AMDGPU: Partially implement getArithmeticInstrCost for FP ops

llvm-svn: 264374

4 files changed, 358 insertions, 0 deletions

diff --git a/llvm/test/Analysis/CostModel/AMDGPU/fadd.ll b/llvm/test/Analysis/CostModel/AMDGPU/fadd.ll
new file mode 100644
index 00000000000..00e91bd6223
--- /dev/null
+++ b/llvm/test/Analysis/CostModel/AMDGPU/fadd.ll
@@ -0,0 +1,88 @@
+; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mattr=+half-rate-64-ops < %s | FileCheck -check-prefix=FASTF64 -check-prefix=ALL %s
+; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mattr=-half-rate-64-ops < %s | FileCheck -check-prefix=SLOWF64 -check-prefix=ALL %s
+
+; ALL: 'fadd_f32'
+; ALL: estimated cost of 1 for {{.*}} fadd float
+define void @fadd_f32(float addrspace(1)* %out, float addrspace(1)* %vaddr, float %b) #0 {
+  %vec = load float, float addrspace(1)* %vaddr
+  %add = fadd float %vec, %b
+  store float %add, float addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fadd_v2f32'
+; ALL: estimated cost of 2 for {{.*}} fadd <2 x float>
+define void @fadd_v2f32(<2 x float> addrspace(1)* %out, <2 x float> addrspace(1)* %vaddr, <2 x float> %b) #0 {
+  %vec = load <2 x float>, <2 x float> addrspace(1)* %vaddr
+  %add = fadd <2 x float> %vec, %b
+  store <2 x float> %add, <2 x float> addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fadd_v3f32'
+; ALL: estimated cost of 3 for {{.*}} fadd <3 x float>
+define void @fadd_v3f32(<3 x float> addrspace(1)* %out, <3 x float> addrspace(1)* %vaddr, <3 x float> %b) #0 {
+  %vec = load <3 x float>, <3 x float> addrspace(1)* %vaddr
+  %add = fadd <3 x float> %vec, %b
+  store <3 x float> %add, <3 x float> addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fadd_f64'
+; FASTF64: estimated cost of 2 for {{.*}} fadd double
+; SLOWF64: estimated cost of 3 for {{.*}} fadd double
+define void @fadd_f64(double addrspace(1)* %out, double addrspace(1)* %vaddr, double %b) #0 {
+  %vec = load double, double addrspace(1)* %vaddr
+  %add = fadd double %vec, %b
+  store double %add, double addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fadd_v2f64'
+; FASTF64: estimated cost of 4 for {{.*}} fadd <2 x double>
+; SLOWF64: estimated cost of 6 for {{.*}} fadd <2 x double>
+define void @fadd_v2f64(<2 x double> addrspace(1)* %out, <2 x double> addrspace(1)* %vaddr, <2 x double> %b) #0 {
+  %vec = load <2 x double>, <2 x double> addrspace(1)* %vaddr
+  %add = fadd <2 x double> %vec, %b
+  store <2 x double> %add, <2 x double> addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fadd_v3f64'
+; FASTF64: estimated cost of 6 for {{.*}} fadd <3 x double>
+; SLOWF64: estimated cost of 9 for {{.*}} fadd <3 x double>
+define void @fadd_v3f64(<3 x double> addrspace(1)* %out, <3 x double> addrspace(1)* %vaddr, <3 x double> %b) #0 {
+  %vec = load <3 x double>, <3 x double> addrspace(1)* %vaddr
+  %add = fadd <3 x double> %vec, %b
+  store <3 x double> %add, <3 x double> addrspace(1)* %out
+  ret void
+}
+
+; ALL 'fadd_f16'
+; ALL estimated cost of 1 for {{.*}} fadd half
+define void @fadd_f16(half addrspace(1)* %out, half addrspace(1)* %vaddr, half %b) #0 {
+  %vec = load half, half addrspace(1)* %vaddr
+  %add = fadd half %vec, %b
+  store half %add, half addrspace(1)* %out
+  ret void
+}
+
+; ALL 'fadd_v2f16'
