From: Matt Arsenault <Matthew.Arsenault@amd.com>
Date: Thu, 13 Nov 2014 19:26:47 +0000 (+0000)
Subject: R600/SI: Allow commuting some 3 op instructions
X-Git-Tag: android-x86-7.1-r4~55534
X-Git-Url: http://git.osdn.net/view?a=commitdiff_plain;h=1aae959de7fc8484b470562d99624879cf5f9396;p=android-x86%2Fexternal-llvm.git

R600/SI: Allow commuting some 3 op instructions

e.g. v_mad_f32 a, b, c -> v_mad_f32 b, a, c

This simplifies matching v_madmk_f32.

This looks somewhat surprising, but it appears to be
OK to do this. We can commute src0 and src1 in all
of these instructions, and that's all that appears
to matter.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@221910 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
index 3842f8489ae..33d2d075eaa 100644
--- a/lib/Target/R600/SIInstructions.td
+++ b/lib/Target/R600/SIInstructions.td
@@ -1374,12 +1374,12 @@ defm V_SUBREV_F32 : VOP2Inst <vop2<0x5>, "v_subrev_f32",
 >;
 } // End isCommutable = 1
 
+let isCommutable = 1 in {
+
 defm V_MAC_LEGACY_F32 : VOP2Inst <vop2<0x6>, "v_mac_legacy_f32",
   VOP_F32_F32_F32
 >;
 
-let isCommutable = 1 in {
-
 defm V_MUL_LEGACY_F32 : VOP2Inst <vop2<0x7>, "v_mul_legacy_f32",
   VOP_F32_F32_F32, int_AMDGPU_mul
 >;
@@ -1388,7 +1388,6 @@ defm V_MUL_F32 : VOP2Inst <vop2<0x8>, "v_mul_f32",
   VOP_F32_F32_F32, fmul
 >;
 
-
 defm V_MUL_I32_I24 : VOP2Inst <vop2<0x9>, "v_mul_i32_i24",
   VOP_I32_I32_I32, AMDGPUmul_i24
 >;
@@ -1449,11 +1448,21 @@ defm V_XOR_B32 : VOP2Inst <vop2<0x1d>, "v_xor_b32",
 
 defm V_BFM_B32 : VOP2Inst <vop2<0x1e>, "v_bfm_b32",
   VOP_I32_I32_I32, AMDGPUbfm>;
+
+let isCommutable = 1 in {
 defm V_MAC_F32 : VOP2Inst <vop2<0x1f>, "v_mac_f32", VOP_F32_F32_F32>;
+} // End isCommutable = 1
+
 defm V_MADMK_F32 : VOP2Inst <vop2<0x20>, "v_madmk_f32", VOP_F32_F32_F32>;
+
+let isCommutable = 1 in {
 defm V_MADAK_F32 : VOP2Inst <vop2<0x21>, "v_madak_f32", VOP_F32_F32_F32>;
+} // End isCommutable = 1
+
+
 defm V_BCNT_U32_B32 : VOP2Inst <vop2<0x22>, "v_bcnt_u32_b32", VOP_I32_I32_I32>;
 defm V_MBCNT_LO_U32_B32 : VOP2Inst <vop2<0x23>, "v_mbcnt_lo_u32_b32",
+
   VOP_I32_I32_I32
 >;
 defm V_MBCNT_HI_U32_B32 : VOP2Inst <vop2<0x24>, "v_mbcnt_hi_u32_b32",
@@ -1503,18 +1512,22 @@ defm V_CVT_PKRTZ_F16_F32 : VOP2Inst <vop2<0x2f>, "v_cvt_pkrtz_f16_f32",
 // VOP3 Instructions
 //===----------------------------------------------------------------------===//
 
+let isCommutable = 1 in {
 defm V_MAD_LEGACY_F32 : VOP3Inst <vop3<0x140>, "v_mad_legacy_f32",
   VOP_F32_F32_F32_F32
 >;
+
 defm V_MAD_F32 : VOP3Inst <vop3<0x141>, "v_mad_f32",
   VOP_F32_F32_F32_F32, fmad
 >;
+
 defm V_MAD_I32_I24 : VOP3Inst <vop3<0x142>, "v_mad_i32_i24",
   VOP_I32_I32_I32_I32, AMDGPUmad_i24
 >;
 defm V_MAD_U32_U24 : VOP3Inst <vop3<0x143>, "v_mad_u32_u24",
   VOP_I32_I32_I32_I32, AMDGPUmad_u24
 >;
+} // End isCommutable = 1
 
