AMDGPU/EG,CM: Implement fsqrt using recip(rsqrt(x)) instead of x * rsqrt(x)

The old version might be faster on EG (RECIP_IEEE is Trans only),
but it'd need extra corner case checks.
This gives correct corner case behaviour and saves a register.
Fixes OCL CTS sqrt test (1-thread, scalar) on Turks.

Reviewer: arsenm
Differential Revision: https://reviews.llvm.org/D74017

Author: jvesely

llvm/lib/Target/AMDGPU/CaymanInstructions.td

@@ -50,6 +50,8 @@ def COS_cm : COS_Common<0x8E>;
 
 def : RsqPat<RECIPSQRT_IEEE_cm, f32>;
 
+def : SqrtPat<RECIPSQRT_IEEE_cm, RECIP_IEEE_cm>;
+
 def : POW_Common <LOG_IEEE_cm, EXP_IEEE_cm, MUL>;
 
 defm DIV_cm : DIV_Common<RECIP_IEEE_cm>;
@@ -70,8 +72,6 @@ def CF_END_CM : CF_CLAUSE_EG<32, (ins), "CF_END"> {
 
 
 
-def : R600Pat<(fsqrt f32:$src), (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_cm $src))>;
-
 class RAT_STORE_DWORD <RegisterClass rc, ValueType vt, bits<4> mask> :
   CF_MEM_RAT_CACHELESS <0x14, 0, mask,
                         (ins rc:$rw_gpr, R600_TReg32_X:$index_gpr),

llvm/lib/Target/AMDGPU/EvergreenInstructions.td

@@ -118,11 +118,12 @@ def LOG_IEEE_eg : LOG_IEEE_Common<0x83>;
 def RECIP_CLAMPED_eg : RECIP_CLAMPED_Common<0x84>;
 def RECIPSQRT_IEEE_eg : RECIPSQRT_IEEE_Common<0x89>;
 def : RsqPat<RECIPSQRT_IEEE_eg, f32>;
+def : SqrtPat<RECIPSQRT_IEEE_eg, RECIP_IEEE_eg>;
+
 def SIN_eg : SIN_Common<0x8D>;
 def COS_eg : COS_Common<0x8E>;
 
 def : POW_Common <LOG_IEEE_eg, EXP_IEEE_eg, MUL>;
-def : EGPat<(fsqrt f32:$src), (MUL $src, (RECIPSQRT_CLAMPED_eg $src))>;
 } // End SubtargetPredicate = isEG
 
 //===----------------------------------------------------------------------===//

llvm/lib/Target/AMDGPU/R600Instructions.td

@@ -1233,6 +1233,11 @@ def : R600Pat<
 def : RcpPat<recip_ieee, f32>;
 }
 
+class SqrtPat<Instruction RsqInst, Instruction RecipInst> : R600Pat <
+  (fsqrt f32:$src),
+  (RecipInst (RsqInst $src))
+>;
+
 //===----------------------------------------------------------------------===//
 // R600 / R700 Instructions
 //===----------------------------------------------------------------------===//
@@ -1272,8 +1277,8 @@ let Predicates = [isR600] in {
   defm DIV_r600 : DIV_Common<RECIP_IEEE_r600>;
   def : POW_Common <LOG_IEEE_r600, EXP_IEEE_r600, MUL>;
 
-  def : R600Pat<(fsqrt f32:$src), (MUL $src, (RECIPSQRT_CLAMPED_r600 $src))>;
   def : RsqPat<RECIPSQRT_IEEE_r600, f32>;
+  def : SqrtPat<RECIPSQRT_IEEE_r600, RECIP_IEEE_r600>;
 
   def R600_ExportSwz : ExportSwzInst {
     let Word1{20-17} = 0; // BURST_COUNT

llvm/test/CodeGen/AMDGPU/fsqrt.ll

@@ -27,8 +27,8 @@ define amdgpu_kernel void @v_unsafe_fsqrt_f32(float addrspace(1)* %out, float ad
 ; FUNC-LABEL: {{^}}s_sqrt_f32:
 ; GCN: v_sqrt_f32_e32
 
-; R600: RECIPSQRT_CLAMPED * T{{[0-9]\.[XYZW]}}, KC0[2].Z
-; R600: MUL NON-IEEE T{{[0-9]\.[XYZW]}}, KC0[2].Z, PS
+; R600: RECIPSQRT_IEEE * T{{[0-9]\.[XYZW]}}, KC0[2].Z
+; R600: RECIP_IEEE * T{{[0-9]\.[XYZW]}}, PS
 define amdgpu_kernel void @s_sqrt_f32(float addrspace(1)* %out, float %in) #1 {
 entry:
   %fdiv = call float @llvm.sqrt.f32(float %in)
@@ -40,10 +40,10 @@ entry:
 ; GCN: v_sqrt_f32_e32
 ; GCN: v_sqrt_f32_e32
 
