[llvm] 387c157 - [NFC][Codegen] Tests with wide scalar shifts, for new potential legalization strategy

Roman Lebedev via llvm-commits llvm-commits at lists.llvm.org
Fri Dec 23 14:30:12 PST 2022


Author: Roman Lebedev
Date: 2022-12-24T00:47:25+03:00
New Revision: 387c1573f89117687f4b964ae3a90ea7c91a4f90

URL: https://github.com/llvm/llvm-project/commit/387c1573f89117687f4b964ae3a90ea7c91a4f90
DIFF: https://github.com/llvm/llvm-project/commit/387c1573f89117687f4b964ae3a90ea7c91a4f90.diff

LOG: [NFC][Codegen] Tests with wide scalar shifts, for new potential legalization strategy

Added: 
    llvm/test/CodeGen/AArch64/wide-scalar-shift-legalization.ll
    llvm/test/CodeGen/PowerPC/wide-scalar-shift-legalization.ll
    llvm/test/CodeGen/X86/wide-scalar-shift-legalization.ll

Modified: 
    

Removed: 
    


################################################################################
diff  --git a/llvm/test/CodeGen/AArch64/wide-scalar-shift-legalization.ll b/llvm/test/CodeGen/AArch64/wide-scalar-shift-legalization.ll
new file mode 100644
index 0000000000000..b909fd3229f70
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/wide-scalar-shift-legalization.ll
@@ -0,0 +1,380 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=aarch64-- | FileCheck %s --check-prefixes=ALL
+
+define void @lshr_4bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; ALL-LABEL: lshr_4bytes:
+; ALL:       // %bb.0:
+; ALL-NEXT:    ldr w8, [x1]
+; ALL-NEXT:    ldr w9, [x0]
+; ALL-NEXT:    lsl w8, w8, #3
+; ALL-NEXT:    lsr w8, w9, w8
+; ALL-NEXT:    str w8, [x2]
+; ALL-NEXT:    ret
+  %src = load i32, ptr %src.ptr, align 1
+  %byteOff = load i32, ptr %byteOff.ptr, align 1
+  %bitOff = shl i32 %byteOff, 3
+  %res = lshr i32 %src, %bitOff
+  store i32 %res, ptr %dst, align 1
+  ret void
+}
+define void @shl_4bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; ALL-LABEL: shl_4bytes:
+; ALL:       // %bb.0:
+; ALL-NEXT:    ldr w8, [x1]
+; ALL-NEXT:    ldr w9, [x0]
+; ALL-NEXT:    lsl w8, w8, #3
+; ALL-NEXT:    lsl w8, w9, w8
+; ALL-NEXT:    str w8, [x2]
+; ALL-NEXT:    ret
+  %src = load i32, ptr %src.ptr, align 1
+  %byteOff = load i32, ptr %byteOff.ptr, align 1
+  %bitOff = shl i32 %byteOff, 3
+  %res = shl i32 %src, %bitOff
+  store i32 %res, ptr %dst, align 1
+  ret void
+}
+define void @ashr_4bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; ALL-LABEL: ashr_4bytes:
+; ALL:       // %bb.0:
+; ALL-NEXT:    ldr w8, [x1]
+; ALL-NEXT:    ldr w9, [x0]
+; ALL-NEXT:    lsl w8, w8, #3
+; ALL-NEXT:    asr w8, w9, w8
+; ALL-NEXT:    str w8, [x2]
+; ALL-NEXT:    ret
+  %src = load i32, ptr %src.ptr, align 1
+  %byteOff = load i32, ptr %byteOff.ptr, align 1
+  %bitOff = shl i32 %byteOff, 3
+  %res = ashr i32 %src, %bitOff
+  store i32 %res, ptr %dst, align 1
+  ret void
+}
+
+define void @lshr_8bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; ALL-LABEL: lshr_8bytes:
+; ALL:       // %bb.0:
+; ALL-NEXT:    ldr x8, [x1]
+; ALL-NEXT:    ldr x9, [x0]
+; ALL-NEXT:    lsl x8, x8, #3
+; ALL-NEXT:    lsr x8, x9, x8
+; ALL-NEXT:    str x8, [x2]
+; ALL-NEXT:    ret
+  %src = load i64, ptr %src.ptr, align 1
+  %byteOff = load i64, ptr %byteOff.ptr, align 1
+  %bitOff = shl i64 %byteOff, 3
+  %res = lshr i64 %src, %bitOff
+  store i64 %res, ptr %dst, align 1
+  ret void
+}
+define void @shl_8bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; ALL-LABEL: shl_8bytes:
+; ALL:       // %bb.0:
+; ALL-NEXT:    ldr x8, [x1]
+; ALL-NEXT:    ldr x9, [x0]
+; ALL-NEXT:    lsl x8, x8, #3
+; ALL-NEXT:    lsl x8, x9, x8
+; ALL-NEXT:    str x8, [x2]
+; ALL-NEXT:    ret
+  %src = load i64, ptr %src.ptr, align 1
+  %byteOff = load i64, ptr %byteOff.ptr, align 1
+  %bitOff = shl i64 %byteOff, 3
+  %res = shl i64 %src, %bitOff
+  store i64 %res, ptr %dst, align 1
+  ret void
+}
+define void @ashr_8bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; ALL-LABEL: ashr_8bytes:
+; ALL:       // %bb.0:
+; ALL-NEXT:    ldr x8, [x1]
+; ALL-NEXT:    ldr x9, [x0]
+; ALL-NEXT:    lsl x8, x8, #3
+; ALL-NEXT:    asr x8, x9, x8
+; ALL-NEXT:    str x8, [x2]
+; ALL-NEXT:    ret
+  %src = load i64, ptr %src.ptr, align 1
+  %byteOff = load i64, ptr %byteOff.ptr, align 1
+  %bitOff = shl i64 %byteOff, 3
+  %res = ashr i64 %src, %bitOff
+  store i64 %res, ptr %dst, align 1
+  ret void
+}
+
+define void @lshr_16bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; ALL-LABEL: lshr_16bytes:
+; ALL:       // %bb.0:
+; ALL-NEXT:    ldr x8, [x1]
+; ALL-NEXT:    ldp x10, x9, [x0]
+; ALL-NEXT:    lsl x8, x8, #3
+; ALL-NEXT:    and x11, x8, #0x38
+; ALL-NEXT:    mvn w12, w8
+; ALL-NEXT:    tst x8, #0x40
+; ALL-NEXT:    lsl x13, x9, #1
+; ALL-NEXT:    lsr x10, x10, x11
+; ALL-NEXT:    lsl x12, x13, x12
+; ALL-NEXT:    lsr x9, x9, x11
+; ALL-NEXT:    orr x8, x12, x10
+; ALL-NEXT:    csel x10, xzr, x9, ne
+; ALL-NEXT:    csel x8, x9, x8, ne
+; ALL-NEXT:    stp x8, x10, [x2]
+; ALL-NEXT:    ret
+  %src = load i128, ptr %src.ptr, align 1
+  %byteOff = load i128, ptr %byteOff.ptr, align 1
+  %bitOff = shl i128 %byteOff, 3
+  %res = lshr i128 %src, %bitOff
+  store i128 %res, ptr %dst, align 1
+  ret void
+}
+define void @shl_16bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; ALL-LABEL: shl_16bytes:
+; ALL:       // %bb.0:
+; ALL-NEXT:    ldr x8, [x1]
+; ALL-NEXT:    ldp x9, x10, [x0]
+; ALL-NEXT:    lsl x8, x8, #3
+; ALL-NEXT:    and x11, x8, #0x38
+; ALL-NEXT:    mvn w12, w8
+; ALL-NEXT:    lsr x13, x9, #1
+; ALL-NEXT:    tst x8, #0x40
+; ALL-NEXT:    lsl x10, x10, x11
+; ALL-NEXT:    lsr x12, x13, x12
+; ALL-NEXT:    lsl x9, x9, x11
+; ALL-NEXT:    orr x8, x10, x12
+; ALL-NEXT:    csel x10, xzr, x9, ne
+; ALL-NEXT:    csel x8, x9, x8, ne
+; ALL-NEXT:    stp x10, x8, [x2]
+; ALL-NEXT:    ret
+  %src = load i128, ptr %src.ptr, align 1
+  %byteOff = load i128, ptr %byteOff.ptr, align 1
+  %bitOff = shl i128 %byteOff, 3
+  %res = shl i128 %src, %bitOff
+  store i128 %res, ptr %dst, align 1
+  ret void
+}
+define void @ashr_16bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; ALL-LABEL: ashr_16bytes:
+; ALL:       // %bb.0:
+; ALL-NEXT:    ldr x8, [x1]
+; ALL-NEXT:    ldp x10, x9, [x0]
+; ALL-NEXT:    lsl x8, x8, #3
+; ALL-NEXT:    and x11, x8, #0x38
+; ALL-NEXT:    mvn w12, w8
+; ALL-NEXT:    tst x8, #0x40
+; ALL-NEXT:    lsl x13, x9, #1
+; ALL-NEXT:    asr x8, x9, #63
+; ALL-NEXT:    lsr x10, x10, x11
+; ALL-NEXT:    lsl x12, x13, x12
+; ALL-NEXT:    asr x11, x9, x11
+; ALL-NEXT:    orr x9, x12, x10
+; ALL-NEXT:    csel x8, x8, x11, ne
+; ALL-NEXT:    csel x9, x11, x9, ne
+; ALL-NEXT:    stp x9, x8, [x2]
+; ALL-NEXT:    ret
+  %src = load i128, ptr %src.ptr, align 1
+  %byteOff = load i128, ptr %byteOff.ptr, align 1
+  %bitOff = shl i128 %byteOff, 3
+  %res = ashr i128 %src, %bitOff
+  store i128 %res, ptr %dst, align 1
+  ret void
+}
+
+define void @lshr_32bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; ALL-LABEL: lshr_32bytes:
+; ALL:       // %bb.0:
+; ALL-NEXT:    ldr x9, [x1]
+; ALL-NEXT:    mov w8, #128
+; ALL-NEXT:    ldp x11, x10, [x0, #8]
+; ALL-NEXT:    lsl x9, x9, #3
+; ALL-NEXT:    ldr x12, [x0]
+; ALL-NEXT:    sub x8, x8, x9
+; ALL-NEXT:    ldr x13, [x0, #24]
+; ALL-NEXT:    and x17, x8, #0x38
+; ALL-NEXT:    mvn w0, w8
+; ALL-NEXT:    lsr x14, x10, #1
+; ALL-NEXT:    and x15, x9, #0x38
+; ALL-NEXT:    mvn w16, w9
+; ALL-NEXT:    tst x8, #0x40
+; ALL-NEXT:    lsl x3, x13, x17
+; ALL-NEXT:    lsr x14, x14, x0
+; ALL-NEXT:    lsl x17, x10, x17
+; ALL-NEXT:    orr x14, x3, x14
+; ALL-NEXT:    lsl x18, x13, #1
+; ALL-NEXT:    csel x0, xzr, x17, ne
+; ALL-NEXT:    csel x14, x17, x14, ne
+; ALL-NEXT:    lsl x17, x11, #1
+; ALL-NEXT:    lsr x8, x10, x15
+; ALL-NEXT:    lsl x1, x18, x16
+; ALL-NEXT:    lsr x3, x12, x15
+; ALL-NEXT:    lsl x16, x17, x16
+; ALL-NEXT:    orr x8, x1, x8
+; ALL-NEXT:    lsr x1, x13, x15
+; ALL-NEXT:    tst x9, #0x40
+; ALL-NEXT:    orr x16, x16, x3
+; ALL-NEXT:    lsr x15, x11, x15
+; ALL-NEXT:    csel x8, x1, x8, ne
+; ALL-NEXT:    csel x16, x15, x16, ne
+; ALL-NEXT:    csel x15, xzr, x15, ne
+; ALL-NEXT:    csel x17, xzr, x1, ne
+; ALL-NEXT:    subs x1, x9, #128
+; ALL-NEXT:    and x3, x1, #0x38
+; ALL-NEXT:    mvn w4, w1
+; ALL-NEXT:    csel x17, x17, xzr, lo
+; ALL-NEXT:    tst x1, #0x40
+; ALL-NEXT:    orr x16, x16, x0
+; ALL-NEXT:    orr x14, x15, x14
+; ALL-NEXT:    lsr x10, x10, x3
+; ALL-NEXT:    lsl x18, x18, x4
+; ALL-NEXT:    orr x10, x18, x10
+; ALL-NEXT:    lsr x13, x13, x3
+; ALL-NEXT:    csel x10, x13, x10, ne
+; ALL-NEXT:    csel x13, xzr, x13, ne
+; ALL-NEXT:    cmp x9, #128
+; ALL-NEXT:    csel x10, x16, x10, lo
+; ALL-NEXT:    csel x8, x8, xzr, lo
+; ALL-NEXT:    csel x13, x14, x13, lo
+; ALL-NEXT:    cmp x9, #0
+; ALL-NEXT:    csel x9, x12, x10, eq
+; ALL-NEXT:    csel x10, x11, x13, eq
+; ALL-NEXT:    stp x8, x17, [x2, #16]
+; ALL-NEXT:    stp x9, x10, [x2]
+; ALL-NEXT:    ret
+  %src = load i256, ptr %src.ptr, align 1
+  %byteOff = load i256, ptr %byteOff.ptr, align 1
+  %bitOff = shl i256 %byteOff, 3
+  %res = lshr i256 %src, %bitOff
+  store i256 %res, ptr %dst, align 1
+  ret void
+}
+define void @shl_32bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; ALL-LABEL: shl_32bytes:
+; ALL:       // %bb.0:
+; ALL-NEXT:    ldr x9, [x1]
+; ALL-NEXT:    mov w8, #128
+; ALL-NEXT:    ldp x10, x11, [x0, #8]
+; ALL-NEXT:    lsl x9, x9, #3
+; ALL-NEXT:    ldr x12, [x0, #24]
+; ALL-NEXT:    sub x8, x8, x9
+; ALL-NEXT:    ldr x13, [x0]
+; ALL-NEXT:    and x17, x8, #0x38
+; ALL-NEXT:    mvn w0, w8
+; ALL-NEXT:    lsl x14, x10, #1
+; ALL-NEXT:    and x15, x9, #0x38
+; ALL-NEXT:    mvn w16, w9
+; ALL-NEXT:    tst x8, #0x40
+; ALL-NEXT:    lsr x3, x13, x17
+; ALL-NEXT:    lsl x14, x14, x0
+; ALL-NEXT:    lsr x17, x10, x17
+; ALL-NEXT:    orr x14, x14, x3
+; ALL-NEXT:    lsr x18, x13, #1
+; ALL-NEXT:    csel x0, xzr, x17, ne
+; ALL-NEXT:    csel x14, x17, x14, ne
+; ALL-NEXT:    lsr x17, x11, #1
+; ALL-NEXT:    lsl x8, x10, x15
+; ALL-NEXT:    lsr x1, x18, x16
+; ALL-NEXT:    lsl x3, x12, x15
+; ALL-NEXT:    lsr x16, x17, x16
+; ALL-NEXT:    orr x8, x8, x1
+; ALL-NEXT:    lsl x1, x13, x15
+; ALL-NEXT:    tst x9, #0x40
+; ALL-NEXT:    orr x16, x3, x16
+; ALL-NEXT:    lsl x15, x11, x15
+; ALL-NEXT:    csel x8, x1, x8, ne
+; ALL-NEXT:    csel x16, x15, x16, ne
+; ALL-NEXT:    csel x15, xzr, x15, ne
+; ALL-NEXT:    csel x17, xzr, x1, ne
+; ALL-NEXT:    subs x1, x9, #128
+; ALL-NEXT:    and x3, x1, #0x38
+; ALL-NEXT:    mvn w4, w1
+; ALL-NEXT:    csel x17, x17, xzr, lo
+; ALL-NEXT:    tst x1, #0x40
+; ALL-NEXT:    orr x16, x16, x0
+; ALL-NEXT:    orr x14, x15, x14
+; ALL-NEXT:    lsl x10, x10, x3
+; ALL-NEXT:    lsr x18, x18, x4
+; ALL-NEXT:    orr x10, x10, x18
+; ALL-NEXT:    lsl x13, x13, x3
+; ALL-NEXT:    csel x10, x13, x10, ne
+; ALL-NEXT:    csel x13, xzr, x13, ne
+; ALL-NEXT:    cmp x9, #128
+; ALL-NEXT:    csel x10, x16, x10, lo
+; ALL-NEXT:    csel x8, x8, xzr, lo
+; ALL-NEXT:    csel x13, x14, x13, lo
+; ALL-NEXT:    cmp x9, #0
+; ALL-NEXT:    csel x9, x12, x10, eq
+; ALL-NEXT:    csel x10, x11, x13, eq
+; ALL-NEXT:    stp x17, x8, [x2]
+; ALL-NEXT:    stp x10, x9, [x2, #16]
+; ALL-NEXT:    ret
+  %src = load i256, ptr %src.ptr, align 1
+  %byteOff = load i256, ptr %byteOff.ptr, align 1
+  %bitOff = shl i256 %byteOff, 3
+  %res = shl i256 %src, %bitOff
+  store i256 %res, ptr %dst, align 1
+  ret void
+}
+define void @ashr_32bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; ALL-LABEL: ashr_32bytes:
+; ALL:       // %bb.0:
+; ALL-NEXT:    ldr x9, [x1]
+; ALL-NEXT:    mov w8, #128
+; ALL-NEXT:    ldp x11, x10, [x0, #8]
+; ALL-NEXT:    lsl x9, x9, #3
+; ALL-NEXT:    ldr x12, [x0]
+; ALL-NEXT:    sub x8, x8, x9
+; ALL-NEXT:    ldr x13, [x0, #24]
+; ALL-NEXT:    and x18, x8, #0x38
+; ALL-NEXT:    mvn w0, w8
+; ALL-NEXT:    lsr x14, x10, #1
+; ALL-NEXT:    and x15, x9, #0x38
+; ALL-NEXT:    mvn w16, w9
+; ALL-NEXT:    lsl x17, x13, #1
+; ALL-NEXT:    lsl x4, x13, x18
+; ALL-NEXT:    lsr x14, x14, x0
+; ALL-NEXT:    tst x8, #0x40
+; ALL-NEXT:    lsl x18, x10, x18
+; ALL-NEXT:    orr x14, x4, x14
+; ALL-NEXT:    lsr x8, x10, x15
+; ALL-NEXT:    lsl x1, x17, x16
+; ALL-NEXT:    csel x0, xzr, x18, ne
+; ALL-NEXT:    csel x14, x18, x14, ne
+; ALL-NEXT:    lsl x18, x11, #1
+; ALL-NEXT:    orr x8, x1, x8
+; ALL-NEXT:    lsr x1, x12, x15
+; ALL-NEXT:    lsl x16, x18, x16
+; ALL-NEXT:    asr x3, x13, x15
+; ALL-NEXT:    tst x9, #0x40
+; ALL-NEXT:    orr x16, x16, x1
+; ALL-NEXT:    lsr x15, x11, x15
+; ALL-NEXT:    asr x18, x13, #63
+; ALL-NEXT:    csel x8, x3, x8, ne
+; ALL-NEXT:    csel x16, x15, x16, ne
+; ALL-NEXT:    csel x15, xzr, x15, ne
+; ALL-NEXT:    csel x1, x18, x3, ne
+; ALL-NEXT:    subs x3, x9, #128
+; ALL-NEXT:    orr x16, x16, x0
+; ALL-NEXT:    and x4, x3, #0x38
+; ALL-NEXT:    mvn w5, w3
+; ALL-NEXT:    orr x14, x15, x14
+; ALL-NEXT:    lsr x10, x10, x4
+; ALL-NEXT:    lsl x17, x17, x5
+; ALL-NEXT:    orr x10, x17, x10
+; ALL-NEXT:    csel x17, x1, x18, lo
+; ALL-NEXT:    asr x13, x13, x4
+; ALL-NEXT:    tst x3, #0x40
+; ALL-NEXT:    csel x10, x13, x10, ne
+; ALL-NEXT:    csel x13, x18, x13, ne
+; ALL-NEXT:    cmp x9, #128
+; ALL-NEXT:    csel x10, x16, x10, lo
+; ALL-NEXT:    csel x8, x8, x18, lo
+; ALL-NEXT:    csel x13, x14, x13, lo
+; ALL-NEXT:    cmp x9, #0
+; ALL-NEXT:    csel x9, x12, x10, eq
+; ALL-NEXT:    csel x10, x11, x13, eq
+; ALL-NEXT:    stp x8, x17, [x2, #16]
+; ALL-NEXT:    stp x9, x10, [x2]
+; ALL-NEXT:    ret
+  %src = load i256, ptr %src.ptr, align 1
+  %byteOff = load i256, ptr %byteOff.ptr, align 1
+  %bitOff = shl i256 %byteOff, 3
+  %res = ashr i256 %src, %bitOff
+  store i256 %res, ptr %dst, align 1
+  ret void
+}

diff  --git a/llvm/test/CodeGen/PowerPC/wide-scalar-shift-legalization.ll b/llvm/test/CodeGen/PowerPC/wide-scalar-shift-legalization.ll
new file mode 100644
index 0000000000000..c6dfa12db0405
--- /dev/null
+++ b/llvm/test/CodeGen/PowerPC/wide-scalar-shift-legalization.ll
@@ -0,0 +1,2385 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=powerpc64le-unknown-linux-gnu | FileCheck %s --check-prefixes=ALL,LE,LE-64BIT
+; RUN: llc < %s -mtriple=powerpc64-unknown-linux-gnu   | FileCheck %s --check-prefixes=ALL,BE
+; RUN: llc < %s -mtriple=ppc32--                       | FileCheck %s --check-prefixes=ALL,LE,LE-32BIT
+
+define void @lshr_4bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; ALL-LABEL: lshr_4bytes:
+; ALL:       # %bb.0:
+; ALL-NEXT:    lwz 4, 0(4)
+; ALL-NEXT:    lwz 3, 0(3)
+; ALL-NEXT:    slwi 4, 4, 3
+; ALL-NEXT:    srw 3, 3, 4
+; ALL-NEXT:    stw 3, 0(5)
+; ALL-NEXT:    blr
+  %src = load i32, ptr %src.ptr, align 1
+  %byteOff = load i32, ptr %byteOff.ptr, align 1
+  %bitOff = shl i32 %byteOff, 3
+  %res = lshr i32 %src, %bitOff
+  store i32 %res, ptr %dst, align 1
+  ret void
+}
+define void @shl_4bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; ALL-LABEL: shl_4bytes:
+; ALL:       # %bb.0:
+; ALL-NEXT:    lwz 4, 0(4)
+; ALL-NEXT:    lwz 3, 0(3)
+; ALL-NEXT:    slwi 4, 4, 3
+; ALL-NEXT:    slw 3, 3, 4
+; ALL-NEXT:    stw 3, 0(5)
+; ALL-NEXT:    blr
+  %src = load i32, ptr %src.ptr, align 1
+  %byteOff = load i32, ptr %byteOff.ptr, align 1
+  %bitOff = shl i32 %byteOff, 3
+  %res = shl i32 %src, %bitOff
+  store i32 %res, ptr %dst, align 1
+  ret void
+}
+define void @ashr_4bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; ALL-LABEL: ashr_4bytes:
+; ALL:       # %bb.0:
+; ALL-NEXT:    lwz 4, 0(4)
+; ALL-NEXT:    lwz 3, 0(3)
+; ALL-NEXT:    slwi 4, 4, 3
+; ALL-NEXT:    sraw 3, 3, 4
+; ALL-NEXT:    stw 3, 0(5)
+; ALL-NEXT:    blr
+  %src = load i32, ptr %src.ptr, align 1
+  %byteOff = load i32, ptr %byteOff.ptr, align 1
+  %bitOff = shl i32 %byteOff, 3
+  %res = ashr i32 %src, %bitOff
+  store i32 %res, ptr %dst, align 1
+  ret void
+}
+
+define void @lshr_8bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; LE-64BIT-LABEL: lshr_8bytes:
+; LE-64BIT:       # %bb.0:
+; LE-64BIT-NEXT:    lwz 4, 0(4)
+; LE-64BIT-NEXT:    ld 3, 0(3)
+; LE-64BIT-NEXT:    slwi 4, 4, 3
+; LE-64BIT-NEXT:    srd 3, 3, 4
+; LE-64BIT-NEXT:    std 3, 0(5)
+; LE-64BIT-NEXT:    blr
+;
+; BE-LABEL: lshr_8bytes:
+; BE:       # %bb.0:
+; BE-NEXT:    lwz 4, 4(4)
+; BE-NEXT:    ld 3, 0(3)
+; BE-NEXT:    slwi 4, 4, 3
+; BE-NEXT:    srd 3, 3, 4
+; BE-NEXT:    std 3, 0(5)
+; BE-NEXT:    blr
+;
+; LE-32BIT-LABEL: lshr_8bytes:
+; LE-32BIT:       # %bb.0:
+; LE-32BIT-NEXT:    lwz 4, 4(4)
+; LE-32BIT-NEXT:    lwz 6, 4(3)
+; LE-32BIT-NEXT:    lwz 3, 0(3)
+; LE-32BIT-NEXT:    slwi 4, 4, 3
+; LE-32BIT-NEXT:    subfic 7, 4, 32
+; LE-32BIT-NEXT:    srw 6, 6, 4
+; LE-32BIT-NEXT:    addi 8, 4, -32
+; LE-32BIT-NEXT:    slw 7, 3, 7
+; LE-32BIT-NEXT:    srw 4, 3, 4
+; LE-32BIT-NEXT:    srw 3, 3, 8
+; LE-32BIT-NEXT:    or 6, 6, 7
+; LE-32BIT-NEXT:    or 3, 6, 3
+; LE-32BIT-NEXT:    stw 4, 0(5)
+; LE-32BIT-NEXT:    stw 3, 4(5)
+; LE-32BIT-NEXT:    blr
+  %src = load i64, ptr %src.ptr, align 1
+  %byteOff = load i64, ptr %byteOff.ptr, align 1
+  %bitOff = shl i64 %byteOff, 3
+  %res = lshr i64 %src, %bitOff
+  store i64 %res, ptr %dst, align 1
+  ret void
+}
+define void @shl_8bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; LE-64BIT-LABEL: shl_8bytes:
+; LE-64BIT:       # %bb.0:
+; LE-64BIT-NEXT:    lwz 4, 0(4)
+; LE-64BIT-NEXT:    ld 3, 0(3)
+; LE-64BIT-NEXT:    slwi 4, 4, 3
+; LE-64BIT-NEXT:    sld 3, 3, 4
+; LE-64BIT-NEXT:    std 3, 0(5)
+; LE-64BIT-NEXT:    blr
+;
+; BE-LABEL: shl_8bytes:
+; BE:       # %bb.0:
+; BE-NEXT:    lwz 4, 4(4)
+; BE-NEXT:    ld 3, 0(3)
+; BE-NEXT:    slwi 4, 4, 3
+; BE-NEXT:    sld 3, 3, 4
+; BE-NEXT:    std 3, 0(5)
+; BE-NEXT:    blr
+;
+; LE-32BIT-LABEL: shl_8bytes:
+; LE-32BIT:       # %bb.0:
+; LE-32BIT-NEXT:    lwz 4, 4(4)
+; LE-32BIT-NEXT:    lwz 6, 0(3)
+; LE-32BIT-NEXT:    lwz 3, 4(3)
+; LE-32BIT-NEXT:    slwi 4, 4, 3
+; LE-32BIT-NEXT:    subfic 7, 4, 32
+; LE-32BIT-NEXT:    slw 6, 6, 4
+; LE-32BIT-NEXT:    addi 8, 4, -32
+; LE-32BIT-NEXT:    srw 7, 3, 7
+; LE-32BIT-NEXT:    slw 4, 3, 4
+; LE-32BIT-NEXT:    slw 3, 3, 8
+; LE-32BIT-NEXT:    or 6, 6, 7
+; LE-32BIT-NEXT:    or 3, 6, 3
+; LE-32BIT-NEXT:    stw 4, 4(5)
+; LE-32BIT-NEXT:    stw 3, 0(5)
+; LE-32BIT-NEXT:    blr
+  %src = load i64, ptr %src.ptr, align 1
+  %byteOff = load i64, ptr %byteOff.ptr, align 1
+  %bitOff = shl i64 %byteOff, 3
+  %res = shl i64 %src, %bitOff
+  store i64 %res, ptr %dst, align 1
+  ret void
+}
+define void @ashr_8bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; LE-64BIT-LABEL: ashr_8bytes:
+; LE-64BIT:       # %bb.0:
+; LE-64BIT-NEXT:    lwz 4, 0(4)
+; LE-64BIT-NEXT:    ld 3, 0(3)
+; LE-64BIT-NEXT:    slwi 4, 4, 3
+; LE-64BIT-NEXT:    srad 3, 3, 4
+; LE-64BIT-NEXT:    std 3, 0(5)
+; LE-64BIT-NEXT:    blr
+;
+; BE-LABEL: ashr_8bytes:
+; BE:       # %bb.0:
+; BE-NEXT:    lwz 4, 4(4)
+; BE-NEXT:    ld 3, 0(3)
+; BE-NEXT:    slwi 4, 4, 3
+; BE-NEXT:    srad 3, 3, 4
+; BE-NEXT:    std 3, 0(5)
+; BE-NEXT:    blr
+;
+; LE-32BIT-LABEL: ashr_8bytes:
+; LE-32BIT:       # %bb.0:
+; LE-32BIT-NEXT:    lwz 4, 4(4)
+; LE-32BIT-NEXT:    lwz 6, 4(3)
+; LE-32BIT-NEXT:    lwz 3, 0(3)
+; LE-32BIT-NEXT:    slwi 4, 4, 3
+; LE-32BIT-NEXT:    subfic 7, 4, 32
+; LE-32BIT-NEXT:    srw 6, 6, 4
+; LE-32BIT-NEXT:    addi 8, 4, -32
+; LE-32BIT-NEXT:    slw 7, 3, 7
+; LE-32BIT-NEXT:    sraw 4, 3, 4
+; LE-32BIT-NEXT:    sraw 3, 3, 8
+; LE-32BIT-NEXT:    cmpwi 8, 1
+; LE-32BIT-NEXT:    or 6, 6, 7
+; LE-32BIT-NEXT:    bc 12, 0, .LBB5_1
+; LE-32BIT-NEXT:    b .LBB5_2
+; LE-32BIT-NEXT:  .LBB5_1:
+; LE-32BIT-NEXT:    addi 3, 6, 0
+; LE-32BIT-NEXT:  .LBB5_2:
+; LE-32BIT-NEXT:    stw 4, 0(5)
+; LE-32BIT-NEXT:    stw 3, 4(5)
+; LE-32BIT-NEXT:    blr
+  %src = load i64, ptr %src.ptr, align 1
+  %byteOff = load i64, ptr %byteOff.ptr, align 1
+  %bitOff = shl i64 %byteOff, 3
+  %res = ashr i64 %src, %bitOff
+  store i64 %res, ptr %dst, align 1
+  ret void
+}
+
+define void @lshr_16bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; LE-64BIT-LABEL: lshr_16bytes:
+; LE-64BIT:       # %bb.0:
+; LE-64BIT-NEXT:    lwz 4, 0(4)
+; LE-64BIT-NEXT:    ld 6, 0(3)
+; LE-64BIT-NEXT:    ld 3, 8(3)
+; LE-64BIT-NEXT:    slwi 4, 4, 3
+; LE-64BIT-NEXT:    subfic 7, 4, 64
+; LE-64BIT-NEXT:    srd 6, 6, 4
+; LE-64BIT-NEXT:    addi 8, 4, -64
+; LE-64BIT-NEXT:    sld 7, 3, 7
+; LE-64BIT-NEXT:    or 6, 6, 7
+; LE-64BIT-NEXT:    srd 7, 3, 8
+; LE-64BIT-NEXT:    or 6, 6, 7
+; LE-64BIT-NEXT:    srd 3, 3, 4
+; LE-64BIT-NEXT:    std 3, 8(5)
+; LE-64BIT-NEXT:    std 6, 0(5)
+; LE-64BIT-NEXT:    blr
+;
+; BE-LABEL: lshr_16bytes:
+; BE:       # %bb.0:
+; BE-NEXT:    lwz 4, 12(4)
+; BE-NEXT:    ld 6, 0(3)
+; BE-NEXT:    ld 3, 8(3)
+; BE-NEXT:    slwi 4, 4, 3
+; BE-NEXT:    subfic 7, 4, 64
+; BE-NEXT:    srd 3, 3, 4
+; BE-NEXT:    sld 7, 6, 7
+; BE-NEXT:    addi 8, 4, -64
+; BE-NEXT:    or 3, 3, 7
+; BE-NEXT:    srd 7, 6, 8
+; BE-NEXT:    srd 4, 6, 4
+; BE-NEXT:    or 3, 3, 7
+; BE-NEXT:    std 4, 0(5)
+; BE-NEXT:    std 3, 8(5)
+; BE-NEXT:    blr
+;
+; LE-32BIT-LABEL: lshr_16bytes:
+; LE-32BIT:       # %bb.0:
+; LE-32BIT-NEXT:    stwu 1, -32(1)
+; LE-32BIT-NEXT:    lwz 4, 12(4)
+; LE-32BIT-NEXT:    li 8, 0
+; LE-32BIT-NEXT:    lwz 6, 8(3)
+; LE-32BIT-NEXT:    lwz 7, 12(3)
+; LE-32BIT-NEXT:    rlwinm. 4, 4, 3, 0, 28
+; LE-32BIT-NEXT:    lwz 9, 4(3)
+; LE-32BIT-NEXT:    subfic 10, 4, 96
+; LE-32BIT-NEXT:    lwz 3, 0(3)
+; LE-32BIT-NEXT:    addi 11, 4, -64
+; LE-32BIT-NEXT:    stw 27, 12(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    cmplwi 1, 4, 64
+; LE-32BIT-NEXT:    stw 28, 16(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    srw 28, 3, 4
+; LE-32BIT-NEXT:    stw 30, 24(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    subfic 30, 4, 32
+; LE-32BIT-NEXT:    slw 10, 3, 10
+; LE-32BIT-NEXT:    srw 27, 9, 11
+; LE-32BIT-NEXT:    addi 12, 4, -96
+; LE-32BIT-NEXT:    srw 0, 7, 4
+; LE-32BIT-NEXT:    or 10, 27, 10
+; LE-32BIT-NEXT:    slw 27, 6, 30
+; LE-32BIT-NEXT:    bc 12, 4, .LBB6_2
+; LE-32BIT-NEXT:  # %bb.1:
+; LE-32BIT-NEXT:    ori 28, 8, 0
+; LE-32BIT-NEXT:    b .LBB6_2
+; LE-32BIT-NEXT:  .LBB6_2:
+; LE-32BIT-NEXT:    stw 29, 20(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    srw 29, 9, 4
+; LE-32BIT-NEXT:    or 0, 0, 27
+; LE-32BIT-NEXT:    slw 27, 3, 30
+; LE-32BIT-NEXT:    stw 28, 0(5)
+; LE-32BIT-NEXT:    subfic 28, 4, 64
+; LE-32BIT-NEXT:    srw 12, 3, 12
+; LE-32BIT-NEXT:    or 29, 29, 27
+; LE-32BIT-NEXT:    addi 27, 4, -32
+; LE-32BIT-NEXT:    or 10, 10, 12
+; LE-32BIT-NEXT:    subfic 12, 28, 32
+; LE-32BIT-NEXT:    slw 30, 9, 30
+; LE-32BIT-NEXT:    srw 12, 9, 12
+; LE-32BIT-NEXT:    slw 9, 9, 28
+; LE-32BIT-NEXT:    slw 28, 3, 28
+; LE-32BIT-NEXT:    srw 11, 3, 11
+; LE-32BIT-NEXT:    srw 3, 3, 27
+; LE-32BIT-NEXT:    srw 27, 6, 27
+; LE-32BIT-NEXT:    or 0, 0, 27
+; LE-32BIT-NEXT:    or 12, 28, 12
+; LE-32BIT-NEXT:    srw 4, 6, 4
+; LE-32BIT-NEXT:    or 3, 29, 3
+; LE-32BIT-NEXT:    or 9, 0, 9
+; LE-32BIT-NEXT:    or 12, 12, 30
+; LE-32BIT-NEXT:    bc 12, 4, .LBB6_4
+; LE-32BIT-NEXT:  # %bb.3:
+; LE-32BIT-NEXT:    ori 3, 8, 0
+; LE-32BIT-NEXT:    ori 8, 10, 0
+; LE-32BIT-NEXT:    b .LBB6_5
+; LE-32BIT-NEXT:  .LBB6_4:
+; LE-32BIT-NEXT:    addi 8, 9, 0
+; LE-32BIT-NEXT:  .LBB6_5:
+; LE-32BIT-NEXT:    or 4, 4, 12
+; LE-32BIT-NEXT:    stw 3, 4(5)
+; LE-32BIT-NEXT:    bc 12, 2, .LBB6_7
+; LE-32BIT-NEXT:  # %bb.6:
+; LE-32BIT-NEXT:    ori 3, 8, 0
+; LE-32BIT-NEXT:    b .LBB6_8
+; LE-32BIT-NEXT:  .LBB6_7:
+; LE-32BIT-NEXT:    addi 3, 7, 0
+; LE-32BIT-NEXT:  .LBB6_8:
+; LE-32BIT-NEXT:    bc 12, 4, .LBB6_10
+; LE-32BIT-NEXT:  # %bb.9:
+; LE-32BIT-NEXT:    ori 4, 11, 0
+; LE-32BIT-NEXT:    b .LBB6_10
+; LE-32BIT-NEXT:  .LBB6_10:
+; LE-32BIT-NEXT:    stw 3, 12(5)
+; LE-32BIT-NEXT:    bc 12, 2, .LBB6_12
+; LE-32BIT-NEXT:  # %bb.11:
+; LE-32BIT-NEXT:    ori 3, 4, 0
+; LE-32BIT-NEXT:    b .LBB6_13
+; LE-32BIT-NEXT:  .LBB6_12:
+; LE-32BIT-NEXT:    addi 3, 6, 0
+; LE-32BIT-NEXT:  .LBB6_13:
+; LE-32BIT-NEXT:    stw 3, 8(5)
+; LE-32BIT-NEXT:    lwz 30, 24(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 29, 20(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 28, 16(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 27, 12(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    addi 1, 1, 32
+; LE-32BIT-NEXT:    blr
+  %src = load i128, ptr %src.ptr, align 1
+  %byteOff = load i128, ptr %byteOff.ptr, align 1
+  %bitOff = shl i128 %byteOff, 3
+  %res = lshr i128 %src, %bitOff
+  store i128 %res, ptr %dst, align 1
+  ret void
+}
+define void @shl_16bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; LE-64BIT-LABEL: shl_16bytes:
+; LE-64BIT:       # %bb.0:
+; LE-64BIT-NEXT:    lwz 4, 0(4)
+; LE-64BIT-NEXT:    ld 6, 8(3)
+; LE-64BIT-NEXT:    ld 3, 0(3)
+; LE-64BIT-NEXT:    slwi 4, 4, 3
+; LE-64BIT-NEXT:    subfic 7, 4, 64
+; LE-64BIT-NEXT:    sld 6, 6, 4
+; LE-64BIT-NEXT:    addi 8, 4, -64
+; LE-64BIT-NEXT:    srd 7, 3, 7
+; LE-64BIT-NEXT:    or 6, 6, 7
+; LE-64BIT-NEXT:    sld 7, 3, 8
+; LE-64BIT-NEXT:    or 6, 6, 7
+; LE-64BIT-NEXT:    sld 3, 3, 4
+; LE-64BIT-NEXT:    std 3, 0(5)
+; LE-64BIT-NEXT:    std 6, 8(5)
+; LE-64BIT-NEXT:    blr
+;
+; BE-LABEL: shl_16bytes:
+; BE:       # %bb.0:
+; BE-NEXT:    lwz 4, 12(4)
+; BE-NEXT:    ld 6, 8(3)
+; BE-NEXT:    ld 3, 0(3)
+; BE-NEXT:    slwi 4, 4, 3
+; BE-NEXT:    subfic 7, 4, 64
+; BE-NEXT:    sld 3, 3, 4
+; BE-NEXT:    srd 7, 6, 7
+; BE-NEXT:    addi 8, 4, -64
+; BE-NEXT:    or 3, 3, 7
+; BE-NEXT:    sld 7, 6, 8
+; BE-NEXT:    sld 4, 6, 4
+; BE-NEXT:    or 3, 3, 7
+; BE-NEXT:    std 4, 8(5)
+; BE-NEXT:    std 3, 0(5)
+; BE-NEXT:    blr
+;
+; LE-32BIT-LABEL: shl_16bytes:
+; LE-32BIT:       # %bb.0:
+; LE-32BIT-NEXT:    stwu 1, -32(1)
+; LE-32BIT-NEXT:    lwz 4, 12(4)
+; LE-32BIT-NEXT:    li 8, 0
+; LE-32BIT-NEXT:    lwz 6, 4(3)
+; LE-32BIT-NEXT:    lwz 7, 0(3)
+; LE-32BIT-NEXT:    rlwinm. 4, 4, 3, 0, 28
+; LE-32BIT-NEXT:    lwz 9, 8(3)
+; LE-32BIT-NEXT:    subfic 10, 4, 96
+; LE-32BIT-NEXT:    lwz 3, 12(3)
+; LE-32BIT-NEXT:    addi 11, 4, -64
+; LE-32BIT-NEXT:    stw 27, 12(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    cmplwi 1, 4, 64
+; LE-32BIT-NEXT:    stw 28, 16(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    slw 28, 3, 4
+; LE-32BIT-NEXT:    stw 30, 24(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    subfic 30, 4, 32
+; LE-32BIT-NEXT:    srw 10, 3, 10
+; LE-32BIT-NEXT:    slw 27, 9, 11
+; LE-32BIT-NEXT:    addi 12, 4, -96
+; LE-32BIT-NEXT:    slw 0, 7, 4
+; LE-32BIT-NEXT:    or 10, 27, 10
+; LE-32BIT-NEXT:    srw 27, 6, 30
+; LE-32BIT-NEXT:    bc 12, 4, .LBB7_2
+; LE-32BIT-NEXT:  # %bb.1:
+; LE-32BIT-NEXT:    ori 28, 8, 0
+; LE-32BIT-NEXT:    b .LBB7_2
+; LE-32BIT-NEXT:  .LBB7_2:
+; LE-32BIT-NEXT:    stw 29, 20(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    slw 29, 9, 4
+; LE-32BIT-NEXT:    or 0, 0, 27
+; LE-32BIT-NEXT:    srw 27, 3, 30
+; LE-32BIT-NEXT:    stw 28, 12(5)
+; LE-32BIT-NEXT:    subfic 28, 4, 64
+; LE-32BIT-NEXT:    slw 12, 3, 12
+; LE-32BIT-NEXT:    or 29, 29, 27
+; LE-32BIT-NEXT:    addi 27, 4, -32
+; LE-32BIT-NEXT:    or 10, 10, 12
+; LE-32BIT-NEXT:    subfic 12, 28, 32
+; LE-32BIT-NEXT:    srw 30, 9, 30
+; LE-32BIT-NEXT:    slw 12, 9, 12
+; LE-32BIT-NEXT:    srw 9, 9, 28
+; LE-32BIT-NEXT:    srw 28, 3, 28
+; LE-32BIT-NEXT:    slw 11, 3, 11
+; LE-32BIT-NEXT:    slw 3, 3, 27
+; LE-32BIT-NEXT:    slw 27, 6, 27
+; LE-32BIT-NEXT:    or 0, 0, 27
+; LE-32BIT-NEXT:    or 12, 28, 12
+; LE-32BIT-NEXT:    slw 4, 6, 4
+; LE-32BIT-NEXT:    or 3, 29, 3
+; LE-32BIT-NEXT:    or 9, 0, 9
+; LE-32BIT-NEXT:    or 12, 12, 30
+; LE-32BIT-NEXT:    bc 12, 4, .LBB7_4
+; LE-32BIT-NEXT:  # %bb.3:
+; LE-32BIT-NEXT:    ori 3, 8, 0
+; LE-32BIT-NEXT:    ori 8, 10, 0
+; LE-32BIT-NEXT:    b .LBB7_5
+; LE-32BIT-NEXT:  .LBB7_4:
+; LE-32BIT-NEXT:    addi 8, 9, 0
+; LE-32BIT-NEXT:  .LBB7_5:
+; LE-32BIT-NEXT:    or 4, 4, 12
+; LE-32BIT-NEXT:    stw 3, 8(5)
+; LE-32BIT-NEXT:    bc 12, 2, .LBB7_7
+; LE-32BIT-NEXT:  # %bb.6:
+; LE-32BIT-NEXT:    ori 3, 8, 0
+; LE-32BIT-NEXT:    b .LBB7_8
+; LE-32BIT-NEXT:  .LBB7_7:
+; LE-32BIT-NEXT:    addi 3, 7, 0
+; LE-32BIT-NEXT:  .LBB7_8:
+; LE-32BIT-NEXT:    bc 12, 4, .LBB7_10
+; LE-32BIT-NEXT:  # %bb.9:
+; LE-32BIT-NEXT:    ori 4, 11, 0
+; LE-32BIT-NEXT:    b .LBB7_10
+; LE-32BIT-NEXT:  .LBB7_10:
+; LE-32BIT-NEXT:    stw 3, 0(5)
+; LE-32BIT-NEXT:    bc 12, 2, .LBB7_12
+; LE-32BIT-NEXT:  # %bb.11:
+; LE-32BIT-NEXT:    ori 3, 4, 0
+; LE-32BIT-NEXT:    b .LBB7_13
+; LE-32BIT-NEXT:  .LBB7_12:
+; LE-32BIT-NEXT:    addi 3, 6, 0
+; LE-32BIT-NEXT:  .LBB7_13:
+; LE-32BIT-NEXT:    stw 3, 4(5)
+; LE-32BIT-NEXT:    lwz 30, 24(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 29, 20(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 28, 16(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 27, 12(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    addi 1, 1, 32
+; LE-32BIT-NEXT:    blr
+  %src = load i128, ptr %src.ptr, align 1
+  %byteOff = load i128, ptr %byteOff.ptr, align 1
+  %bitOff = shl i128 %byteOff, 3
+  %res = shl i128 %src, %bitOff
+  store i128 %res, ptr %dst, align 1
+  ret void
+}
+define void @ashr_16bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; LE-64BIT-LABEL: ashr_16bytes:
+; LE-64BIT:       # %bb.0:
+; LE-64BIT-NEXT:    lwz 4, 0(4)
+; LE-64BIT-NEXT:    ld 6, 0(3)
+; LE-64BIT-NEXT:    ld 3, 8(3)
+; LE-64BIT-NEXT:    slwi 4, 4, 3
+; LE-64BIT-NEXT:    subfic 7, 4, 64
+; LE-64BIT-NEXT:    srd 6, 6, 4
+; LE-64BIT-NEXT:    addi 8, 4, -64
+; LE-64BIT-NEXT:    sld 7, 3, 7
+; LE-64BIT-NEXT:    cmpwi 8, 1
+; LE-64BIT-NEXT:    or 6, 6, 7
+; LE-64BIT-NEXT:    srad 7, 3, 8
+; LE-64BIT-NEXT:    isellt 6, 6, 7
+; LE-64BIT-NEXT:    srad 3, 3, 4
+; LE-64BIT-NEXT:    std 3, 8(5)
+; LE-64BIT-NEXT:    std 6, 0(5)
+; LE-64BIT-NEXT:    blr
+;
+; BE-LABEL: ashr_16bytes:
+; BE:       # %bb.0:
+; BE-NEXT:    lwz 4, 12(4)
+; BE-NEXT:    ld 6, 8(3)
+; BE-NEXT:    ld 3, 0(3)
+; BE-NEXT:    slwi 4, 4, 3
+; BE-NEXT:    subfic 7, 4, 64
+; BE-NEXT:    srd 6, 6, 4
+; BE-NEXT:    addi 8, 4, -64
+; BE-NEXT:    sld 7, 3, 7
+; BE-NEXT:    cmpwi 8, 1
+; BE-NEXT:    or 6, 6, 7
+; BE-NEXT:    srad 7, 3, 8
+; BE-NEXT:    srad 3, 3, 4
+; BE-NEXT:    bc 12, 0, .LBB8_2
+; BE-NEXT:  # %bb.1:
+; BE-NEXT:    ori 6, 7, 0
+; BE-NEXT:    b .LBB8_2
+; BE-NEXT:  .LBB8_2:
+; BE-NEXT:    std 3, 0(5)
+; BE-NEXT:    std 6, 8(5)
+; BE-NEXT:    blr
+;
+; LE-32BIT-LABEL: ashr_16bytes:
+; LE-32BIT:       # %bb.0:
+; LE-32BIT-NEXT:    stwu 1, -32(1)
+; LE-32BIT-NEXT:    lwz 4, 12(4)
+; LE-32BIT-NEXT:    lwz 8, 0(3)
+; LE-32BIT-NEXT:    lwz 9, 4(3)
+; LE-32BIT-NEXT:    lwz 6, 8(3)
+; LE-32BIT-NEXT:    lwz 7, 12(3)
+; LE-32BIT-NEXT:    rlwinm. 3, 4, 3, 0, 28
+; LE-32BIT-NEXT:    subfic 10, 3, 96
+; LE-32BIT-NEXT:    addi 11, 3, -64
+; LE-32BIT-NEXT:    stw 27, 12(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    addi 12, 3, -96
+; LE-32BIT-NEXT:    stw 30, 24(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    subfic 30, 3, 32
+; LE-32BIT-NEXT:    slw 10, 8, 10
+; LE-32BIT-NEXT:    srw 27, 9, 11
+; LE-32BIT-NEXT:    stw 26, 8(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    sraw 26, 8, 12
+; LE-32BIT-NEXT:    stw 28, 16(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    srw 28, 9, 3
+; LE-32BIT-NEXT:    stw 29, 20(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    addi 29, 3, -32
+; LE-32BIT-NEXT:    cmpwi 1, 12, 1
+; LE-32BIT-NEXT:    slw 12, 8, 30
+; LE-32BIT-NEXT:    or 10, 27, 10
+; LE-32BIT-NEXT:    srw 0, 7, 3
+; LE-32BIT-NEXT:    sraw 27, 8, 29
+; LE-32BIT-NEXT:    bc 12, 4, .LBB8_2
+; LE-32BIT-NEXT:  # %bb.1:
+; LE-32BIT-NEXT:    ori 10, 26, 0
+; LE-32BIT-NEXT:    b .LBB8_2
+; LE-32BIT-NEXT:  .LBB8_2:
+; LE-32BIT-NEXT:    cmpwi 1, 29, 1
+; LE-32BIT-NEXT:    or 12, 28, 12
+; LE-32BIT-NEXT:    subfic 28, 3, 64
+; LE-32BIT-NEXT:    slw 26, 6, 30
+; LE-32BIT-NEXT:    srawi 4, 8, 31
+; LE-32BIT-NEXT:    bc 12, 4, .LBB8_4
+; LE-32BIT-NEXT:  # %bb.3:
+; LE-32BIT-NEXT:    ori 12, 27, 0
+; LE-32BIT-NEXT:    b .LBB8_4
+; LE-32BIT-NEXT:  .LBB8_4:
+; LE-32BIT-NEXT:    sraw 27, 8, 3
+; LE-32BIT-NEXT:    or 0, 0, 26
+; LE-32BIT-NEXT:    slw 26, 9, 28
+; LE-32BIT-NEXT:    sraw 11, 8, 11
+; LE-32BIT-NEXT:    slw 8, 8, 28
+; LE-32BIT-NEXT:    subfic 28, 28, 32
+; LE-32BIT-NEXT:    slw 30, 9, 30
+; LE-32BIT-NEXT:    srw 9, 9, 28
+; LE-32BIT-NEXT:    srw 29, 6, 29
+; LE-32BIT-NEXT:    or 8, 8, 9
+; LE-32BIT-NEXT:    cmplwi 1, 3, 64
+; LE-32BIT-NEXT:    or 0, 0, 29
+; LE-32BIT-NEXT:    srw 3, 6, 3
+; LE-32BIT-NEXT:    or 8, 8, 30
+; LE-32BIT-NEXT:    or 9, 0, 26
+; LE-32BIT-NEXT:    or 3, 3, 8
+; LE-32BIT-NEXT:    bc 12, 4, .LBB8_6
+; LE-32BIT-NEXT:  # %bb.5:
+; LE-32BIT-NEXT:    ori 28, 4, 0
+; LE-32BIT-NEXT:    ori 9, 10, 0
+; LE-32BIT-NEXT:    ori 3, 11, 0
+; LE-32BIT-NEXT:    b .LBB8_7
+; LE-32BIT-NEXT:  .LBB8_6:
+; LE-32BIT-NEXT:    addi 28, 27, 0
+; LE-32BIT-NEXT:    addi 4, 12, 0
+; LE-32BIT-NEXT:  .LBB8_7:
+; LE-32BIT-NEXT:    lwz 30, 24(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    bc 12, 2, .LBB8_8
+; LE-32BIT-NEXT:    b .LBB8_9
+; LE-32BIT-NEXT:  .LBB8_8:
+; LE-32BIT-NEXT:    addi 3, 6, 0
+; LE-32BIT-NEXT:  .LBB8_9:
+; LE-32BIT-NEXT:    stw 4, 4(5)
+; LE-32BIT-NEXT:    bc 12, 2, .LBB8_11
+; LE-32BIT-NEXT:  # %bb.10:
+; LE-32BIT-NEXT:    ori 4, 9, 0
+; LE-32BIT-NEXT:    b .LBB8_12
+; LE-32BIT-NEXT:  .LBB8_11:
+; LE-32BIT-NEXT:    addi 4, 7, 0
+; LE-32BIT-NEXT:  .LBB8_12:
+; LE-32BIT-NEXT:    stw 28, 0(5)
+; LE-32BIT-NEXT:    stw 4, 12(5)
+; LE-32BIT-NEXT:    stw 3, 8(5)
+; LE-32BIT-NEXT:    lwz 29, 20(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 28, 16(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 27, 12(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 26, 8(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    addi 1, 1, 32
+; LE-32BIT-NEXT:    blr
+  %src = load i128, ptr %src.ptr, align 1
+  %byteOff = load i128, ptr %byteOff.ptr, align 1
+  %bitOff = shl i128 %byteOff, 3
+  %res = ashr i128 %src, %bitOff
+  store i128 %res, ptr %dst, align 1
+  ret void
+}
+
+define void @lshr_32bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; LE-64BIT-LABEL: lshr_32bytes:
+; LE-64BIT:       # %bb.0:
+; LE-64BIT-NEXT:    lwz 4, 0(4)
+; LE-64BIT-NEXT:    ld 7, 0(3)
+; LE-64BIT-NEXT:    ld 8, 8(3)
+; LE-64BIT-NEXT:    ld 9, 16(3)
+; LE-64BIT-NEXT:    li 6, 0
+; LE-64BIT-NEXT:    ld 3, 24(3)
+; LE-64BIT-NEXT:    std 28, -32(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    std 21, -88(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    std 24, -64(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    rlwinm. 4, 4, 3, 0, 28
+; LE-64BIT-NEXT:    std 25, -56(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    subfic 28, 4, 64
+; LE-64BIT-NEXT:    subfic 11, 4, 192
+; LE-64BIT-NEXT:    std 27, -40(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    std 29, -24(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    addi 0, 4, -128
+; LE-64BIT-NEXT:    srd 29, 9, 4
+; LE-64BIT-NEXT:    addi 27, 4, -64
+; LE-64BIT-NEXT:    subfic 25, 4, 128
+; LE-64BIT-NEXT:    sld 24, 8, 28
+; LE-64BIT-NEXT:    std 22, -80(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    std 26, -48(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    sld 21, 9, 28
+; LE-64BIT-NEXT:    sld 28, 3, 28
+; LE-64BIT-NEXT:    std 30, -16(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    srd 10, 7, 4
+; LE-64BIT-NEXT:    addi 30, 4, -192
+; LE-64BIT-NEXT:    subfic 22, 25, 64
+; LE-64BIT-NEXT:    sld 11, 3, 11
+; LE-64BIT-NEXT:    srd 26, 9, 0
+; LE-64BIT-NEXT:    or 29, 29, 28
+; LE-64BIT-NEXT:    std 23, -72(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    or 10, 10, 24
+; LE-64BIT-NEXT:    srd 28, 3, 27
+; LE-64BIT-NEXT:    srd 30, 3, 30
+; LE-64BIT-NEXT:    or 11, 26, 11
+; LE-64BIT-NEXT:    ld 26, -48(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    srd 23, 8, 27
+; LE-64BIT-NEXT:    srd 27, 9, 22
+; LE-64BIT-NEXT:    or 29, 29, 28
+; LE-64BIT-NEXT:    or 11, 11, 30
+; LE-64BIT-NEXT:    ld 24, -64(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    sld 28, 3, 25
+; LE-64BIT-NEXT:    or 10, 10, 23
+; LE-64BIT-NEXT:    ld 23, -72(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    ld 22, -80(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    sld 9, 9, 25
+; LE-64BIT-NEXT:    or 30, 28, 27
+; LE-64BIT-NEXT:    ld 28, -32(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    ld 27, -40(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    ld 25, -56(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    cmplwi 1, 4, 128
+; LE-64BIT-NEXT:    srd 12, 8, 4
+; LE-64BIT-NEXT:    or 9, 10, 9
+; LE-64BIT-NEXT:    or 30, 30, 21
+; LE-64BIT-NEXT:    ld 21, -88(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    srd 10, 3, 0
+; LE-64BIT-NEXT:    isel 9, 9, 11, 4
+; LE-64BIT-NEXT:    or 11, 12, 30
+; LE-64BIT-NEXT:    ld 30, -16(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    iseleq 7, 7, 9
+; LE-64BIT-NEXT:    srd 3, 3, 4
+; LE-64BIT-NEXT:    isel 9, 11, 10, 4
+; LE-64BIT-NEXT:    std 7, 0(5)
+; LE-64BIT-NEXT:    isel 0, 29, 6, 4
+; LE-64BIT-NEXT:    ld 29, -24(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    iseleq 4, 8, 9
+; LE-64BIT-NEXT:    std 0, 16(5)
+; LE-64BIT-NEXT:    isel 3, 3, 6, 4
+; LE-64BIT-NEXT:    std 4, 8(5)
+; LE-64BIT-NEXT:    std 3, 24(5)
+; LE-64BIT-NEXT:    blr
+;
+; BE-LABEL: lshr_32bytes:
+; BE:       # %bb.0:
+; BE-NEXT:    lwz 4, 28(4)
+; BE-NEXT:    ld 7, 16(3)
+; BE-NEXT:    ld 8, 24(3)
+; BE-NEXT:    ld 9, 8(3)
+; BE-NEXT:    ld 3, 0(3)
+; BE-NEXT:    std 27, -40(1) # 8-byte Folded Spill
+; BE-NEXT:    std 30, -16(1) # 8-byte Folded Spill
+; BE-NEXT:    std 28, -32(1) # 8-byte Folded Spill
+; BE-NEXT:    std 29, -24(1) # 8-byte Folded Spill
+; BE-NEXT:    li 6, 0
+; BE-NEXT:    rlwinm. 4, 4, 3, 0, 28
+; BE-NEXT:    subfic 10, 4, 192
+; BE-NEXT:    addi 11, 4, -128
+; BE-NEXT:    addi 12, 4, -192
+; BE-NEXT:    subfic 30, 4, 64
+; BE-NEXT:    sld 10, 3, 10
+; BE-NEXT:    srd 27, 9, 11
+; BE-NEXT:    srd 0, 8, 4
+; BE-NEXT:    addi 29, 4, -64
+; BE-NEXT:    subfic 28, 4, 128
+; BE-NEXT:    srd 12, 3, 12
+; BE-NEXT:    or 10, 27, 10
+; BE-NEXT:    sld 27, 7, 30
+; BE-NEXT:    or 10, 10, 12
+; BE-NEXT:    or 0, 0, 27
+; BE-NEXT:    srd 27, 7, 29
+; BE-NEXT:    subfic 12, 28, 64
+; BE-NEXT:    or 0, 0, 27
+; BE-NEXT:    sld 27, 3, 28
+; BE-NEXT:    srd 12, 9, 12
+; BE-NEXT:    sld 28, 9, 28
+; BE-NEXT:    cmplwi 1, 4, 128
+; BE-NEXT:    or 12, 27, 12
+; BE-NEXT:    or 28, 0, 28
+; BE-NEXT:    sld 0, 9, 30
+; BE-NEXT:    srd 9, 9, 4
+; BE-NEXT:    srd 11, 3, 11
+; BE-NEXT:    bc 12, 4, .LBB9_1
+; BE-NEXT:    b .LBB9_2
+; BE-NEXT:  .LBB9_1:
+; BE-NEXT:    addi 10, 28, 0
+; BE-NEXT:  .LBB9_2:
+; BE-NEXT:    ld 28, -32(1) # 8-byte Folded Reload
+; BE-NEXT:    ld 27, -40(1) # 8-byte Folded Reload
+; BE-NEXT:    or 12, 12, 0
+; BE-NEXT:    srd 0, 7, 4
+; BE-NEXT:    or 12, 0, 12
+; BE-NEXT:    sld 0, 3, 30
+; BE-NEXT:    srd 30, 3, 29
+; BE-NEXT:    bc 12, 4, .LBB9_3
+; BE-NEXT:    b .LBB9_4
+; BE-NEXT:  .LBB9_3:
+; BE-NEXT:    addi 11, 12, 0
+; BE-NEXT:  .LBB9_4:
+; BE-NEXT:    srd 3, 3, 4
+; BE-NEXT:    bc 12, 2, .LBB9_6
+; BE-NEXT:  # %bb.5:
+; BE-NEXT:    ori 4, 10, 0
+; BE-NEXT:    b .LBB9_7
+; BE-NEXT:  .LBB9_6:
+; BE-NEXT:    addi 4, 8, 0
+; BE-NEXT:  .LBB9_7:
+; BE-NEXT:    ld 29, -24(1) # 8-byte Folded Reload
+; BE-NEXT:    or 9, 9, 0
+; BE-NEXT:    or 9, 9, 30
+; BE-NEXT:    bc 12, 2, .LBB9_9
+; BE-NEXT:  # %bb.8:
+; BE-NEXT:    ori 7, 11, 0
+; BE-NEXT:    b .LBB9_9
+; BE-NEXT:  .LBB9_9:
+; BE-NEXT:    bc 12, 4, .LBB9_11
+; BE-NEXT:  # %bb.10:
+; BE-NEXT:    ori 8, 6, 0
+; BE-NEXT:    ori 3, 6, 0
+; BE-NEXT:    b .LBB9_12
+; BE-NEXT:  .LBB9_11:
+; BE-NEXT:    addi 8, 9, 0
+; BE-NEXT:  .LBB9_12:
+; BE-NEXT:    std 4, 24(5)
+; BE-NEXT:    ld 30, -16(1) # 8-byte Folded Reload
+; BE-NEXT:    std 3, 0(5)
+; BE-NEXT:    std 8, 8(5)
+; BE-NEXT:    std 7, 16(5)
+; BE-NEXT:    blr
+;
+; LE-32BIT-LABEL: lshr_32bytes:
+; LE-32BIT:       # %bb.0:
+; LE-32BIT-NEXT:    stwu 1, -144(1)
+; LE-32BIT-NEXT:    mfcr 12