+; ALL estimated cost of 2 for {{.*}} fadd <2 x half>
+define void @fadd_v2f16(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %vaddr, <2 x half> %b) #0 {
+  %vec = load <2 x half>, <2 x half> addrspace(1)* %vaddr
+  %add = fadd <2 x half> %vec, %b
+  store <2 x half> %add, <2 x half> addrspace(1)* %out
+  ret void
+}
+
+; ALL 'fadd_v4f16'
+; ALL estimated cost of 4 for {{.*}} fadd <4 x half>
+define void @fadd_v4f16(<4 x half> addrspace(1)* %out, <4 x half> addrspace(1)* %vaddr, <4 x half> %b) #0 {
+  %vec = load <4 x half>, <4 x half> addrspace(1)* %vaddr
+  %add = fadd <4 x half> %vec, %b
+  store <4 x half> %add, <4 x half> addrspace(1)* %out
+  ret void
+}
+
+attributes #0 = { nounwind }
diff --git a/llvm/test/Analysis/CostModel/AMDGPU/fdiv.ll b/llvm/test/Analysis/CostModel/AMDGPU/fdiv.ll
new file mode 100644
index 00000000000..3f374422ad9
--- /dev/null
+++ b/llvm/test/Analysis/CostModel/AMDGPU/fdiv.ll
@@ -0,0 +1,96 @@
+; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mcpu=hawaii -mattr=+half-rate-64-ops < %s | FileCheck -check-prefix=ALL -check-prefix=CIFASTF64 %s
+; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mcpu=kaveri -mattr=-half-rate-64-ops < %s | FileCheck -check-prefix=ALL -check-prefix=CISLOWF64 %s
+; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mcpu=tahiti -mattr=+half-rate-64-ops < %s | FileCheck -check-prefix=ALL -check-prefix=SIFASTF64 %s
+; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mcpu=verde -mattr=-half-rate-64-ops < %s | FileCheck -check-prefix=ALL -check-prefix=SISLOWF64 %s
+
+; CHECK: 'fdiv_f32'
+; ALL: estimated cost of 10 for {{.*}} fdiv float
+define void @fdiv_f32(float addrspace(1)* %out, float addrspace(1)* %vaddr, float %b) #0 {
+  %vec = load float, float addrspace(1)* %vaddr
+  %add = fdiv float %vec, %b
+  store float %add, float addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fdiv_v2f32'
+; ALL: estimated cost of 20 for {{.*}} fdiv <2 x float>
+define void @fdiv_v2f32(<2 x float> addrspace(1)* %out, <2 x float> addrspace(1)* %vaddr, <2 x float> %b) #0 {
+  %vec = load <2 x float>, <2 x float> addrspace(1)* %vaddr
+  %add = fdiv <2 x float> %vec, %b
+  store <2 x float> %add, <2 x float> addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fdiv_v3f32'
+; ALL: estimated cost of 30 for {{.*}} fdiv <3 x float>
+define void @fdiv_v3f32(<3 x float> addrspace(1)* %out, <3 x float> addrspace(1)* %vaddr, <3 x float> %b) #0 {
+  %vec = load <3 x float>, <3 x float> addrspace(1)* %vaddr
+  %add = fdiv <3 x float> %vec, %b
+  store <3 x float> %add, <3 x float> addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fdiv_f64'
+; CIFASTF64: estimated cost of 29 for {{.*}} fdiv double
+; CISLOWF64: estimated cost of 33 for {{.*}} fdiv double
+; SIFASTF64: estimated cost of 32 for {{.*}} fdiv double
+; SISLOWF64: estimated cost of 36 for {{.*}} fdiv double
+define void @fdiv_f64(double addrspace(1)* %out, double addrspace(1)* %vaddr, double %b) #0 {
+  %vec = load double, double addrspace(1)* %vaddr
+  %add = fdiv double %vec, %b
+  store double %add, double addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fdiv_v2f64'
+; CIFASTF64: estimated cost of 58 for {{.*}} fdiv <2 x double>
+; CISLOWF64: estimated cost of 66 for {{.*}} fdiv <2 x double>