 defm V_CUBEID_F32 : VOP3Inst <vop3<0x144>, "v_cubeid_f32",
   VOP_F32_F32_F32_F32
@@ -1537,12 +1550,16 @@ defm V_BFE_I32 : VOP3Inst <vop3<0x149>, "v_bfe_i32",
 defm V_BFI_B32 : VOP3Inst <vop3<0x14a>, "v_bfi_b32",
   VOP_I32_I32_I32_I32, AMDGPUbfi
 >;
+
+let isCommutable = 1 in {
 defm V_FMA_F32 : VOP3Inst <vop3<0x14b>, "v_fma_f32",
   VOP_F32_F32_F32_F32, fma
 >;
 defm V_FMA_F64 : VOP3Inst <vop3<0x14c>, "v_fma_f64",
   VOP_F64_F64_F64_F64, fma
 >;
+} // End isCommutable = 1
+
 //def V_LERP_U8 : VOP3_U8 <0x0000014d, "v_lerp_u8", []>;
 defm V_ALIGNBIT_B32 : VOP3Inst <vop3<0x14e>, "v_alignbit_b32",
   VOP_I32_I32_I32_I32
@@ -1629,15 +1646,19 @@ defm V_DIV_SCALE_F32 : VOP3b_32 <vop3<0x16d>, "v_div_scale_f32", []>;
 // Double precision division pre-scale.
 defm V_DIV_SCALE_F64 : VOP3b_64 <vop3<0x16e>, "v_div_scale_f64", []>;
 
+let isCommutable = 1 in {
 defm V_DIV_FMAS_F32 : VOP3Inst <vop3<0x16f>, "v_div_fmas_f32",
   VOP_F32_F32_F32_F32, AMDGPUdiv_fmas
 >;
 defm V_DIV_FMAS_F64 : VOP3Inst <vop3<0x170>, "v_div_fmas_f64",
   VOP_F64_F64_F64_F64, AMDGPUdiv_fmas
 >;
+} // End isCommutable = 1
+
 //def V_MSAD_U8 : VOP3_U8 <0x00000171, "v_msad_u8", []>;
 //def V_QSAD_U8 : VOP3_U8 <0x00000172, "v_qsad_u8", []>;
 //def V_MQSAD_U8 : VOP3_U8 <0x00000173, "v_mqsad_u8", []>;
+
 defm V_TRIG_PREOP_F64 : VOP3Inst <
   vop3<0x174>, "v_trig_preop_f64", VOP_F64_F64_I32, AMDGPUtrig_preop
 >;
@@ -2840,6 +2861,8 @@ defm V_MQSAD_U16_U8 : VOP3Inst <vop3<0x172>, "v_mqsad_u16_u8",
 defm V_MQSAD_U32_U8 : VOP3Inst <vop3<0x175>, "v_mqsad_u32_u8",
   VOP_I32_I32_I32
 >;
+
+let isCommutable = 1 in {
 defm V_MAD_U64_U32 : VOP3Inst <vop3<0x176>, "v_mad_u64_u32",
   VOP_I64_I32_I32_I64
 >;
@@ -2848,6 +2871,7 @@ defm V_MAD_U64_U32 : VOP3Inst <vop3<0x176>, "v_mad_u64_u32",
 defm V_MAD_I64_I32 : VOP3Inst <vop3<0x177>, "v_mad_i64_i32",
   VOP_I64_I32_I32_I64
 >;
+} // End isCommutable = 1
 
 // Remaining instructions:
 // FLAT_*
diff --git a/test/CodeGen/R600/commute_modifiers.ll b/test/CodeGen/R600/commute_modifiers.ll
index 2504688bbc4..f2885a759a9 100644
--- a/test/CodeGen/R600/commute_modifiers.ll
+++ b/test/CodeGen/R600/commute_modifiers.ll
@@ -46,5 +46,111 @@ define void @commute_mul_imm_fneg_f32(float addrspace(1)* %out, float addrspace(
   ret void
 }
 
