT: [[VAL1]]
-; SI: v_min_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[VAL0]], [[VAL1]]
-; SI-NOT: [[RESULT]]
-; SI: ds_write_b64 v{{[0-9]+}}, [[RESULT]]
+; GCN-LABEL: {{^}}test_fmin_f64_ieee_flush:
+; GCN: s_load_dwordx2 [[A:s\[[0-9]+:[0-9]+\]]]
+; GCN: s_load_dwordx2 [[B:s\[[0-9]+:[0-9]+\]]]
+; GFX678-DAG: v_mul_f64 [[QUIETA:v\[[0-9]+:[0-9]+\]]], 1.0, [[A]]
+; GFX678-DAG: v_mul_f64 [[QUIETB:v\[[0-9]+:[0-9]+\]]], 1.0, [[B]]
+
+; GFX9-DAG: v_max_f64 [[QUIETA:v\[[0-9]+:[0-9]+\]]], [[A]], [[A]]
+; GFX9-DAG: v_max_f64 [[QUIETB:v\[[0-9]+:[0-9]+\]]], [[B]], [[B]]
+
+; GCN: v_min_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[QUIETA]], [[QUIETB]]
+define amdgpu_kernel void @test_fmin_f64_ieee_flush([8 x i32], double %a, [8 x i32], double %b) #2 {
+  %val = call double @llvm.minnum.f64(double %a, double %b) #0
+  store double %val, double addrspace(1)* undef, align 8
+  ret void
+}
+
+; GCN-LABEL: {{^}}test_fmin_f64_no_ieee:
+; GCN: ds_read_b64 [[VAL0:v\[[0-9]+:[0-9]+\]]]
+; GCN: ds_read_b64 [[VAL1:v\[[0-9]+:[0-9]+\]]]
+; GCN-NOT: [[VAL0]]
+; GCN-NOT: [[VAL1]]
+; GCN: v_min_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[VAL0]], [[VAL1]]
+; GCN-NOT: [[RESULT]]
+; GCN: ds_write_b64 v{{[0-9]+}}, [[RESULT]]
 define amdgpu_ps void @test_fmin_f64_no_ieee() nounwind {
   %a = load volatile double, double addrspace(3)* undef
   %b = load volatile double, double addrspace(3)* undef
@@ -35,58 +54,58 @@ define amdgpu_ps void @test_fmin_f64_no_ieee() nounwind {
   ret void
 }
 
-; FUNC-LABEL: {{^}}test_fmin_v2f64:
-; SI: v_min_f64
-; SI: v_min_f64
+; GCN-LABEL: {{^}}test_fmin_v2f64:
+; GCN: v_min_f64
+; GCN: v_min_f64
 define amdgpu_kernel void @test_fmin_v2f64(<2 x double> addrspace(1)* %out, <2 x double> %a, <2 x double> %b) nounwind {
   %val = call <2 x double> @llvm.minnum.v2f64(<2 x double> %a, <2 x double> %b) #0
   store <2 x double> %val, <2 x double> addrspace(1)* %out, align 16
   ret void
 }
 
-; FUNC-LABEL: {{^}}test_fmin_v4f64:
-; SI: v_min_f64
-; SI: v_min_f64
-; SI: v_min_f64
-; SI: v_min_f64
+; GCN-LABEL: {{^}}test_fmin_v4f64:
+; GCN: v_min_f64
+; GCN: v_min_f64
+; GCN: v_min_f64
+; GCN: v_min_f64
 define amdgpu_kernel void @test_fmin_v4f64(<4 x double> addrspace(1)* %out, <4 x double> %a, <4 x double> %b) nounwind {
   %val = call <4 x double> @llvm.minnum.v4f64(<4 x double> %a, <4 x double> %b) #0
   store <4 x double> %val, <4 x double> addrspace(1)* %out, align 32
   ret void
 }
 
-; FUNC-LABEL: {{^}}test_fmin_v8f64:
-; SI: v_min_f64
-; SI: v_min_f64
-; SI: v_min_f64
-; SI: v_min_f64
-; SI: v_min_f64
-; SI: v_min_f64
-; SI: v_min_f64
-; SI: v_min_f64
+; GCN-LABEL: {{^}}test_fmin_v8f64:
+; GCN: v_min_f64
+; GCN: v_min_f64
+; GCN: v_min_f64
+; GCN: v_min_f64
+; GCN: v_min_f64
+; GCN: v_min_f64
+; GCN: v_min_f64
+; GCN: v_min_f64
 define amdgpu_kernel void @test_fmin_v8f64(<8 x double> addrspace(1)* %out, <8 x double> %a, <8 x double> %b) nounwind {
   %val = call <8 x double> @llvm.minnum.v8f64(<8 x double> %a, <8 x double> %b) #0
   store <8 x double> %val, <8 x double> addrspace(1)* %out, align 64
   ret void
 }
 
-; FUNC-LABEL: {{^}}test_fmin_v16f64:
-; SI: v_min_f64
-; SI: v_min_f64
-; SI: v_min_f64
-; SI: v_min_f64
-; SI: v_min_f64
-; SI: v_min_f64
-; SI: v_min_f64
-; SI: v_min_f64
-; SI: v_min_f64
-; SI: v_min_f64
-; SI: v_min_f64
-; SI: v_min_f64
-; SI: v_min_f64
-; SI: v_min_f64
-; SI: v_min_f64
-; SI: v_min_f64
+; GCN-LABEL: {{^}}test_fmin_v16f64:
+; GCN: v_min_f64
+; GCN: v_min_f64
+; GCN: v_min_f64
+; GCN: v_min_f64
+; GCN: v_min_f64
+; GCN: v_min_f64
+; GCN: v_min_f64
+; GCN: v_min_f64
+; GCN: v_min_f64
+; GCN: v_min_f64
+; GCN: v_min_f64
+; GCN: v_min_f64
+; GCN: v_min_f64
+; GCN: v_min_f64
+; GCN: v_min_f64
+; GCN: v_min_f64
 define amdgpu_kernel void @test_fmin_v16f64(<16 x double> addrspace(1)* %out, <16 x double> %a, <16 x double> %b) nounwind {
   %val = call <16 x double> @llvm.minnum.v16f64(<16 x double> %a, <16 x double> %b) #0
   store <16 x double> %val, <16 x double> addrspace(1)* %out, align 128
@@ -94,3 +113,5 @@ define amdgpu_kernel void @test_fmin_v16f64(<16 x double> addrspace(1)* %out, <1
 }
 
 attributes #0 = { nounwind readnone }
+attributes #1 = { nounwind "denormal-fp-math"="ieee,ieee" }
+attributes #2 = { nounwind "denormal-fp-math"="preserve-sign,preserve-sign" }