+; LE-32BIT-NEXT:    stw 14, 72(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 15, 76(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 16, 80(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 17, 84(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 18, 88(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 19, 92(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 20, 96(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 21, 100(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 22, 104(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 23, 108(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 24, 112(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 25, 116(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 26, 120(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 27, 124(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 28, 128(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 29, 132(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 30, 136(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 31, 140(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 12, 68(1)
+; LE-32BIT-NEXT:    lwz 0, 28(4)
+; LE-32BIT-NEXT:    lwz 11, 4(3)
+; LE-32BIT-NEXT:    lwz 6, 0(3)
+; LE-32BIT-NEXT:    rlwinm. 30, 0, 3, 0, 28
+; LE-32BIT-NEXT:    stw 5, 64(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    subfic 21, 30, 224
+; LE-32BIT-NEXT:    lwz 5, 24(3)
+; LE-32BIT-NEXT:    subfic 4, 30, 160
+; LE-32BIT-NEXT:    lwz 7, 28(3)
+; LE-32BIT-NEXT:    addi 0, 30, -128
+; LE-32BIT-NEXT:    lwz 10, 20(3)
+; LE-32BIT-NEXT:    subfic 28, 30, 96
+; LE-32BIT-NEXT:    lwz 8, 16(3)
+; LE-32BIT-NEXT:    addi 29, 30, -64
+; LE-32BIT-NEXT:    lwz 27, 12(3)
+; LE-32BIT-NEXT:    subfic 12, 30, 32
+; LE-32BIT-NEXT:    lwz 9, 8(3)
+; LE-32BIT-NEXT:    addi 3, 30, -192
+; LE-32BIT-NEXT:    slw 21, 6, 21
+; LE-32BIT-NEXT:    srw 16, 11, 3
+; LE-32BIT-NEXT:    stw 3, 56(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    srw 20, 7, 30
+; LE-32BIT-NEXT:    slw 15, 9, 4
+; LE-32BIT-NEXT:    srw 14, 27, 0
+; LE-32BIT-NEXT:    slw 31, 8, 28
+; LE-32BIT-NEXT:    srw 3, 10, 29
+; LE-32BIT-NEXT:    or 21, 16, 21
+; LE-32BIT-NEXT:    slw 16, 5, 12
+; LE-32BIT-NEXT:    srw 19, 10, 30
+; LE-32BIT-NEXT:    or 15, 14, 15
+; LE-32BIT-NEXT:    slw 14, 8, 12
+; LE-32BIT-NEXT:    or 3, 3, 31
+; LE-32BIT-NEXT:    slw 31, 6, 4
+; LE-32BIT-NEXT:    or 20, 20, 16
+; LE-32BIT-NEXT:    srw 16, 11, 0
+; LE-32BIT-NEXT:    stw 7, 60(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    addi 26, 30, -224
+; LE-32BIT-NEXT:    mr 7, 10
+; LE-32BIT-NEXT:    mr 10, 12
+; LE-32BIT-NEXT:    or 19, 19, 14
+; LE-32BIT-NEXT:    slw 14, 6, 28
+; LE-32BIT-NEXT:    or 16, 16, 31
+; LE-32BIT-NEXT:    srw 31, 11, 29
+; LE-32BIT-NEXT:    addi 23, 30, -160
+; LE-32BIT-NEXT:    srw 18, 27, 30
+; LE-32BIT-NEXT:    stw 0, 40(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    mr 12, 28
+; LE-32BIT-NEXT:    or 14, 31, 14
+; LE-32BIT-NEXT:    mr 28, 9
+; LE-32BIT-NEXT:    slw 31, 9, 10
+; LE-32BIT-NEXT:    srw 0, 6, 26
+; LE-32BIT-NEXT:    addi 25, 30, -96
+; LE-32BIT-NEXT:    srw 17, 11, 30
+; LE-32BIT-NEXT:    stw 4, 36(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    or 18, 18, 31
+; LE-32BIT-NEXT:    slw 31, 6, 10
+; LE-32BIT-NEXT:    or 4, 21, 0
+; LE-32BIT-NEXT:    srw 0, 28, 23
+; LE-32BIT-NEXT:    or 17, 17, 31
+; LE-32BIT-NEXT:    addi 31, 30, -32
+; LE-32BIT-NEXT:    or 0, 15, 0
+; LE-32BIT-NEXT:    srw 15, 8, 25
+; LE-32BIT-NEXT:    or 3, 3, 15
+; LE-32BIT-NEXT:    srw 15, 5, 31
+; LE-32BIT-NEXT:    or 20, 20, 15
+; LE-32BIT-NEXT:    srw 15, 8, 31
+; LE-32BIT-NEXT:    stw 3, 24(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    or 3, 19, 15
+; LE-32BIT-NEXT:    srw 23, 6, 23
+; LE-32BIT-NEXT:    stw 3, 48(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    subfic 15, 30, 64
+; LE-32BIT-NEXT:    or 3, 16, 23
+; LE-32BIT-NEXT:    stw 3, 44(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    subfic 3, 15, 32
+; LE-32BIT-NEXT:    slw 16, 28, 15
+; LE-32BIT-NEXT:    srw 22, 27, 3
+; LE-32BIT-NEXT:    stw 4, 32(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    or 4, 16, 22
+; LE-32BIT-NEXT:    subfic 16, 30, 128
+; LE-32BIT-NEXT:    stw 5, 28(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    subfic 5, 16, 32
+; LE-32BIT-NEXT:    stw 4, 20(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    slw 4, 6, 16
+; LE-32BIT-NEXT:    srw 24, 11, 5
+; LE-32BIT-NEXT:    stw 29, 52(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    mr 29, 27
+; LE-32BIT-NEXT:    or 22, 4, 24
+; LE-32BIT-NEXT:    slw 24, 28, 16
+; LE-32BIT-NEXT:    srw 27, 27, 5
+; LE-32BIT-NEXT:    or 27, 24, 27
+; LE-32BIT-NEXT:    slw 24, 8, 15
+; LE-32BIT-NEXT:    srw 26, 7, 3
+; LE-32BIT-NEXT:    or 26, 24, 26
+; LE-32BIT-NEXT:    subfic 24, 30, 192
+; LE-32BIT-NEXT:    mr 9, 10
+; LE-32BIT-NEXT:    mr 10, 28
+; LE-32BIT-NEXT:    subfic 28, 24, 32
+; LE-32BIT-NEXT:    srw 28, 11, 28
+; LE-32BIT-NEXT:    slw 19, 6, 24
+; LE-32BIT-NEXT:    or 28, 19, 28
+; LE-32BIT-NEXT:    srw 19, 6, 25
+; LE-32BIT-NEXT:    or 19, 14, 19
+; LE-32BIT-NEXT:    srw 14, 10, 31
+; LE-32BIT-NEXT:    lwz 4, 64(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    or 18, 18, 14
+; LE-32BIT-NEXT:    srw 3, 11, 3
+; LE-32BIT-NEXT:    slw 14, 6, 15
+; LE-32BIT-NEXT:    cmplwi 5, 30, 64
+; LE-32BIT-NEXT:    cmplwi 1, 30, 128
+; LE-32BIT-NEXT:    slw 24, 11, 24
+; LE-32BIT-NEXT:    mr 21, 8
+; LE-32BIT-NEXT:    or 8, 14, 3
+; LE-32BIT-NEXT:    srw 14, 6, 31
+; LE-32BIT-NEXT:    crnand 28, 4, 20
+; LE-32BIT-NEXT:    srw 31, 6, 30
+; LE-32BIT-NEXT:    or 24, 0, 24
+; LE-32BIT-NEXT:    slw 0, 7, 15
+; LE-32BIT-NEXT:    mr 23, 7
+; LE-32BIT-NEXT:    or 17, 17, 14
+; LE-32BIT-NEXT:    bc 12, 28, .LBB9_2
+; LE-32BIT-NEXT:  # %bb.1:
+; LE-32BIT-NEXT:    ori 14, 31, 0
+; LE-32BIT-NEXT:    b .LBB9_3
+; LE-32BIT-NEXT:  .LBB9_2:
+; LE-32BIT-NEXT:    li 14, 0
+; LE-32BIT-NEXT:  .LBB9_3:
+; LE-32BIT-NEXT:    or 20, 20, 0
+; LE-32BIT-NEXT:    subfic 0, 16, 64
+; LE-32BIT-NEXT:    lwz 7, 20(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    srw 31, 29, 0
+; LE-32BIT-NEXT:    stw 14, 0(4)
+; LE-32BIT-NEXT:    subfic 14, 0, 32
+; LE-32BIT-NEXT:    slw 14, 10, 14
+; LE-32BIT-NEXT:    or 14, 31, 14
+; LE-32BIT-NEXT:    slw 31, 29, 9
+; LE-32BIT-NEXT:    lwz 3, 36(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    or 7, 7, 31
+; LE-32BIT-NEXT:    slw 31, 11, 12
+; LE-32BIT-NEXT:    stw 7, 20(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    or 7, 22, 31
+; LE-32BIT-NEXT:    slw 31, 29, 12
+; LE-32BIT-NEXT:    or 27, 27, 31
+; LE-32BIT-NEXT:    slw 31, 23, 9
+; LE-32BIT-NEXT:    or 26, 26, 31
+; LE-32BIT-NEXT:    slw 31, 11, 3
+; LE-32BIT-NEXT:    or 28, 28, 31
+; LE-32BIT-NEXT:    slw 31, 11, 15
+; LE-32BIT-NEXT:    lwz 22, 28(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    or 18, 18, 31
+; LE-32BIT-NEXT:    lwz 31, 40(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    srw 0, 10, 0
+; LE-32BIT-NEXT:    or 7, 7, 0
+; LE-32BIT-NEXT:    srw 0, 22, 30
+; LE-32BIT-NEXT:    slw 25, 11, 9
+; LE-32BIT-NEXT:    or 26, 0, 26
+; LE-32BIT-NEXT:    srw 0, 10, 31
+; LE-32BIT-NEXT:    or 3, 8, 25
+; LE-32BIT-NEXT:    or 28, 0, 28
+; LE-32BIT-NEXT:    srw 0, 10, 30
+; LE-32BIT-NEXT:    srw 5, 10, 5
+; LE-32BIT-NEXT:    or 3, 0, 3
+; LE-32BIT-NEXT:    bc 12, 28, .LBB9_5
+; LE-32BIT-NEXT:  # %bb.4:
+; LE-32BIT-NEXT:    ori 0, 17, 0
+; LE-32BIT-NEXT:    b .LBB9_6
+; LE-32BIT-NEXT:  .LBB9_5:
+; LE-32BIT-NEXT:    li 0, 0
+; LE-32BIT-NEXT:  .LBB9_6:
+; LE-32BIT-NEXT:    lwz 8, 32(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    or 5, 14, 5
+; LE-32BIT-NEXT:    mr 14, 4
+; LE-32BIT-NEXT:    stw 0, 4(4)
+; LE-32BIT-NEXT:    slw 0, 11, 16
+; LE-32BIT-NEXT:    lwz 4, 52(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    cmplwi 6, 31, 64
+; LE-32BIT-NEXT:    mr 9, 21
+; LE-32BIT-NEXT:    or 5, 0, 5
+; LE-32BIT-NEXT:    lwz 0, 56(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    bc 12, 24, .LBB9_8
+; LE-32BIT-NEXT:  # %bb.7:
+; LE-32BIT-NEXT:    ori 25, 8, 0
+; LE-32BIT-NEXT:    b .LBB9_9
+; LE-32BIT-NEXT:  .LBB9_8:
+; LE-32BIT-NEXT:    addi 25, 24, 0
+; LE-32BIT-NEXT:  .LBB9_9:
+; LE-32BIT-NEXT:    bc 12, 20, .LBB9_11
+; LE-32BIT-NEXT:  # %bb.10:
+; LE-32BIT-NEXT:    ori 24, 19, 0
+; LE-32BIT-NEXT:    b .LBB9_12
+; LE-32BIT-NEXT:  .LBB9_11:
+; LE-32BIT-NEXT:    addi 24, 18, 0
+; LE-32BIT-NEXT:  .LBB9_12:
+; LE-32BIT-NEXT:    srw 19, 9, 4
+; LE-32BIT-NEXT:    srw 17, 6, 4
+; LE-32BIT-NEXT:    lwz 4, 20(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    srw 30, 21, 30
+; LE-32BIT-NEXT:    lwz 8, 24(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    slw 21, 29, 16
+; LE-32BIT-NEXT:    cmplwi 7, 16, 64
+; LE-32BIT-NEXT:    cmplwi 3, 16, 0
+; LE-32BIT-NEXT:    li 16, 0
+; LE-32BIT-NEXT:    srw 18, 6, 0
+; LE-32BIT-NEXT:    bc 12, 28, .LBB9_14
+; LE-32BIT-NEXT:  # %bb.13:
+; LE-32BIT-NEXT:    ori 0, 16, 0
+; LE-32BIT-NEXT:    b .LBB9_15
+; LE-32BIT-NEXT:  .LBB9_14:
+; LE-32BIT-NEXT:    addi 0, 21, 0
+; LE-32BIT-NEXT:  .LBB9_15:
+; LE-32BIT-NEXT:    lwz 21, 60(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    bc 12, 28, .LBB9_16
+; LE-32BIT-NEXT:    b .LBB9_17
+; LE-32BIT-NEXT:  .LBB9_16:
+; LE-32BIT-NEXT:    addi 4, 7, 0
+; LE-32BIT-NEXT:  .LBB9_17:
+; LE-32BIT-NEXT:    bc 12, 20, .LBB9_18
+; LE-32BIT-NEXT:    b .LBB9_19
+; LE-32BIT-NEXT:  .LBB9_18:
+; LE-32BIT-NEXT:    addi 8, 20, 0
+; LE-32BIT-NEXT:  .LBB9_19:
+; LE-32BIT-NEXT:    mr 12, 29
+; LE-32BIT-NEXT:    lwz 7, 48(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    slw 20, 29, 15
+; LE-32BIT-NEXT:    srw 29, 6, 31
+; LE-32BIT-NEXT:    bc 12, 2, .LBB9_20
+; LE-32BIT-NEXT:    b .LBB9_21
+; LE-32BIT-NEXT:  .LBB9_20:
+; LE-32BIT-NEXT:    addi 8, 21, 0
+; LE-32BIT-NEXT:  .LBB9_21:
+; LE-32BIT-NEXT:    cmplwi 2, 31, 0
+; LE-32BIT-NEXT:    bc 12, 20, .LBB9_23
+; LE-32BIT-NEXT:  # %bb.22:
+; LE-32BIT-NEXT:    ori 26, 19, 0
+; LE-32BIT-NEXT:    ori 3, 17, 0
+; LE-32BIT-NEXT:    b .LBB9_23
+; LE-32BIT-NEXT:  .LBB9_23:
+; LE-32BIT-NEXT:    or 8, 8, 0
+; LE-32BIT-NEXT:    bc 12, 20, .LBB9_25
+; LE-32BIT-NEXT:  # %bb.24:
+; LE-32BIT-NEXT:    ori 0, 16, 0
+; LE-32BIT-NEXT:    b .LBB9_26
+; LE-32BIT-NEXT:  .LBB9_25:
+; LE-32BIT-NEXT:    addi 0, 30, 0
+; LE-32BIT-NEXT:  .LBB9_26:
+; LE-32BIT-NEXT:    bc 12, 24, .LBB9_28
+; LE-32BIT-NEXT:  # %bb.27:
+; LE-32BIT-NEXT:    ori 30, 16, 0
+; LE-32BIT-NEXT:    b .LBB9_29
+; LE-32BIT-NEXT:  .LBB9_28:
+; LE-32BIT-NEXT:    addi 30, 29, 0
+; LE-32BIT-NEXT:  .LBB9_29:
+; LE-32BIT-NEXT:    bc 12, 20, .LBB9_31
+; LE-32BIT-NEXT:  # %bb.30:
+; LE-32BIT-NEXT:    ori 29, 16, 0
+; LE-32BIT-NEXT:    b .LBB9_32
+; LE-32BIT-NEXT:  .LBB9_31:
+; LE-32BIT-NEXT:    addi 29, 7, 0
+; LE-32BIT-NEXT:  .LBB9_32:
+; LE-32BIT-NEXT:    lwz 7, 44(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    bc 12, 10, .LBB9_33
+; LE-32BIT-NEXT:    b .LBB9_34
+; LE-32BIT-NEXT:  .LBB9_33:
+; LE-32BIT-NEXT:    addi 25, 12, 0
+; LE-32BIT-NEXT:  .LBB9_34:
+; LE-32BIT-NEXT:    bc 12, 14, .LBB9_35
+; LE-32BIT-NEXT:    b .LBB9_36
+; LE-32BIT-NEXT:  .LBB9_35:
+; LE-32BIT-NEXT:    addi 4, 6, 0
+; LE-32BIT-NEXT:  .LBB9_36:
+; LE-32BIT-NEXT:    bc 12, 2, .LBB9_38
+; LE-32BIT-NEXT:  # %bb.37:
+; LE-32BIT-NEXT:    ori 6, 26, 0
+; LE-32BIT-NEXT:    b .LBB9_39
+; LE-32BIT-NEXT:  .LBB9_38:
+; LE-32BIT-NEXT:    addi 6, 22, 0
+; LE-32BIT-NEXT:  .LBB9_39:
+; LE-32BIT-NEXT:    li 26, 0
+; LE-32BIT-NEXT:    bc 12, 2, .LBB9_40
+; LE-32BIT-NEXT:    b .LBB9_41
+; LE-32BIT-NEXT:  .LBB9_40:
+; LE-32BIT-NEXT:    addi 3, 10, 0
+; LE-32BIT-NEXT:  .LBB9_41:
+; LE-32BIT-NEXT:    bc 12, 28, .LBB9_43
+; LE-32BIT-NEXT:  # %bb.42:
+; LE-32BIT-NEXT:    ori 5, 20, 0
+; LE-32BIT-NEXT:    b .LBB9_43
+; LE-32BIT-NEXT:  .LBB9_43:
+; LE-32BIT-NEXT:    bc 12, 4, .LBB9_45
+; LE-32BIT-NEXT:  # %bb.44:
+; LE-32BIT-NEXT:    ori 8, 25, 0
+; LE-32BIT-NEXT:    b .LBB9_45
+; LE-32BIT-NEXT:  .LBB9_45:
+; LE-32BIT-NEXT:    bc 12, 24, .LBB9_47
+; LE-32BIT-NEXT:  # %bb.46:
+; LE-32BIT-NEXT:    ori 28, 18, 0
+; LE-32BIT-NEXT:    b .LBB9_47
+; LE-32BIT-NEXT:  .LBB9_47:
+; LE-32BIT-NEXT:    bc 12, 28, .LBB9_49
+; LE-32BIT-NEXT:  # %bb.48:
+; LE-32BIT-NEXT:    ori 27, 16, 0
+; LE-32BIT-NEXT:    b .LBB9_49
+; LE-32BIT-NEXT:  .LBB9_49:
+; LE-32BIT-NEXT:    bc 12, 2, .LBB9_51
+; LE-32BIT-NEXT:  # %bb.50:
+; LE-32BIT-NEXT:    ori 12, 24, 0
+; LE-32BIT-NEXT:    b .LBB9_51
+; LE-32BIT-NEXT:  .LBB9_51:
+; LE-32BIT-NEXT:    bc 12, 4, .LBB9_53
+; LE-32BIT-NEXT:  # %bb.52:
+; LE-32BIT-NEXT:    ori 3, 26, 0
+; LE-32BIT-NEXT:    b .LBB9_53
+; LE-32BIT-NEXT:  .LBB9_53:
+; LE-32BIT-NEXT:    bc 12, 14, .LBB9_54
+; LE-32BIT-NEXT:    b .LBB9_55
+; LE-32BIT-NEXT:  .LBB9_54:
+; LE-32BIT-NEXT:    addi 5, 11, 0
+; LE-32BIT-NEXT:  .LBB9_55:
+; LE-32BIT-NEXT:    bc 12, 10, .LBB9_56
+; LE-32BIT-NEXT:    b .LBB9_57
+; LE-32BIT-NEXT:  .LBB9_56:
+; LE-32BIT-NEXT:    addi 28, 10, 0
+; LE-32BIT-NEXT:  .LBB9_57:
+; LE-32BIT-NEXT:    or 6, 6, 27
+; LE-32BIT-NEXT:    stw 3, 8(14)
+; LE-32BIT-NEXT:    or 3, 0, 4
+; LE-32BIT-NEXT:    bc 12, 2, .LBB9_59
+; LE-32BIT-NEXT:  # %bb.58:
+; LE-32BIT-NEXT:    ori 4, 8, 0
+; LE-32BIT-NEXT:    b .LBB9_60
+; LE-32BIT-NEXT:  .LBB9_59:
+; LE-32BIT-NEXT:    addi 4, 21, 0
+; LE-32BIT-NEXT:  .LBB9_60:
+; LE-32BIT-NEXT:    bc 12, 24, .LBB9_62
+; LE-32BIT-NEXT:  # %bb.61:
+; LE-32BIT-NEXT:    ori 24, 16, 0
+; LE-32BIT-NEXT:    b .LBB9_63
+; LE-32BIT-NEXT:  .LBB9_62:
+; LE-32BIT-NEXT:    addi 24, 7, 0
+; LE-32BIT-NEXT:  .LBB9_63:
+; LE-32BIT-NEXT:    bc 12, 4, .LBB9_65
+; LE-32BIT-NEXT:  # %bb.64:
+; LE-32BIT-NEXT:    ori 3, 30, 0
+; LE-32BIT-NEXT:    ori 6, 28, 0
+; LE-32BIT-NEXT:    ori 12, 16, 0
+; LE-32BIT-NEXT:    b .LBB9_65
+; LE-32BIT-NEXT:  .LBB9_65:
+; LE-32BIT-NEXT:    stw 4, 28(14)
+; LE-32BIT-NEXT:    or 4, 29, 5
+; LE-32BIT-NEXT:    bc 12, 4, .LBB9_67
+; LE-32BIT-NEXT:  # %bb.66:
+; LE-32BIT-NEXT:    ori 4, 24, 0
+; LE-32BIT-NEXT:    b .LBB9_67
+; LE-32BIT-NEXT:  .LBB9_67:
+; LE-32BIT-NEXT:    bc 12, 2, .LBB9_69
+; LE-32BIT-NEXT:  # %bb.68:
+; LE-32BIT-NEXT:    ori 5, 6, 0
+; LE-32BIT-NEXT:    b .LBB9_70
+; LE-32BIT-NEXT:  .LBB9_69:
+; LE-32BIT-NEXT:    addi 3, 9, 0
+; LE-32BIT-NEXT:    addi 5, 22, 0
+; LE-32BIT-NEXT:  .LBB9_70:
+; LE-32BIT-NEXT:    stw 12, 12(14)
+; LE-32BIT-NEXT:    stw 3, 16(14)
+; LE-32BIT-NEXT:    bc 12, 2, .LBB9_72
+; LE-32BIT-NEXT:  # %bb.71:
+; LE-32BIT-NEXT:    ori 3, 4, 0
+; LE-32BIT-NEXT:    b .LBB9_73
+; LE-32BIT-NEXT:  .LBB9_72:
+; LE-32BIT-NEXT:    addi 3, 23, 0
+; LE-32BIT-NEXT:  .LBB9_73:
+; LE-32BIT-NEXT:    stw 5, 24(14)
+; LE-32BIT-NEXT:    stw 3, 20(14)
+; LE-32BIT-NEXT:    lwz 12, 68(1)
+; LE-32BIT-NEXT:    lwz 31, 140(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    mtcrf 32, 12 # cr2
+; LE-32BIT-NEXT:    mtcrf 16, 12 # cr3
+; LE-32BIT-NEXT:    lwz 30, 136(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 29, 132(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 28, 128(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 27, 124(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 26, 120(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 25, 116(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 24, 112(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 23, 108(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 22, 104(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 21, 100(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 20, 96(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 19, 92(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 18, 88(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 17, 84(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 16, 80(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 15, 76(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 14, 72(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    addi 1, 1, 144
+; LE-32BIT-NEXT:    blr
+  %src = load i256, ptr %src.ptr, align 1
+  %byteOff = load i256, ptr %byteOff.ptr, align 1
+  %bitOff = shl i256 %byteOff, 3
+  %res = lshr i256 %src, %bitOff
+  store i256 %res, ptr %dst, align 1
+  ret void
+}
+define void @shl_32bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; LE-64BIT-LABEL: shl_32bytes:
+; LE-64BIT:       # %bb.0:
+; LE-64BIT-NEXT:    lwz 4, 0(4)
+; LE-64BIT-NEXT:    ld 7, 24(3)
+; LE-64BIT-NEXT:    ld 8, 16(3)
+; LE-64BIT-NEXT:    ld 9, 8(3)
+; LE-64BIT-NEXT:    li 6, 0
+; LE-64BIT-NEXT:    ld 3, 0(3)
+; LE-64BIT-NEXT:    std 28, -32(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    std 21, -88(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    std 24, -64(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    rlwinm. 4, 4, 3, 0, 28
+; LE-64BIT-NEXT:    std 25, -56(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    subfic 28, 4, 64
+; LE-64BIT-NEXT:    subfic 11, 4, 192
+; LE-64BIT-NEXT:    std 27, -40(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    std 29, -24(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    addi 0, 4, -128
+; LE-64BIT-NEXT:    sld 29, 9, 4
+; LE-64BIT-NEXT:    addi 27, 4, -64
+; LE-64BIT-NEXT:    subfic 25, 4, 128
+; LE-64BIT-NEXT:    srd 24, 8, 28
+; LE-64BIT-NEXT:    std 22, -80(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    std 26, -48(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    srd 21, 9, 28
+; LE-64BIT-NEXT:    srd 28, 3, 28
+; LE-64BIT-NEXT:    std 30, -16(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    sld 10, 7, 4
+; LE-64BIT-NEXT:    addi 30, 4, -192
+; LE-64BIT-NEXT:    subfic 22, 25, 64
+; LE-64BIT-NEXT:    srd 11, 3, 11
+; LE-64BIT-NEXT:    sld 26, 9, 0
+; LE-64BIT-NEXT:    or 29, 29, 28
+; LE-64BIT-NEXT:    std 23, -72(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    or 10, 10, 24
+; LE-64BIT-NEXT:    sld 28, 3, 27
+; LE-64BIT-NEXT:    sld 30, 3, 30
+; LE-64BIT-NEXT:    or 11, 26, 11
+; LE-64BIT-NEXT:    ld 26, -48(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    sld 23, 8, 27
+; LE-64BIT-NEXT:    sld 27, 9, 22
+; LE-64BIT-NEXT:    or 29, 29, 28
+; LE-64BIT-NEXT:    or 11, 11, 30
+; LE-64BIT-NEXT:    ld 24, -64(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    srd 28, 3, 25
+; LE-64BIT-NEXT:    or 10, 10, 23
+; LE-64BIT-NEXT:    ld 23, -72(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    ld 22, -80(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    srd 9, 9, 25
+; LE-64BIT-NEXT:    or 30, 28, 27
+; LE-64BIT-NEXT:    ld 28, -32(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    ld 27, -40(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    ld 25, -56(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    cmplwi 1, 4, 128
+; LE-64BIT-NEXT:    sld 12, 8, 4
+; LE-64BIT-NEXT:    or 9, 10, 9
+; LE-64BIT-NEXT:    or 30, 30, 21
+; LE-64BIT-NEXT:    ld 21, -88(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    sld 10, 3, 0
+; LE-64BIT-NEXT:    isel 9, 9, 11, 4
+; LE-64BIT-NEXT:    or 11, 12, 30
+; LE-64BIT-NEXT:    ld 30, -16(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    iseleq 7, 7, 9
+; LE-64BIT-NEXT:    sld 3, 3, 4
+; LE-64BIT-NEXT:    isel 9, 11, 10, 4
+; LE-64BIT-NEXT:    std 7, 24(5)
+; LE-64BIT-NEXT:    isel 0, 29, 6, 4
+; LE-64BIT-NEXT:    ld 29, -24(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    iseleq 4, 8, 9
+; LE-64BIT-NEXT:    std 0, 8(5)
+; LE-64BIT-NEXT:    isel 3, 3, 6, 4
+; LE-64BIT-NEXT:    std 4, 16(5)
+; LE-64BIT-NEXT:    std 3, 0(5)
+; LE-64BIT-NEXT:    blr
+;
+; BE-LABEL: shl_32bytes:
+; BE:       # %bb.0:
+; BE-NEXT:    lwz 4, 28(4)
+; BE-NEXT:    ld 7, 8(3)
+; BE-NEXT:    ld 8, 0(3)
+; BE-NEXT:    ld 9, 16(3)
+; BE-NEXT:    ld 3, 24(3)
+; BE-NEXT:    std 27, -40(1) # 8-byte Folded Spill
+; BE-NEXT:    std 30, -16(1) # 8-byte Folded Spill
+; BE-NEXT:    std 28, -32(1) # 8-byte Folded Spill
+; BE-NEXT:    std 29, -24(1) # 8-byte Folded Spill
+; BE-NEXT:    li 6, 0
+; BE-NEXT:    rlwinm. 4, 4, 3, 0, 28
+; BE-NEXT:    subfic 10, 4, 192
+; BE-NEXT:    addi 11, 4, -128
+; BE-NEXT:    addi 12, 4, -192
+; BE-NEXT:    subfic 30, 4, 64
+; BE-NEXT:    srd 10, 3, 10
+; BE-NEXT:    sld 27, 9, 11
+; BE-NEXT:    sld 0, 8, 4
+; BE-NEXT:    addi 29, 4, -64
+; BE-NEXT:    subfic 28, 4, 128
+; BE-NEXT:    sld 12, 3, 12
+; BE-NEXT:    or 10, 27, 10
+; BE-NEXT:    srd 27, 7, 30
+; BE-NEXT:    or 10, 10, 12
+; BE-NEXT:    or 0, 0, 27
+; BE-NEXT:    sld 27, 7, 29
+; BE-NEXT:    subfic 12, 28, 64
+; BE-NEXT:    or 0, 0, 27
+; BE-NEXT:    srd 27, 3, 28
+; BE-NEXT:    sld 12, 9, 12
+; BE-NEXT:    srd 28, 9, 28
+; BE-NEXT:    cmplwi 1, 4, 128
+; BE-NEXT:    or 12, 27, 12
+; BE-NEXT:    or 28, 0, 28
+; BE-NEXT:    srd 0, 9, 30
+; BE-NEXT:    sld 9, 9, 4
+; BE-NEXT:    sld 11, 3, 11
+; BE-NEXT:    bc 12, 4, .LBB10_1
+; BE-NEXT:    b .LBB10_2
+; BE-NEXT:  .LBB10_1:
+; BE-NEXT:    addi 10, 28, 0
+; BE-NEXT:  .LBB10_2:
+; BE-NEXT:    ld 28, -32(1) # 8-byte Folded Reload
+; BE-NEXT:    ld 27, -40(1) # 8-byte Folded Reload
+; BE-NEXT:    or 12, 12, 0
+; BE-NEXT:    sld 0, 7, 4
+; BE-NEXT:    or 12, 0, 12
+; BE-NEXT:    srd 0, 3, 30
+; BE-NEXT:    sld 30, 3, 29
+; BE-NEXT:    bc 12, 4, .LBB10_3
+; BE-NEXT:    b .LBB10_4
+; BE-NEXT:  .LBB10_3:
+; BE-NEXT:    addi 11, 12, 0
+; BE-NEXT:  .LBB10_4:
+; BE-NEXT:    sld 3, 3, 4
+; BE-NEXT:    bc 12, 2, .LBB10_6
+; BE-NEXT:  # %bb.5:
+; BE-NEXT:    ori 4, 10, 0
+; BE-NEXT:    b .LBB10_7
+; BE-NEXT:  .LBB10_6:
+; BE-NEXT:    addi 4, 8, 0
+; BE-NEXT:  .LBB10_7:
+; BE-NEXT:    ld 29, -24(1) # 8-byte Folded Reload
+; BE-NEXT:    or 9, 9, 0
+; BE-NEXT:    or 9, 9, 30
+; BE-NEXT:    bc 12, 2, .LBB10_9
+; BE-NEXT:  # %bb.8:
+; BE-NEXT:    ori 7, 11, 0
+; BE-NEXT:    b .LBB10_9
+; BE-NEXT:  .LBB10_9:
+; BE-NEXT:    bc 12, 4, .LBB10_11
+; BE-NEXT:  # %bb.10:
+; BE-NEXT:    ori 8, 6, 0
+; BE-NEXT:    ori 3, 6, 0
+; BE-NEXT:    b .LBB10_12
+; BE-NEXT:  .LBB10_11:
+; BE-NEXT:    addi 8, 9, 0
+; BE-NEXT:  .LBB10_12:
+; BE-NEXT:    std 4, 0(5)
+; BE-NEXT:    ld 30, -16(1) # 8-byte Folded Reload
+; BE-NEXT:    std 3, 24(5)
+; BE-NEXT:    std 8, 16(5)
+; BE-NEXT:    std 7, 8(5)
+; BE-NEXT:    blr
+;
+; LE-32BIT-LABEL: shl_32bytes:
+; LE-32BIT:       # %bb.0:
+; LE-32BIT-NEXT:    stwu 1, -144(1)
+; LE-32BIT-NEXT:    mfcr 12
+; LE-32BIT-NEXT:    stw 14, 72(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 15, 76(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 16, 80(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 17, 84(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 18, 88(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 19, 92(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 20, 96(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 21, 100(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 22, 104(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 23, 108(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 24, 112(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 25, 116(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 26, 120(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 27, 124(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 28, 128(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 29, 132(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 30, 136(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 31, 140(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 12, 68(1)
+; LE-32BIT-NEXT:    lwz 0, 28(4)
+; LE-32BIT-NEXT:    stw 5, 64(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    lwz 6, 24(3)
+; LE-32BIT-NEXT:    rlwinm. 30, 0, 3, 0, 28
+; LE-32BIT-NEXT:    lwz 5, 28(3)
+; LE-32BIT-NEXT:    subfic 21, 30, 224
+; LE-32BIT-NEXT:    lwz 7, 4(3)
+; LE-32BIT-NEXT:    subfic 0, 30, 160
+; LE-32BIT-NEXT:    lwz 9, 0(3)
+; LE-32BIT-NEXT:    addi 4, 30, -128
+; LE-32BIT-NEXT:    lwz 10, 8(3)
+; LE-32BIT-NEXT:    subfic 28, 30, 96
+; LE-32BIT-NEXT:    lwz 8, 12(3)
+; LE-32BIT-NEXT:    addi 29, 30, -64
+; LE-32BIT-NEXT:    lwz 12, 16(3)
+; LE-32BIT-NEXT:    subfic 25, 30, 32
+; LE-32BIT-NEXT:    lwz 11, 20(3)
+; LE-32BIT-NEXT:    addi 3, 30, -192
+; LE-32BIT-NEXT:    srw 21, 5, 21
+; LE-32BIT-NEXT:    slw 16, 6, 3
+; LE-32BIT-NEXT:    stw 3, 56(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    slw 20, 9, 30
+; LE-32BIT-NEXT:    srw 15, 11, 0
+; LE-32BIT-NEXT:    slw 14, 12, 4
+; LE-32BIT-NEXT:    srw 31, 8, 28
+; LE-32BIT-NEXT:    slw 3, 10, 29
+; LE-32BIT-NEXT:    or 21, 16, 21
+; LE-32BIT-NEXT:    srw 16, 7, 25
+; LE-32BIT-NEXT:    slw 19, 10, 30
+; LE-32BIT-NEXT:    or 15, 14, 15
+; LE-32BIT-NEXT:    srw 14, 8, 25
+; LE-32BIT-NEXT:    or 3, 3, 31
+; LE-32BIT-NEXT:    srw 31, 5, 0
+; LE-32BIT-NEXT:    or 20, 20, 16
+; LE-32BIT-NEXT:    slw 16, 6, 4
+; LE-32BIT-NEXT:    addi 27, 30, -224
+; LE-32BIT-NEXT:    or 19, 19, 14
+; LE-32BIT-NEXT:    srw 14, 5, 28
+; LE-32BIT-NEXT:    or 16, 16, 31
+; LE-32BIT-NEXT:    slw 31, 6, 29
+; LE-32BIT-NEXT:    addi 23, 30, -160
+; LE-32BIT-NEXT:    slw 18, 12, 30
+; LE-32BIT-NEXT:    stw 0, 40(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    or 14, 31, 14
+; LE-32BIT-NEXT:    srw 31, 11, 25
+; LE-32BIT-NEXT:    slw 0, 5, 27
+; LE-32BIT-NEXT:    addi 26, 30, -96
+; LE-32BIT-NEXT:    slw 17, 6, 30
+; LE-32BIT-NEXT:    or 18, 18, 31
+; LE-32BIT-NEXT:    srw 31, 5, 25
+; LE-32BIT-NEXT:    or 21, 21, 0
+; LE-32BIT-NEXT:    slw 0, 11, 23
+; LE-32BIT-NEXT:    or 17, 17, 31
+; LE-32BIT-NEXT:    addi 31, 30, -32
+; LE-32BIT-NEXT:    or 0, 15, 0
+; LE-32BIT-NEXT:    slw 15, 8, 26
+; LE-32BIT-NEXT:    stw 29, 52(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    or 29, 3, 15
+; LE-32BIT-NEXT:    slw 15, 7, 31
+; LE-32BIT-NEXT:    or 20, 20, 15
+; LE-32BIT-NEXT:    slw 15, 8, 31
+; LE-32BIT-NEXT:    or 3, 19, 15
+; LE-32BIT-NEXT:    subfic 15, 30, 128
+; LE-32BIT-NEXT:    slw 23, 5, 23
+; LE-32BIT-NEXT:    stw 3, 48(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    or 3, 16, 23
+; LE-32BIT-NEXT:    subfic 16, 15, 32
+; LE-32BIT-NEXT:    stw 3, 44(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    srw 3, 11, 15
+; LE-32BIT-NEXT:    slw 22, 12, 16
+; LE-32BIT-NEXT:    or 23, 3, 22
+; LE-32BIT-NEXT:    subfic 22, 30, 64
+; LE-32BIT-NEXT:    stw 9, 60(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    mr 9, 10
+; LE-32BIT-NEXT:    subfic 3, 22, 32
+; LE-32BIT-NEXT:    stw 4, 36(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    srw 4, 8, 22
+; LE-32BIT-NEXT:    slw 24, 9, 3
+; LE-32BIT-NEXT:    or 4, 4, 24
+; LE-32BIT-NEXT:    subfic 24, 30, 192
+; LE-32BIT-NEXT:    subfic 27, 24, 32
+; LE-32BIT-NEXT:    mr 10, 26
+; LE-32BIT-NEXT:    slw 27, 6, 27
+; LE-32BIT-NEXT:    srw 26, 5, 24
+; LE-32BIT-NEXT:    stw 28, 24(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    or 27, 26, 27
+; LE-32BIT-NEXT:    srw 26, 11, 22
+; LE-32BIT-NEXT:    slw 28, 12, 3
+; LE-32BIT-NEXT:    or 28, 26, 28
+; LE-32BIT-NEXT:    srw 26, 5, 15
+; LE-32BIT-NEXT:    slw 19, 6, 16
+; LE-32BIT-NEXT:    or 26, 26, 19
+; LE-32BIT-NEXT:    slw 19, 5, 10
+; LE-32BIT-NEXT:    stw 7, 32(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    mr 7, 9
+; LE-32BIT-NEXT:    or 19, 14, 19
+; LE-32BIT-NEXT:    slw 14, 11, 31
+; LE-32BIT-NEXT:    lwz 9, 64(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    or 18, 18, 14
+; LE-32BIT-NEXT:    slw 3, 6, 3
+; LE-32BIT-NEXT:    srw 14, 5, 22
+; LE-32BIT-NEXT:    cmplwi 5, 30, 64
+; LE-32BIT-NEXT:    cmplwi 1, 30, 128
+; LE-32BIT-NEXT:    srw 24, 6, 24
+; LE-32BIT-NEXT:    or 10, 14, 3
+; LE-32BIT-NEXT:    slw 14, 5, 31
+; LE-32BIT-NEXT:    crnand 28, 4, 20
+; LE-32BIT-NEXT:    slw 31, 5, 30
+; LE-32BIT-NEXT:    or 24, 0, 24
+; LE-32BIT-NEXT:    mr 3, 7
+; LE-32BIT-NEXT:    stw 7, 28(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    srw 0, 7, 22
+; LE-32BIT-NEXT:    lwz 7, 24(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    or 17, 17, 14
+; LE-32BIT-NEXT:    bc 12, 28, .LBB10_2
+; LE-32BIT-NEXT:  # %bb.1:
+; LE-32BIT-NEXT:    ori 14, 31, 0
+; LE-32BIT-NEXT:    b .LBB10_3
+; LE-32BIT-NEXT:  .LBB10_2:
+; LE-32BIT-NEXT:    li 14, 0
+; LE-32BIT-NEXT:  .LBB10_3:
+; LE-32BIT-NEXT:    or 20, 20, 0
+; LE-32BIT-NEXT:    subfic 0, 15, 64
+; LE-32BIT-NEXT:    stw 14, 28(9)
+; LE-32BIT-NEXT:    subfic 14, 0, 32
+; LE-32BIT-NEXT:    srw 14, 11, 14
+; LE-32BIT-NEXT:    slw 31, 12, 0
+; LE-32BIT-NEXT:    or 14, 31, 14
+; LE-32BIT-NEXT:    srw 31, 12, 7
+; LE-32BIT-NEXT:    or 23, 23, 31
+; LE-32BIT-NEXT:    srw 31, 3, 25
+; LE-32BIT-NEXT:    lwz 3, 40(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    or 4, 4, 31
+; LE-32BIT-NEXT:    slw 0, 11, 0
+; LE-32BIT-NEXT:    cmplwi 3, 15, 0
+; LE-32BIT-NEXT:    srw 31, 6, 3
+; LE-32BIT-NEXT:    or 27, 27, 31
+; LE-32BIT-NEXT:    srw 31, 12, 25
+; LE-32BIT-NEXT:    or 28, 28, 31
+; LE-32BIT-NEXT:    srw 31, 6, 7
+; LE-32BIT-NEXT:    or 26, 26, 31
+; LE-32BIT-NEXT:    srw 31, 6, 22
+; LE-32BIT-NEXT:    or 18, 18, 31
+; LE-32BIT-NEXT:    lwz 31, 36(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    srw 25, 6, 25
+; LE-32BIT-NEXT:    or 3, 10, 25
+; LE-32BIT-NEXT:    or 26, 26, 0
+; LE-32BIT-NEXT:    cmplwi 6, 31, 64
+; LE-32BIT-NEXT:    slw 0, 11, 30
+; LE-32BIT-NEXT:    bc 12, 24, .LBB10_5
+; LE-32BIT-NEXT:  # %bb.4:
+; LE-32BIT-NEXT:    ori 25, 21, 0
+; LE-32BIT-NEXT:    b .LBB10_6
+; LE-32BIT-NEXT:  .LBB10_5:
+; LE-32BIT-NEXT:    addi 25, 24, 0
+; LE-32BIT-NEXT:  .LBB10_6:
+; LE-32BIT-NEXT:    slw 24, 11, 16
+; LE-32BIT-NEXT:    lwz 10, 32(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    or 3, 0, 3
+; LE-32BIT-NEXT:    bc 12, 28, .LBB10_8
+; LE-32BIT-NEXT:  # %bb.7:
+; LE-32BIT-NEXT:    ori 0, 17, 0
+; LE-32BIT-NEXT:    b .LBB10_9
+; LE-32BIT-NEXT:  .LBB10_8:
+; LE-32BIT-NEXT:    li 0, 0
+; LE-32BIT-NEXT:  .LBB10_9:
+; LE-32BIT-NEXT:    or 24, 14, 24
+; LE-32BIT-NEXT:    stw 0, 24(9)
+; LE-32BIT-NEXT:    srw 0, 6, 15
+; LE-32BIT-NEXT:    or 24, 0, 24
+; LE-32BIT-NEXT:    lwz 0, 56(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    slw 21, 10, 30
+; LE-32BIT-NEXT:    bc 12, 20, .LBB10_11
+; LE-32BIT-NEXT:  # %bb.10:
+; LE-32BIT-NEXT:    ori 7, 29, 0
+; LE-32BIT-NEXT:    b .LBB10_12
+; LE-32BIT-NEXT:  .LBB10_11:
+; LE-32BIT-NEXT:    addi 7, 20, 0
+; LE-32BIT-NEXT:  .LBB10_12:
+; LE-32BIT-NEXT:    or 4, 21, 4
+; LE-32BIT-NEXT:    slw 21, 11, 31
+; LE-32BIT-NEXT:    srw 20, 12, 15
+; LE-32BIT-NEXT:    cmplwi 7, 15, 64
+; LE-32BIT-NEXT:    li 15, 0
+; LE-32BIT-NEXT:    or 27, 21, 27
+; LE-32BIT-NEXT:    bc 12, 20, .LBB10_14
+; LE-32BIT-NEXT:  # %bb.13:
+; LE-32BIT-NEXT:    ori 21, 19, 0
+; LE-32BIT-NEXT:    b .LBB10_15
+; LE-32BIT-NEXT:  .LBB10_14:
+; LE-32BIT-NEXT:    addi 21, 18, 0
+; LE-32BIT-NEXT:  .LBB10_15:
+; LE-32BIT-NEXT:    mr 16, 9
+; LE-32BIT-NEXT:    lwz 9, 52(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    slw 18, 5, 0
+; LE-32BIT-NEXT:    bc 12, 28, .LBB10_17
+; LE-32BIT-NEXT:  # %bb.16:
+; LE-32BIT-NEXT:    ori 0, 15, 0
+; LE-32BIT-NEXT:    b .LBB10_18
+; LE-32BIT-NEXT:  .LBB10_17:
+; LE-32BIT-NEXT:    addi 0, 20, 0
+; LE-32BIT-NEXT:  .LBB10_18:
+; LE-32BIT-NEXT:    lwz 20, 60(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    slw 30, 8, 30
+; LE-32BIT-NEXT:    slw 19, 8, 9
+; LE-32BIT-NEXT:    slw 17, 5, 9
+; LE-32BIT-NEXT:    bc 12, 2, .LBB10_20
+; LE-32BIT-NEXT:  # %bb.19:
+; LE-32BIT-NEXT:    ori 9, 7, 0
+; LE-32BIT-NEXT:    b .LBB10_21
+; LE-32BIT-NEXT:  .LBB10_20:
+; LE-32BIT-NEXT:    addi 9, 20, 0
+; LE-32BIT-NEXT:  .LBB10_21:
+; LE-32BIT-NEXT:    lwz 7, 48(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    slw 29, 5, 31
+; LE-32BIT-NEXT:    or 9, 9, 0
+; LE-32BIT-NEXT:    bc 12, 20, .LBB10_23
+; LE-32BIT-NEXT:  # %bb.22:
+; LE-32BIT-NEXT:    ori 0, 15, 0
+; LE-32BIT-NEXT:    b .LBB10_24
+; LE-32BIT-NEXT:  .LBB10_23:
+; LE-32BIT-NEXT:    addi 0, 30, 0
+; LE-32BIT-NEXT:  .LBB10_24:
+; LE-32BIT-NEXT:    bc 12, 24, .LBB10_26
+; LE-32BIT-NEXT:  # %bb.25:
+; LE-32BIT-NEXT:    ori 30, 15, 0
+; LE-32BIT-NEXT:    b .LBB10_27
+; LE-32BIT-NEXT:  .LBB10_26:
+; LE-32BIT-NEXT:    addi 30, 29, 0
+; LE-32BIT-NEXT:  .LBB10_27:
+; LE-32BIT-NEXT:    bc 12, 28, .LBB10_28
+; LE-32BIT-NEXT:    b .LBB10_29
+; LE-32BIT-NEXT:  .LBB10_28:
+; LE-32BIT-NEXT:    addi 28, 26, 0
+; LE-32BIT-NEXT:  .LBB10_29:
+; LE-32BIT-NEXT:    bc 12, 20, .LBB10_31
+; LE-32BIT-NEXT:  # %bb.30:
+; LE-32BIT-NEXT:    ori 3, 17, 0
+; LE-32BIT-NEXT:    b .LBB10_31
+; LE-32BIT-NEXT:  .LBB10_31:
+; LE-32BIT-NEXT:    srw 22, 12, 22
+; LE-32BIT-NEXT:    bc 12, 20, .LBB10_33
+; LE-32BIT-NEXT:  # %bb.32:
+; LE-32BIT-NEXT:    ori 29, 15, 0
+; LE-32BIT-NEXT:    b .LBB10_34
+; LE-32BIT-NEXT:  .LBB10_33:
+; LE-32BIT-NEXT:    addi 29, 7, 0
+; LE-32BIT-NEXT:  .LBB10_34:
+; LE-32BIT-NEXT:    lwz 7, 44(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    bc 12, 20, .LBB10_36
+; LE-32BIT-NEXT:  # %bb.35:
+; LE-32BIT-NEXT:    ori 4, 19, 0
+; LE-32BIT-NEXT:    b .LBB10_36
+; LE-32BIT-NEXT:  .LBB10_36:
+; LE-32BIT-NEXT:    bc 12, 14, .LBB10_38
+; LE-32BIT-NEXT:  # %bb.37:
+; LE-32BIT-NEXT:    ori 5, 28, 0
+; LE-32BIT-NEXT:    b .LBB10_38
+; LE-32BIT-NEXT:  .LBB10_38:
+; LE-32BIT-NEXT:    li 28, 0
+; LE-32BIT-NEXT:    bc 12, 2, .LBB10_39
+; LE-32BIT-NEXT:    b .LBB10_40
+; LE-32BIT-NEXT:  .LBB10_39:
+; LE-32BIT-NEXT:    addi 3, 11, 0
+; LE-32BIT-NEXT:  .LBB10_40:
+; LE-32BIT-NEXT:    cmplwi 2, 31, 0
+; LE-32BIT-NEXT:    bc 12, 24, .LBB10_42
+; LE-32BIT-NEXT:  # %bb.41:
+; LE-32BIT-NEXT:    ori 27, 18, 0
+; LE-32BIT-NEXT:    b .LBB10_42
+; LE-32BIT-NEXT:  .LBB10_42:
+; LE-32BIT-NEXT:    bc 12, 28, .LBB10_44
+; LE-32BIT-NEXT:  # %bb.43:
+; LE-32BIT-NEXT:    ori 26, 22, 0
+; LE-32BIT-NEXT:    b .LBB10_45
+; LE-32BIT-NEXT:  .LBB10_44:
+; LE-32BIT-NEXT:    addi 26, 24, 0
+; LE-32BIT-NEXT:  .LBB10_45:
+; LE-32BIT-NEXT:    bc 12, 2, .LBB10_46
+; LE-32BIT-NEXT:    b .LBB10_47
+; LE-32BIT-NEXT:  .LBB10_46:
+; LE-32BIT-NEXT:    addi 4, 10, 0
+; LE-32BIT-NEXT:  .LBB10_47:
+; LE-32BIT-NEXT:    bc 12, 4, .LBB10_49
+; LE-32BIT-NEXT:  # %bb.48:
+; LE-32BIT-NEXT:    ori 3, 28, 0
+; LE-32BIT-NEXT:    b .LBB10_49
+; LE-32BIT-NEXT:  .LBB10_49:
+; LE-32BIT-NEXT:    bc 12, 10, .LBB10_50
+; LE-32BIT-NEXT:    b .LBB10_51
+; LE-32BIT-NEXT:  .LBB10_50:
+; LE-32BIT-NEXT:    addi 25, 12, 0
+; LE-32BIT-NEXT:  .LBB10_51:
+; LE-32BIT-NEXT:    or 5, 0, 5
+; LE-32BIT-NEXT:    bc 12, 24, .LBB10_53
+; LE-32BIT-NEXT:  # %bb.52:
+; LE-32BIT-NEXT:    ori 24, 15, 0
+; LE-32BIT-NEXT:    b .LBB10_54
+; LE-32BIT-NEXT:  .LBB10_53:
+; LE-32BIT-NEXT:    addi 24, 7, 0
+; LE-32BIT-NEXT:  .LBB10_54:
+; LE-32BIT-NEXT:    bc 12, 28, .LBB10_56
+; LE-32BIT-NEXT:  # %bb.55:
+; LE-32BIT-NEXT:    ori 7, 15, 0
+; LE-32BIT-NEXT:    b .LBB10_57
+; LE-32BIT-NEXT:  .LBB10_56:
+; LE-32BIT-NEXT:    addi 7, 23, 0
+; LE-32BIT-NEXT:  .LBB10_57:
+; LE-32BIT-NEXT:    bc 12, 10, .LBB10_58
+; LE-32BIT-NEXT:    b .LBB10_59
+; LE-32BIT-NEXT:  .LBB10_58:
+; LE-32BIT-NEXT:    addi 27, 11, 0
+; LE-32BIT-NEXT:  .LBB10_59:
+; LE-32BIT-NEXT:    stw 3, 20(16)
+; LE-32BIT-NEXT:    or 3, 4, 7
+; LE-32BIT-NEXT:    bc 12, 4, .LBB10_61
+; LE-32BIT-NEXT:  # %bb.60:
+; LE-32BIT-NEXT:    ori 3, 27, 0
+; LE-32BIT-NEXT:    ori 9, 25, 0
+; LE-32BIT-NEXT:    b .LBB10_61
+; LE-32BIT-NEXT:  .LBB10_61:
+; LE-32BIT-NEXT:    bc 12, 14, .LBB10_63
+; LE-32BIT-NEXT:  # %bb.62:
+; LE-32BIT-NEXT:    ori 6, 26, 0
+; LE-32BIT-NEXT:    b .LBB10_63
+; LE-32BIT-NEXT:  .LBB10_63:
+; LE-32BIT-NEXT:    bc 12, 2, .LBB10_65
+; LE-32BIT-NEXT:  # %bb.64:
+; LE-32BIT-NEXT:    ori 12, 21, 0
+; LE-32BIT-NEXT:    b .LBB10_65
+; LE-32BIT-NEXT:  .LBB10_65:
+; LE-32BIT-NEXT:    bc 12, 4, .LBB10_67
+; LE-32BIT-NEXT:  # %bb.66:
+; LE-32BIT-NEXT:    ori 5, 30, 0
+; LE-32BIT-NEXT:    b .LBB10_67
+; LE-32BIT-NEXT:  .LBB10_67:
+; LE-32BIT-NEXT:    bc 12, 2, .LBB10_69
+; LE-32BIT-NEXT:  # %bb.68:
+; LE-32BIT-NEXT:    ori 4, 9, 0
+; LE-32BIT-NEXT:    b .LBB10_70
+; LE-32BIT-NEXT:  .LBB10_69:
+; LE-32BIT-NEXT:    addi 3, 10, 0
+; LE-32BIT-NEXT:    addi 4, 20, 0
+; LE-32BIT-NEXT:  .LBB10_70:
+; LE-32BIT-NEXT:    bc 12, 4, .LBB10_72
+; LE-32BIT-NEXT:  # %bb.71:
+; LE-32BIT-NEXT:    ori 12, 15, 0
+; LE-32BIT-NEXT:    b .LBB10_72
+; LE-32BIT-NEXT:  .LBB10_72:
+; LE-32BIT-NEXT:    bc 12, 2, .LBB10_73
+; LE-32BIT-NEXT:    b .LBB10_74
+; LE-32BIT-NEXT:  .LBB10_73:
+; LE-32BIT-NEXT:    addi 5, 8, 0
+; LE-32BIT-NEXT:  .LBB10_74:
+; LE-32BIT-NEXT:    stw 3, 4(16)
+; LE-32BIT-NEXT:    lwz 3, 28(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    stw 4, 0(16)
+; LE-32BIT-NEXT:    or 4, 29, 6
+; LE-32BIT-NEXT:    bc 12, 4, .LBB10_76
+; LE-32BIT-NEXT:  # %bb.75:
+; LE-32BIT-NEXT:    ori 4, 24, 0
+; LE-32BIT-NEXT:    b .LBB10_76
+; LE-32BIT-NEXT:  .LBB10_76:
+; LE-32BIT-NEXT:    stw 12, 16(16)
+; LE-32BIT-NEXT:    bc 12, 2, .LBB10_78
+; LE-32BIT-NEXT:  # %bb.77:
+; LE-32BIT-NEXT:    ori 3, 4, 0
+; LE-32BIT-NEXT:    b .LBB10_78
+; LE-32BIT-NEXT:  .LBB10_78:
+; LE-32BIT-NEXT:    stw 5, 12(16)
+; LE-32BIT-NEXT:    stw 3, 8(16)
+; LE-32BIT-NEXT:    lwz 12, 68(1)
+; LE-32BIT-NEXT:    lwz 31, 140(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    mtcrf 32, 12 # cr2
+; LE-32BIT-NEXT:    mtcrf 16, 12 # cr3
+; LE-32BIT-NEXT:    lwz 30, 136(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 29, 132(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 28, 128(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 27, 124(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 26, 120(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 25, 116(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 24, 112(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 23, 108(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 22, 104(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 21, 100(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 20, 96(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 19, 92(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 18, 88(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 17, 84(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 16, 80(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 15, 76(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 14, 72(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    addi 1, 1, 144
+; LE-32BIT-NEXT:    blr
+  %src = load i256, ptr %src.ptr, align 1
+  %byteOff = load i256, ptr %byteOff.ptr, align 1
+  %bitOff = shl i256 %byteOff, 3
+  %res = shl i256 %src, %bitOff
+  store i256 %res, ptr %dst, align 1
+  ret void
+}
+define void @ashr_32bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; LE-64BIT-LABEL: ashr_32bytes:
+; LE-64BIT:       # %bb.0:
+; LE-64BIT-NEXT:    lwz 4, 0(4)
+; LE-64BIT-NEXT:    ld 6, 24(3)
+; LE-64BIT-NEXT:    ld 8, 16(3)
+; LE-64BIT-NEXT:    std 28, -32(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    std 29, -24(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    std 30, -16(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    std 26, -48(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    rlwinm. 4, 4, 3, 0, 28
+; LE-64BIT-NEXT:    sradi 9, 6, 63
+; LE-64BIT-NEXT:    subfic 10, 4, 192
+; LE-64BIT-NEXT:    addi 11, 4, -128
+; LE-64BIT-NEXT:    addi 30, 4, -192
+; LE-64BIT-NEXT:    sld 10, 6, 10
+; LE-64BIT-NEXT:    srd 29, 8, 11
+; LE-64BIT-NEXT:    subfic 28, 4, 64
+; LE-64BIT-NEXT:    std 27, -40(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    ld 7, 0(3)
+; LE-64BIT-NEXT:    ld 3, 8(3)
+; LE-64BIT-NEXT:    srd 0, 8, 4
+; LE-64BIT-NEXT:    srad 27, 6, 30
+; LE-64BIT-NEXT:    or 10, 29, 10
+; LE-64BIT-NEXT:    std 25, -56(1) # 8-byte Folded Spill
+; LE-64BIT-NEXT:    cmpwi 1, 30, 1
+; LE-64BIT-NEXT:    sld 26, 6, 28
+; LE-64BIT-NEXT:    addi 30, 4, -64
+; LE-64BIT-NEXT:    isel 10, 10, 27, 4
+; LE-64BIT-NEXT:    or 27, 0, 26
+; LE-64BIT-NEXT:    subfic 0, 4, 128
+; LE-64BIT-NEXT:    srd 12, 7, 4
+; LE-64BIT-NEXT:    sld 26, 3, 28
+; LE-64BIT-NEXT:    subfic 25, 0, 64
+; LE-64BIT-NEXT:    srad 29, 6, 30
+; LE-64BIT-NEXT:    cmpwi 1, 30, 1
+; LE-64BIT-NEXT:    or 12, 12, 26
+; LE-64BIT-NEXT:    srd 30, 3, 30
+; LE-64BIT-NEXT:    sld 28, 8, 28
+; LE-64BIT-NEXT:    srd 26, 8, 25
+; LE-64BIT-NEXT:    sld 8, 8, 0
+; LE-64BIT-NEXT:    or 12, 12, 30
+; LE-64BIT-NEXT:    ld 30, -16(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    ld 25, -56(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    sld 0, 6, 0
+; LE-64BIT-NEXT:    isel 29, 27, 29, 4
+; LE-64BIT-NEXT:    or 8, 12, 8
+; LE-64BIT-NEXT:    or 0, 0, 26
+; LE-64BIT-NEXT:    cmplwi 1, 4, 128
+; LE-64BIT-NEXT:    ld 26, -48(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    srd 27, 3, 4
+; LE-64BIT-NEXT:    or 0, 0, 28
+; LE-64BIT-NEXT:    ld 28, -32(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    srad 11, 6, 11
+; LE-64BIT-NEXT:    isel 8, 8, 10, 4
+; LE-64BIT-NEXT:    or 10, 27, 0
+; LE-64BIT-NEXT:    ld 27, -40(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    iseleq 7, 7, 8