+; SIFASTF64: estimated cost of 64 for {{.*}} fdiv <2 x double>
+; SISLOWF64: estimated cost of 72 for {{.*}} fdiv <2 x double>
+define void @fdiv_v2f64(<2 x double> addrspace(1)* %out, <2 x double> addrspace(1)* %vaddr, <2 x double> %b) #0 {
+  %vec = load <2 x double>, <2 x double> addrspace(1)* %vaddr
+  %add = fdiv <2 x double> %vec, %b
+  store <2 x double> %add, <2 x double> addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fdiv_v3f64'
+; CIFASTF64: estimated cost of 87 for {{.*}} fdiv <3 x double>
+; CISLOWF64: estimated cost of 99 for {{.*}} fdiv <3 x double>
+; SIFASTF64: estimated cost of 96 for {{.*}} fdiv <3 x double>
+; SISLOWF64: estimated cost of 108 for {{.*}} fdiv <3 x double>
+define void @fdiv_v3f64(<3 x double> addrspace(1)* %out, <3 x double> addrspace(1)* %vaddr, <3 x double> %b) #0 {
+  %vec = load <3 x double>, <3 x double> addrspace(1)* %vaddr
+  %add = fdiv <3 x double> %vec, %b
+  store <3 x double> %add, <3 x double> addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fdiv_f16'
+; ALL: estimated cost of 10 for {{.*}} fdiv half
+define void @fdiv_f16(half addrspace(1)* %out, half addrspace(1)* %vaddr, half %b) #0 {
+  %vec = load half, half addrspace(1)* %vaddr
+  %add = fdiv half %vec, %b
+  store half %add, half addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fdiv_v2f16'
+; ALL: estimated cost of 20 for {{.*}} fdiv <2 x half>
+define void @fdiv_v2f16(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %vaddr, <2 x half> %b) #0 {
+  %vec = load <2 x half>, <2 x half> addrspace(1)* %vaddr
+  %add = fdiv <2 x half> %vec, %b
+  store <2 x half> %add, <2 x half> addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fdiv_v4f16'
+; ALL: estimated cost of 40 for {{.*}} fdiv <4 x half>
+define void @fdiv_v4f16(<4 x half> addrspace(1)* %out, <4 x half> addrspace(1)* %vaddr, <4 x half> %b) #0 {
+  %vec = load <4 x half>, <4 x half> addrspace(1)* %vaddr
+  %add = fdiv <4 x half> %vec, %b
+  store <4 x half> %add, <4 x half> addrspace(1)* %out
+  ret void
+}
+
+attributes #0 = { nounwind }
diff --git a/llvm/test/Analysis/CostModel/AMDGPU/fmul.ll b/llvm/test/Analysis/CostModel/AMDGPU/fmul.ll
new file mode 100644
index 00000000000..6303bb7988c
--- /dev/null
+++ b/llvm/test/Analysis/CostModel/AMDGPU/fmul.ll
@@ -0,0 +1,88 @@
+; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mattr=+half-rate-64-ops < %s | FileCheck -check-prefix=FASTF64 -check-prefix=ALL %s
+; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mattr=-half-rate-64-ops < %s | FileCheck -check-prefix=SLOWF64 -check-prefix=ALL %s
+
+; ALL: 'fmul_f32'
+; ALL: estimated cost of 1 for {{.*}} fmul float
+define void @fmul_f32(float addrspace(1)* %out, float addrspace(1)* %vaddr, float %b) #0 {
+  %vec = load float, float addrspace(1)* %vaddr
+  %add = fmul float %vec, %b
+  store float %add, float addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fmul_v2f32'
+; ALL: estimated cost of 2 for {{.*}} fmul <2 x float>
+define void @fmul_v2f32(<2 x float> addrspace(1)* %out, <2 x float> addrspace(1)* %vaddr, <2 x float> %b) #0 {
+  %vec = load <2 x float>, <2 x float> addrspace(1)* %vaddr
+  %add = fmul <2 x float> %vec, %b
+  store <2 x float> %add, <2 x float> addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fmul_v3f32'
