[PATCH][x86] Add more patterns for SSE/AVX scalar single/double-precision fp arithmetic instructions.
Andrea Di Biagio
andrea.dibiagio at gmail.com
Wed Dec 11 05:11:11 PST 2013
Hi,
This patch teaches ISel how to select scalar fp arithmetic
instructions from a packed fp vector instruction followed by a vector
insert.
Example, the following function in IR:
define <4 x float> @foo(<4 x float> %a, <4 x float> %b) {
%1 = fadd <4 x float> %a, %b
%2 = shufflevector <4 x float> %1, <4 x float> %a, <4 x i32> <i32
0, i32 5, i32 6, i32 7>
ret <4 x float> %2
}
currently produces the following assembly
addps %xmm0, %xmm1
movss %xmm1, %xmm0
However it could be optimized into:
addss %xmm1, %xmm0
This patch adds new patterns to X86InstrSSE.td specifically for
selecting single/double-precision scalar fp arithmetic instructions
from a SSE/AVX packed fp instruction plus vector insert.
I also added tests to verify that when the new patterns are matched,
the generated code does not contain any redundant insert
(movss/movsd).
Please let me know what you think.
Thanks,
Andrea Di Biagio
SN Systems - Sony Computer Entertainment
-------------- next part --------------
Index: test/CodeGen/X86/sse-scalar-fp-arith-2.ll
===================================================================
--- test/CodeGen/X86/sse-scalar-fp-arith-2.ll (revision 0)
+++ test/CodeGen/X86/sse-scalar-fp-arith-2.ll (revision 0)
@@ -0,0 +1,215 @@
+; RUN: llc -mtriple=x86_64-pc-linux -mcpu=corei7 < %s | FileCheck -check-prefix=CHECK -check-prefix=SSE2 %s
+; RUN: llc -mtriple=x86_64-pc-linux -mattr=-sse4.1 -mcpu=corei7 < %s | FileCheck -check-prefix=CHECK -check-prefix=SSE2 %s
+; RUN: llc -mtriple=x86_64-pc-linux -mcpu=corei7-avx < %s | FileCheck -check-prefix=CHECK -check-prefix=AVX %s
+
+; Ensure that the backend selects SSE/AVX scalar fp instructions
+; from a packed fp instrution plus a vector insert.
+
+
+define <4 x float> @test_add_ss(<4 x float> %a, <4 x float> %b) {
+ %1 = fadd <4 x float> %a, %b
+ %2 = shufflevector <4 x float> %1, <4 x float> %a, <4 x i32> <i32 0, i32 5, i32 6, i32 7>
+ ret <4 x float> %2
+}
+
+; CHECK-LABEL: test_add_ss
+; SSE2: addss %xmm1, %xmm0
+; AVX: vaddss %xmm1, %xmm0, %xmm0
+; CHECK-NOT: movss
+; CHECK: ret
+
+
+define <4 x float> @test_sub_ss(<4 x float> %a, <4 x float> %b) {
+ %1 = fsub <4 x float> %a, %b
+ %2 = shufflevector <4 x float> %1, <4 x float> %a, <4 x i32> <i32 0, i32 5, i32 6, i32 7>
+ ret <4 x float> %2
+}
+
+; CHECK-LABEL: test_sub_ss
+; SSE2: subss %xmm1, %xmm0
+; AVX: vsubss %xmm1, %xmm0, %xmm0
+; CHECK-NOT: movss
+; CHECK: ret
+
+
+define <4 x float> @test_mul_ss(<4 x float> %a, <4 x float> %b) {
+ %1 = fmul <4 x float> %a, %b
+ %2 = shufflevector <4 x float> %1, <4 x float> %a, <4 x i32> <i32 0, i32 5, i32 6, i32 7>
+ ret <4 x float> %2
+}
+
+; CHECK-LABEL: test_mul_ss
+; SSE2: mulss %xmm1, %xmm0
+; AVX: vmulss %xmm1, %xmm0, %xmm0
+; CHECK-NOT: movss
+; CHECK: ret
+
+
+define <4 x float> @test_div_ss(<4 x float> %a, <4 x float> %b) {
+ %1 = fdiv <4 x float> %a, %b
+ %2 = shufflevector <4 x float> %1, <4 x float> %a, <4 x i32> <i32 0, i32 5, i32 6, i32 7>
+ ret <4 x float> %2
+}
+
+; CHECK-LABEL: test_div_ss
+; SSE2: divss %xmm1, %xmm0
+; AVX: vdivss %xmm1, %xmm0, %xmm0
+; CHECK-NOT: movss
+; CHECK: ret
+
+
+define <2 x double> @test_add_sd(<2 x double> %a, <2 x double> %b) {
+ %1 = fadd <2 x double> %a, %b
+ %2 = shufflevector <2 x double> %1, <2 x double> %a, <2 x i32> <i32 0, i32 3>
+ ret <2 x double> %2
+}
+
+; CHECK-LABEL: test_add_sd
+; SSE2: addsd %xmm1, %xmm0
+; AVX: vaddsd %xmm1, %xmm0, %xmm0
+; CHECK-NOT: movsd
+; CHECK: ret
+
+
+define <2 x double> @test_sub_sd(<2 x double> %a, <2 x double> %b) {
+ %1 = fsub <2 x double> %a, %b
+ %2 = shufflevector <2 x double> %1, <2 x double> %a, <2 x i32> <i32 0, i32 3>
+ ret <2 x double> %2
+}
+
+; CHECK-LABEL: test_sub_sd
+; SSE2: subsd %xmm1, %xmm0