+; LE-64BIT-NEXT:    srad 4, 6, 4
+; LE-64BIT-NEXT:    isel 8, 10, 11, 4
+; LE-64BIT-NEXT:    std 7, 0(5)
+; LE-64BIT-NEXT:    isel 12, 29, 9, 4
+; LE-64BIT-NEXT:    ld 29, -24(1) # 8-byte Folded Reload
+; LE-64BIT-NEXT:    iseleq 3, 3, 8
+; LE-64BIT-NEXT:    std 12, 16(5)
+; LE-64BIT-NEXT:    isel 4, 4, 9, 4
+; LE-64BIT-NEXT:    std 3, 8(5)
+; LE-64BIT-NEXT:    std 4, 24(5)
+; LE-64BIT-NEXT:    blr
+;
+; BE-LABEL: ashr_32bytes:
+; BE:       # %bb.0:
+; BE-NEXT:    lwz 4, 28(4)
+; BE-NEXT:    ld 6, 16(3)
+; BE-NEXT:    ld 7, 24(3)
+; BE-NEXT:    ld 8, 8(3)
+; BE-NEXT:    ld 3, 0(3)
+; BE-NEXT:    std 27, -40(1) # 8-byte Folded Spill
+; BE-NEXT:    std 29, -24(1) # 8-byte Folded Spill
+; BE-NEXT:    std 28, -32(1) # 8-byte Folded Spill
+; BE-NEXT:    std 30, -16(1) # 8-byte Folded Spill
+; BE-NEXT:    rlwinm. 4, 4, 3, 0, 28
+; BE-NEXT:    subfic 9, 4, 192
+; BE-NEXT:    addi 10, 4, -128
+; BE-NEXT:    addi 11, 4, -192
+; BE-NEXT:    subfic 0, 4, 64
+; BE-NEXT:    sld 9, 3, 9
+; BE-NEXT:    srd 27, 8, 10
+; BE-NEXT:    srd 12, 7, 4
+; BE-NEXT:    subfic 29, 4, 128
+; BE-NEXT:    cmpwi 1, 11, 1
+; BE-NEXT:    srad 11, 3, 11
+; BE-NEXT:    or 9, 27, 9
+; BE-NEXT:    sld 27, 6, 0
+; BE-NEXT:    addi 30, 4, -64
+; BE-NEXT:    srd 28, 8, 4
+; BE-NEXT:    or 12, 12, 27
+; BE-NEXT:    sld 27, 3, 0
+; BE-NEXT:    bc 12, 4, .LBB11_2
+; BE-NEXT:  # %bb.1:
+; BE-NEXT:    ori 9, 11, 0
+; BE-NEXT:    b .LBB11_2
+; BE-NEXT:  .LBB11_2:
+; BE-NEXT:    subfic 11, 29, 64
+; BE-NEXT:    or 28, 28, 27
+; BE-NEXT:    srd 27, 6, 30
+; BE-NEXT:    sld 0, 8, 0
+; BE-NEXT:    srd 11, 8, 11
+; BE-NEXT:    sld 8, 8, 29
+; BE-NEXT:    sld 29, 3, 29
+; BE-NEXT:    cmplwi 1, 4, 128
+; BE-NEXT:    or 12, 12, 27
+; BE-NEXT:    or 11, 29, 11
+; BE-NEXT:    or 8, 12, 8
+; BE-NEXT:    srd 12, 6, 4
+; BE-NEXT:    or 11, 11, 0
+; BE-NEXT:    srad 10, 3, 10
+; BE-NEXT:    srad 29, 3, 30
+; BE-NEXT:    or 11, 12, 11
+; BE-NEXT:    cmpwi 5, 30, 1
+; BE-NEXT:    bc 12, 20, .LBB11_4
+; BE-NEXT:  # %bb.3:
+; BE-NEXT:    ori 12, 29, 0
+; BE-NEXT:    b .LBB11_5
+; BE-NEXT:  .LBB11_4:
+; BE-NEXT:    addi 12, 28, 0
+; BE-NEXT:  .LBB11_5:
+; BE-NEXT:    bc 12, 4, .LBB11_7
+; BE-NEXT:  # %bb.6:
+; BE-NEXT:    ori 8, 9, 0
+; BE-NEXT:    ori 9, 10, 0
+; BE-NEXT:    b .LBB11_8
+; BE-NEXT:  .LBB11_7:
+; BE-NEXT:    addi 9, 11, 0
+; BE-NEXT:  .LBB11_8:
+; BE-NEXT:    sradi 10, 3, 63
+; BE-NEXT:    srad 3, 3, 4
+; BE-NEXT:    ld 30, -16(1) # 8-byte Folded Reload
+; BE-NEXT:    ld 29, -24(1) # 8-byte Folded Reload
+; BE-NEXT:    ld 28, -32(1) # 8-byte Folded Reload
+; BE-NEXT:    ld 27, -40(1) # 8-byte Folded Reload
+; BE-NEXT:    bc 12, 2, .LBB11_10
+; BE-NEXT:  # %bb.9:
+; BE-NEXT:    ori 4, 8, 0
+; BE-NEXT:    ori 6, 9, 0
+; BE-NEXT:    b .LBB11_11
+; BE-NEXT:  .LBB11_10:
+; BE-NEXT:    addi 4, 7, 0
+; BE-NEXT:  .LBB11_11:
+; BE-NEXT:    bc 12, 4, .LBB11_13
+; BE-NEXT:  # %bb.12:
+; BE-NEXT:    ori 7, 10, 0
+; BE-NEXT:    ori 3, 10, 0
+; BE-NEXT:    b .LBB11_14
+; BE-NEXT:  .LBB11_13:
+; BE-NEXT:    addi 7, 12, 0
+; BE-NEXT:  .LBB11_14:
+; BE-NEXT:    std 3, 0(5)
+; BE-NEXT:    std 7, 8(5)
+; BE-NEXT:    std 4, 24(5)
+; BE-NEXT:    std 6, 16(5)
+; BE-NEXT:    blr
+;
+; LE-32BIT-LABEL: ashr_32bytes:
+; LE-32BIT:       # %bb.0:
+; LE-32BIT-NEXT:    stwu 1, -144(1)
+; LE-32BIT-NEXT:    mfcr 12
+; LE-32BIT-NEXT:    stw 14, 72(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 15, 76(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 16, 80(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 17, 84(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 18, 88(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 19, 92(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 20, 96(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 21, 100(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 22, 104(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 23, 108(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 24, 112(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 25, 116(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 26, 120(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 27, 124(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 28, 128(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 29, 132(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 30, 136(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 31, 140(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    stw 12, 68(1)
+; LE-32BIT-NEXT:    lwz 0, 28(4)
+; LE-32BIT-NEXT:    lwz 29, 4(3)
+; LE-32BIT-NEXT:    lwz 12, 0(3)
+; LE-32BIT-NEXT:    rlwinm. 30, 0, 3, 0, 28
+; LE-32BIT-NEXT:    stw 5, 64(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    subfic 23, 30, 224
+; LE-32BIT-NEXT:    lwz 5, 24(3)
+; LE-32BIT-NEXT:    addi 21, 30, -224
+; LE-32BIT-NEXT:    lwz 8, 28(3)
+; LE-32BIT-NEXT:    subfic 4, 30, 160
+; LE-32BIT-NEXT:    lwz 10, 20(3)
+; LE-32BIT-NEXT:    addi 11, 30, -128
+; LE-32BIT-NEXT:    lwz 9, 16(3)
+; LE-32BIT-NEXT:    subfic 25, 30, 96
+; LE-32BIT-NEXT:    lwz 26, 12(3)
+; LE-32BIT-NEXT:    addi 0, 30, -64
+; LE-32BIT-NEXT:    lwz 7, 8(3)
+; LE-32BIT-NEXT:    addi 3, 30, -192
+; LE-32BIT-NEXT:    subfic 27, 30, 32
+; LE-32BIT-NEXT:    slw 23, 12, 23
+; LE-32BIT-NEXT:    srw 16, 29, 3
+; LE-32BIT-NEXT:    stw 3, 56(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    srw 20, 8, 30
+; LE-32BIT-NEXT:    sraw 15, 12, 21
+; LE-32BIT-NEXT:    cmpwi 1, 21, 1
+; LE-32BIT-NEXT:    slw 21, 7, 4
+; LE-32BIT-NEXT:    srw 14, 26, 11
+; LE-32BIT-NEXT:    slw 31, 9, 25
+; LE-32BIT-NEXT:    srw 3, 10, 0
+; LE-32BIT-NEXT:    or 23, 16, 23
+; LE-32BIT-NEXT:    slw 16, 5, 27
+; LE-32BIT-NEXT:    srw 19, 10, 30
+; LE-32BIT-NEXT:    or 21, 14, 21
+; LE-32BIT-NEXT:    slw 14, 9, 27
+; LE-32BIT-NEXT:    or 3, 3, 31
+; LE-32BIT-NEXT:    slw 31, 12, 4
+; LE-32BIT-NEXT:    or 20, 20, 16
+; LE-32BIT-NEXT:    srw 16, 29, 11
+; LE-32BIT-NEXT:    or 19, 19, 14
+; LE-32BIT-NEXT:    slw 14, 12, 25
+; LE-32BIT-NEXT:    or 16, 16, 31
+; LE-32BIT-NEXT:    srw 31, 29, 0
+; LE-32BIT-NEXT:    addi 24, 30, -160
+; LE-32BIT-NEXT:    srw 18, 26, 30
+; LE-32BIT-NEXT:    or 14, 31, 14
+; LE-32BIT-NEXT:    slw 31, 7, 27
+; LE-32BIT-NEXT:    addi 28, 30, -96
+; LE-32BIT-NEXT:    srw 17, 29, 30
+; LE-32BIT-NEXT:    stw 4, 32(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    or 18, 18, 31
+; LE-32BIT-NEXT:    slw 31, 12, 27
+; LE-32BIT-NEXT:    bc 12, 4, .LBB11_2
+; LE-32BIT-NEXT:  # %bb.1:
+; LE-32BIT-NEXT:    ori 4, 15, 0
+; LE-32BIT-NEXT:    b .LBB11_3
+; LE-32BIT-NEXT:  .LBB11_2:
+; LE-32BIT-NEXT:    addi 4, 23, 0
+; LE-32BIT-NEXT:  .LBB11_3:
+; LE-32BIT-NEXT:    srw 15, 7, 24
+; LE-32BIT-NEXT:    or 17, 17, 31
+; LE-32BIT-NEXT:    addi 31, 30, -32
+; LE-32BIT-NEXT:    or 21, 21, 15
+; LE-32BIT-NEXT:    srw 15, 9, 28
+; LE-32BIT-NEXT:    or 3, 3, 15
+; LE-32BIT-NEXT:    srw 15, 5, 31
+; LE-32BIT-NEXT:    or 20, 20, 15
+; LE-32BIT-NEXT:    srw 15, 9, 31
+; LE-32BIT-NEXT:    stw 3, 28(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    or 3, 19, 15
+; LE-32BIT-NEXT:    subfic 15, 30, 64
+; LE-32BIT-NEXT:    stw 4, 24(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    cmpwi 1, 24, 1
+; LE-32BIT-NEXT:    sraw 24, 12, 24
+; LE-32BIT-NEXT:    subfic 4, 15, 32
+; LE-32BIT-NEXT:    stw 0, 52(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    srw 0, 26, 4
+; LE-32BIT-NEXT:    stw 3, 48(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    bc 12, 4, .LBB11_5
+; LE-32BIT-NEXT:  # %bb.4:
+; LE-32BIT-NEXT:    ori 3, 24, 0
+; LE-32BIT-NEXT:    b .LBB11_6
+; LE-32BIT-NEXT:  .LBB11_5:
+; LE-32BIT-NEXT:    addi 3, 16, 0
+; LE-32BIT-NEXT:  .LBB11_6:
+; LE-32BIT-NEXT:    slw 16, 7, 15
+; LE-32BIT-NEXT:    or 0, 16, 0
+; LE-32BIT-NEXT:    subfic 16, 30, 128
+; LE-32BIT-NEXT:    stw 5, 36(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    subfic 5, 16, 32
+; LE-32BIT-NEXT:    stw 3, 44(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    slw 3, 12, 16
+; LE-32BIT-NEXT:    srw 22, 29, 5
+; LE-32BIT-NEXT:    stw 8, 60(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    mr 8, 10
+; LE-32BIT-NEXT:    mr 10, 27
+; LE-32BIT-NEXT:    or 23, 3, 22
+; LE-32BIT-NEXT:    slw 22, 7, 16
+; LE-32BIT-NEXT:    srw 27, 26, 5
+; LE-32BIT-NEXT:    stw 11, 40(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    mr 6, 26
+; LE-32BIT-NEXT:    or 11, 22, 27
+; LE-32BIT-NEXT:    slw 22, 9, 15
+; LE-32BIT-NEXT:    srw 26, 8, 4
+; LE-32BIT-NEXT:    subfic 3, 30, 192
+; LE-32BIT-NEXT:    or 26, 22, 26
+; LE-32BIT-NEXT:    cmpwi 1, 28, 1
+; LE-32BIT-NEXT:    sraw 22, 12, 28
+; LE-32BIT-NEXT:    subfic 19, 3, 32
+; LE-32BIT-NEXT:    srw 4, 29, 4
+; LE-32BIT-NEXT:    slw 28, 12, 15
+; LE-32BIT-NEXT:    stw 9, 20(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    srw 19, 29, 19
+; LE-32BIT-NEXT:    slw 24, 12, 3
+; LE-32BIT-NEXT:    or 9, 28, 4
+; LE-32BIT-NEXT:    lwz 4, 64(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    or 24, 24, 19
+; LE-32BIT-NEXT:    bc 12, 4, .LBB11_7
+; LE-32BIT-NEXT:    b .LBB11_8
+; LE-32BIT-NEXT:  .LBB11_7:
+; LE-32BIT-NEXT:    addi 22, 14, 0
+; LE-32BIT-NEXT:  .LBB11_8:
+; LE-32BIT-NEXT:    srw 19, 7, 31
+; LE-32BIT-NEXT:    cmplwi 5, 30, 64
+; LE-32BIT-NEXT:    cmplwi 1, 30, 128
+; LE-32BIT-NEXT:    slw 3, 29, 3
+; LE-32BIT-NEXT:    or 19, 18, 19
+; LE-32BIT-NEXT:    cmpwi 6, 31, 1
+; LE-32BIT-NEXT:    sraw 18, 12, 31
+; LE-32BIT-NEXT:    crand 28, 4, 20
+; LE-32BIT-NEXT:    srawi 14, 12, 31
+; LE-32BIT-NEXT:    sraw 31, 12, 30
+; LE-32BIT-NEXT:    or 3, 21, 3
+; LE-32BIT-NEXT:    slw 21, 8, 15
+; LE-32BIT-NEXT:    bc 12, 24, .LBB11_10
+; LE-32BIT-NEXT:  # %bb.9:
+; LE-32BIT-NEXT:    ori 28, 18, 0
+; LE-32BIT-NEXT:    b .LBB11_11
+; LE-32BIT-NEXT:  .LBB11_10:
+; LE-32BIT-NEXT:    addi 28, 17, 0
+; LE-32BIT-NEXT:  .LBB11_11:
+; LE-32BIT-NEXT:    bc 12, 28, .LBB11_13
+; LE-32BIT-NEXT:  # %bb.12:
+; LE-32BIT-NEXT:    ori 18, 14, 0
+; LE-32BIT-NEXT:    b .LBB11_14
+; LE-32BIT-NEXT:  .LBB11_13:
+; LE-32BIT-NEXT:    addi 18, 31, 0
+; LE-32BIT-NEXT:  .LBB11_14:
+; LE-32BIT-NEXT:    or 21, 20, 21
+; LE-32BIT-NEXT:    subfic 20, 16, 64
+; LE-32BIT-NEXT:    stw 18, 0(4)
+; LE-32BIT-NEXT:    subfic 18, 20, 32
+; LE-32BIT-NEXT:    slw 18, 7, 18
+; LE-32BIT-NEXT:    srw 17, 6, 20
+; LE-32BIT-NEXT:    or 18, 17, 18
+; LE-32BIT-NEXT:    slw 17, 6, 10
+; LE-32BIT-NEXT:    or 27, 0, 17
+; LE-32BIT-NEXT:    slw 0, 29, 25
+; LE-32BIT-NEXT:    mr 31, 8
+; LE-32BIT-NEXT:    or 8, 23, 0
+; LE-32BIT-NEXT:    slw 0, 6, 25
+; LE-32BIT-NEXT:    or 11, 11, 0
+; LE-32BIT-NEXT:    stw 11, 16(1) # 4-byte Folded Spill
+; LE-32BIT-NEXT:    slw 0, 31, 10
+; LE-32BIT-NEXT:    lwz 11, 32(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    or 0, 26, 0
+; LE-32BIT-NEXT:    slw 25, 29, 10
+; LE-32BIT-NEXT:    or 23, 9, 25
+; LE-32BIT-NEXT:    slw 26, 29, 11
+; LE-32BIT-NEXT:    or 26, 24, 26
+; LE-32BIT-NEXT:    slw 24, 29, 15
+; LE-32BIT-NEXT:    or 24, 19, 24
+; LE-32BIT-NEXT:    lwz 19, 40(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    srw 25, 7, 20
+; LE-32BIT-NEXT:    lwz 9, 24(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    or 8, 8, 25
+; LE-32BIT-NEXT:    cmplwi 6, 19, 64
+; LE-32BIT-NEXT:    srw 5, 7, 5
+; LE-32BIT-NEXT:    bc 12, 24, .LBB11_16
+; LE-32BIT-NEXT:  # %bb.15:
+; LE-32BIT-NEXT:    ori 3, 9, 0
+; LE-32BIT-NEXT:    b .LBB11_16
+; LE-32BIT-NEXT:  .LBB11_16:
+; LE-32BIT-NEXT:    lwz 9, 28(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    or 5, 18, 5
+; LE-32BIT-NEXT:    lwz 17, 20(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    mr 18, 4
+; LE-32BIT-NEXT:    bc 12, 20, .LBB11_18
+; LE-32BIT-NEXT:  # %bb.17:
+; LE-32BIT-NEXT:    ori 10, 9, 0
+; LE-32BIT-NEXT:    b .LBB11_19
+; LE-32BIT-NEXT:  .LBB11_18:
+; LE-32BIT-NEXT:    addi 10, 21, 0
+; LE-32BIT-NEXT:  .LBB11_19:
+; LE-32BIT-NEXT:    lwz 9, 36(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    bc 12, 20, .LBB11_21
+; LE-32BIT-NEXT:  # %bb.20:
+; LE-32BIT-NEXT:    ori 24, 22, 0
+; LE-32BIT-NEXT:    b .LBB11_21
+; LE-32BIT-NEXT:  .LBB11_21:
+; LE-32BIT-NEXT:    cmplwi 2, 19, 0
+; LE-32BIT-NEXT:    bc 12, 10, .LBB11_22
+; LE-32BIT-NEXT:    b .LBB11_23
+; LE-32BIT-NEXT:  .LBB11_22:
+; LE-32BIT-NEXT:    addi 3, 6, 0
+; LE-32BIT-NEXT:  .LBB11_23:
+; LE-32BIT-NEXT:    cmplwi 3, 16, 0
+; LE-32BIT-NEXT:    srw 25, 9, 30
+; LE-32BIT-NEXT:    or 25, 25, 0
+; LE-32BIT-NEXT:    srw 0, 7, 19
+; LE-32BIT-NEXT:    or 26, 0, 26
+; LE-32BIT-NEXT:    srw 0, 7, 30
+; LE-32BIT-NEXT:    or 11, 0, 23
+; LE-32BIT-NEXT:    bc 12, 28, .LBB11_25
+; LE-32BIT-NEXT:  # %bb.24:
+; LE-32BIT-NEXT:    ori 0, 14, 0
+; LE-32BIT-NEXT:    b .LBB11_26
+; LE-32BIT-NEXT:  .LBB11_25:
+; LE-32BIT-NEXT:    addi 0, 28, 0
+; LE-32BIT-NEXT:  .LBB11_26:
+; LE-32BIT-NEXT:    slw 28, 6, 16
+; LE-32BIT-NEXT:    stw 0, 4(4)
+; LE-32BIT-NEXT:    slw 0, 29, 16
+; LE-32BIT-NEXT:    lwz 4, 52(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    or 5, 0, 5
+; LE-32BIT-NEXT:    lwz 0, 56(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    cmplwi 7, 16, 64
+; LE-32BIT-NEXT:    slw 23, 6, 15
+; LE-32BIT-NEXT:    srw 22, 17, 4
+; LE-32BIT-NEXT:    li 15, 0
+; LE-32BIT-NEXT:    sraw 21, 12, 0
+; LE-32BIT-NEXT:    bc 12, 28, .LBB11_28
+; LE-32BIT-NEXT:  # %bb.27:
+; LE-32BIT-NEXT:    ori 0, 15, 0
+; LE-32BIT-NEXT:    b .LBB11_29
+; LE-32BIT-NEXT:  .LBB11_28:
+; LE-32BIT-NEXT:    addi 0, 28, 0
+; LE-32BIT-NEXT:  .LBB11_29:
+; LE-32BIT-NEXT:    bc 12, 20, .LBB11_31
+; LE-32BIT-NEXT:  # %bb.30:
+; LE-32BIT-NEXT:    ori 28, 22, 0
+; LE-32BIT-NEXT:    b .LBB11_32
+; LE-32BIT-NEXT:  .LBB11_31:
+; LE-32BIT-NEXT:    addi 28, 25, 0
+; LE-32BIT-NEXT:  .LBB11_32:
+; LE-32BIT-NEXT:    bc 12, 2, .LBB11_34
+; LE-32BIT-NEXT:  # %bb.33:
+; LE-32BIT-NEXT:    ori 22, 24, 0
+; LE-32BIT-NEXT:    b .LBB11_35
+; LE-32BIT-NEXT:  .LBB11_34:
+; LE-32BIT-NEXT:    addi 22, 6, 0
+; LE-32BIT-NEXT:  .LBB11_35:
+; LE-32BIT-NEXT:    lwz 6, 48(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    sraw 20, 12, 4
+; LE-32BIT-NEXT:    lwz 16, 60(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    bc 12, 20, .LBB11_37
+; LE-32BIT-NEXT:  # %bb.36:
+; LE-32BIT-NEXT:    ori 4, 20, 0
+; LE-32BIT-NEXT:    b .LBB11_38
+; LE-32BIT-NEXT:  .LBB11_37:
+; LE-32BIT-NEXT:    addi 4, 11, 0
+; LE-32BIT-NEXT:  .LBB11_38:
+; LE-32BIT-NEXT:    srw 30, 17, 30
+; LE-32BIT-NEXT:    bc 12, 20, .LBB11_40
+; LE-32BIT-NEXT:  # %bb.39:
+; LE-32BIT-NEXT:    ori 25, 15, 0
+; LE-32BIT-NEXT:    b .LBB11_41
+; LE-32BIT-NEXT:  .LBB11_40:
+; LE-32BIT-NEXT:    addi 25, 6, 0
+; LE-32BIT-NEXT:  .LBB11_41:
+; LE-32BIT-NEXT:    lwz 6, 44(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    bc 12, 28, .LBB11_43
+; LE-32BIT-NEXT:  # %bb.42:
+; LE-32BIT-NEXT:    ori 8, 27, 0
+; LE-32BIT-NEXT:    ori 5, 23, 0
+; LE-32BIT-NEXT:    b .LBB11_43
+; LE-32BIT-NEXT:  .LBB11_43:
+; LE-32BIT-NEXT:    bc 12, 2, .LBB11_44
+; LE-32BIT-NEXT:    b .LBB11_45
+; LE-32BIT-NEXT:  .LBB11_44:
+; LE-32BIT-NEXT:    addi 4, 7, 0
+; LE-32BIT-NEXT:  .LBB11_45:
+; LE-32BIT-NEXT:    sraw 19, 12, 19
+; LE-32BIT-NEXT:    bc 12, 2, .LBB11_46
+; LE-32BIT-NEXT:    b .LBB11_47
+; LE-32BIT-NEXT:  .LBB11_46:
+; LE-32BIT-NEXT:    addi 10, 16, 0
+; LE-32BIT-NEXT:  .LBB11_47:
+; LE-32BIT-NEXT:    bc 12, 24, .LBB11_49
+; LE-32BIT-NEXT:  # %bb.48:
+; LE-32BIT-NEXT:    ori 26, 21, 0
+; LE-32BIT-NEXT:    b .LBB11_49
+; LE-32BIT-NEXT:  .LBB11_49:
+; LE-32BIT-NEXT:    bc 12, 14, .LBB11_50
+; LE-32BIT-NEXT:    b .LBB11_51
+; LE-32BIT-NEXT:  .LBB11_50:
+; LE-32BIT-NEXT:    addi 5, 29, 0
+; LE-32BIT-NEXT:  .LBB11_51:
+; LE-32BIT-NEXT:    bc 12, 4, .LBB11_53
+; LE-32BIT-NEXT:  # %bb.52:
+; LE-32BIT-NEXT:    ori 4, 14, 0
+; LE-32BIT-NEXT:    b .LBB11_53
+; LE-32BIT-NEXT:  .LBB11_53:
+; LE-32BIT-NEXT:    or 10, 10, 0
+; LE-32BIT-NEXT:    bc 12, 24, .LBB11_55
+; LE-32BIT-NEXT:  # %bb.54:
+; LE-32BIT-NEXT:    ori 24, 14, 0
+; LE-32BIT-NEXT:    b .LBB11_56
+; LE-32BIT-NEXT:  .LBB11_55:
+; LE-32BIT-NEXT:    addi 24, 6, 0
+; LE-32BIT-NEXT:  .LBB11_56:
+; LE-32BIT-NEXT:    lwz 6, 16(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    bc 12, 4, .LBB11_57
+; LE-32BIT-NEXT:    b .LBB11_58
+; LE-32BIT-NEXT:  .LBB11_57:
+; LE-32BIT-NEXT:    addi 3, 10, 0
+; LE-32BIT-NEXT:  .LBB11_58:
+; LE-32BIT-NEXT:    bc 12, 20, .LBB11_60
+; LE-32BIT-NEXT:  # %bb.59:
+; LE-32BIT-NEXT:    ori 0, 15, 0
+; LE-32BIT-NEXT:    b .LBB11_61
+; LE-32BIT-NEXT:  .LBB11_60:
+; LE-32BIT-NEXT:    addi 0, 30, 0
+; LE-32BIT-NEXT:  .LBB11_61:
+; LE-32BIT-NEXT:    bc 12, 24, .LBB11_63
+; LE-32BIT-NEXT:  # %bb.62:
+; LE-32BIT-NEXT:    ori 30, 14, 0
+; LE-32BIT-NEXT:    b .LBB11_64
+; LE-32BIT-NEXT:  .LBB11_63:
+; LE-32BIT-NEXT:    addi 30, 19, 0
+; LE-32BIT-NEXT:  .LBB11_64:
+; LE-32BIT-NEXT:    bc 12, 2, .LBB11_65
+; LE-32BIT-NEXT:    b .LBB11_66
+; LE-32BIT-NEXT:  .LBB11_65:
+; LE-32BIT-NEXT:    addi 3, 16, 0
+; LE-32BIT-NEXT:  .LBB11_66:
+; LE-32BIT-NEXT:    stw 4, 8(18)
+; LE-32BIT-NEXT:    bc 12, 28, .LBB11_68
+; LE-32BIT-NEXT:  # %bb.67:
+; LE-32BIT-NEXT:    ori 27, 15, 0
+; LE-32BIT-NEXT:    b .LBB11_69
+; LE-32BIT-NEXT:  .LBB11_68:
+; LE-32BIT-NEXT:    addi 27, 6, 0
+; LE-32BIT-NEXT:  .LBB11_69:
+; LE-32BIT-NEXT:    bc 12, 14, .LBB11_71
+; LE-32BIT-NEXT:  # %bb.70:
+; LE-32BIT-NEXT:    ori 6, 8, 0
+; LE-32BIT-NEXT:    b .LBB11_72
+; LE-32BIT-NEXT:  .LBB11_71:
+; LE-32BIT-NEXT:    addi 6, 12, 0
+; LE-32BIT-NEXT:  .LBB11_72:
+; LE-32BIT-NEXT:    bc 12, 2, .LBB11_74
+; LE-32BIT-NEXT:  # %bb.73:
+; LE-32BIT-NEXT:    ori 8, 28, 0
+; LE-32BIT-NEXT:    b .LBB11_75
+; LE-32BIT-NEXT:  .LBB11_74:
+; LE-32BIT-NEXT:    addi 8, 9, 0
+; LE-32BIT-NEXT:  .LBB11_75:
+; LE-32BIT-NEXT:    bc 12, 10, .LBB11_77
+; LE-32BIT-NEXT:  # %bb.76:
+; LE-32BIT-NEXT:    ori 28, 26, 0
+; LE-32BIT-NEXT:    b .LBB11_78
+; LE-32BIT-NEXT:  .LBB11_77:
+; LE-32BIT-NEXT:    addi 28, 7, 0
+; LE-32BIT-NEXT:  .LBB11_78:
+; LE-32BIT-NEXT:    stw 3, 28(18)
+; LE-32BIT-NEXT:    or 7, 8, 27
+; LE-32BIT-NEXT:    or 4, 0, 6
+; LE-32BIT-NEXT:    or 3, 25, 5
+; LE-32BIT-NEXT:    bc 12, 4, .LBB11_80
+; LE-32BIT-NEXT:  # %bb.79:
+; LE-32BIT-NEXT:    ori 6, 28, 0
+; LE-32BIT-NEXT:    ori 4, 30, 0
+; LE-32BIT-NEXT:    ori 3, 24, 0
+; LE-32BIT-NEXT:    ori 12, 14, 0
+; LE-32BIT-NEXT:    b .LBB11_81
+; LE-32BIT-NEXT:  .LBB11_80:
+; LE-32BIT-NEXT:    addi 6, 7, 0
+; LE-32BIT-NEXT:    addi 12, 22, 0
+; LE-32BIT-NEXT:  .LBB11_81:
+; LE-32BIT-NEXT:    bc 12, 2, .LBB11_83
+; LE-32BIT-NEXT:  # %bb.82:
+; LE-32BIT-NEXT:    ori 5, 6, 0
+; LE-32BIT-NEXT:    b .LBB11_84
+; LE-32BIT-NEXT:  .LBB11_83:
+; LE-32BIT-NEXT:    addi 5, 9, 0
+; LE-32BIT-NEXT:    addi 4, 17, 0
+; LE-32BIT-NEXT:    addi 3, 31, 0
+; LE-32BIT-NEXT:  .LBB11_84:
+; LE-32BIT-NEXT:    stw 12, 12(18)
+; LE-32BIT-NEXT:    stw 5, 24(18)
+; LE-32BIT-NEXT:    stw 4, 16(18)
+; LE-32BIT-NEXT:    stw 3, 20(18)
+; LE-32BIT-NEXT:    lwz 12, 68(1)
+; LE-32BIT-NEXT:    lwz 31, 140(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    mtcrf 32, 12 # cr2
+; LE-32BIT-NEXT:    mtcrf 16, 12 # cr3
+; LE-32BIT-NEXT:    lwz 30, 136(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 29, 132(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 28, 128(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 27, 124(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 26, 120(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 25, 116(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 24, 112(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 23, 108(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 22, 104(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 21, 100(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 20, 96(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 19, 92(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 18, 88(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 17, 84(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 16, 80(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 15, 76(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    lwz 14, 72(1) # 4-byte Folded Reload
+; LE-32BIT-NEXT:    addi 1, 1, 144
+; LE-32BIT-NEXT:    blr
+  %src = load i256, ptr %src.ptr, align 1
+  %byteOff = load i256, ptr %byteOff.ptr, align 1
+  %bitOff = shl i256 %byteOff, 3
+  %res = ashr i256 %src, %bitOff
+  store i256 %res, ptr %dst, align 1
+  ret void
+}
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; LE: {{.*}}

diff  --git a/llvm/test/CodeGen/X86/wide-scalar-shift-legalization.ll b/llvm/test/CodeGen/X86/wide-scalar-shift-legalization.ll
new file mode 100644
index 0000000000000..c6d3a5f7a90bd
--- /dev/null
+++ b/llvm/test/CodeGen/X86/wide-scalar-shift-legalization.ll
@@ -0,0 +1,8030 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mattr=+sse2,-bmi2,+slow-shld | FileCheck %s --check-prefixes=ALL,X64,X64-NO-BMI2,X64-NO-SHLD,X64-NO-BMI2-NO-SHLD
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mattr=+sse2,-bmi2,-slow-shld | FileCheck %s --check-prefixes=ALL,X64,X64-NO-BMI2,X64-SHLD,X64-NO-BMI2-HAVE-SHLD
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mattr=+sse2,+bmi2,+slow-shld | FileCheck %s --check-prefixes=ALL,X64,X64-BMI2,X64-NO-SHLD,X64-HAVE-BMI2-NO-SHLD
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mattr=+sse2,+bmi2,-slow-shld | FileCheck %s --check-prefixes=ALL,X64,X64-BMI2,X64-SHLD,X64-HAVE-BMI2-HAVE-SHLD
+; RUN: llc < %s -mtriple=i686-unknown-linux-gnu -mattr=+sse2,-bmi2,+slow-shld | FileCheck %s --check-prefixes=ALL,X32,X32-NO-BMI2,X32-NO-SHLD,X32-NO-BMI2-NO-SHLD
+; RUN: llc < %s -mtriple=i686-unknown-linux-gnu -mattr=+sse2,-bmi2,-slow-shld | FileCheck %s --check-prefixes=ALL,X32,X32-NO-BMI2,X32-SHLD,X32-NO-BMI2-HAVE-SHLD
+; RUN: llc < %s -mtriple=i686-unknown-linux-gnu -mattr=+sse2,+bmi2,+slow-shld | FileCheck %s --check-prefixes=ALL,X32,X32-BMI2,X32-NO-SHLD,X32-HAVE-BMI2-NO-SHLD
+; RUN: llc < %s -mtriple=i686-unknown-linux-gnu -mattr=+sse2,+bmi2,-slow-shld | FileCheck %s --check-prefixes=ALL,X32,X32-BMI2,X32-SHLD,X32-HAVE-BMI2-HAVE-SHLD
+
+define void @lshr_4bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; X64-NO-BMI2-LABEL: lshr_4bytes:
+; X64-NO-BMI2:       # %bb.0:
+; X64-NO-BMI2-NEXT:    movl (%rdi), %eax
+; X64-NO-BMI2-NEXT:    movzbl (%rsi), %ecx
+; X64-NO-BMI2-NEXT:    shlb $3, %cl
+; X64-NO-BMI2-NEXT:    shrl %cl, %eax
+; X64-NO-BMI2-NEXT:    movl %eax, (%rdx)
+; X64-NO-BMI2-NEXT:    retq
+;
+; X64-BMI2-LABEL: lshr_4bytes:
+; X64-BMI2:       # %bb.0:
+; X64-BMI2-NEXT:    movzbl (%rsi), %eax
+; X64-BMI2-NEXT:    shlb $3, %al
+; X64-BMI2-NEXT:    shrxl %eax, (%rdi), %eax
+; X64-BMI2-NEXT:    movl %eax, (%rdx)
+; X64-BMI2-NEXT:    retq
+;
+; X32-NO-BMI2-LABEL: lshr_4bytes:
+; X32-NO-BMI2:       # %bb.0:
+; X32-NO-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X32-NO-BMI2-NEXT:    movl (%edx), %edx
+; X32-NO-BMI2-NEXT:    movzbl (%ecx), %ecx
+; X32-NO-BMI2-NEXT:    shlb $3, %cl
+; X32-NO-BMI2-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-NEXT:    movl %edx, (%eax)
+; X32-NO-BMI2-NEXT:    retl
+;
+; X32-BMI2-LABEL: lshr_4bytes:
+; X32-BMI2:       # %bb.0:
+; X32-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X32-BMI2-NEXT:    movzbl (%edx), %edx
+; X32-BMI2-NEXT:    shlb $3, %dl
+; X32-BMI2-NEXT:    shrxl %edx, (%ecx), %ecx
+; X32-BMI2-NEXT:    movl %ecx, (%eax)
+; X32-BMI2-NEXT:    retl
+  %src = load i32, ptr %src.ptr, align 1
+  %byteOff = load i32, ptr %byteOff.ptr, align 1
+  %bitOff = shl i32 %byteOff, 3
+  %res = lshr i32 %src, %bitOff
+  store i32 %res, ptr %dst, align 1
+  ret void
+}
+define void @shl_4bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; X64-NO-BMI2-LABEL: shl_4bytes:
+; X64-NO-BMI2:       # %bb.0:
+; X64-NO-BMI2-NEXT:    movl (%rdi), %eax
+; X64-NO-BMI2-NEXT:    movzbl (%rsi), %ecx
+; X64-NO-BMI2-NEXT:    shlb $3, %cl
+; X64-NO-BMI2-NEXT:    shll %cl, %eax
+; X64-NO-BMI2-NEXT:    movl %eax, (%rdx)
+; X64-NO-BMI2-NEXT:    retq
+;
+; X64-BMI2-LABEL: shl_4bytes:
+; X64-BMI2:       # %bb.0:
+; X64-BMI2-NEXT:    movzbl (%rsi), %eax
+; X64-BMI2-NEXT:    shlb $3, %al
+; X64-BMI2-NEXT:    shlxl %eax, (%rdi), %eax
+; X64-BMI2-NEXT:    movl %eax, (%rdx)
+; X64-BMI2-NEXT:    retq
+;
+; X32-NO-BMI2-LABEL: shl_4bytes:
+; X32-NO-BMI2:       # %bb.0:
+; X32-NO-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X32-NO-BMI2-NEXT:    movl (%edx), %edx
+; X32-NO-BMI2-NEXT:    movzbl (%ecx), %ecx
+; X32-NO-BMI2-NEXT:    shlb $3, %cl
+; X32-NO-BMI2-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-NEXT:    movl %edx, (%eax)
+; X32-NO-BMI2-NEXT:    retl
+;
+; X32-BMI2-LABEL: shl_4bytes:
+; X32-BMI2:       # %bb.0:
+; X32-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X32-BMI2-NEXT:    movzbl (%edx), %edx
+; X32-BMI2-NEXT:    shlb $3, %dl
+; X32-BMI2-NEXT:    shlxl %edx, (%ecx), %ecx
+; X32-BMI2-NEXT:    movl %ecx, (%eax)
+; X32-BMI2-NEXT:    retl
+  %src = load i32, ptr %src.ptr, align 1
+  %byteOff = load i32, ptr %byteOff.ptr, align 1
+  %bitOff = shl i32 %byteOff, 3
+  %res = shl i32 %src, %bitOff
+  store i32 %res, ptr %dst, align 1
+  ret void
+}
+define void @ashr_4bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; X64-NO-BMI2-LABEL: ashr_4bytes:
+; X64-NO-BMI2:       # %bb.0:
+; X64-NO-BMI2-NEXT:    movl (%rdi), %eax
+; X64-NO-BMI2-NEXT:    movzbl (%rsi), %ecx
+; X64-NO-BMI2-NEXT:    shlb $3, %cl
+; X64-NO-BMI2-NEXT:    sarl %cl, %eax
+; X64-NO-BMI2-NEXT:    movl %eax, (%rdx)
+; X64-NO-BMI2-NEXT:    retq
+;
+; X64-BMI2-LABEL: ashr_4bytes:
+; X64-BMI2:       # %bb.0:
+; X64-BMI2-NEXT:    movzbl (%rsi), %eax
+; X64-BMI2-NEXT:    shlb $3, %al
+; X64-BMI2-NEXT:    sarxl %eax, (%rdi), %eax
+; X64-BMI2-NEXT:    movl %eax, (%rdx)
+; X64-BMI2-NEXT:    retq
+;
+; X32-NO-BMI2-LABEL: ashr_4bytes:
+; X32-NO-BMI2:       # %bb.0:
+; X32-NO-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X32-NO-BMI2-NEXT:    movl (%edx), %edx
+; X32-NO-BMI2-NEXT:    movzbl (%ecx), %ecx
+; X32-NO-BMI2-NEXT:    shlb $3, %cl
+; X32-NO-BMI2-NEXT:    sarl %cl, %edx
+; X32-NO-BMI2-NEXT:    movl %edx, (%eax)
+; X32-NO-BMI2-NEXT:    retl
+;
+; X32-BMI2-LABEL: ashr_4bytes:
+; X32-BMI2:       # %bb.0:
+; X32-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-BMI2-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X32-BMI2-NEXT:    movzbl (%edx), %edx
+; X32-BMI2-NEXT:    shlb $3, %dl
+; X32-BMI2-NEXT:    sarxl %edx, (%ecx), %ecx
+; X32-BMI2-NEXT:    movl %ecx, (%eax)
+; X32-BMI2-NEXT:    retl
+  %src = load i32, ptr %src.ptr, align 1
+  %byteOff = load i32, ptr %byteOff.ptr, align 1
+  %bitOff = shl i32 %byteOff, 3
+  %res = ashr i32 %src, %bitOff
+  store i32 %res, ptr %dst, align 1
+  ret void
+}
+
+define void @lshr_8bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; X64-NO-BMI2-LABEL: lshr_8bytes:
+; X64-NO-BMI2:       # %bb.0:
+; X64-NO-BMI2-NEXT:    movq (%rdi), %rax
+; X64-NO-BMI2-NEXT:    movzbl (%rsi), %ecx
+; X64-NO-BMI2-NEXT:    shlb $3, %cl
+; X64-NO-BMI2-NEXT:    shrq %cl, %rax
+; X64-NO-BMI2-NEXT:    movq %rax, (%rdx)
+; X64-NO-BMI2-NEXT:    retq
+;
+; X64-BMI2-LABEL: lshr_8bytes:
+; X64-BMI2:       # %bb.0:
+; X64-BMI2-NEXT:    movzbl (%rsi), %eax
+; X64-BMI2-NEXT:    shlb $3, %al
+; X64-BMI2-NEXT:    shrxq %rax, (%rdi), %rax
+; X64-BMI2-NEXT:    movq %rax, (%rdx)
+; X64-BMI2-NEXT:    retq
+;
+; X32-NO-BMI2-NO-SHLD-LABEL: lshr_8bytes:
+; X32-NO-BMI2-NO-SHLD:       # %bb.0:
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl (%ecx), %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 4(%ecx), %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl (%eax), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    shlb $3, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    leal (%esi,%esi), %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %ebx, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    xorl %ecx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel %esi, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, 4(%edx)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, (%edx)
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    retl
+;
+; X32-NO-BMI2-HAVE-SHLD-LABEL: lshr_8bytes:
+; X32-NO-BMI2-HAVE-SHLD:       # %bb.0:
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl (%esi), %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 4(%esi), %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movzbl (%ecx), %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shlb $3, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %esi, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    xorl %esi, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel %edi, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, 4(%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, (%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    retl
+;
+; X32-HAVE-BMI2-NO-SHLD-LABEL: lshr_8bytes:
+; X32-HAVE-BMI2-NO-SHLD:       # %bb.0:
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 4(%edx), %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movzbl (%ecx), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlb $3, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ecx, (%edx), %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    leal (%esi,%esi), %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ebx, %edi, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %edx, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ecx, %esi, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    xorl %esi, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edx, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel %edx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, 4(%eax)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, (%eax)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    retl
+;
+; X32-HAVE-BMI2-HAVE-SHLD-LABEL: lshr_8bytes:
+; X32-HAVE-BMI2-HAVE-SHLD:       # %bb.0:
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl (%esi), %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 4(%esi), %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movzbl (%ecx), %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlb $3, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %esi, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %ecx, %esi, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    xorl %edi, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel %esi, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, 4(%eax)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, (%eax)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    retl
+  %src = load i64, ptr %src.ptr, align 1
+  %byteOff = load i64, ptr %byteOff.ptr, align 1
+  %bitOff = shl i64 %byteOff, 3
+  %res = lshr i64 %src, %bitOff
+  store i64 %res, ptr %dst, align 1
+  ret void
+}
+define void @shl_8bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; X64-NO-BMI2-LABEL: shl_8bytes:
+; X64-NO-BMI2:       # %bb.0:
+; X64-NO-BMI2-NEXT:    movq (%rdi), %rax
+; X64-NO-BMI2-NEXT:    movzbl (%rsi), %ecx
+; X64-NO-BMI2-NEXT:    shlb $3, %cl
+; X64-NO-BMI2-NEXT:    shlq %cl, %rax
+; X64-NO-BMI2-NEXT:    movq %rax, (%rdx)
+; X64-NO-BMI2-NEXT:    retq
+;
+; X64-BMI2-LABEL: shl_8bytes:
+; X64-BMI2:       # %bb.0:
+; X64-BMI2-NEXT:    movzbl (%rsi), %eax
+; X64-BMI2-NEXT:    shlb $3, %al
+; X64-BMI2-NEXT:    shlxq %rax, (%rdi), %rax
+; X64-BMI2-NEXT:    movq %rax, (%rdx)
+; X64-BMI2-NEXT:    retq
+;
+; X32-NO-BMI2-NO-SHLD-LABEL: shl_8bytes:
+; X32-NO-BMI2-NO-SHLD:       # %bb.0:
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl (%ecx), %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 4(%ecx), %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl (%eax), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    shlb $3, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %ebx, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    xorl %ecx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel %esi, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, (%edx)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, 4(%edx)
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    retl
+;
+; X32-NO-BMI2-HAVE-SHLD-LABEL: shl_8bytes:
+; X32-NO-BMI2-HAVE-SHLD:       # %bb.0:
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl (%edx), %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 4(%edx), %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movzbl (%ecx), %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shlb $3, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %esi, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    xorl %esi, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel %edi, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, 4(%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, (%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    retl
+;
+; X32-HAVE-BMI2-NO-SHLD-LABEL: shl_8bytes:
+; X32-HAVE-BMI2-NO-SHLD:       # %bb.0:
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl (%edx), %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movzbl (%ecx), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlb $3, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ecx, 4(%edx), %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ecx, %esi, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrl %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %esi, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %edx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    xorl %edx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel %edi, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, (%eax)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, 4(%eax)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    retl
+;
+; X32-HAVE-BMI2-HAVE-SHLD-LABEL: shl_8bytes:
+; X32-HAVE-BMI2-HAVE-SHLD:       # %bb.0:
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl (%edx), %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 4(%edx), %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movzbl (%ecx), %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlb $3, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %esi, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %ecx, %esi, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    xorl %edi, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel %esi, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, 4(%eax)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, (%eax)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    retl
+  %src = load i64, ptr %src.ptr, align 1
+  %byteOff = load i64, ptr %byteOff.ptr, align 1
+  %bitOff = shl i64 %byteOff, 3
+  %res = shl i64 %src, %bitOff
+  store i64 %res, ptr %dst, align 1
+  ret void
+}
+define void @ashr_8bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; X64-NO-BMI2-LABEL: ashr_8bytes:
+; X64-NO-BMI2:       # %bb.0:
+; X64-NO-BMI2-NEXT:    movq (%rdi), %rax
+; X64-NO-BMI2-NEXT:    movzbl (%rsi), %ecx
+; X64-NO-BMI2-NEXT:    shlb $3, %cl
+; X64-NO-BMI2-NEXT:    sarq %cl, %rax
+; X64-NO-BMI2-NEXT:    movq %rax, (%rdx)
+; X64-NO-BMI2-NEXT:    retq
+;
+; X64-BMI2-LABEL: ashr_8bytes:
+; X64-BMI2:       # %bb.0:
+; X64-BMI2-NEXT:    movzbl (%rsi), %eax
+; X64-BMI2-NEXT:    shlb $3, %al
+; X64-BMI2-NEXT:    sarxq %rax, (%rdi), %rax
+; X64-BMI2-NEXT:    movq %rax, (%rdx)
+; X64-BMI2-NEXT:    retq
+;
+; X32-NO-BMI2-NO-SHLD-LABEL: ashr_8bytes:
+; X32-NO-BMI2-NO-SHLD:       # %bb.0:
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl (%ecx), %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 4(%ecx), %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl (%eax), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    shlb $3, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    leal (%esi,%esi), %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %ebx, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    sarl %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    sarl $31, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %ebx, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel %ebx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, 4(%edx)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, (%edx)
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    retl
+;
+; X32-NO-BMI2-HAVE-SHLD-LABEL: ashr_8bytes:
+; X32-NO-BMI2-HAVE-SHLD:       # %bb.0:
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl (%esi), %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 4(%esi), %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movzbl (%ecx), %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shlb $3, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    sarl %cl, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %esi, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    sarl $31, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel %edi, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, 4(%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, (%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    retl
+;
+; X32-HAVE-BMI2-NO-SHLD-LABEL: ashr_8bytes:
+; X32-HAVE-BMI2-NO-SHLD:       # %bb.0:
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 4(%esi), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movzbl (%edx), %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlb $3, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %edx, (%esi), %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    leal (%ecx,%ecx), %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ebx, %edi, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %esi, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    sarxl %edx, %ecx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    sarl $31, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel %esi, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, 4(%eax)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, (%eax)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    retl
+;
+; X32-HAVE-BMI2-HAVE-SHLD-LABEL: ashr_8bytes:
+; X32-HAVE-BMI2-HAVE-SHLD:       # %bb.0:
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl (%esi), %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 4(%esi), %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movzbl (%ecx), %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlb $3, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %esi, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    sarxl %ecx, %esi, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    sarl $31, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel %edi, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, 4(%eax)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, (%eax)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    retl
+  %src = load i64, ptr %src.ptr, align 1
+  %byteOff = load i64, ptr %byteOff.ptr, align 1
+  %bitOff = shl i64 %byteOff, 3
+  %res = ashr i64 %src, %bitOff
+  store i64 %res, ptr %dst, align 1
+  ret void
+}
+
+define void @lshr_16bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; X64-NO-BMI2-NO-SHLD-LABEL: lshr_16bytes:
+; X64-NO-BMI2-NO-SHLD:       # %bb.0:
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq (%rdi), %r8
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq 8(%rdi), %rdi
+; X64-NO-BMI2-NO-SHLD-NEXT:    movzbl (%rsi), %eax
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlb $3, %al
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %cl, %r8
+; X64-NO-BMI2-NO-SHLD-NEXT:    leaq (%rdi,%rdi), %rsi
+; X64-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlq %cl, %rsi
+; X64-NO-BMI2-NO-SHLD-NEXT:    orq %r8, %rsi
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %cl, %rdi
+; X64-NO-BMI2-NO-SHLD-NEXT:    xorl %ecx, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    testb $64, %al
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %rdi, %rsi
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmoveq %rdi, %rcx
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rcx, 8(%rdx)
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rsi, (%rdx)
+; X64-NO-BMI2-NO-SHLD-NEXT:    retq
+;
+; X64-NO-BMI2-HAVE-SHLD-LABEL: lshr_16bytes:
+; X64-NO-BMI2-HAVE-SHLD:       # %bb.0:
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq (%rdi), %rax
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq 8(%rdi), %rdi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movzbl (%rsi), %ecx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shlb $3, %cl