+; ALL: estimated cost of 3 for {{.*}} fmul <3 x float>
+define void @fmul_v3f32(<3 x float> addrspace(1)* %out, <3 x float> addrspace(1)* %vaddr, <3 x float> %b) #0 {
+  %vec = load <3 x float>, <3 x float> addrspace(1)* %vaddr
+  %add = fmul <3 x float> %vec, %b
+  store <3 x float> %add, <3 x float> addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fmul_f64'
+; FASTF64: estimated cost of 2 for {{.*}} fmul double
+; SLOWF64: estimated cost of 3 for {{.*}} fmul double
+define void @fmul_f64(double addrspace(1)* %out, double addrspace(1)* %vaddr, double %b) #0 {
+  %vec = load double, double addrspace(1)* %vaddr
+  %add = fmul double %vec, %b
+  store double %add, double addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fmul_v2f64'
+; FASTF64: estimated cost of 4 for {{.*}} fmul <2 x double>
+; SLOWF64: estimated cost of 6 for {{.*}} fmul <2 x double>
+define void @fmul_v2f64(<2 x double> addrspace(1)* %out, <2 x double> addrspace(1)* %vaddr, <2 x double> %b) #0 {
+  %vec = load <2 x double>, <2 x double> addrspace(1)* %vaddr
+  %add = fmul <2 x double> %vec, %b
+  store <2 x double> %add, <2 x double> addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fmul_v3f64'
+; FASTF64: estimated cost of 6 for {{.*}} fmul <3 x double>
+; SLOWF64: estimated cost of 9 for {{.*}} fmul <3 x double>
+define void @fmul_v3f64(<3 x double> addrspace(1)* %out, <3 x double> addrspace(1)* %vaddr, <3 x double> %b) #0 {
+  %vec = load <3 x double>, <3 x double> addrspace(1)* %vaddr
+  %add = fmul <3 x double> %vec, %b
+  store <3 x double> %add, <3 x double> addrspace(1)* %out
+  ret void
+}
+
+; ALL 'fmul_f16'
+; ALL estimated cost of 1 for {{.*}} fmul half
+define void @fmul_f16(half addrspace(1)* %out, half addrspace(1)* %vaddr, half %b) #0 {
+  %vec = load half, half addrspace(1)* %vaddr
+  %add = fmul half %vec, %b
+  store half %add, half addrspace(1)* %out
+  ret void
+}
+
+; ALL 'fmul_v2f16'
+; ALL estimated cost of 2 for {{.*}} fmul <2 x half>
+define void @fmul_v2f16(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %vaddr, <2 x half> %b) #0 {
+  %vec = load <2 x half>, <2 x half> addrspace(1)* %vaddr
+  %add = fmul <2 x half> %vec, %b
+  store <2 x half> %add, <2 x half> addrspace(1)* %out
+  ret void
+}
+
+; ALL 'fmul_v4f16'
+; ALL estimated cost of 4 for {{.*}} fmul <4 x half>
+define void @fmul_v4f16(<4 x half> addrspace(1)* %out, <4 x half> addrspace(1)* %vaddr, <4 x half> %b) #0 {
+  %vec = load <4 x half>, <4 x half> addrspace(1)* %vaddr
+  %add = fmul <4 x half> %vec, %b
+  store <4 x half> %add, <4 x half> addrspace(1)* %out
+  ret void
+}
+
+attributes #0 = { nounwind }
diff --git a/llvm/test/Analysis/CostModel/AMDGPU/fsub.ll b/llvm/test/Analysis/CostModel/AMDGPU/fsub.ll
new file mode 100644
index 00000000000..e0850be9867
--- /dev/null
+++ b/llvm/test/Analysis/CostModel/AMDGPU/fsub.ll
@@ -0,0 +1,86 @@
+; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mattr=+half-rate-64-ops < %s | FileCheck -check-prefix=FASTF64 -check-prefix=ALL %s
+; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mattr=-half-rate-64-ops < %s | FileCheck -check-prefix=SLOWF64 -check-prefix=ALL %s
+
+; ALL: 'fsub_f32'
+; ALL: estimated cost of 1 for {{.*}} fsub float