+; AVX: vsubsd %xmm1, %xmm0, %xmm0
+; CHECK-NOT: movsd
+; CHECK: ret
+
+
+define <2 x double> @test_mul_sd(<2 x double> %a, <2 x double> %b) {
+ %1 = fmul <2 x double> %a, %b
+ %2 = shufflevector <2 x double> %1, <2 x double> %a, <2 x i32> <i32 0, i32 3>
+ ret <2 x double> %2
+}
+
+; CHECK-LABEL: test_mul_sd
+; SSE2: mulsd %xmm1, %xmm0
+; AVX: vmulsd %xmm1, %xmm0, %xmm0
+; CHECK-NOT: movsd
+; CHECK: ret
+
+
+define <2 x double> @test_div_sd(<2 x double> %a, <2 x double> %b) {
+ %1 = fdiv <2 x double> %a, %b
+ %2 = shufflevector <2 x double> %1, <2 x double> %a, <2 x i32> <i32 0, i32 3>
+ ret <2 x double> %2
+}
+
+; CHECK-LABEL: test_div_sd
+; SSE2: divsd %xmm1, %xmm0
+; AVX: vdivsd %xmm1, %xmm0, %xmm0
+; CHECK-NOT: movsd
+; CHECK: ret
+
+
+define <4 x float> @test2_add_ss(<4 x float> %a, <4 x float> %b) {
+ %1 = fadd <4 x float> %b, %a
+ %2 = shufflevector <4 x float> %1, <4 x float> %b, <4 x i32> <i32 0, i32 5, i32 6, i32 7>
+ ret <4 x float> %2
+}
+
+; CHECK-LABEL: test2_add_ss
+; SSE2: addss %xmm0, %xmm1
+; AVX: vaddss %xmm0, %xmm1, %xmm0
+; CHECK-NOT: movss
+; CHECK: ret
+
+
+define <4 x float> @test2_sub_ss(<4 x float> %a, <4 x float> %b) {
+ %1 = fsub <4 x float> %b, %a
+ %2 = shufflevector <4 x float> %1, <4 x float> %b, <4 x i32> <i32 0, i32 5, i32 6, i32 7>
+ ret <4 x float> %2
+}
+
+; CHECK-LABEL: test2_sub_ss
+; SSE2: subss %xmm0, %xmm1
+; AVX: vsubss %xmm0, %xmm1, %xmm0
+; CHECK-NOT: movss
+; CHECK: ret
+
+
+define <4 x float> @test2_mul_ss(<4 x float> %a, <4 x float> %b) {
+ %1 = fmul <4 x float> %b, %a
+ %2 = shufflevector <4 x float> %1, <4 x float> %b, <4 x i32> <i32 0, i32 5, i32 6, i32 7>
+ ret <4 x float> %2
+}
+
+; CHECK-LABEL: test2_mul_ss
+; SSE2: mulss %xmm0, %xmm1
+; AVX: vmulss %xmm0, %xmm1, %xmm0
+; CHECK-NOT: movss
+; CHECK: ret
+
+
+define <4 x float> @test2_div_ss(<4 x float> %a, <4 x float> %b) {
+ %1 = fdiv <4 x float> %b, %a
+ %2 = shufflevector <4 x float> %1, <4 x float> %b, <4 x i32> <i32 0, i32 5, i32 6, i32 7>
+ ret <4 x float> %2
+}
+
+; CHECK-LABEL: test2_div_ss
+; SSE2: divss %xmm0, %xmm1
+; AVX: vdivss %xmm0, %xmm1, %xmm0
+; CHECK-NOT: movss
+; CHECK: ret
+
+
+define <2 x double> @test2_add_sd(<2 x double> %a, <2 x double> %b) {
+ %1 = fadd <2 x double> %b, %a
+ %2 = shufflevector <2 x double> %1, <2 x double> %b, <2 x i32> <i32 0, i32 3>
+ ret <2 x double> %2
+}
+
+; CHECK-LABEL: test2_add_sd
+; SSE2: addsd %xmm0, %xmm1
+; AVX: vaddsd %xmm0, %xmm1, %xmm0
+; CHECK-NOT: movsd
+; CHECK: ret
+
+
+define <2 x double> @test2_sub_sd(<2 x double> %a, <2 x double> %b) {
+ %1 = fsub <2 x double> %b, %a
+ %2 = shufflevector <2 x double> %1, <2 x double> %b, <2 x i32> <i32 0, i32 3>
+ ret <2 x double> %2
+}
+
+; CHECK-LABEL: test2_sub_sd
+; SSE2: subsd %xmm0, %xmm1
+; AVX: vsubsd %xmm0, %xmm1, %xmm0
+; CHECK-NOT: movsd
+; CHECK: ret
+
+
+define <2 x double> @test2_mul_sd(<2 x double> %a, <2 x double> %b) {
+ %1 = fmul <2 x double> %b, %a
+ %2 = shufflevector <2 x double> %1, <2 x double> %b, <2 x i32> <i32 0, i32 3>
+ ret <2 x double> %2
+}
+
+; CHECK-LABEL: test2_mul_sd
+; SSE2: mulsd %xmm0, %xmm1
+; AVX: vmulsd %xmm0, %xmm1, %xmm0
+; CHECK-NOT: movsd
+; CHECK: ret
+
+
+define <2 x double> @test2_div_sd(<2 x double> %a, <2 x double> %b) {
+ %1 = fdiv <2 x double> %b, %a
+ %2 = shufflevector <2 x double> %1, <2 x double> %b, <2 x i32> <i32 0, i32 3>
+ ret <2 x double> %2
+}
+
+; CHECK-LABEL: test2_div_sd
+; SSE2: divsd %xmm0, %xmm1
+; AVX: vdivsd %xmm0, %xmm1, %xmm0
+; CHECK-NOT: movsd
+; CHECK: ret
+
Index: lib/Target/X86/X86InstrSSE.td
===================================================================
--- lib/Target/X86/X86InstrSSE.td (revision 197035)
+++ lib/Target/X86/X86InstrSSE.td (working copy)
@@ -3142,6 +3142,89 @@
(VDIVSSrr_Int v4f32:$dst, (COPY_TO_REGCLASS FR32:$src, VR128))>;
}
+// Patterns used to select SSE scalar fp arithmetic instructions from