+; FIXME: Should use SGPR for literal.
+; FUNC-LABEL: @commute_add_lit_fabs_f32
+; SI: buffer_load_dword [[X:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI: v_mov_b32_e32 [[K:v[0-9]+]], 0x44800000
+; SI: v_add_f32_e64 [[REG:v[0-9]+]], |[[X]]|, [[K]]
+; SI-NEXT: buffer_store_dword [[REG]]
+define void @commute_add_lit_fabs_f32(float addrspace(1)* %out, float addrspace(1)* %in) #0 {
+  %tid = call i32 @llvm.r600.read.tidig.x() #1
+  %gep.0 = getelementptr float addrspace(1)* %in, i32 %tid
+  %x = load float addrspace(1)* %gep.0
+  %x.fabs = call float @llvm.fabs.f32(float %x) #1
+  %z = fadd float 1024.0, %x.fabs
+  store float %z, float addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: @commute_add_fabs_f32
+; SI-DAG: buffer_load_dword [[X:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: buffer_load_dword [[Y:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:0x4
+; SI: v_add_f32_e64 [[REG:v[0-9]+]], [[X]], |[[Y]]|
+; SI-NEXT: buffer_store_dword [[REG]]
+define void @commute_add_fabs_f32(float addrspace(1)* %out, float addrspace(1)* %in) #0 {
+  %tid = call i32 @llvm.r600.read.tidig.x() #1
+  %gep.0 = getelementptr float addrspace(1)* %in, i32 %tid
+  %gep.1 = getelementptr float addrspace(1)* %gep.0, i32 1
+  %x = load float addrspace(1)* %gep.0
+  %y = load float addrspace(1)* %gep.1
+  %y.fabs = call float @llvm.fabs.f32(float %y) #1
+  %z = fadd float %x, %y.fabs
+  store float %z, float addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: @commute_mul_fneg_f32
+; SI-DAG: buffer_load_dword [[X:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: buffer_load_dword [[Y:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:0x4
+; SI: v_mul_f32_e64 [[REG:v[0-9]+]], [[X]], -[[Y]]
+; SI-NEXT: buffer_store_dword [[REG]]
+define void @commute_mul_fneg_f32(float addrspace(1)* %out, float addrspace(1)* %in) #0 {
+  %tid = call i32 @llvm.r600.read.tidig.x() #1
+  %gep.0 = getelementptr float addrspace(1)* %in, i32 %tid
+  %gep.1 = getelementptr float addrspace(1)* %gep.0, i32 1
+  %x = load float addrspace(1)* %gep.0
+  %y = load float addrspace(1)* %gep.1
+  %y.fneg = fsub float -0.000000e+00, %y
+  %z = fmul float %x, %y.fneg
+  store float %z, float addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: @commute_mul_fabs_fneg_f32
+; SI-DAG: buffer_load_dword [[X:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: buffer_load_dword [[Y:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:0x4
+; SI: v_mul_f32_e64 [[REG:v[0-9]+]], [[X]], -|[[Y]]|
+; SI-NEXT: buffer_store_dword [[REG]]
+define void @commute_mul_fabs_fneg_f32(float addrspace(1)* %out, float addrspace(1)* %in) #0 {
+  %tid = call i32 @llvm.r600.read.tidig.x() #1
+  %gep.0 = getelementptr float addrspace(1)* %in, i32 %tid
+  %gep.1 = getelementptr float addrspace(1)* %gep.0, i32 1
+  %x = load float addrspace(1)* %gep.0
+  %y = load float addrspace(1)* %gep.1
+  %y.fabs = call float @llvm.fabs.f32(float %y) #1
+  %y.fabs.fneg = fsub float -0.000000e+00, %y.fabs
+  %z = fmul float %x, %y.fabs.fneg
+  store float %z, float addrspace(1)* %out
+  ret void
+}
+
+; There's no reason to commute this.
+; FUNC-LABEL: @commute_mul_fabs_x_fabs_y_f32
+; SI-DAG: buffer_load_dword [[X:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: buffer_load_dword [[Y:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:0x4
+; SI: v_mul_f32_e64 [[REG:v[0-9]+]], |[[X]]|, |[[Y]]|
+; SI-NEXT: buffer_store_dword [[REG]]
+define void @commute_mul_fabs_x_fabs_y_f32(float addrspace(1)* %out, float addrspace(1)* %in) #0 {
+  %tid = call i32 @llvm.r600.read.tidig.x() #1
+  %gep.0 = getelementptr float addrspace(1)* %in, i32 %tid
+  %gep.1 = getelementptr float addrspace(1)* %gep.0, i32 1
+  %x = load float addrspace(1)* %gep.0
+  %y = load float addrspace(1)* %gep.1
+  %x.fabs = call float @llvm.fabs.f32(float %x) #1
+  %y.fabs = call float @llvm.fabs.f32(float %y) #1
+  %z = fmul float %x.fabs, %y.fabs
+  store float %z, float addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: @commute_mul_fabs_x_fneg_fabs_y_f32
+; SI-DAG: buffer_load_dword [[X:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: buffer_load_dword [[Y:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:0x4
+; SI: v_mul_f32_e64 [[REG:v[0-9]+]], |[[X]]|, -|[[Y]]|
+; SI-NEXT: buffer_store_dword [[REG]]
+define void @commute_mul_fabs_x_fneg_fabs_y_f32(float addrspace(1)* %out, float addrspace(1)* %in) #0 {
+  %tid = call i32 @llvm.r600.read.tidig.x() #1
+  %gep.0 = getelementptr float addrspace(1)* %in, i32 %tid
+  %gep.1 = getelementptr float addrspace(1)* %gep.0, i32 1
+  %x = load float addrspace(1)* %gep.0
+  %y = load float addrspace(1)* %gep.1
+  %x.fabs = call float @llvm.fabs.f32(float %x) #1
+  %y.fabs = call float @llvm.fabs.f32(float %y) #1
+  %y.fabs.fneg = fsub float -0.000000e+00, %y.fabs
+  %z = fmul float %x.fabs, %y.fabs.fneg
+  store float %z, float addrspace(1)* %out
+  ret void
+}
+
 attributes #0 = { nounwind }
 attributes #1 = { nounwind readnone }
diff --git a/test/CodeGen/R600/fma.ll b/test/CodeGen/R600/fma.ll
index d715c99a7dc..637e7995f2c 100644
--- a/test/CodeGen/R600/fma.ll
+++ b/test/CodeGen/R600/fma.ll
@@ -5,6 +5,8 @@ declare float @llvm.fma.f32(float, float, float) nounwind readnone
 declare <2 x float> @llvm.fma.v2f32(<2 x float>, <2 x float>, <2 x float>) nounwind readnone
 declare <4 x float> @llvm.fma.v4f32(<4 x float>, <4 x float>, <4 x float>) nounwind readnone
 