+define void @fsub_f32(float addrspace(1)* %out, float addrspace(1)* %vaddr, float %b) #0 {
+  %vec = load float, float addrspace(1)* %vaddr
+  %add = fsub float %vec, %b
+  store float %add, float addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fsub_v2f32'
+; ALL: estimated cost of 2 for {{.*}} fsub <2 x float>
+define void @fsub_v2f32(<2 x float> addrspace(1)* %out, <2 x float> addrspace(1)* %vaddr, <2 x float> %b) #0 {
+  %vec = load <2 x float>, <2 x float> addrspace(1)* %vaddr
+  %add = fsub <2 x float> %vec, %b
+  store <2 x float> %add, <2 x float> addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fsub_v3f32'
+; ALL: estimated cost of 3 for {{.*}} fsub <3 x float>
+define void @fsub_v3f32(<3 x float> addrspace(1)* %out, <3 x float> addrspace(1)* %vaddr, <3 x float> %b) #0 {
+  %vec = load <3 x float>, <3 x float> addrspace(1)* %vaddr
+  %add = fsub <3 x float> %vec, %b
+  store <3 x float> %add, <3 x float> addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fsub_f64'
+; FASTF64: estimated cost of 2 for {{.*}} fsub double
+; SLOWF64: estimated cost of 3 for {{.*}} fsub double
+define void @fsub_f64(double addrspace(1)* %out, double addrspace(1)* %vaddr, double %b) #0 {
+  %vec = load double, double addrspace(1)* %vaddr
+  %add = fsub double %vec, %b
+  store double %add, double addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fsub_v2f64'
+; FASTF64: estimated cost of 4 for {{.*}} fsub <2 x double>
+; SLOWF64: estimated cost of 6 for {{.*}} fsub <2 x double>
+define void @fsub_v2f64(<2 x double> addrspace(1)* %out, <2 x double> addrspace(1)* %vaddr, <2 x double> %b) #0 {
+  %vec = load <2 x double>, <2 x double> addrspace(1)* %vaddr
+  %add = fsub <2 x double> %vec, %b
+  store <2 x double> %add, <2 x double> addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fsub_v3f64'
+; FASTF64: estimated cost of 6 for {{.*}} fsub <3 x double>
+; SLOWF64: estimated cost of 9 for {{.*}} fsub <3 x double>
+define void @fsub_v3f64(<3 x double> addrspace(1)* %out, <3 x double> addrspace(1)* %vaddr, <3 x double> %b) #0 {
+  %vec = load <3 x double>, <3 x double> addrspace(1)* %vaddr
+  %add = fsub <3 x double> %vec, %b
+  store <3 x double> %add, <3 x double> addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fsub_f16'
+; ALL: estimated cost of 1 for {{.*}} fsub half
+define void @fsub_f16(half addrspace(1)* %out, half addrspace(1)* %vaddr, half %b) #0 {
+  %vec = load half, half addrspace(1)* %vaddr
+  %add = fsub half %vec, %b
+  store half %add, half addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fsub_v2f16'
+; ALL: estimated cost of 2 for {{.*}} fsub <2 x half>
+define void @fsub_v2f16(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %vaddr, <2 x half> %b) #0 {
+  %vec = load <2 x half>, <2 x half> addrspace(1)* %vaddr
+  %add = fsub <2 x half> %vec, %b
+  store <2 x half> %add, <2 x half> addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'fsub_v4f16'
+; ALL: estimated cost of 4 for {{.*}} fsub <4 x half>
+define void @fsub_v4f16(<4 x half> addrspace(1)* %out, <4 x half> addrspace(1)* %vaddr, <4 x half> %b) #0 {
+  %vec = load <4 x half>, <4 x half> addrspace(1)* %vaddr
+  %add = fsub <4 x half> %vec, %b
+  store <4 x half> %add, <4 x half> addrspace(1)* %out
+  ret void
+}