-; R600-DAG: RECIPSQRT_CLAMPED * T{{[0-9]\.[XYZW]}}, KC0[2].W
-; R600-DAG: MUL NON-IEEE T{{[0-9]\.[XYZW]}}, KC0[2].W, PS
-; R600-DAG: RECIPSQRT_CLAMPED * T{{[0-9]\.[XYZW]}}, KC0[3].X
-; R600-DAG: MUL NON-IEEE T{{[0-9]\.[XYZW]}}, KC0[3].X, PS
+; R600-DAG: RECIPSQRT_IEEE * T{{[0-9]\.[XYZW]}}, KC0[2].W
+; R600-DAG: RECIP_IEEE * T{{[0-9]\.[XYZW]}}, PS
+; R600-DAG: RECIPSQRT_IEEE * T{{[0-9]\.[XYZW]}}, KC0[3].X
+; R600-DAG: RECIP_IEEE * T{{[0-9]\.[XYZW]}}, T{{[0-9]\.[XYZW]}}
 define amdgpu_kernel void @s_sqrt_v2f32(<2 x float> addrspace(1)* %out, <2 x float> %in) #1 {
 entry:
   %fdiv = call <2 x float> @llvm.sqrt.v2f32(<2 x float> %in)
@@ -57,14 +57,14 @@ entry:
 ; GCN: v_sqrt_f32_e32
 ; GCN: v_sqrt_f32_e32
 
-; R600-DAG: RECIPSQRT_CLAMPED * T{{[0-9]\.[XYZW]}}, KC0[3].Y
-; R600-DAG: MUL NON-IEEE T{{[0-9]\.[XYZW]}}, KC0[3].Y, PS
-; R600-DAG: RECIPSQRT_CLAMPED * T{{[0-9]\.[XYZW]}}, KC0[3].Z
-; R600-DAG: MUL NON-IEEE T{{[0-9]\.[XYZW]}}, KC0[3].Z, PS
-; R600-DAG: RECIPSQRT_CLAMPED * T{{[0-9]\.[XYZW]}}, KC0[3].W
-; R600-DAG: MUL NON-IEEE T{{[0-9]\.[XYZW]}}, KC0[3].W, PS
-; R600-DAG: RECIPSQRT_CLAMPED * T{{[0-9]\.[XYZW]}}, KC0[4].X
-; R600-DAG: MUL NON-IEEE T{{[0-9]\.[XYZW]}}, KC0[4].X, PS
+; R600-DAG: RECIPSQRT_IEEE * T{{[0-9]\.[XYZW]}}, KC0[3].Y
+; R600-DAG: RECIP_IEEE * T{{[0-9]\.[XYZW]}}, PS
+; R600-DAG: RECIPSQRT_IEEE * T{{[0-9]\.[XYZW]}}, KC0[3].Z
+; R600-DAG: RECIP_IEEE * T{{[0-9]\.[XYZW]}}, T{{[0-9]\.[XYZW]}}
+; R600-DAG: RECIPSQRT_IEEE * T{{[0-9]\.[XYZW]}}, KC0[3].W
+; R600-DAG: RECIP_IEEE * T{{[0-9]\.[XYZW]}}, T{{[0-9]\.[XYZW]}}
+; R600-DAG: RECIPSQRT_IEEE * T{{[0-9]\.[XYZW]}}, KC0[4].X
+; R600-DAG: RECIP_IEEE * T{{[0-9]\.[XYZW]}}, T{{[0-9]\.[XYZW]}}
 define amdgpu_kernel void @s_sqrt_v4f32(<4 x float> addrspace(1)* %out, <4 x float> %in) #1 {
 entry:
   %fdiv = call <4 x float> @llvm.sqrt.v4f32(<4 x float> %in)
@@ -134,6 +134,16 @@ entry:
   ret void
 }
 
+; FUNC-LABEL: {{^}}recip_sqrt:
+; R600: RECIPSQRT_IEEE
+; R600-NOT: RECIP_IEEE
+define amdgpu_kernel void @recip_sqrt(float addrspace(1)* %out, float %src) nounwind {
+  %sqrt = call float @llvm.sqrt.f32(float %src)
+  %recipsqrt = fdiv fast float 1.0, %sqrt
+  store float %recipsqrt, float addrspace(1)* %out, align 4
+  ret void
+}
+
 declare float @llvm.sqrt.f32(float %in) #0
 declare <2 x float> @llvm.sqrt.v2f32(<2 x float> %in) #0
 declare <4 x float> @llvm.sqrt.v4f32(<4 x float> %in) #0