+declare i32 @llvm.r600.read.tidig.x() nounwind readnone
+
 ; FUNC-LABEL: {{^}}fma_f32:
 ; SI: v_fma_f32 {{v[0-9]+, v[0-9]+, v[0-9]+, v[0-9]+}}
 
@@ -12,12 +14,12 @@ declare <4 x float> @llvm.fma.v4f32(<4 x float>, <4 x float>, <4 x float>) nounw
 ; EG: FMA {{\*? *}}[[RES]]
 define void @fma_f32(float addrspace(1)* %out, float addrspace(1)* %in1,
                      float addrspace(1)* %in2, float addrspace(1)* %in3) {
-   %r0 = load float addrspace(1)* %in1
-   %r1 = load float addrspace(1)* %in2
-   %r2 = load float addrspace(1)* %in3
-   %r3 = tail call float @llvm.fma.f32(float %r0, float %r1, float %r2)
-   store float %r3, float addrspace(1)* %out
-   ret void
+  %r0 = load float addrspace(1)* %in1
+  %r1 = load float addrspace(1)* %in2
+  %r2 = load float addrspace(1)* %in3
+  %r3 = tail call float @llvm.fma.f32(float %r0, float %r1, float %r2)
+  store float %r3, float addrspace(1)* %out
+  ret void
 }
 
 ; FUNC-LABEL: {{^}}fma_v2f32:
@@ -29,12 +31,12 @@ define void @fma_f32(float addrspace(1)* %out, float addrspace(1)* %in1,
 ; EG-DAG: FMA {{\*? *}}[[RES]].[[CHHI]]
 define void @fma_v2f32(<2 x float> addrspace(1)* %out, <2 x float> addrspace(1)* %in1,
                        <2 x float> addrspace(1)* %in2, <2 x float> addrspace(1)* %in3) {
-   %r0 = load <2 x float> addrspace(1)* %in1
-   %r1 = load <2 x float> addrspace(1)* %in2
-   %r2 = load <2 x float> addrspace(1)* %in3
-   %r3 = tail call <2 x float> @llvm.fma.v2f32(<2 x float> %r0, <2 x float> %r1, <2 x float> %r2)
-   store <2 x float> %r3, <2 x float> addrspace(1)* %out
-   ret void
+  %r0 = load <2 x float> addrspace(1)* %in1
+  %r1 = load <2 x float> addrspace(1)* %in2
+  %r2 = load <2 x float> addrspace(1)* %in3
+  %r3 = tail call <2 x float> @llvm.fma.v2f32(<2 x float> %r0, <2 x float> %r1, <2 x float> %r2)
+  store <2 x float> %r3, <2 x float> addrspace(1)* %out
+  ret void
 }
 