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %rdi, %rsi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shrq %cl, %rsi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shrdq %cl, %rdi, %rax
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    xorl %edi, %edi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    testb $64, %cl
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %rsi, %rax
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmoveq %rsi, %rdi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %rdi, 8(%rdx)
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %rax, (%rdx)
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    retq
+;
+; X64-HAVE-BMI2-NO-SHLD-LABEL: lshr_16bytes:
+; X64-HAVE-BMI2-NO-SHLD:       # %bb.0:
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq 8(%rdi), %rax
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movzbl (%rsi), %ecx
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlb $3, %cl
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrxq %rcx, (%rdi), %rsi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %edi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    notb %dil
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    leaq (%rax,%rax), %r8
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlxq %rdi, %r8, %rdi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    orq %rsi, %rdi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrxq %rcx, %rax, %rax
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    xorl %esi, %esi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    testb $64, %cl
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %rax, %rdi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmoveq %rax, %rsi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %rsi, 8(%rdx)
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %rdi, (%rdx)
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    retq
+;
+; X64-HAVE-BMI2-HAVE-SHLD-LABEL: lshr_16bytes:
+; X64-HAVE-BMI2-HAVE-SHLD:       # %bb.0:
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq (%rdi), %rax
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq 8(%rdi), %rdi
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movzbl (%rsi), %ecx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shlb $3, %cl
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdq %cl, %rdi, %rax
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxq %rcx, %rdi, %rsi
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    xorl %edi, %edi
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $64, %cl
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %rsi, %rax
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmoveq %rsi, %rdi
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %rdi, 8(%rdx)
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %rax, (%rdx)
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    retq
+;
+; X32-NO-BMI2-NO-SHLD-LABEL: lshr_16bytes:
+; X32-NO-BMI2-NO-SHLD:       # %bb.0:
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    subl $32, %esp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 12(%ecx), %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 8(%ecx), %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl (%eax), %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shlb $3, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    leal (%edx,%edx), %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, (%esp) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bl, %bh
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %bh
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bh, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %eax, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 4(%eax), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %edx, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %ecx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %ecx, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    subb $64, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %ecx, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    decb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    negb %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    andb $24, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl (%esp), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %ebp, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %esi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bl, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    addb $-64, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl (%esp), %edi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, (%esp) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovbl %edx, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %esi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl (%eax), %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bl, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    addl %eax, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bh, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %esi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel (%esp), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %ebp, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %edx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %bl, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, 12(%ecx)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, 8(%ecx)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, (%ecx)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, 4(%ecx)
+; X32-NO-BMI2-NO-SHLD-NEXT:    addl $32, %esp
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    retl
+;
+; X32-NO-BMI2-HAVE-SHLD-LABEL: lshr_16bytes:
+; X32-NO-BMI2-HAVE-SHLD:       # %bb.0:
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    subl $32, %esp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 4(%ecx), %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 12(%ecx), %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 8(%ecx), %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movzbl (%eax), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shlb $3, %al
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %ebx, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebx, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, (%esp) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ecx, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ecx, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    subb $64, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %ecx, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    negb %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %ebp, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    andb $24, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl %edi, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %al, %ah
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    addb $-64, %ah
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ah, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ah
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ecx, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %al
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovbl %esi, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl (%esp), %ecx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %esi, %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ecx, (%esp) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl (%ecx), %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %esi, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ah, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %ebx, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ah
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl %ebp, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %al
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %esi, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %al, %al
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl (%esp), %ecx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ecx, 8(%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ecx, 12(%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, (%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, 4(%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    addl $32, %esp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    retl
+;
+; X32-HAVE-BMI2-NO-SHLD-LABEL: lshr_16bytes:
+; X32-HAVE-BMI2-NO-SHLD:       # %bb.0:
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    subl $36, %esp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 12(%ecx), %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 8(%ecx), %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movzbl (%eax), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlb $3, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ecx, %edx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    leal (%ebp,%ebp), %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ebx, %edx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %eax, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ecx, %ebp, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, (%esp) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 4(%esi), %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ecx, %edx, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    subb $64, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %esi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    decb %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrl %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %eax, %edi, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    negb %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    andb $24, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ecx, %ebp, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ecx, %esi, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ebx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    addb $-64, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %ebp, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ebp, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovbl %eax, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl (%esp), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %eax, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, (%esp) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    addl %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ecx, %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl (%ecx), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %edx, %ecx, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ecx, {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %edx, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl (%esp), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %dl, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, 12(%ecx)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, 8(%ecx)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, (%ecx)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, 4(%ecx)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    addl $36, %esp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    retl
+;
+; X32-HAVE-BMI2-HAVE-SHLD-LABEL: lshr_16bytes:
+; X32-HAVE-BMI2-HAVE-SHLD:       # %bb.0:
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    subl $32, %esp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 12(%ecx), %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 8(%ecx), %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movzbl (%eax), %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlb $3, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %esi, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %edx, %esi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 4(%ebp), %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %edx, %ecx, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebx, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    subb $64, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %ebx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    negb %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %edi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    andb $24, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %ebx, %edi, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, (%esp) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl %ebp, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    addb $-64, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %ebx, %esi, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovbl %eax, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %ebp, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    xorl %ebp, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl (%esp), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, (%esp) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl (%eax), %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %ebp, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %ebp, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl (%esp), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %edi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %dl, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, 8(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, 12(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, (%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, 4(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    addl $32, %esp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    retl
+  %src = load i128, ptr %src.ptr, align 1
+  %byteOff = load i128, ptr %byteOff.ptr, align 1
+  %bitOff = shl i128 %byteOff, 3
+  %res = lshr i128 %src, %bitOff
+  store i128 %res, ptr %dst, align 1
+  ret void
+}
+define void @shl_16bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; X64-NO-BMI2-NO-SHLD-LABEL: shl_16bytes:
+; X64-NO-BMI2-NO-SHLD:       # %bb.0:
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq (%rdi), %r8
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq 8(%rdi), %rdi
+; X64-NO-BMI2-NO-SHLD-NEXT:    movzbl (%rsi), %eax
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlb $3, %al
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlq %cl, %rdi
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r8, %rsi
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %rsi
+; X64-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %cl, %rsi
+; X64-NO-BMI2-NO-SHLD-NEXT:    orq %rdi, %rsi
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlq %cl, %r8
+; X64-NO-BMI2-NO-SHLD-NEXT:    xorl %ecx, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    testb $64, %al
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %r8, %rsi
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmoveq %r8, %rcx
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rcx, (%rdx)
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rsi, 8(%rdx)
+; X64-NO-BMI2-NO-SHLD-NEXT:    retq
+;
+; X64-NO-BMI2-HAVE-SHLD-LABEL: shl_16bytes:
+; X64-NO-BMI2-HAVE-SHLD:       # %bb.0:
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq (%rdi), %rax
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq 8(%rdi), %rdi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movzbl (%rsi), %ecx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shlb $3, %cl
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %rax, %rsi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shlq %cl, %rsi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shldq %cl, %rax, %rdi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    xorl %eax, %eax
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    testb $64, %cl
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %rsi, %rdi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmoveq %rsi, %rax
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %rdi, 8(%rdx)
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %rax, (%rdx)
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    retq
+;
+; X64-HAVE-BMI2-NO-SHLD-LABEL: shl_16bytes:
+; X64-HAVE-BMI2-NO-SHLD:       # %bb.0:
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq (%rdi), %rax
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movzbl (%rsi), %ecx
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlb $3, %cl
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlxq %rcx, 8(%rdi), %rsi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %edi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    notb %dil
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlxq %rcx, %rax, %r8
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrq %rax
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrxq %rdi, %rax, %rax
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    orq %rsi, %rax
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    xorl %esi, %esi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    testb $64, %cl
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %r8, %rax
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmoveq %r8, %rsi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %rsi, (%rdx)
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %rax, 8(%rdx)
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    retq
+;
+; X64-HAVE-BMI2-HAVE-SHLD-LABEL: shl_16bytes:
+; X64-HAVE-BMI2-HAVE-SHLD:       # %bb.0:
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq (%rdi), %rax
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq 8(%rdi), %rdi
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movzbl (%rsi), %ecx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shlb $3, %cl
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shldq %cl, %rax, %rdi
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxq %rcx, %rax, %rax
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    xorl %esi, %esi
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $64, %cl
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %rax, %rdi
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmoveq %rax, %rsi
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %rdi, 8(%rdx)
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %rsi, (%rdx)
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    retq
+;
+; X32-NO-BMI2-NO-SHLD-LABEL: shl_16bytes:
+; X32-NO-BMI2-NO-SHLD:       # %bb.0:
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    subl $36, %esp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl (%esi), %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 4(%esi), %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl (%eax), %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shlb $3, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bl, %bh
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %bh
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bh, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %eax, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 8(%edi), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %ecx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %ecx, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    subb $64, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %ecx, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    decb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    leal (%esi,%esi), %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    negb %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    andb $24, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %ebp, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bl, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    addb $-64, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovbl %edx, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %esi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    xorl %esi, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 12(%eax), %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bl, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bh, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edx, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %ebp, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %ebx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %cl, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel %esi, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, (%ecx)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, 4(%ecx)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, 12(%ecx)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, 8(%ecx)
+; X32-NO-BMI2-NO-SHLD-NEXT:    addl $36, %esp
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    retl
+;
+; X32-NO-BMI2-HAVE-SHLD-LABEL: shl_16bytes:
+; X32-NO-BMI2-HAVE-SHLD:       # %bb.0:
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    subl $32, %esp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl (%ecx), %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 4(%ecx), %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 8(%ecx), %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movzbl (%eax), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shlb $3, %al
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %esi, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, (%esp) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edx, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edx, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    xorl %ebx, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    subb $64, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %ebx, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    negb %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %ebp, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    andb $24, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl %esi, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %al, %ah
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    addb $-64, %ah
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ah, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ah
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %al
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovbl %ebx, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl (%esp), %ecx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %ebp, %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    xorl %ebx, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ecx, (%esp) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebx, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 12(%ecx), %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %ebx, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ah, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %ebp, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ah
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl %edi, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %al
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %ecx, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %al, %al
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ecx, (%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl (%esp), %ecx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ecx, 4(%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, 8(%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, 12(%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    addl $32, %esp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    retl
+;
+; X32-HAVE-BMI2-NO-SHLD-LABEL: shl_16bytes:
+; X32-HAVE-BMI2-NO-SHLD:       # %bb.0:
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    subl $36, %esp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl (%ecx), %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 4(%ecx), %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movzbl (%eax), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlb $3, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ecx, %edx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrl %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %edi, %edx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %eax, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ecx, %ebp, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, (%esp) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 8(%ebx), %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ecx, %edx, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    subb $64, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    decb %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    addl %esi, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, %esi, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    negb %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    andb $24, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ecx, %ebp, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %esi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ecx, %edi, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ebx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    addb $-64, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ebx, %ebp, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovbl %eax, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl (%esp), %edi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %eax, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, (%esp) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrl %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ecx, %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 12(%ecx), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %edx, %ecx, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ecx, {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ebx, {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %edx, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl (%esp), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %dl, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, (%ecx)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, 4(%ecx)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, 12(%ecx)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, 8(%ecx)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    addl $36, %esp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    retl
+;
+; X32-HAVE-BMI2-HAVE-SHLD-LABEL: shl_16bytes:
+; X32-HAVE-BMI2-HAVE-SHLD:       # %bb.0:
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    subl $32, %esp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl (%ecx), %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 4(%ecx), %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movzbl (%eax), %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlb $3, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %esi, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %edx, %esi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 8(%ebp), %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %edx, %ecx, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebx, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    subb $64, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %ebx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    negb %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %edi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    andb $24, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %ebx, %edi, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, (%esp) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl %ebp, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    addb $-64, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %ebx, %esi, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovbl %eax, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %ebp, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    xorl %ebp, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl (%esp), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, (%esp) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 12(%eax), %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %ebp, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %ebp, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl (%esp), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %edi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %dl, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, (%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, 4(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, 8(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, 12(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    addl $32, %esp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    retl
+  %src = load i128, ptr %src.ptr, align 1
+  %byteOff = load i128, ptr %byteOff.ptr, align 1
+  %bitOff = shl i128 %byteOff, 3
+  %res = shl i128 %src, %bitOff
+  store i128 %res, ptr %dst, align 1
+  ret void
+}
+define void @ashr_16bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; X64-NO-BMI2-NO-SHLD-LABEL: ashr_16bytes:
+; X64-NO-BMI2-NO-SHLD:       # %bb.0:
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq (%rdi), %r8
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq 8(%rdi), %rdi
+; X64-NO-BMI2-NO-SHLD-NEXT:    movzbl (%rsi), %eax
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlb $3, %al
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %cl, %r8
+; X64-NO-BMI2-NO-SHLD-NEXT:    leaq (%rdi,%rdi), %rsi
+; X64-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlq %cl, %rsi
+; X64-NO-BMI2-NO-SHLD-NEXT:    orq %r8, %rsi
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rdi, %r8
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    sarq %cl, %r8
+; X64-NO-BMI2-NO-SHLD-NEXT:    sarq $63, %rdi
+; X64-NO-BMI2-NO-SHLD-NEXT:    testb $64, %al
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %r8, %rsi
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmoveq %r8, %rdi
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rdi, 8(%rdx)
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rsi, (%rdx)
+; X64-NO-BMI2-NO-SHLD-NEXT:    retq
+;
+; X64-NO-BMI2-HAVE-SHLD-LABEL: ashr_16bytes:
+; X64-NO-BMI2-HAVE-SHLD:       # %bb.0:
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq (%rdi), %rax
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq 8(%rdi), %rdi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movzbl (%rsi), %ecx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shlb $3, %cl
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %rdi, %rsi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    sarq %cl, %rsi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shrdq %cl, %rdi, %rax
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    sarq $63, %rdi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    testb $64, %cl
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %rsi, %rax
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmoveq %rsi, %rdi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %rdi, 8(%rdx)
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %rax, (%rdx)
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    retq
+;
+; X64-HAVE-BMI2-NO-SHLD-LABEL: ashr_16bytes:
+; X64-HAVE-BMI2-NO-SHLD:       # %bb.0:
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq 8(%rdi), %rax
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movzbl (%rsi), %ecx
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlb $3, %cl
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrxq %rcx, (%rdi), %rsi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %edi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    notb %dil
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    leaq (%rax,%rax), %r8
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlxq %rdi, %r8, %rdi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    orq %rsi, %rdi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    sarxq %rcx, %rax, %rsi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    sarq $63, %rax
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    testb $64, %cl
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %rsi, %rdi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmoveq %rsi, %rax
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %rax, 8(%rdx)
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %rdi, (%rdx)
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    retq
+;
+; X64-HAVE-BMI2-HAVE-SHLD-LABEL: ashr_16bytes:
+; X64-HAVE-BMI2-HAVE-SHLD:       # %bb.0:
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq (%rdi), %rax
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq 8(%rdi), %rdi
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movzbl (%rsi), %ecx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shlb $3, %cl
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdq %cl, %rdi, %rax
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    sarxq %rcx, %rdi, %rsi
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    sarq $63, %rdi
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $64, %cl
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %rsi, %rax
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmoveq %rsi, %rdi
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %rdi, 8(%rdx)
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %rax, (%rdx)
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    retq
+;
+; X32-NO-BMI2-NO-SHLD-LABEL: ashr_16bytes:
+; X32-NO-BMI2-NO-SHLD:       # %bb.0:
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    subl $36, %esp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 12(%edx), %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 8(%edx), %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl (%eax), %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shlb $3, %dl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    leal (%esi,%esi), %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %eax, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 4(%edi), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    sarl %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    sarl $31, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    xorl %eax, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %dl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %dl, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    subb $64, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %esi, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    decb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    negb %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    andb $24, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %ebx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %ebp, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %dl, %dh
+; X32-NO-BMI2-NO-SHLD-NEXT:    addb $-64, %dh
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %dh, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    sarl %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %dh
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %dl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %esi, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovbl %eax, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl (%eax), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    leal (%ecx,%ecx), %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %eax, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %dl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %dh, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edi, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %dh
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %ebp, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %dl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %eax, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %dl, %dl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, 12(%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, 8(%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, (%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, 4(%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    addl $36, %esp
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    retl
+;
+; X32-NO-BMI2-HAVE-SHLD-LABEL: ashr_16bytes:
+; X32-NO-BMI2-HAVE-SHLD:       # %bb.0:
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    subl $32, %esp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 4(%ecx), %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 12(%ecx), %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 8(%ecx), %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movzbl (%eax), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shlb $3, %al
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %edi, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    sarl %cl, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    sarl $31, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebx, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, (%esp) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edx, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    subb $64, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %ebp, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    negb %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %ebx, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    andb $24, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebx, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl %edi, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %al, %ah
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    addb $-64, %ah
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ah, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    sarl %cl, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ah
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %al
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl (%esp), %ecx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %ebp, %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ecx, (%esp) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovbl %esi, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ecx, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl (%ecx), %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %ebp, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ah, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %esi, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ah
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl %ebx, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %al
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %ebp, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %al, %al
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl (%esp), %ecx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ecx, 8(%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ecx, 12(%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, (%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, 4(%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    addl $32, %esp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    retl
+;
+; X32-HAVE-BMI2-NO-SHLD-LABEL: ashr_16bytes:
+; X32-HAVE-BMI2-NO-SHLD:       # %bb.0:
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    subl $44, %esp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 12(%edx), %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 8(%edx), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movzbl (%eax), %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlb $3, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    leal (%edi,%edi), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %edx, %ecx, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %eax, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    sarl $31, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    sarxl %ebx, %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 4(%esi), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %ecx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ecx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, (%esp) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    subb $64, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    decb %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrl %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %eax, %ebp, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    negb %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    andb $24, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ecx, %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ebp, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ecx, %esi, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl (%esp), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    addb $-64, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    sarxl %ecx, %edi, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, (%esp) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %esi, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovbl %eax, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    leal (%ebp,%ebp), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %edx, %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl (%edx), %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %edx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %edx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %edx, {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ecx, {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %esi, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel (%esp), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %edx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %bl, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel %ebp, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, 12(%ecx)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, 8(%ecx)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, (%ecx)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, 4(%ecx)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    addl $44, %esp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    retl
+;
+; X32-HAVE-BMI2-HAVE-SHLD-LABEL: ashr_16bytes:
+; X32-HAVE-BMI2-HAVE-SHLD:       # %bb.0:
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    subl $36, %esp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 12(%ecx), %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 8(%ecx), %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movzbl (%eax), %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlb $3, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %esi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    sarl $31, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    xorl %edi, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    sarxl %edx, %esi, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 4(%eax), %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %edx, %eax, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, (%esp) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    subb $64, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %ebp, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    negb %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %esi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    andb $24, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %ebx, %esi, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl (%esp), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    addb $-64, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    sarxl %ebx, %edi, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, (%esp) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %ebp, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovbl %eax, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl (%eax), %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %edi, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %eax, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel (%esp), %edi # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %edi, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %dl, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, 8(%eax)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, 12(%eax)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, (%eax)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, 4(%eax)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    addl $36, %esp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    retl
+  %src = load i128, ptr %src.ptr, align 1
+  %byteOff = load i128, ptr %byteOff.ptr, align 1
+  %bitOff = shl i128 %byteOff, 3
+  %res = ashr i128 %src, %bitOff
+  store i128 %res, ptr %dst, align 1
+  ret void
+}
+
+define void @lshr_32bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; X64-NO-BMI2-NO-SHLD-LABEL: lshr_32bytes:
+; X64-NO-BMI2-NO-SHLD:       # %bb.0:
+; X64-NO-BMI2-NO-SHLD-NEXT:    pushq %rbp
+; X64-NO-BMI2-NO-SHLD-NEXT:    pushq %r15
+; X64-NO-BMI2-NO-SHLD-NEXT:    pushq %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    pushq %r13
+; X64-NO-BMI2-NO-SHLD-NEXT:    pushq %r12
+; X64-NO-BMI2-NO-SHLD-NEXT:    pushq %rbx
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rdx, %r9
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq 24(%rdi), %r8
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq (%rdi), %rdx
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rdx, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq 8(%rdi), %r10
+; X64-NO-BMI2-NO-SHLD-NEXT:    movzbl (%rsi), %eax
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlb $3, %al
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r10, %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %cl, %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r8, %rbp
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %cl, %rbp
+; X64-NO-BMI2-NO-SHLD-NEXT:    xorl %r11d, %r11d
+; X64-NO-BMI2-NO-SHLD-NEXT:    testb $64, %al
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rbp, %rbx
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %r11, %rbx
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r14, %r15
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %r11, %r15
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %cl, %rdx
+; X64-NO-BMI2-NO-SHLD-NEXT:    leaq (%r10,%r10), %r13
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %esi
+; X64-NO-BMI2-NO-SHLD-NEXT:    notb %sil
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlq %cl, %r13
+; X64-NO-BMI2-NO-SHLD-NEXT:    orq %rdx, %r13
+; X64-NO-BMI2-NO-SHLD-NEXT:    testb $64, %al
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq 16(%rdi), %rdx
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %r14, %r13
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rdx, %r12
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %cl, %r12
+; X64-NO-BMI2-NO-SHLD-NEXT:    leaq (%r8,%r8), %rdi
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rdi, %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlq %cl, %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    orq %r12, %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    testb $64, %al
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %rbp, %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    movb $-128, %sil
+; X64-NO-BMI2-NO-SHLD-NEXT:    subb %al, %sil
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r8, %r12
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlq %cl, %r12
+; X64-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rdx, %rbp
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %rbp
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %cl, %rbp
+; X64-NO-BMI2-NO-SHLD-NEXT:    orq %r12, %rbp
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rdx, %r12
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlq %cl, %r12
+; X64-NO-BMI2-NO-SHLD-NEXT:    testb $64, %sil
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %r12, %rbp
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %r11, %r12
+; X64-NO-BMI2-NO-SHLD-NEXT:    orq %r13, %r12
+; X64-NO-BMI2-NO-SHLD-NEXT:    orq %r15, %rbp
+; X64-NO-BMI2-NO-SHLD-NEXT:    leal -128(%rax), %esi
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %cl, %rdx
+; X64-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlq %cl, %rdi
+; X64-NO-BMI2-NO-SHLD-NEXT:    orq %rdx, %rdi
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %cl, %r8
+; X64-NO-BMI2-NO-SHLD-NEXT:    testb $64, %sil
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %r8, %rdi
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %r11, %r8
+; X64-NO-BMI2-NO-SHLD-NEXT:    testb %al, %al
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovnsq %r12, %rdi
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmoveq {{[-0-9]+}}(%r{{[sb]}}p), %rdi # 8-byte Folded Reload
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovnsq %rbp, %r8
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmoveq %r10, %r8
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovsq %r11, %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovsq %r11, %rbx
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rbx, 24(%r9)
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r14, 16(%r9)
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rdi, (%r9)
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r8, 8(%r9)
+; X64-NO-BMI2-NO-SHLD-NEXT:    popq %rbx
+; X64-NO-BMI2-NO-SHLD-NEXT:    popq %r12
+; X64-NO-BMI2-NO-SHLD-NEXT:    popq %r13
+; X64-NO-BMI2-NO-SHLD-NEXT:    popq %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    popq %r15
+; X64-NO-BMI2-NO-SHLD-NEXT:    popq %rbp
+; X64-NO-BMI2-NO-SHLD-NEXT:    retq
+;
+; X64-NO-BMI2-HAVE-SHLD-LABEL: lshr_32bytes:
+; X64-NO-BMI2-HAVE-SHLD:       # %bb.0:
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    pushq %r15
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    pushq %r14
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    pushq %r13
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    pushq %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    pushq %rbx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq 16(%rdi), %r8
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq (%rdi), %r9
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq 8(%rdi), %r10
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq 24(%rdi), %rdi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movzbl (%rsi), %eax
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shlb $3, %al
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %rdi, %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ecx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shrq %cl, %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %r10, %r14
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shrq %cl, %r14
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %r9, %r15
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shrdq %cl, %r10, %r15
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    xorl %esi, %esi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    testb $64, %al
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %r14, %r15
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %r12, %r11
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %rsi, %r11
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %rsi, %r14
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %r8, %rbx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shrdq %cl, %rdi, %rbx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    testb $64, %al
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %r12, %rbx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movb $-128, %cl
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    subb %al, %cl
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %rdi, %r13
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shldq %cl, %r8, %r13
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %r8, %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shlq %cl, %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    testb $64, %cl
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %r12, %r13
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %rsi, %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    orq %r15, %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    orq %r14, %r13
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    leal -128(%rax), %ecx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shrdq %cl, %rdi, %r8
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shrq %cl, %rdi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    testb $64, %cl
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %rdi, %r8
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %rsi, %rdi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    testb %al, %al
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovnsq %r13, %rdi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmoveq %r10, %rdi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovnsq %r12, %r8
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmoveq %r9, %r8
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovsq %rsi, %rbx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovsq %rsi, %r11
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %rbx, 16(%rdx)
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %r11, 24(%rdx)
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %r8, (%rdx)
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %rdi, 8(%rdx)
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    popq %rbx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    popq %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    popq %r13
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    popq %r14
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    popq %r15
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    retq
+;
+; X64-HAVE-BMI2-NO-SHLD-LABEL: lshr_32bytes:
+; X64-HAVE-BMI2-NO-SHLD:       # %bb.0:
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    pushq %rbp
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    pushq %r15
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    pushq %r14
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    pushq %r13
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    pushq %r12
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    pushq %rbx
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq (%rdi), %rcx
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq 8(%rdi), %rax
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq 24(%rdi), %r10
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movzbl (%rsi), %r9d
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlb $3, %r9b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrxq %r9, %r10, %r15
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    xorl %esi, %esi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    testb $64, %r9b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %r15, %r8
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %rsi, %r8
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrxq %r9, %rax, %r14
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %r14, %r11
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %rsi, %r11
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrxq %r9, %rcx, %rbx
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movl %r9d, %r13d
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    notb %r13b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    leaq (%rax,%rax), %r12
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlxq %r13, %r12, %r12
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    orq %rbx, %r12
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    testb $64, %r9b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq 16(%rdi), %rbx
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %r14, %r12
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrxq %r9, %rbx, %rbp
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    leaq (%r10,%r10), %r14
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlxq %r13, %r14, %rdi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    orq %rbp, %rdi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    testb $64, %r9b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %r15, %rdi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movb $-128, %r13b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    subb %r9b, %r13b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movl %r13d, %r15d
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    notb %r15b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %rbx, %rbp
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrq %rbp
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrxq %r15, %rbp, %r15
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlxq %r13, %r10, %rbp
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    orq %rbp, %r15
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    testb $64, %r13b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlxq %r13, %rbx, %r13
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %r13, %r15
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %rsi, %r13
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    orq %r12, %r13
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    orq %r11, %r15
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    leal -128(%r9), %r11d
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movl %r11d, %r12d
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    notb %r12b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlxq %r12, %r14, %r14
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrxq %r11, %rbx, %rbx
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    orq %rbx, %r14
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrxq %r11, %r10, %r10
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    testb $64, %r11b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %r10, %r14
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %rsi, %r10
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    testb %r9b, %r9b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovnsq %r13, %r14
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmoveq %rcx, %r14
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovnsq %r15, %r10
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmoveq %rax, %r10
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovsq %rsi, %rdi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovsq %rsi, %r8
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %r8, 24(%rdx)
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %rdi, 16(%rdx)
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %r14, (%rdx)
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %r10, 8(%rdx)
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    popq %rbx
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    popq %r12
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    popq %r13
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    popq %r14
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    popq %r15
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    popq %rbp
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    retq
+;
+; X64-HAVE-BMI2-HAVE-SHLD-LABEL: lshr_32bytes:
+; X64-HAVE-BMI2-HAVE-SHLD:       # %bb.0:
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    pushq %r15
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    pushq %r14
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    pushq %r13
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    pushq %r12
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    pushq %rbx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq 16(%rdi), %r8
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq 24(%rdi), %r11
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq (%rdi), %r9
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq 8(%rdi), %rdi
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movzbl (%rsi), %eax
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shlb $3, %al
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %r9, %r14
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ecx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdq %cl, %rdi, %r14
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxq %rax, %rdi, %r15
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    xorl %esi, %esi
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $64, %al
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %r15, %r14
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxq %rax, %r11, %r12
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %r12, %r10
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %rsi, %r10
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %rsi, %r15
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %r8, %rbx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdq %cl, %r11, %rbx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $64, %al
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %r12, %rbx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movb $-128, %cl
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    subb %al, %cl
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %r11, %r12
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shldq %cl, %r8, %r12
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxq %rcx, %r8, %r13
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $64, %cl
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %r13, %r12
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %rsi, %r13
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    orq %r14, %r13
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    orq %r15, %r12
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    leal -128(%rax), %ecx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdq %cl, %r11, %r8
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxq %rcx, %r11, %r11
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $64, %cl
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %r11, %r8
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %rsi, %r11
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %al, %al
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnsq %r12, %r11
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmoveq %rdi, %r11
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnsq %r13, %r8
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmoveq %r9, %r8
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsq %rsi, %rbx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsq %rsi, %r10
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %rbx, 16(%rdx)
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %r10, 24(%rdx)
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %r8, (%rdx)
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %r11, 8(%rdx)
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    popq %rbx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    popq %r12
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    popq %r13
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    popq %r14
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    popq %r15
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    retq
+;
+; X32-NO-BMI2-NO-SHLD-LABEL: lshr_32bytes:
+; X32-NO-BMI2-NO-SHLD:       # %bb.0:
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    subl $136, %esp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 12(%ecx), %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 28(%ecx), %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb (%eax), %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    shlb $3, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    xorl %esi, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 20(%ebx), %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 4(%ebx), %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 24(%ebx), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    leal (%eax,%eax), %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %ebx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 8(%esi), %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    addl %esi, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %al, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edi, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %edx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %eax, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    addb $-128, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    subb $64, %dl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %eax, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    decb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    negb %dl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %dl, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %dl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    andb $24, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %al, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    addb $-64, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %ebx, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovbl %edx, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %edi, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bl, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    addb $64, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %ebx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %ebx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %ebx, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %ecx, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %bh # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %bh
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %ecx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 16(%eax), %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    leal (%ecx,%ecx), %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edx, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %ch # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bl, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %esi, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %eax, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl (%eax), %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %bh
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bh, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %cl, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    addl %esi, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bl, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edx, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edi, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %ebx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %ch # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %esi, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %edx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb $-128, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    subb {{[-0-9]+}}(%e{{[sb]}}p), %ch # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %ebx, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb $-64, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    subb %bl, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %al, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edx, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %al, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %edx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %bl, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel %eax, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovsl %edx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovsl %edx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovsl %edx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    subb $64, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %edx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bl, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    decb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    negb %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bl, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    andb $24, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %esi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %bh # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %bh
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %esi, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %ch, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $0, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel %ebp, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovsl %esi, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %ebp, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %bh
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %esi, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %ebx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %ebx, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %esi, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %eax, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %ch, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %ah # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %ah, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovsl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %ebx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %al # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %al, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %ah, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnsl %ebx, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %ah, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovsl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %al, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %ah, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovsl %ebx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, 28(%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, 24(%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, 16(%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, 20(%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, (%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, 12(%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, 4(%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, 8(%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    addl $136, %esp
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    retl
+;
+; X32-NO-BMI2-HAVE-SHLD-LABEL: lshr_32bytes:
+; X32-NO-BMI2-HAVE-SHLD:       # %bb.0:
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    subl $120, %esp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 12(%ecx), %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 28(%ecx), %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb (%eax), %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shlb $3, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    xorl %eax, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    xorl %eax, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 20(%ebp), %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 4(%ebp), %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 8(%edx), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %eax, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 24(%ebp), %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %ebp, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %eax, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    addb $-128, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    subb $64, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %edi, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %edi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    negb %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %eax, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %cl, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %cl, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    andb $24, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %dh, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %dh, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ch, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    addb $-64, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %dh, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovbl %edi, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %ebp, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    xorl %ebp, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %dl, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %esi, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %edi, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %dl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 16(%esi), %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %esi, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %dh, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %ebx, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl %eax, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %esi, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl (%eax), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %eax, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb $-128, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    subb %ch, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %cl, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %dh, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %eax, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %eax, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %dl, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    subb $64, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %ebp, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    negb %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %dh, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    andb $24, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %edi, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %ebx, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %dl, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %edx, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl %esi, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %ebp, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %cl, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel %edx, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    addb $64, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %esi, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %ch, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel %ebx, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovsl %ebp, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovsl %ebp, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovsl %ebp, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovsl %ebp, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 1-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %bl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %ebp, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    xorl %ebp, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %ch, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnsl %eax, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %bl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %ebp, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $0, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel %esi, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %ch, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovsl %eax, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb $-64, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    subb %ch, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %bl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %bl, %bl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %ch, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovsl %ebp, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %ebp, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edx, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %edx, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %edx, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %cl, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel %ebp, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %ch, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovsl %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, 24(%ecx)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, 28(%ecx)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, 16(%ecx)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, 20(%ecx)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, (%ecx)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, 4(%ecx)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, 8(%ecx)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, 12(%ecx)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    addl $120, %esp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    retl
+;
+; X32-HAVE-BMI2-NO-SHLD-LABEL: lshr_32bytes:
+; X32-HAVE-BMI2-NO-SHLD:       # %bb.0:
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    subl $160, %esp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 20(%esi), %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 4(%esi), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movzbl (%eax), %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlb $3, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    xorl %edi, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %edx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 12(%esi), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 28(%esi), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %eax, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 24(%ecx), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %ecx, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    leal (%eax,%eax), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %edx, %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 8(%eax), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %eax, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    addl %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %edx, %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %esi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %ecx, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, (%esp) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    addb $-128, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %eax, %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    xorl %ecx, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    # kill: def $bl killed $bl killed $ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    subb $64, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    decb %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrl %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    negb %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movb %bl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    andb $24, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %esi, %edx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, %edi, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %edx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, %ecx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl (%esp), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    addb $-64, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %edx, %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovbl %esi, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %eax, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %edx, %ecx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrl %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %edx, %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ebp, %ecx, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %eax, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ebp, %esi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl (%esp), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    addb $64, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, %esi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    addl %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ecx, %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    leal (%eax,%eax), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ecx, %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ecx, %esi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ecx, {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %edx, %esi, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ebx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ecx, {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ecx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %edx, %ebp, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ecx, %esi, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ebp, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl (%esp), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %esi, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 16(%esi), %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ecx, %edi, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ebp, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ebp, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %edx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %eax, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl (%eax), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edx, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ecx, %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %eax, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebp, {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %eax, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ebx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, (%esp) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %ebp, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ebp, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %edx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ebp, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    xorl %ebp, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %edi, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ecx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ebp, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movb $-128, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl (%esp), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    subb %cl, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrl %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, %edx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %ebp, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %esi, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, %edi, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ebx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, %ebp, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movb $-64, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    subb %cl, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %edx, {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %edi, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, %ebp, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %cl, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel %edx, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovsl %eax, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovsl %eax, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovsl %eax, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    # kill: def $cl killed $cl killed $ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    subb $64, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %eax, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    decb %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    negb %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    andb $24, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %edx, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %eax, {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %eax, {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ebp, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %cl, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $0, (%esp) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovsl %edi, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %edi, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, %edi, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ebp, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %ecx, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %al, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel %edi, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $0, (%esp) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovsl %edx, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %eax, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %al, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl (%esp), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %dl, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnsl %edi, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %dl, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovsl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %al, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %dl, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovsl %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, 28(%eax)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, 24(%eax)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, 16(%eax)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, 20(%eax)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, (%eax)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, 12(%eax)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, 4(%eax)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, 8(%eax)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    addl $160, %esp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    retl
+;
+; X32-HAVE-BMI2-HAVE-SHLD-LABEL: lshr_32bytes:
+; X32-HAVE-BMI2-HAVE-SHLD:       # %bb.0:
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    subl $128, %esp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 28(%ecx), %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 24(%ecx), %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 12(%ecx), %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 8(%ecx), %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movzbl (%eax), %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlb $3, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %edx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %esi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %ebx, %esi, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ecx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %ebx, %edx, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 4(%edx), %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %ebx, %eax, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, (%esp) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 20(%edx), %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %ebx, %edx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %ebp, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    addb $-128, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %ecx, %edi, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %ecx, %edi, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    subb $64, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %ebp, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %ebp, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    negb %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %edi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    andb $24, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %ebx, %edi, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl %edx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    addb $-64, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %ecx, %esi, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovbl %eax, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %esi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    xorl %edx, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %ecx, %eax, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %edx, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %eax, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %edi, %eax, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl (%esp), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %esi, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 16(%eax), %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, (%esp) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %eax, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    # kill: def $cl killed $cl killed $ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %ebx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl %ebp, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %eax, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl (%eax), %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl (%esp), %edi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %esi, %edi, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %eax, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movb $-128, %al
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    subb %dl, %al
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %edx, %eax, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ecx, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %edx, %edi, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ecx, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %edx, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %esi, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %al
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl (%esp), %esi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %edx, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %al
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    subb $64, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %ebx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    negb %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %edx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    andb $24, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %edx, %ebx, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl %ebp, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %edi, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl (%esp), %edi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %edi, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %ebx, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl %esi, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %edx, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %cl, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel %ebx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    addb $64, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %ebp, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %ecx, %ebp, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ecx, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %dl, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel (%esp), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsl %esi, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsl %esi, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsl %esi, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsl %esi, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %esi, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    xorl %esi, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %bl, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnsl %eax, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %esi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $0, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel %edx, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %bl, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsl %eax, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movb $-64, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    subb %bl, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %ecx, {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %al
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %al, %al
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel (%esp), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %bl, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsl %esi, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %eax, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl (%esp), %ebx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %ebx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %ebx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %cl, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel %esi, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $0, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsl %edx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, 24(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, 28(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, 16(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, 20(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, (%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, 4(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, 8(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, 12(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    addl $128, %esp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    retl
+  %src = load i256, ptr %src.ptr, align 1
+  %byteOff = load i256, ptr %byteOff.ptr, align 1
+  %bitOff = shl i256 %byteOff, 3
+  %res = lshr i256 %src, %bitOff
+  store i256 %res, ptr %dst, align 1
+  ret void
+}
+define void @shl_32bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; X64-NO-BMI2-NO-SHLD-LABEL: shl_32bytes:
+; X64-NO-BMI2-NO-SHLD:       # %bb.0:
+; X64-NO-BMI2-NO-SHLD-NEXT:    pushq %rbp
+; X64-NO-BMI2-NO-SHLD-NEXT:    pushq %r15
+; X64-NO-BMI2-NO-SHLD-NEXT:    pushq %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    pushq %r13
+; X64-NO-BMI2-NO-SHLD-NEXT:    pushq %r12
+; X64-NO-BMI2-NO-SHLD-NEXT:    pushq %rbx
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rdx, %r9
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq 24(%rdi), %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r14, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq (%rdi), %r8
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq 8(%rdi), %rdx
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq 16(%rdi), %rdi
+; X64-NO-BMI2-NO-SHLD-NEXT:    movzbl (%rsi), %eax
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlb $3, %al
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rdi, %rbx
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlq %cl, %rbx
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r8, %rbp
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlq %cl, %rbp
+; X64-NO-BMI2-NO-SHLD-NEXT:    xorl %r10d, %r10d
+; X64-NO-BMI2-NO-SHLD-NEXT:    testb $64, %al
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rbp, %r11
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %r10, %r11
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rbx, %r12
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %r10, %r12
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlq %cl, %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rdi, %r13
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %r13
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %esi
+; X64-NO-BMI2-NO-SHLD-NEXT:    notb %sil
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %cl, %r13
+; X64-NO-BMI2-NO-SHLD-NEXT:    orq %r14, %r13
+; X64-NO-BMI2-NO-SHLD-NEXT:    testb $64, %al
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %rbx, %r13
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rdx, %rbx
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlq %cl, %rbx
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r8, %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r14, %r15
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %cl, %r15
+; X64-NO-BMI2-NO-SHLD-NEXT:    orq %rbx, %r15
+; X64-NO-BMI2-NO-SHLD-NEXT:    testb $64, %al
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %rbp, %r15
+; X64-NO-BMI2-NO-SHLD-NEXT:    movb $-128, %sil
+; X64-NO-BMI2-NO-SHLD-NEXT:    subb %al, %sil
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r8, %rbx
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %cl, %rbx
+; X64-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X64-NO-BMI2-NO-SHLD-NEXT:    leaq (%rdx,%rdx), %rbp
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlq %cl, %rbp
+; X64-NO-BMI2-NO-SHLD-NEXT:    orq %rbx, %rbp
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rdx, %rbx
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %cl, %rbx
+; X64-NO-BMI2-NO-SHLD-NEXT:    testb $64, %sil
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %rbx, %rbp
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %r10, %rbx
+; X64-NO-BMI2-NO-SHLD-NEXT:    orq %r13, %rbx
+; X64-NO-BMI2-NO-SHLD-NEXT:    orq %r12, %rbp
+; X64-NO-BMI2-NO-SHLD-NEXT:    leal -128(%rax), %esi
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlq %cl, %rdx
+; X64-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %cl, %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    orq %rdx, %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlq %cl, %r8
+; X64-NO-BMI2-NO-SHLD-NEXT:    testb $64, %sil
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %r8, %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %r10, %r8
+; X64-NO-BMI2-NO-SHLD-NEXT:    testb %al, %al
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovnsq %rbx, %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmoveq {{[-0-9]+}}(%r{{[sb]}}p), %r14 # 8-byte Folded Reload
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovnsq %rbp, %r8
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmoveq %rdi, %r8
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovsq %r10, %r15
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovsq %r10, %r11
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r11, (%r9)
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r15, 8(%r9)
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r14, 24(%r9)
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r8, 16(%r9)
+; X64-NO-BMI2-NO-SHLD-NEXT:    popq %rbx
+; X64-NO-BMI2-NO-SHLD-NEXT:    popq %r12
+; X64-NO-BMI2-NO-SHLD-NEXT:    popq %r13
+; X64-NO-BMI2-NO-SHLD-NEXT:    popq %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    popq %r15
+; X64-NO-BMI2-NO-SHLD-NEXT:    popq %rbp
+; X64-NO-BMI2-NO-SHLD-NEXT:    retq
+;
+; X64-NO-BMI2-HAVE-SHLD-LABEL: shl_32bytes:
+; X64-NO-BMI2-HAVE-SHLD:       # %bb.0:
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    pushq %rbp
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    pushq %r15
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    pushq %r14
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    pushq %r13
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    pushq %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    pushq %rbx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq 16(%rdi), %r8
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq 24(%rdi), %r9
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq (%rdi), %r10
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq 8(%rdi), %rdi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movzbl (%rsi), %eax
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shlb $3, %al
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movb $-128, %cl
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    subb %al, %cl
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %rdi, %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shrq %cl, %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %r10, %r13
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shrdq %cl, %rdi, %r13
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    xorl %esi, %esi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    testb $64, %cl
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %r12, %r13
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %rsi, %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %r9, %r11
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ecx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shldq %cl, %r8, %r11
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %r10, %rbp
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shlq %cl, %rbp
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %r8, %r14
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shlq %cl, %r14
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    testb $64, %al
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %r14, %r11
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %rbp, %rbx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %rsi, %rbx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %rsi, %r14
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %rdi, %r15
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shldq %cl, %r10, %r15
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    testb $64, %al
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %rbp, %r15
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    orq %r13, %r14
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    orq %r12, %r11
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    leal -128(%rax), %ecx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %r10, %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shlq %cl, %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shldq %cl, %r10, %rdi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    testb $64, %cl
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %r12, %rdi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %rsi, %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    testb %al, %al
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovnsq %r11, %rdi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmoveq %r9, %rdi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovnsq %r14, %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmoveq %r8, %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovsq %rsi, %r15
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovsq %rsi, %rbx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %rbx, (%rdx)
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %r15, 8(%rdx)
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %r12, 16(%rdx)
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %rdi, 24(%rdx)
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    popq %rbx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    popq %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    popq %r13
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    popq %r14
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    popq %r15
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    popq %rbp
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    retq
+;
+; X64-HAVE-BMI2-NO-SHLD-LABEL: shl_32bytes:
+; X64-HAVE-BMI2-NO-SHLD:       # %bb.0:
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    pushq %rbp
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    pushq %r15
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    pushq %r14
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    pushq %r13
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    pushq %r12
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    pushq %rbx
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq 24(%rdi), %rcx
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq 16(%rdi), %rax
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq (%rdi), %r9
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq 8(%rdi), %r10
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movzbl (%rsi), %r8d
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlb $3, %r8b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlxq %r8, %r9, %r15
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    xorl %esi, %esi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    testb $64, %r8b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %r15, %rdi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %rsi, %rdi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlxq %r8, %rax, %r11
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %r11, %rbx
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %rsi, %rbx
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlxq %r8, %rcx, %r14
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movl %r8d, %r13d
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    notb %r13b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %rax, %r12
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrq %r12
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrxq %r13, %r12, %r12
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    orq %r14, %r12
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    testb $64, %r8b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %r11, %r12
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlxq %r8, %r10, %rbp
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %r9, %r14
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrq %r14
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrxq %r13, %r14, %r11
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    orq %rbp, %r11
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    testb $64, %r8b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %r15, %r11
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movb $-128, %r13b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    subb %r8b, %r13b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movl %r13d, %r15d
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    notb %r15b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    leaq (%r10,%r10), %rbp
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlxq %r15, %rbp, %r15
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrxq %r13, %r9, %rbp
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    orq %rbp, %r15
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    testb $64, %r13b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrxq %r13, %r10, %r13
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %r13, %r15
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %rsi, %r13
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    orq %r12, %r13
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    orq %rbx, %r15
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    leal -128(%r8), %ebx
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %r12d
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    notb %r12b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrxq %r12, %r14, %r14
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlxq %rbx, %r10, %r10
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    orq %r10, %r14
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlxq %rbx, %r9, %r9
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    testb $64, %bl
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %r9, %r14
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %rsi, %r9
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    testb %r8b, %r8b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovnsq %r13, %r14
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmoveq %rcx, %r14
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovnsq %r15, %r9
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmoveq %rax, %r9
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovsq %rsi, %r11
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovsq %rsi, %rdi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %rdi, (%rdx)
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %r11, 8(%rdx)
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %r14, 24(%rdx)
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %r9, 16(%rdx)
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    popq %rbx
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    popq %r12
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    popq %r13
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    popq %r14
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    popq %r15
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    popq %rbp
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    retq
+;
+; X64-HAVE-BMI2-HAVE-SHLD-LABEL: shl_32bytes:
+; X64-HAVE-BMI2-HAVE-SHLD:       # %bb.0:
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    pushq %rbp
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    pushq %r15
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    pushq %r14
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    pushq %r13
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    pushq %r12
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    pushq %rbx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq 16(%rdi), %r8
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq 24(%rdi), %r9
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq (%rdi), %r10
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq 8(%rdi), %rdi
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movzbl (%rsi), %eax
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shlb $3, %al
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movb $-128, %cl
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    subb %al, %cl
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %r10, %r12
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdq %cl, %rdi, %r12
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxq %rcx, %rdi, %r13
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    xorl %esi, %esi
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $64, %cl
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %r13, %r12
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %rsi, %r13
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %r9, %r14
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ecx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shldq %cl, %r8, %r14
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxq %rax, %r8, %rbx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $64, %al
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %rbx, %r14
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxq %rax, %r10, %rbp
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %rbp, %r11
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %rsi, %r11
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %rsi, %rbx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %rdi, %r15
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shldq %cl, %r10, %r15
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $64, %al
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %rbp, %r15
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    orq %r12, %rbx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    orq %r13, %r14
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    leal -128(%rax), %ecx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shldq %cl, %r10, %rdi
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxq %rcx, %r10, %r10
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $64, %cl
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %r10, %rdi
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %rsi, %r10
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %al, %al
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnsq %r14, %rdi
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmoveq %r9, %rdi
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnsq %rbx, %r10
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmoveq %r8, %r10
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsq %rsi, %r15
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsq %rsi, %r11
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %r11, (%rdx)
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %r15, 8(%rdx)
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %r10, 16(%rdx)
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %rdi, 24(%rdx)
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    popq %rbx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    popq %r12
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    popq %r13
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    popq %r14
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    popq %r15
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    popq %rbp
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    retq
+;
+; X32-NO-BMI2-NO-SHLD-LABEL: shl_32bytes:
+; X32-NO-BMI2-NO-SHLD:       # %bb.0:
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    subl $140, %esp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 16(%ecx), %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl (%ecx), %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl (%eax), %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shlb $3, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    xorl %ebp, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %ebp, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %ebp, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 8(%ebx), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %ebp, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 24(%ebx), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %ebp, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 4(%ebx), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %cl, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %cl, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ah, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ah, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %ebx, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 20(%edi), %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ah, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %esi, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %edx, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %al, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %al, (%esp) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    addb $-128, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 12(%edi), %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edx, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %ebp, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %ebp, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %esi, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %edx, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb (%esp), %ah # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ah, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    subb $64, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %edx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    decb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    leal (%ebp,%ebp), %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    negb %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    andb $24, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %ebp, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ah, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    addb $-64, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ah, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %edx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovbl %esi, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %edx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    xorl %edi, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ah, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    leal (%edi,%edi), %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edx, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %al, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl (%esp), %edx # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %dl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %esi, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %esi, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %esi, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %dl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %dl, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ah, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ah, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ah, %dl
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %dl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %dl, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %esi, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb $-128, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %dh
+; X32-NO-BMI2-NO-SHLD-NEXT:    subb %ch, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %al, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    subb $64, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    negb %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ah, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ah, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    andb $24, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %al, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %al, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %al, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %al, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %al, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %esi, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %ebx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %dh
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %ebp, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %ch # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 28(%esi), %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %dh, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %esi, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %dh
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %dh, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %ah # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ah, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %dl, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %esi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edi, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %edx, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    xorl %edi, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    leal (%ebp,%ebp), %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edx, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb (%esp), %ah # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    addb $64, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ah, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %ebp, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ah, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    subb $64, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %esi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    decb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %esi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %edx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb $-64, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl (%esp), %edx # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    subb %dl, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ah, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %ecx, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %esi, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %esi, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %al, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel %edi, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel %esi, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ah, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %esi, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %dl, %dl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovsl %ecx, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovsl %ecx, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovsl %ecx, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel %edi, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovsl %edx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %edx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %dl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %esi, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %dl, %dl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %dl, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb (%esp), %al # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %al, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnsl %esi, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel %edi, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %al, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovsl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %al, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovsl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %ah, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %al, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovsl %ebp, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, (%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, 4(%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, 12(%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, 8(%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, 28(%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, 24(%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, 16(%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, 20(%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    addl $140, %esp