+// a vector packed single/double fp operation followed by a vector insert.
+//
+// The effect is that the backend convert the packed fp instruction
+// followed by a vector insert into a single SSE scalar fp instruction.
+//
+// For example, given the following code:
+// __m128 foo(__m128 A, __m128 B) {
+// __m128 C = A + B;
+// return (__m128) {c[0], a[1], a[2], a[3]};
+// }
+//
+// previously we generated:
+// addps %xmm0, %xmm1
+// movss %xmm1, %xmm0
+//
+// we now generate:
+// addss %xmm1, %xmm0
+
+def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
+ (fadd (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
+ (ADDSSrr_Int v4f32:$dst, v4f32:$src)>;
+def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
+ (fsub (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
+ (SUBSSrr_Int v4f32:$dst, v4f32:$src)>;
+def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
+ (fmul (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
+ (MULSSrr_Int v4f32:$dst, v4f32:$src)>;
+def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
+ (fdiv (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
+ (DIVSSrr_Int v4f32:$dst, v4f32:$src)>;
+
+let Predicates = [HasSSE2] in {
+ // SSE2 patterns to select scalar double-precision fp arithmetic instructions
+ // from a packed double-precision fp instruction plus movsd.
+
+ def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
+ (fadd (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
+ (ADDSDrr_Int v2f64:$dst, v2f64:$src)>;
+ def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
+ (fsub (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
+ (SUBSDrr_Int v2f64:$dst, v2f64:$src)>;
+ def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
+ (fmul (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
+ (MULSDrr_Int v2f64:$dst, v2f64:$src)>;
+ def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
+ (fdiv (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
+ (DIVSDrr_Int v2f64:$dst, v2f64:$src)>;
+}
+
+let AddedComplexity = 20, Predicates = [HasAVX] in {
+ // The following patterns select AVX Scalar single/double precision fp
+ // arithmetic instructions from a packed single precision fp instruction
+ // plus movss/movsd.
+ // The 'AddedComplexity' is required to give them higher priority over
+ // the equivalent SSE/SSE2 patterns.
+
+ def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
+ (fadd (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
+ (VADDSSrr_Int v4f32:$dst, v4f32:$src)>;
+ def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
+ (fsub (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
+ (VSUBSSrr_Int v4f32:$dst, v4f32:$src)>;
+ def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
+ (fmul (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
+ (VMULSSrr_Int v4f32:$dst, v4f32:$src)>;
+ def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
+ (fdiv (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
+ (VDIVSSrr_Int v4f32:$dst, v4f32:$src)>;
+ def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
+ (fadd (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
+ (VADDSDrr_Int v2f64:$dst, v2f64:$src)>;
+ def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
+ (fsub (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
+ (VSUBSDrr_Int v2f64:$dst, v2f64:$src)>;
+ def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
+ (fmul (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
+ (VMULSDrr_Int v2f64:$dst, v2f64:$src)>;
+ def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
+ (fdiv (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
+ (VDIVSDrr_Int v2f64:$dst, v2f64:$src)>;
+}
+
/// Unop Arithmetic
/// In addition, we also have a special variant of the scalar form here to
/// represent the associated intrinsic operation. This form is unlike the
More information about the llvm-commits
mailing list