 ; FUNC-LABEL: {{^}}fma_v4f32:
@@ -50,10 +52,41 @@ define void @fma_v2f32(<2 x float> addrspace(1)* %out, <2 x float> addrspace(1)*
 ; EG-DAG: FMA {{\*? *}}[[RES]].W
 define void @fma_v4f32(<4 x float> addrspace(1)* %out, <4 x float> addrspace(1)* %in1,
                        <4 x float> addrspace(1)* %in2, <4 x float> addrspace(1)* %in3) {
-   %r0 = load <4 x float> addrspace(1)* %in1
-   %r1 = load <4 x float> addrspace(1)* %in2
-   %r2 = load <4 x float> addrspace(1)* %in3
-   %r3 = tail call <4 x float> @llvm.fma.v4f32(<4 x float> %r0, <4 x float> %r1, <4 x float> %r2)
-   store <4 x float> %r3, <4 x float> addrspace(1)* %out
-   ret void
+  %r0 = load <4 x float> addrspace(1)* %in1
+  %r1 = load <4 x float> addrspace(1)* %in2
+  %r2 = load <4 x float> addrspace(1)* %in3
+  %r3 = tail call <4 x float> @llvm.fma.v4f32(<4 x float> %r0, <4 x float> %r1, <4 x float> %r2)
+  store <4 x float> %r3, <4 x float> addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: @fma_commute_mul_inline_imm_f32
+; SI: v_fma_f32 {{v[0-9]+}}, 2.0, {{v[0-9]+}}, {{v[0-9]+}}
+define void @fma_commute_mul_inline_imm_f32(float addrspace(1)* noalias %out, float addrspace(1)* noalias %in.a, float addrspace(1)* noalias %in.b) nounwind {
+  %tid = tail call i32 @llvm.r600.read.tidig.x() nounwind readnone
+  %in.a.gep = getelementptr float addrspace(1)* %in.a, i32 %tid
+  %in.b.gep = getelementptr float addrspace(1)* %in.b, i32 %tid
+  %out.gep = getelementptr float addrspace(1)* %out, i32 %tid
+
+  %a = load float addrspace(1)* %in.a.gep, align 4
+  %b = load float addrspace(1)* %in.b.gep, align 4
+
+  %fma = call float @llvm.fma.f32(float %a, float 2.0, float %b)
+  store float %fma, float addrspace(1)* %out.gep, align 4
+  ret void
+}
+
+; FUNC-LABEL: @fma_commute_mul_s_f32
+define void @fma_commute_mul_s_f32(float addrspace(1)* noalias %out, float addrspace(1)* noalias %in.a, float addrspace(1)* noalias %in.b, float %b) nounwind {
+  %tid = tail call i32 @llvm.r600.read.tidig.x() nounwind readnone
+  %in.a.gep = getelementptr float addrspace(1)* %in.a, i32 %tid
+  %in.b.gep = getelementptr float addrspace(1)* %in.b, i32 %tid
+  %out.gep = getelementptr float addrspace(1)* %out, i32 %tid
+
+  %a = load float addrspace(1)* %in.a.gep, align 4
+  %c = load float addrspace(1)* %in.b.gep, align 4
+
+  %fma = call float @llvm.fma.f32(float %a, float %b, float %c)
+  store float %fma, float addrspace(1)* %out.gep, align 4
+  ret void
 }
diff --git a/test/CodeGen/R600/fmuladd.ll b/test/CodeGen/R600/fmuladd.ll
index becc3e441ca..16003a52237 100644
--- a/test/CodeGen/R600/fmuladd.ll
+++ b/test/CodeGen/R600/fmuladd.ll
@@ -116,7 +116,7 @@ define void @fadd_b_a_a_f32(float addrspace(1)* %out,
 ; CHECK-LABEL: {{^}}fmuladd_neg_2.0_a_b_f32
 ; CHECK-DAG: buffer_load_dword [[R1:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
 ; CHECK-DAG: buffer_load_dword [[R2:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:0x4
-; CHECK: v_mad_f32 [[RESULT:v[0-9]+]], [[R1]], -2.0, [[R2]]
+; CHECK: v_mad_f32 [[RESULT:v[0-9]+]], -2.0, [[R1]], [[R2]]
 ; CHECK: buffer_store_dword [[RESULT]]
 define void @fmuladd_neg_2.0_a_b_f32(float addrspace(1)* %out, float addrspace(1)* %in) {
   %tid = call i32 @llvm.r600.read.tidig.x() nounwind readnone
@@ -158,7 +158,7 @@ define void @fmuladd_neg_2.0_neg_a_b_f32(float addrspace(1)* %out, float addrspa
 ; CHECK-LABEL: {{^}}fmuladd_2.0_neg_a_b_f32
 ; CHECK-DAG: buffer_load_dword [[R1:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
 ; CHECK-DAG: buffer_load_dword [[R2:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:0x4
-; CHECK: v_mad_f32 [[RESULT:v[0-9]+]], [[R1]], -2.0, [[R2]]
+; CHECK: v_mad_f32 [[RESULT:v[0-9]+]], -2.0, [[R1]], [[R2]]
 ; CHECK: buffer_store_dword [[RESULT]]
 define void @fmuladd_2.0_neg_a_b_f32(float addrspace(1)* %out, float addrspace(1)* %in) {
   %tid = call i32 @llvm.r600.read.tidig.x() nounwind readnone
diff --git a/test/CodeGen/R600/llvm.AMDGPU.umad24.ll b/test/CodeGen/R600/llvm.AMDGPU.umad24.ll
index 4de1f75db69..59d62489ca7 100644
--- a/test/CodeGen/R600/llvm.AMDGPU.umad24.ll
+++ b/test/CodeGen/R600/llvm.AMDGPU.umad24.ll
@@ -5,6 +5,7 @@
 ; XUN: llc -march=r600 -mcpu=rv770 -verify-machineinstrs < %s | FileCheck -check-prefix=R600 -check-prefix=FUNC %s
 