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    retl
+;
+; X32-NO-BMI2-HAVE-SHLD-LABEL: shl_32bytes:
+; X32-NO-BMI2-HAVE-SHLD:       # %bb.0:
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    subl $116, %esp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 16(%edx), %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl (%edx), %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movzbl (%eax), %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shlb $3, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    xorl %ebp, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 8(%edx), %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 24(%edx), %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 4(%edx), %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %ebx, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 20(%edx), %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %edx, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %edi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %cl, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    addb $-128, %al
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %al, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %al, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ch, %bh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    subb $64, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %edi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    negb %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %cl, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %eax, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %cl, %bl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    andb $24, %bl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %bl, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %bl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %bh, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %bh, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    addb $-64, %bh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %bh, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %bh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovbl %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %bh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %ebp, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %bl, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ch, %bl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %ch # 1-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    addb $64, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %ebp, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %ebp, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 12(%esi), %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %eax, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %bh, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %ebp, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %bh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %eax, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %bl, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %bl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 28(%esi), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %eax, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %bh, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %bh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl %ebp, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ch, %bl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %eax, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 1-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 1-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %esi, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb $-128, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    subb %bl, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %bl, %bl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %bl, %bh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel %ebp, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovsl %esi, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovsl %esi, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovsl %esi, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ch, %bl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    subb $64, %bl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    negb %bl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %bl, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %ebp, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    andb $24, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl %edi, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb $-64, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    subb %bh, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %bh # 1-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %bh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %edi, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    xorl %edi, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %ch, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel %eax, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $0, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel %ebp, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovsl %edi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %ebp, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %bh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %edi, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %ebx, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl %esi, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %eax, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %ch, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel %eax, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %cl, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovsl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %cl, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovsl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %ebx, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %ebx, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %cl, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %ch # 1-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %ch, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovsl %eax, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %cl, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %ch, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovsl %edi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ecx, (%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ecx, 4(%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ecx, 8(%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ecx, 12(%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, 24(%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, 28(%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, 16(%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, 20(%eax)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    addl $116, %esp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    retl
+;
+; X32-HAVE-BMI2-NO-SHLD-LABEL: shl_32bytes:
+; X32-HAVE-BMI2-NO-SHLD:       # %bb.0:
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    subl $164, %esp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 8(%edx), %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 24(%edx), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movzbl (%eax), %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlb $3, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ebx, %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    xorl %ecx, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ebx, %esi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 16(%edx), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ebx, %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl (%edx), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ebx, %eax, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ecx, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrl %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 4(%esi), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ebx, %eax, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %edx, %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %esi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrl %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 20(%ecx), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ebx, %ecx, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %edx, %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    addb $-128, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrl %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 12(%esi), %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, %esi, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %ecx, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %esi, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %ecx, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, %edx, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ecx, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edx, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, %ebp, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %edi, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    subb $64, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %edx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    decb %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    negb %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    andb $24, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    leal (%ebp,%ebp), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %edx, %eax, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %eax, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %esi, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %edx, %ebp, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, (%esp) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ebp, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    addb $-64, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ecx, %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ebp, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovbl %ebx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %ebp, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    xorl %edx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ecx, %eax, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edx, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    leal (%edx,%edx), %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %edi, %ebp, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %esi, %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ebp, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %esi, %edx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %ebx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %esi, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl (%esp), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, (%esp) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ecx, {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %edx, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, %edi, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %edx, {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ebx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movb $-128, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    subb %cl, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    subb $64, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    negb %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movb %al, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    andb $24, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ecx, {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %edi, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl (%esp), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %esi, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, (%esp) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrl %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %eax, %edx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %eax, {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 28(%eax), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %edi, %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %eax, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ecx, {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %eax, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    leal (%ecx,%ecx), %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ebx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %edx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %eax, %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %edx, %ecx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    # kill: def $bl killed $bl killed $ebx def $ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    addb $64, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ecx, {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %edx, {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ebx, {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ebp, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ebx, {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %edx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    subb $64, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %edi, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    decb %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %eax, {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %edx, {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %eax, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %eax, {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %edx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %eax, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movb $-64, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    subb {{[-0-9]+}}(%e{{[sb]}}p), %dl # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %edx, %esi, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %edi, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %edi, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %bl, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel %esi, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel %ebp, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %edx, %ebp, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ebp, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $0, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl (%esp), %esi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovsl %edi, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, (%esp) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel %ebx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovsl %edi, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovsl %edi, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovsl %edi, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    xorl %edi, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %edi, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %dl, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel %ebx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %bl, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnsl %esi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %bl, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovsl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %bl, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovsl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %dl, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %bl, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovsl %ecx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, (%ecx)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, 4(%ecx)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl (%esp), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, 12(%ecx)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, 8(%ecx)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, 28(%ecx)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, 24(%ecx)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, 16(%ecx)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, 20(%ecx)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    addl $164, %esp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    retl
+;
+; X32-HAVE-BMI2-HAVE-SHLD-LABEL: shl_32bytes:
+; X32-HAVE-BMI2-HAVE-SHLD:       # %bb.0:
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    subl $128, %esp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 8(%ecx), %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 24(%ecx), %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movzbl (%eax), %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlb $3, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %ecx, %edx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    xorl %ebx, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %ecx, %esi, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebx, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 16(%ebp), %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %ecx, %edi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebx, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl (%ebp), %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %ecx, %ebp, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebx, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 4(%ebx), %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %ebp, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %edx, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 20(%edx), %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %edi, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %edi, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    addb $-128, %al
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %al
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %eax, %ebp, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    xorl %eax, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    subb $64, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %eax, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    negb %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %esi, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    andb $24, %al
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %eax, %esi, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ecx, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    addb $-64, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %ebx, %edi, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ecx, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %al
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovbl %edx, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %ecx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %ebx, %eax, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ecx, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %edi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %ebp, %edi, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ecx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    # kill: def $al killed $al killed $eax def $eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    addb $64, %al
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %ebp, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %ecx, %ebp, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %ebp, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %esi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %eax, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 12(%esi), %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %edx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %ebp, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl %edi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %edx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 28(%esi), %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %edx, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %eax, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl %edi, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %edx, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %eax, %esi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %ecx, %edx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebx, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %edx, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movb $-128, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    subb %al, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %ebx, {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %ebx, %esi, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %al, %al
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel %eax, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsl %ebp, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsl %ebp, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsl %ebp, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movb %bl, %ch
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    subb $64, %ch
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %ebp, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    negb %ch
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movb %ch, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %eax, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movb %ch, %al
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    andb $24, %al
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %eax, %esi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %ch
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl %edi, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movb $-64, %al
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    subb {{[-0-9]+}}(%e{{[sb]}}p), %al # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %al
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %esi, %eax, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %esi, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %bl, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel %eax, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $0, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel %ebp, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsl %esi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %ebp, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %esi, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %ch
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %edi, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %eax, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %bl, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel %esi, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %al, %al
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %al, %al
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    # kill: def $cl killed $cl killed $ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %esi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %esi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %esi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %bl, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel %edi, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $0, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsl %ecx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %bl, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $0, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsl %esi, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, (%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, 4(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, 8(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, 12(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, 24(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, 28(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, 16(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, 20(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    addl $128, %esp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    retl
+  %src = load i256, ptr %src.ptr, align 1
+  %byteOff = load i256, ptr %byteOff.ptr, align 1
+  %bitOff = shl i256 %byteOff, 3
+  %res = shl i256 %src, %bitOff
+  store i256 %res, ptr %dst, align 1
+  ret void
+}
+define void @ashr_32bytes(ptr %src.ptr, ptr %byteOff.ptr, ptr %dst) nounwind {
+; X64-NO-BMI2-NO-SHLD-LABEL: ashr_32bytes:
+; X64-NO-BMI2-NO-SHLD:       # %bb.0:
+; X64-NO-BMI2-NO-SHLD-NEXT:    pushq %rbp
+; X64-NO-BMI2-NO-SHLD-NEXT:    pushq %r15
+; X64-NO-BMI2-NO-SHLD-NEXT:    pushq %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    pushq %r13
+; X64-NO-BMI2-NO-SHLD-NEXT:    pushq %r12
+; X64-NO-BMI2-NO-SHLD-NEXT:    pushq %rbx
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rdx, %r10
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq 24(%rdi), %r8
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq (%rdi), %r9
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r9, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq 8(%rdi), %r12
+; X64-NO-BMI2-NO-SHLD-NEXT:    movzbl (%rsi), %eax
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlb $3, %al
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r12, %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %cl, %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    xorl %ecx, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    testb $64, %al
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r14, %r15
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %rcx, %r15
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r8, %rdx
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    sarq %cl, %rdx
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r8, %r11
+; X64-NO-BMI2-NO-SHLD-NEXT:    sarq $63, %r11
+; X64-NO-BMI2-NO-SHLD-NEXT:    testb $64, %al
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rdx, %rbx
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %r11, %rbx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %cl, %r9
+; X64-NO-BMI2-NO-SHLD-NEXT:    leaq (%r12,%r12), %rbp
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %esi
+; X64-NO-BMI2-NO-SHLD-NEXT:    notb %sil
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlq %cl, %rbp
+; X64-NO-BMI2-NO-SHLD-NEXT:    orq %r9, %rbp
+; X64-NO-BMI2-NO-SHLD-NEXT:    testb $64, %al
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq 16(%rdi), %r9
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %r14, %rbp
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r9, %r13
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %cl, %r13
+; X64-NO-BMI2-NO-SHLD-NEXT:    leaq (%r8,%r8), %rdi
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rdi, %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlq %cl, %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    orq %r13, %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    testb $64, %al
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %rdx, %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    movb $-128, %sil
+; X64-NO-BMI2-NO-SHLD-NEXT:    subb %al, %sil
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r8, %rdx
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlq %cl, %rdx
+; X64-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r9, %r13
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %r13
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %cl, %r13
+; X64-NO-BMI2-NO-SHLD-NEXT:    orq %rdx, %r13
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r9, %rdx
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlq %cl, %rdx
+; X64-NO-BMI2-NO-SHLD-NEXT:    testb $64, %sil
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %rdx, %r13
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl $0, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %rcx, %rdx
+; X64-NO-BMI2-NO-SHLD-NEXT:    orq %rbp, %rdx
+; X64-NO-BMI2-NO-SHLD-NEXT:    orq %r15, %r13
+; X64-NO-BMI2-NO-SHLD-NEXT:    leal -128(%rax), %esi
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    shrq %cl, %r9
+; X64-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X64-NO-BMI2-NO-SHLD-NEXT:    shlq %cl, %rdi
+; X64-NO-BMI2-NO-SHLD-NEXT:    orq %r9, %rdi
+; X64-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ecx
+; X64-NO-BMI2-NO-SHLD-NEXT:    sarq %cl, %r8
+; X64-NO-BMI2-NO-SHLD-NEXT:    testb $64, %sil
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %r8, %rdi
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovneq %r11, %r8
+; X64-NO-BMI2-NO-SHLD-NEXT:    testb %al, %al
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovnsq %rdx, %rdi
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmoveq {{[-0-9]+}}(%r{{[sb]}}p), %rdi # 8-byte Folded Reload
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovnsq %r13, %r8
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmoveq %r12, %r8
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovsq %r11, %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    cmovsq %r11, %rbx
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rbx, 24(%r10)
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r14, 16(%r10)
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %rdi, (%r10)
+; X64-NO-BMI2-NO-SHLD-NEXT:    movq %r8, 8(%r10)
+; X64-NO-BMI2-NO-SHLD-NEXT:    popq %rbx
+; X64-NO-BMI2-NO-SHLD-NEXT:    popq %r12
+; X64-NO-BMI2-NO-SHLD-NEXT:    popq %r13
+; X64-NO-BMI2-NO-SHLD-NEXT:    popq %r14
+; X64-NO-BMI2-NO-SHLD-NEXT:    popq %r15
+; X64-NO-BMI2-NO-SHLD-NEXT:    popq %rbp
+; X64-NO-BMI2-NO-SHLD-NEXT:    retq
+;
+; X64-NO-BMI2-HAVE-SHLD-LABEL: ashr_32bytes:
+; X64-NO-BMI2-HAVE-SHLD:       # %bb.0:
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    pushq %rbp
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    pushq %r15
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    pushq %r14
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    pushq %r13
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    pushq %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    pushq %rbx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq 16(%rdi), %r8
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq (%rdi), %r9
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq 8(%rdi), %r10
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq 24(%rdi), %rdi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movzbl (%rsi), %eax
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shlb $3, %al
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %rdi, %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ecx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    sarq %cl, %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %r10, %r14
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shrq %cl, %r14
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %r9, %r15
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shrdq %cl, %r10, %r15
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %rdi, %rsi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    sarq $63, %rsi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    xorl %ebp, %ebp
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    testb $64, %al
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %r14, %r15
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %r12, %r11
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %rsi, %r11
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %rbp, %r14
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %r8, %rbx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shrdq %cl, %rdi, %rbx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    testb $64, %al
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %r12, %rbx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movb $-128, %cl
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    subb %al, %cl
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %rdi, %r13
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shldq %cl, %r8, %r13
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %r8, %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shlq %cl, %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    testb $64, %cl
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %r12, %r13
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %rbp, %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    orq %r15, %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    orq %r14, %r13
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    leal -128(%rax), %ecx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    shrdq %cl, %rdi, %r8
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    sarq %cl, %rdi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    testb $64, %cl
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %rdi, %r8
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovneq %rsi, %rdi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    testb %al, %al
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovnsq %r13, %rdi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmoveq %r10, %rdi
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovnsq %r12, %r8
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmoveq %r9, %r8
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovsq %rsi, %rbx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    cmovsq %rsi, %r11
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %rbx, 16(%rdx)
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %r11, 24(%rdx)
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %r8, (%rdx)
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    movq %rdi, 8(%rdx)
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    popq %rbx
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    popq %r12
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    popq %r13
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    popq %r14
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    popq %r15
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    popq %rbp
+; X64-NO-BMI2-HAVE-SHLD-NEXT:    retq
+;
+; X64-HAVE-BMI2-NO-SHLD-LABEL: ashr_32bytes:
+; X64-HAVE-BMI2-NO-SHLD:       # %bb.0:
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    pushq %rbp
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    pushq %r15
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    pushq %r14
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    pushq %r13
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    pushq %r12
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    pushq %rbx
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq 24(%rdi), %r10
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq (%rdi), %rbx
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq 8(%rdi), %rcx
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movzbl (%rsi), %r8d
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlb $3, %r8b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrxq %r8, %rcx, %r15
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    xorl %eax, %eax
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    testb $64, %r8b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %r15, %r11
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %rax, %r11
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %r10, %rsi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    sarq $63, %rsi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    testb $64, %r8b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    sarxq %r8, %r10, %r12
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %r12, %r9
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %rsi, %r9
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrxq %r8, %rbx, %r14
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movl %r8d, %ebp
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    notb %bpl
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    leaq (%rcx,%rcx), %r13
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlxq %rbp, %r13, %r13
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    orq %r14, %r13
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    testb $64, %r8b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq 16(%rdi), %r14
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %r15, %r13
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrxq %r8, %r14, %rax
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    leaq (%r10,%r10), %r15
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlxq %rbp, %r15, %rdi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    orq %rax, %rdi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    testb $64, %r8b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %r12, %rdi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movb $-128, %al
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    subb %r8b, %al
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, %r12d
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    notb %r12b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %r14, %rbp
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrq %rbp
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrxq %r12, %rbp, %r12
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlxq %rax, %r10, %rbp
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    orq %rbp, %r12
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    testb $64, %al
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlxq %rax, %r14, %rax
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %rax, %r12
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %ebp
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %rbp, %rax
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    orq %r13, %rax
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    orq %r11, %r12
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    leal -128(%r8), %r11d
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movl %r11d, %r13d
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    notb %r13b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shlxq %r13, %r15, %r15
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    shrxq %r11, %r14, %r14
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    orq %r14, %r15