 declare i32 @llvm.AMDGPU.umad24(i32, i32, i32) nounwind readnone
+declare i32 @llvm.r600.read.tidig.x() nounwind readnone
 
 ; FUNC-LABEL: {{^}}test_umad24:
 ; SI: v_mad_u32_u24
@@ -17,3 +18,21 @@ define void @test_umad24(i32 addrspace(1)* %out, i32 %src0, i32 %src1, i32 %src2
   ret void
 }
 
+; FUNC-LABEL: {{^}}commute_umad24:
+; SI-DAG: buffer_load_dword [[SRC0:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: buffer_load_dword [[SRC2:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:0x4
+; SI: v_mad_u32_u24 [[RESULT:v[0-9]+]], 4, [[SRC0]], [[SRC2]]
+; SI: buffer_store_dword [[RESULT]]
+define void @commute_umad24(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %tid = call i32 @llvm.r600.read.tidig.x() nounwind readnone
+  %out.gep = getelementptr i32 addrspace(1)* %out, i32 %tid
+  %src0.gep = getelementptr i32 addrspace(1)* %out, i32 %tid
+  %src2.gep = getelementptr i32 addrspace(1)* %src0.gep, i32 1
+
+  %src0 = load i32 addrspace(1)* %src0.gep, align 4
+  %src2 = load i32 addrspace(1)* %src2.gep, align 4
+  %mad = call i32 @llvm.AMDGPU.umad24(i32 %src0, i32 4, i32 %src2) nounwind readnone
+  store i32 %mad, i32 addrspace(1)* %out.gep, align 4
+  ret void
+}
+
diff --git a/test/CodeGen/R600/use-sgpr-multiple-times.ll b/test/CodeGen/R600/use-sgpr-multiple-times.ll
index 2e67fbf5e2a..aa94a0e8c23 100644
--- a/test/CodeGen/R600/use-sgpr-multiple-times.ll
+++ b/test/CodeGen/R600/use-sgpr-multiple-times.ll
@@ -73,7 +73,7 @@ define void @test_sgpr_use_twice_ternary_op_a_a_imm(float addrspace(1)* %out, fl
 
 ; SI-LABEL: {{^}}test_sgpr_use_twice_ternary_op_a_imm_a:
 ; SI: s_load_dword [[SGPR:s[0-9]+]]
-; SI: v_fma_f32 [[RESULT:v[0-9]+]], [[SGPR]], 2.0, [[SGPR]]
+; SI: v_fma_f32 [[RESULT:v[0-9]+]], 2.0, [[SGPR]], [[SGPR]]
 ; SI: buffer_store_dword [[RESULT]]
 define void @test_sgpr_use_twice_ternary_op_a_imm_a(float addrspace(1)* %out, float %a) #0 {
   %fma = call float @llvm.fma.f32(float %a, float 2.0, float %a) #1