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    sarxq %r11, %r10, %r10
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    testb $64, %r11b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %r10, %r15
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovneq %rsi, %r10
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    testb %r8b, %r8b
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovnsq %rax, %r15
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmoveq %rbx, %r15
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovnsq %r12, %r10
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmoveq %rcx, %r10
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovsq %rsi, %rdi
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    cmovsq %rsi, %r9
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %r9, 24(%rdx)
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %rdi, 16(%rdx)
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %r15, (%rdx)
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    movq %r10, 8(%rdx)
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    popq %rbx
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    popq %r12
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    popq %r13
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    popq %r14
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    popq %r15
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    popq %rbp
+; X64-HAVE-BMI2-NO-SHLD-NEXT:    retq
+;
+; X64-HAVE-BMI2-HAVE-SHLD-LABEL: ashr_32bytes:
+; X64-HAVE-BMI2-HAVE-SHLD:       # %bb.0:
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    pushq %rbp
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    pushq %r15
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    pushq %r14
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    pushq %r13
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    pushq %r12
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    pushq %rbx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq 16(%rdi), %r8
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq 24(%rdi), %r11
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq (%rdi), %r9
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq 8(%rdi), %rdi
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movzbl (%rsi), %eax
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shlb $3, %al
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %r9, %r14
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ecx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdq %cl, %rdi, %r14
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxq %rax, %rdi, %r15
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %r11, %rsi
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    sarq $63, %rsi
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    xorl %r12d, %r12d
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $64, %al
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %r15, %r14
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    sarxq %rax, %r11, %r13
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %r13, %r10
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %rsi, %r10
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %r12, %r15
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %r8, %rbx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdq %cl, %r11, %rbx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $64, %al
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %r13, %rbx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movb $-128, %cl
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    subb %al, %cl
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %r11, %r13
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shldq %cl, %r8, %r13
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxq %rcx, %r8, %rbp
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $64, %cl
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %rbp, %r13
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %r12, %rbp
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    orq %r14, %rbp
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    orq %r15, %r13
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    leal -128(%rax), %ecx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdq %cl, %r11, %r8
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    sarxq %rcx, %r11, %r11
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $64, %cl
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %r11, %r8
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovneq %rsi, %r11
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %al, %al
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnsq %r13, %r11
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmoveq %rdi, %r11
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnsq %rbp, %r8
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmoveq %r9, %r8
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsq %rsi, %rbx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsq %rsi, %r10
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %rbx, 16(%rdx)
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %r10, 24(%rdx)
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %r8, (%rdx)
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    movq %r11, 8(%rdx)
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    popq %rbx
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    popq %r12
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    popq %r13
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    popq %r14
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    popq %r15
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    popq %rbp
+; X64-HAVE-BMI2-HAVE-SHLD-NEXT:    retq
+;
+; X32-NO-BMI2-NO-SHLD-LABEL: ashr_32bytes:
+; X32-NO-BMI2-NO-SHLD:       # %bb.0:
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    pushl %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    subl $144, %esp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 12(%ecx), %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 28(%ecx), %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb (%eax), %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    shlb $3, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    sarl %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    sarl $31, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    xorl %ebx, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %ebx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 20(%ebp), %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 4(%ebp), %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 24(%ebp), %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    leal (%eax,%eax), %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %ebx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 8(%esi), %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    addl %esi, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %al, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edi, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %edx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %edi, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    addb $-128, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    sarl %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bl, %bh
+; X32-NO-BMI2-NO-SHLD-NEXT:    subb $64, %bh
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %eax, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bh, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    decb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    negb %bh
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bh, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    andb $24, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %bh
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bh, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bl, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    addb $-64, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    sarl %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovbl %edx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %eax, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %bh
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bl, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    addb $64, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %ebx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    sarl %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %ebx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %edx, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %bh # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %bh
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %edx, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl 16(%eax), %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %cl, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    leal (%eax,%eax), %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %ah # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ah, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ah, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edx, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edi, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %ebp, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl (%ecx), %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %bh
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bh, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %al, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    addl %esi, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ah, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bl, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edx, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edi, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %ebx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %ch # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %esi, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %edx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb $-128, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    subb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %edx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %edx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %cl, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edx, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %ch # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %ebx, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb $-64, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %ah # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    subb %ah, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %al, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    notb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %edx, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %al, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %edx, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %ebp, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %ah, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel %eax, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovsl %edx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovsl %edx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovsl %esi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %bh
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    subb $64, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %ebp, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bl, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    decb %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    negb %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %bl, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    andb $24, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %esi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %edx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %ecx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %bh, %bh
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $0, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel %ebp, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovsl %ecx, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %ebp, %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel %ebx, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %ch # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-NO-SHLD-NEXT:    shll %cl, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    shrl %cl, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %ebx, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl %esi, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %eax, %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebp, %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %ch, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %ah # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %ah, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovsl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %ch
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl $0, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael %ebx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %al # 1-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmpb $64, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %al, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %ah, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovnsl %ebx, %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %ah, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovsl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %al, %al
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    testb %ah, %ah
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovsl %ebx, %ecx
+; X32-NO-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, 28(%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, 24(%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, 16(%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ecx, 20(%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %ebx, (%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %esi, 12(%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edi, 4(%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    movl %edx, 8(%eax)
+; X32-NO-BMI2-NO-SHLD-NEXT:    addl $144, %esp
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %esi
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %edi
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %ebx
+; X32-NO-BMI2-NO-SHLD-NEXT:    popl %ebp
+; X32-NO-BMI2-NO-SHLD-NEXT:    retl
+;
+; X32-NO-BMI2-HAVE-SHLD-LABEL: ashr_32bytes:
+; X32-NO-BMI2-HAVE-SHLD:       # %bb.0:
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    pushl %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    subl $124, %esp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 12(%ecx), %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 28(%ecx), %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb (%eax), %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shlb $3, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    sarl %cl, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    sarl $31, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    xorl %edx, %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edx, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebx, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 20(%edx), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 4(%edx), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 8(%edx), %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %ebp, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 24(%edx), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %ebp, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %edx, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    addb $-128, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    sarl %cl, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edx, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ch, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    subb $64, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %edi, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %edi, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    negb %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %dl, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %edi, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    andb $24, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ch, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    addb $-64, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %dh, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    sarl %cl, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovbl %ebp, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %eax, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %dl, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %dl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %ebp, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl 16(%eax), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %dh, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %ebp, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebx, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl %edi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl (%eax), %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %eax, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb $-128, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    subb %ch, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %cl, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %dh, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %eax, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %eax, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %dl, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    subb $64, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %edi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    xorl %edi, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    negb %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %dh, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    andb $24, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrl %cl, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %dl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %ebx, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %edi, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %dl, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %edx, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %cl # 1-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl %esi, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %ebp, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %cl, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel %edx, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ch, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    addb $64, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %edi, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    sarl %cl, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %ch, %ch
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb %ch, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovsl %ebp, %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovsl %ebp, %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovsl %ebp, %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovsl %ecx, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb {{[-0-9]+}}(%e{{[sb]}}p), %dh # 1-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %ecx, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %dl, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnsl %eax, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %ecx, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $0, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %dl, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovsl %eax, %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movb $-64, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    subb %dl, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shll %cl, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %esi, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %dh, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %dl, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovsl %ebp, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %edi, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 1-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %ebp, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %cl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl $0, %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovael %ebp, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %dh, %dh
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel %edi, %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    testb %dl, %dl
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovsl %esi, %eax
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, 24(%ecx)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, 28(%ecx)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, 16(%ecx)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, 20(%ecx)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %edx, (%ecx)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, 4(%ecx)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %eax, 8(%ecx)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    movl %ebx, 12(%ecx)
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    addl $124, %esp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %esi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %edi
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %ebx
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    popl %ebp
+; X32-NO-BMI2-HAVE-SHLD-NEXT:    retl
+;
+; X32-HAVE-BMI2-NO-SHLD-LABEL: ashr_32bytes:
+; X32-HAVE-BMI2-NO-SHLD:       # %bb.0:
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    pushl %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    subl $168, %esp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 28(%edx), %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 4(%edx), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movzbl (%eax), %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlb $3, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    sarl $31, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    xorl %ecx, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 20(%edx), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 12(%edx), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    sarxl %ebx, %esi, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 24(%ecx), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %ecx, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    leal (%eax,%eax), %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %edx, %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 8(%eax), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %eax, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    addl %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %edx, %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %esi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %edi, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    addb $-128, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    sarxl %eax, {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    subb $64, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    decb %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrl %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    negb %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movb %bl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    andb $24, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %esi, %edx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ebp, %eax, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %edx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ebp, %ecx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    addb $-64, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    sarxl %ecx, %eax, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovbl %esi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ecx, %edx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ecx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrl %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %edi, %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ebp, %edx, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %eax, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ebp, %esi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    # kill: def $bl killed $bl killed $ebx def $ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    addb $64, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, %edi, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    addl %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ecx, %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    leal (%eax,%eax), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ecx, %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    # kill: def $cl killed $cl killed $ecx def $ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ecx, %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ecx, {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %edi, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ebx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    sarxl %esi, {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ecx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %edx, %ebp, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ecx, %edi, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ebp, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %ebp, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ebp, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl 16(%edi), %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ecx, %edi, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ebp, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %edx, {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ebp, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %eax, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl (%eax), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edx, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ecx, %eax, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %eax, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebp, {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %eax, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ebx, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %ebp, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ebp, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %edx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ebp, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %edi, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ecx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movb $-128, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    subb %cl, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrl %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ebx, %edx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %esi, %ebp, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %eax, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %ebx, %edi, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edx, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, %ebx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movb $-64, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    subb %cl, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    notb %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %edx, {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %edi, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, %ebx, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %edi, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %esi, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %cl, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel %edx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovsl %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovsl %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovsl %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    subb $64, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %esi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    decb %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    negb %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movb %bl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    # kill: def $bl killed $bl def $ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    andb $24, %bl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %ebx, %edx, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %eax, {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shrxl %eax, {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %ebp, %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %cl, %cl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $0, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovsl %edi, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %edi, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel %eax, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    shlxl %eax, %edi, %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl %ebp, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb $32, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %ecx, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %al, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel %edi, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $0, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovsl %edx, %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl $0, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael %eax, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmpb $64, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %al, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %dl, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovnsl %edi, %ecx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %dl, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovsl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %al, %al
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    testb %dl, %dl
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovsl %edi, %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %eax, %edx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, 28(%eax)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, 24(%eax)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, 16(%eax)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edi, 20(%eax)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %edx, (%eax)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ebx, 12(%eax)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %ecx, 4(%eax)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    movl %esi, 8(%eax)
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    addl $168, %esp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %esi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %edi
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %ebx
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    popl %ebp
+; X32-HAVE-BMI2-NO-SHLD-NEXT:    retl
+;
+; X32-HAVE-BMI2-HAVE-SHLD-LABEL: ashr_32bytes:
+; X32-HAVE-BMI2-HAVE-SHLD:       # %bb.0:
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    pushl %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    subl $132, %esp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 28(%ecx), %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 24(%ecx), %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 12(%ecx), %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 8(%ecx), %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movzbl (%eax), %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlb $3, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %edx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %esi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    sarl $31, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    sarxl %ebx, %esi, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ecx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %ebx, %edx, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 4(%ebp), %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %ebx, %eax, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %edx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 20(%ebp), %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %ebx, %edx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %ebp, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    addb $-128, %al
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %al
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %eax, %edx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %edx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    sarxl %eax, %ebx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    subb $64, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %eax, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %eax, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    negb %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movb %cl, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %edi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    andb $24, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %ebx, %edi, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    addb $-64, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    sarxl %ebx, %esi, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %edx, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovbl %eax, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %edx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %ebx, %eax, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %edx, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %eax, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %edi, %eax, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %esi, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl 16(%eax), %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %eax, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %esi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl %ebp, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %eax, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl (%eax), %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %edx, %esi, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %edx, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movb $-128, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    subb %cl, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %edx, %edi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %edx, %esi, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %eax, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %edi, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %al
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %edx, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %al
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    subb $64, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %ebp, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    negb %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %edx, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    andb $24, %dl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrxl %edx, %ebx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ebp, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %edi, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %edi, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl %eax, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %edx, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %bl, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel %edi, %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    addb $64, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shrdl %cl, %edi, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    sarxl %ecx, %edi, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %ecx, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %edx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %al, %al
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsl %edx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsl %edx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsl %edx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsl %edx, %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %edx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %cl, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnsl %esi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %edx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $0, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel %edi, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %cl, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsl %eax, %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movb $-64, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    subb %dl, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shlxl %ecx, %esi, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edi, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %eax, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %al
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %al, %al
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel %eax, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $0, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsl %edx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %esi, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel %edi, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %esi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    shldl %cl, %ebx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb $32, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ecx, %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $64, %bl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl $0, %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovael %edi, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    testb %cl, %cl
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel %esi, %edx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmpb $0, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovsl %edx, %eax
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    cmovel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, 24(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, 28(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, 16(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, 20(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %edx, (%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, 4(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %eax, 8(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    movl %ebp, 12(%ecx)
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    addl $132, %esp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %esi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %edi
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %ebx
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    popl %ebp
+; X32-HAVE-BMI2-HAVE-SHLD-NEXT:    retl
+  %src = load i256, ptr %src.ptr, align 1
+  %byteOff = load i256, ptr %byteOff.ptr, align 1
+  %bitOff = shl i256 %byteOff, 3
+  %res = ashr i256 %src, %bitOff
+  store i256 %res, ptr %dst, align 1
+  ret void
+}
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; ALL: {{.*}}
+; X32: {{.*}}
+; X32-NO-SHLD: {{.*}}
+; X32-SHLD: {{.*}}
+; X64: {{.*}}
+; X64-NO-SHLD: {{.*}}
+; X64-SHLD: {{.*}}


        


More information about the llvm-commits mailing list